freebsd-dev/lib/libpthread/thread/thr_kern.c

2485 lines
67 KiB
C
Raw Normal View History

/*
* Copyright (C) 2003 Daniel M. Eischen <deischen@freebsd.org>
* Copyright (C) 2002 Jonathon Mini <mini@freebsd.org>
* Copyright (c) 1995-1998 John Birrell <jb@cimlogic.com.au>
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* 3. All advertising materials mentioning features or use of this software
* must display the following acknowledgement:
* This product includes software developed by John Birrell.
* 4. Neither the name of the author nor the names of any co-contributors
* may be used to endorse or promote products derived from this software
* without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY JOHN BIRRELL AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
[ The author's description... ] o Runnable threads are now maintained in priority queues. The implementation requires two things: 1.) The priority queues must be protected during insertion and removal of threads. Since the kernel scheduler must modify the priority queues, a spinlock for protection cannot be used. The functions _thread_kern_sched_defer() and _thread_kern_sched_undefer() were added to {un}defer kernel scheduler activation. 2.) A thread (active) priority change can be performed only when the thread is removed from the priority queue. The implementation uses a threads active priority when inserting it into the queue. A by-product is that thread switches are much faster. A separate queue is used for waiting and/or blocked threads, and it is searched at most 2 times in the kernel scheduler when there are active threads. It should be possible to reduce this to once by combining polling of threads waiting on I/O with the loop that looks for timed out threads and the minimum timeout value. o Functions to defer kernel scheduler activation were added. These are _thread_kern_sched_defer() and _thread_kern_sched_undefer() and may be called recursively. These routines do not block the scheduling signal, but latch its occurrence. The signal handler will not call the kernel scheduler when the running thread has deferred scheduling, but it will be called when running thread undefers scheduling. o Added support for _POSIX_THREAD_PRIORITY_SCHEDULING. All the POSIX routines required by this should now be implemented. One note, SCHED_OTHER, SCHED_FIFO, and SCHED_RR are required to be defined by including pthread.h. These defines are currently in sched.h. I modified pthread.h to include sched.h but don't know if this is the proper thing to do. o Added support for priority protection and inheritence mutexes. This allows definition of _POSIX_THREAD_PRIO_PROTECT and _POSIX_THREAD_PRIO_INHERIT. o Added additional error checks required by POSIX for mutexes and condition variables. o Provided a wrapper for sigpending which is marked as a hidden syscall. o Added a non-portable function as a debugging aid to allow an application to monitor thread context switches. An application can install a routine that gets called everytime a thread (explicitly created by the application) gets context switched. The routine gets passed the pthread IDs of the threads that are being switched in and out. Submitted by: Dan Eischen <eischen@vigrid.com> Changes by me: o Added a PS_SPINBLOCK state to deal with the priority inversion problem most often (I think) seen by threads calling malloc/free/realloc. o Dispatch signals to the running thread directly rather than at a context switch to avoid the situation where the switch never occurs.
1999-03-23 05:07:56 +00:00
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*
*/
#include <sys/cdefs.h>
2003-04-18 07:45:03 +00:00
__FBSDID("$FreeBSD$");
#include <sys/types.h>
#include <sys/kse.h>
#include <sys/signalvar.h>
#include <sys/queue.h>
#include <machine/atomic.h>
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
#include <machine/sigframe.h>
#include <assert.h>
#include <errno.h>
#include <signal.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>
#include <ucontext.h>
#include <unistd.h>
#include "atomic_ops.h"
#include "thr_private.h"
#include "libc_private.h"
/*#define DEBUG_THREAD_KERN */
Implement zero system call thread switching. Performance of thread switches should be on par with that under scheduler activations. o Timing is achieved through the use of a fixed interval timer (ITIMER_PROF) to count scheduling ticks instead of retrieving the time-of-day upon every thread switch and calculating elapsed real time. o Polling for I/O readiness is performed once for each scheduling tick instead of every thread switch. o The non-signal saving/restoring versions of setjmp/longjmp are used to save and restore thread contexts. This may allow the removal of _THREAD_SAFE macros from setjmp() and longjmp() - needs more investigation. Change signal handling so that signals are handled in the context of the thread that is receiving the signal. When signals are dispatched to a thread, a special signal handling frame is created on top of the target threads stack. The frame contains the threads saved state information and a new context in which the thread can run. The applications signal handler is invoked through a wrapper routine that knows how to restore the threads saved state and unwind to previous frames. Fix interruption of threads due to signals. Some states were being improperly interrupted while other states were not being interrupted. This should fix several PRs. Signal handlers, which are invoked as a result of a process signal (not by pthread_kill()), are now called with the code (or siginfo_t if SA_SIGINFO was set in sa_flags) and sigcontext_t as received from the process signal handler. Modify the search for a thread to which a signal is delivered. The search algorithm is now: o First thread found in sigwait() with signal in wait mask. o First thread found sigsuspend()'d on the signal. o Current thread if signal is unmasked. o First thread found with signal unmasked. Collapse machine dependent support into macros defined in pthread_private.h. These should probably eventually be moved into separate MD files. Change the range of settable priorities to be compliant with POSIX (0-31). The threads library uses higher priorities internally for real-time threads (not yet implemented) and threads executing signal handlers. Real-time threads and threads running signal handlers add 64 and 32, respectively, to a threads base priority. Some other small changes and cleanups. PR: 17757 18559 21943 Reviewed by: jasone
2000-10-13 22:12:32 +00:00
#ifdef DEBUG_THREAD_KERN
#define DBG_MSG stdout_debug
#else
#define DBG_MSG(x...)
#endif
/*
* Define a high water mark for the maximum number of threads that
* will be cached. Once this level is reached, any extra threads
* will be free()'d.
*/
#define MAX_CACHED_THREADS 100
/*
* Define high water marks for the maximum number of KSEs and KSE groups
* that will be cached. Because we support 1:1 threading, there could have
* same number of KSEs and KSE groups as threads. Once these levels are
* reached, any extra KSE and KSE groups will be free()'d.
*/
#ifdef SYSTEM_SCOPE_ONLY
#define MAX_CACHED_KSES 100
#define MAX_CACHED_KSEGS 100
#else
#define MAX_CACHED_KSES 50
#define MAX_CACHED_KSEGS 50
#endif
#define KSE_STACKSIZE 16384
#define KSE_SET_MBOX(kse, thrd) \
(kse)->k_kcb->kcb_kmbx.km_curthread = &(thrd)->tcb->tcb_tmbx
In the words of the author: o The polling mechanism for I/O readiness was changed from select() to poll(). In additon, a wrapped version of poll() is now provided. o The wrapped select routine now converts each fd_set to a poll array so that the thread scheduler doesn't have to perform a bitwise search for selected fds each time file descriptors are polled for I/O readiness. o The thread scheduler was modified to use a new queue (_workq) for threads that need work. Threads waiting for I/O readiness and spinblocks are added to the work queue in addition to the waiting queue. This reduces the time spent forming/searching the array of file descriptors being polled. o The waiting queue (_waitingq) is now maintained in order of thread wakeup time. This allows the thread scheduler to find the nearest wakeup time by looking at the first thread in the queue instead of searching the entire queue. o Removed file descriptor locking for select/poll routines. An application should not rely on the threads library for providing this locking; if necessary, the application should use mutexes to protect selecting/polling of file descriptors. o Retrieve and use the kernel clock rate/resolution at startup instead of hardcoding the clock resolution to 10 msec (tested with kernel running at 1000 HZ). o All queues have been changed to use queue.h macros. These include the queues of all threads, dead threads, and threads waiting for file descriptor locks. o Added reinitialization of the GC mutex and condition variable after a fork. Also prevented reallocation of the ready queue after a fork. o Prevented the wrapped close routine from closing the thread kernel pipes. o Initialized file descriptor table for stdio entries at thread init. o Provided additional flags to indicate to what queues threads belong. o Moved TAILQ initialization for statically allocated mutex and condition variables to after the spinlock. o Added dispatching of signals to pthread_kill. Removing the dispatching of signals from thread activation broke sigsuspend when pthread_kill was used to send a signal to a thread. o Temporarily set the state of a thread to PS_SUSPENDED when it is first created and placed in the list of threads so that it will not be accidentally scheduled before becoming a member of one of the scheduling queues. o Change the signal handler to queue signals to the thread kernel pipe if the scheduling queues are protected. When scheduling queues are unprotected, signals are then dequeued and handled. o Ensured that all installed signal handlers block the scheduling signal and that the scheduling signal handler blocks all other signals. This ensures that the signal handler is only interruptible for and by non-scheduling signals. An atomic lock is used to decide which instance of the signal handler will handle pending signals. o Removed _lock_thread_list and _unlock_thread_list as they are no longer used to protect the thread list. o Added missing RCS IDs to modified files. o Added checks for appropriate queue membership and activity when adding, removing, and searching the scheduling queues. These checks add very little overhead and are enabled when compiled with _PTHREADS_INVARIANTS defined. Suggested and implemented by Tor Egge with some modification by me. o Close a race condition in uthread_close. (Tor Egge) o Protect the scheduling queues while modifying them in pthread_cond_signal and _thread_fd_unlock. (Tor Egge) o Ensure that when a thread gets a mutex, the mutex is on that threads list of owned mutexes. (Tor Egge) o Set the kernel-in-scheduler flag in _thread_kern_sched_state and _thread_kern_sched_state_unlock to prevent a scheduling signal from calling the scheduler again. (Tor Egge) o Don't use TAILQ_FOREACH macro while searching the waiting queue for threads in a sigwait state, because a change of state destroys the TAILQ link. It is actually safe to do so, though, because once a sigwaiting thread is found, the loop ends and the function returns. (Tor Egge) o When dispatching signals to threads, make the thread inherit the signal deferral flag of the currently running thread. (Tor Egge) Submitted by: Daniel Eischen <eischen@vigrid.com> and Tor Egge <Tor.Egge@fast.no>
1999-06-20 08:28:48 +00:00
#define KSE_SET_EXITED(kse) (kse)->k_flags |= KF_EXITED
/*
* Macros for manipulating the run queues. The priority queue
* routines use the thread's pqe link and also handle the setting
* and clearing of the thread's THR_FLAGS_IN_RUNQ flag.
*/
#define KSE_RUNQ_INSERT_HEAD(kse, thrd) \
_pq_insert_head(&(kse)->k_schedq->sq_runq, thrd)
#define KSE_RUNQ_INSERT_TAIL(kse, thrd) \
_pq_insert_tail(&(kse)->k_schedq->sq_runq, thrd)
#define KSE_RUNQ_REMOVE(kse, thrd) \
_pq_remove(&(kse)->k_schedq->sq_runq, thrd)
#define KSE_RUNQ_FIRST(kse) _pq_first(&(kse)->k_schedq->sq_runq)
#define KSE_RUNQ_THREADS(kse) ((kse)->k_schedq->sq_runq.pq_threads)
/*
* We've got to keep track of everything that is allocated, not only
* to have a speedy free list, but also so they can be deallocated
* after a fork().
*/
static TAILQ_HEAD(, kse) active_kseq;
static TAILQ_HEAD(, kse) free_kseq;
static TAILQ_HEAD(, kse_group) free_kse_groupq;
static TAILQ_HEAD(, kse_group) active_kse_groupq;
static TAILQ_HEAD(, kse_group) gc_ksegq;
static struct lock kse_lock; /* also used for kseg queue */
static int free_kse_count = 0;
static int free_kseg_count = 0;
static TAILQ_HEAD(, pthread) free_threadq;
static struct lock thread_lock;
static int free_thread_count = 0;
static int inited = 0;
static int active_threads = 1;
static int active_kse_count = 0;
static int active_kseg_count = 0;
static u_int64_t next_uniqueid = 1;
LIST_HEAD(thread_hash_head, pthread);
#define THREAD_HASH_QUEUES 127
static struct thread_hash_head thr_hashtable[THREAD_HASH_QUEUES];
#define THREAD_HASH(thrd) ((unsigned long)thrd % THREAD_HASH_QUEUES)
#ifdef DEBUG_THREAD_KERN
static void dump_queues(struct kse *curkse);
#endif
static void kse_check_completed(struct kse *kse);
static void kse_check_waitq(struct kse *kse);
static void kse_fini(struct kse *curkse);
static void kse_reinit(struct kse *kse, int sys_scope);
static void kse_sched_multi(struct kse_mailbox *kmbx);
static void kse_sched_single(struct kse_mailbox *kmbx);
static void kse_switchout_thread(struct kse *kse, struct pthread *thread);
static void kse_wait(struct kse *kse, struct pthread *td_wait, int sigseq);
static void kse_free_unlocked(struct kse *kse);
static void kse_destroy(struct kse *kse);
static void kseg_free_unlocked(struct kse_group *kseg);
static void kseg_init(struct kse_group *kseg);
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
static void kseg_reinit(struct kse_group *kseg);
static void kseg_destroy(struct kse_group *kseg);
static void kse_waitq_insert(struct pthread *thread);
static void kse_wakeup_multi(struct kse *curkse);
static struct kse_mailbox *kse_wakeup_one(struct pthread *thread);
static void thr_cleanup(struct kse *kse, struct pthread *curthread);
static void thr_link(struct pthread *thread);
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
static void thr_resume_wrapper(int sig, siginfo_t *, ucontext_t *);
static void thr_resume_check(struct pthread *curthread, ucontext_t *ucp,
struct pthread_sigframe *psf);
static int thr_timedout(struct pthread *thread, struct timespec *curtime);
static void thr_unlink(struct pthread *thread);
static void thread_gc(struct pthread *thread);
static void kse_gc(struct pthread *thread);
static void kseg_gc(struct pthread *thread);
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
static void __inline
thr_accounting(struct pthread *thread)
{
if ((thread->slice_usec != -1) &&
(thread->slice_usec <= TIMESLICE_USEC) &&
(thread->attr.sched_policy != SCHED_FIFO)) {
thread->slice_usec += (thread->tcb->tcb_tmbx.tm_uticks
+ thread->tcb->tcb_tmbx.tm_sticks) * _clock_res_usec;
/* Check for time quantum exceeded: */
if (thread->slice_usec > TIMESLICE_USEC)
thread->slice_usec = -1;
}
thread->tcb->tcb_tmbx.tm_uticks = 0;
thread->tcb->tcb_tmbx.tm_sticks = 0;
}
/*
* This is called after a fork().
* No locks need to be taken here since we are guaranteed to be
* single threaded.
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
*
* XXX
* POSIX says for threaded process, fork() function is used
* only to run new programs, and the effects of calling functions
* that require certain resources between the call to fork() and
* the call to an exec function are undefined.
*
* Here it is not safe to reinitialize the library after fork().
* Because memory management may be corrupted, further calling
* malloc()/free() may cause undefined behavior.
*/
Implement zero system call thread switching. Performance of thread switches should be on par with that under scheduler activations. o Timing is achieved through the use of a fixed interval timer (ITIMER_PROF) to count scheduling ticks instead of retrieving the time-of-day upon every thread switch and calculating elapsed real time. o Polling for I/O readiness is performed once for each scheduling tick instead of every thread switch. o The non-signal saving/restoring versions of setjmp/longjmp are used to save and restore thread contexts. This may allow the removal of _THREAD_SAFE macros from setjmp() and longjmp() - needs more investigation. Change signal handling so that signals are handled in the context of the thread that is receiving the signal. When signals are dispatched to a thread, a special signal handling frame is created on top of the target threads stack. The frame contains the threads saved state information and a new context in which the thread can run. The applications signal handler is invoked through a wrapper routine that knows how to restore the threads saved state and unwind to previous frames. Fix interruption of threads due to signals. Some states were being improperly interrupted while other states were not being interrupted. This should fix several PRs. Signal handlers, which are invoked as a result of a process signal (not by pthread_kill()), are now called with the code (or siginfo_t if SA_SIGINFO was set in sa_flags) and sigcontext_t as received from the process signal handler. Modify the search for a thread to which a signal is delivered. The search algorithm is now: o First thread found in sigwait() with signal in wait mask. o First thread found sigsuspend()'d on the signal. o Current thread if signal is unmasked. o First thread found with signal unmasked. Collapse machine dependent support into macros defined in pthread_private.h. These should probably eventually be moved into separate MD files. Change the range of settable priorities to be compliant with POSIX (0-31). The threads library uses higher priorities internally for real-time threads (not yet implemented) and threads executing signal handlers. Real-time threads and threads running signal handlers add 64 and 32, respectively, to a threads base priority. Some other small changes and cleanups. PR: 17757 18559 21943 Reviewed by: jasone
2000-10-13 22:12:32 +00:00
void
_kse_single_thread(struct pthread *curthread)
Implement zero system call thread switching. Performance of thread switches should be on par with that under scheduler activations. o Timing is achieved through the use of a fixed interval timer (ITIMER_PROF) to count scheduling ticks instead of retrieving the time-of-day upon every thread switch and calculating elapsed real time. o Polling for I/O readiness is performed once for each scheduling tick instead of every thread switch. o The non-signal saving/restoring versions of setjmp/longjmp are used to save and restore thread contexts. This may allow the removal of _THREAD_SAFE macros from setjmp() and longjmp() - needs more investigation. Change signal handling so that signals are handled in the context of the thread that is receiving the signal. When signals are dispatched to a thread, a special signal handling frame is created on top of the target threads stack. The frame contains the threads saved state information and a new context in which the thread can run. The applications signal handler is invoked through a wrapper routine that knows how to restore the threads saved state and unwind to previous frames. Fix interruption of threads due to signals. Some states were being improperly interrupted while other states were not being interrupted. This should fix several PRs. Signal handlers, which are invoked as a result of a process signal (not by pthread_kill()), are now called with the code (or siginfo_t if SA_SIGINFO was set in sa_flags) and sigcontext_t as received from the process signal handler. Modify the search for a thread to which a signal is delivered. The search algorithm is now: o First thread found in sigwait() with signal in wait mask. o First thread found sigsuspend()'d on the signal. o Current thread if signal is unmasked. o First thread found with signal unmasked. Collapse machine dependent support into macros defined in pthread_private.h. These should probably eventually be moved into separate MD files. Change the range of settable priorities to be compliant with POSIX (0-31). The threads library uses higher priorities internally for real-time threads (not yet implemented) and threads executing signal handlers. Real-time threads and threads running signal handlers add 64 and 32, respectively, to a threads base priority. Some other small changes and cleanups. PR: 17757 18559 21943 Reviewed by: jasone
2000-10-13 22:12:32 +00:00
{
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
#ifdef NOTYET
struct kse *kse;
struct kse_group *kseg;
struct pthread *thread;
kse_critical_t crit;
int i;
if (__isthreaded) {
_thr_rtld_fini();
_thr_signal_deinit();
}
__isthreaded = 0;
/*
* Restore signal mask early, so any memory problems could
* dump core.
*/
sigprocmask(SIG_SETMASK, &curthread->sigmask, NULL);
active_threads = 1;
2002-10-30 06:07:18 +00:00
/*
* Enter a loop to remove and free all threads other than
* the running thread from the active thread list:
2002-10-30 06:07:18 +00:00
*/
while ((thread = TAILQ_FIRST(&_thread_list)) != NULL) {
THR_GCLIST_REMOVE(thread);
/*
* Remove this thread from the list (the current
* thread will be removed but re-added by libpthread
* initialization.
2002-10-30 06:07:18 +00:00
*/
TAILQ_REMOVE(&_thread_list, thread, tle);
/* Make sure this isn't the running thread: */
if (thread != curthread) {
_thr_stack_free(&thread->attr);
if (thread->specific != NULL)
free(thread->specific);
for (i = 0; i < MAX_THR_LOCKLEVEL; i++) {
_lockuser_destroy(&thread->lockusers[i]);
}
_lock_destroy(&thread->lock);
free(thread);
}
}
TAILQ_INIT(&curthread->mutexq); /* initialize mutex queue */
curthread->joiner = NULL; /* no joining threads yet */
curthread->refcount = 0;
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
SIGEMPTYSET(curthread->sigpend); /* clear pending signals */
if (curthread->specific != NULL) {
free(curthread->specific);
curthread->specific = NULL;
curthread->specific_data_count = 0;
}
/* Free the free KSEs: */
while ((kse = TAILQ_FIRST(&free_kseq)) != NULL) {
TAILQ_REMOVE(&free_kseq, kse, k_qe);
for (i = 0; i < MAX_KSE_LOCKLEVEL; i++) {
_lockuser_destroy(&kse->k_lockusers[i]);
}
_lock_destroy(&kse->k_lock);
_kcb_dtor(kse->k_kcb);
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
if (kse->k_stack.ss_sp != NULL)
free(kse->k_stack.ss_sp);
free(kse);
}
free_kse_count = 0;
/* Free the active KSEs: */
while ((kse = TAILQ_FIRST(&active_kseq)) != NULL) {
TAILQ_REMOVE(&active_kseq, kse, k_qe);
for (i = 0; i < MAX_KSE_LOCKLEVEL; i++) {
_lockuser_destroy(&kse->k_lockusers[i]);
2002-10-30 06:07:18 +00:00
}
_lock_destroy(&kse->k_lock);
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
if (kse->k_stack.ss_sp != NULL)
free(kse->k_stack.ss_sp);
free(kse);
}
active_kse_count = 0;
/* Free the free KSEGs: */
while ((kseg = TAILQ_FIRST(&free_kse_groupq)) != NULL) {
TAILQ_REMOVE(&free_kse_groupq, kseg, kg_qe);
_lock_destroy(&kseg->kg_lock);
_pq_free(&kseg->kg_schedq.sq_runq);
free(kseg);
}
free_kseg_count = 0;
/* Free the active KSEGs: */
while ((kseg = TAILQ_FIRST(&active_kse_groupq)) != NULL) {
TAILQ_REMOVE(&active_kse_groupq, kseg, kg_qe);
_lock_destroy(&kseg->kg_lock);
_pq_free(&kseg->kg_schedq.sq_runq);
free(kseg);
}
active_kseg_count = 0;
/* Free the free threads. */
while ((thread = TAILQ_FIRST(&free_threadq)) != NULL) {
TAILQ_REMOVE(&free_threadq, thread, tle);
if (thread->specific != NULL)
free(thread->specific);
for (i = 0; i < MAX_THR_LOCKLEVEL; i++) {
_lockuser_destroy(&thread->lockusers[i]);
}
_lock_destroy(&thread->lock);
free(thread);
}
free_thread_count = 0;
/* Free the to-be-gc'd threads. */
while ((thread = TAILQ_FIRST(&_thread_gc_list)) != NULL) {
TAILQ_REMOVE(&_thread_gc_list, thread, gcle);
for (i = 0; i < MAX_THR_LOCKLEVEL; i++) {
_lockuser_destroy(&thread->lockusers[i]);
}
_lock_destroy(&thread->lock);
free(thread);
}
TAILQ_INIT(&gc_ksegq);
_gc_count = 0;
if (inited != 0) {
/*
* Destroy these locks; they'll be recreated to assure they
* are in the unlocked state.
*/
_lock_destroy(&kse_lock);
_lock_destroy(&thread_lock);
_lock_destroy(&_thread_list_lock);
inited = 0;
}
/*
* After a fork(), the leftover thread goes back to being
* scope process.
*/
curthread->attr.flags &= ~PTHREAD_SCOPE_SYSTEM;
curthread->attr.flags |= PTHREAD_SCOPE_PROCESS;
/*
* After a fork, we are still operating on the thread's original
* stack. Don't clear the THR_FLAGS_USER from the thread's
* attribute flags.
*/
/* Initialize the threads library. */
curthread->kse = NULL;
curthread->kseg = NULL;
_kse_initial = NULL;
_libpthread_init(curthread);
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
#else
if (__isthreaded) {
_thr_rtld_fini();
_thr_signal_deinit();
}
__isthreaded = 0;
/*
* Restore signal mask early, so any memory problems could
* dump core.
*/
sigprocmask(SIG_SETMASK, &curthread->sigmask, NULL);
curthread->kse->k_kcb->kcb_kmbx.km_curthread = NULL;
active_threads = 1;
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
#endif
}
/*
* This is used to initialize housekeeping and to initialize the
* KSD for the KSE.
*/
void
_kse_init(void)
{
if (inited == 0) {
TAILQ_INIT(&active_kseq);
TAILQ_INIT(&active_kse_groupq);
TAILQ_INIT(&free_kseq);
TAILQ_INIT(&free_kse_groupq);
TAILQ_INIT(&free_threadq);
TAILQ_INIT(&gc_ksegq);
if (_lock_init(&kse_lock, LCK_ADAPTIVE,
_kse_lock_wait, _kse_lock_wakeup) != 0)
PANIC("Unable to initialize free KSE queue lock");
if (_lock_init(&thread_lock, LCK_ADAPTIVE,
_kse_lock_wait, _kse_lock_wakeup) != 0)
PANIC("Unable to initialize free thread queue lock");
if (_lock_init(&_thread_list_lock, LCK_ADAPTIVE,
_kse_lock_wait, _kse_lock_wakeup) != 0)
PANIC("Unable to initialize thread list lock");
active_kse_count = 0;
active_kseg_count = 0;
_gc_count = 0;
inited = 1;
2002-10-30 06:07:18 +00:00
}
}
int
_kse_isthreaded(void)
{
return (__isthreaded != 0);
}
/*
* This is called when the first thread (other than the initial
* thread) is created.
*/
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
int
_kse_setthreaded(int threaded)
{
sigset_t sigset;
if ((threaded != 0) && (__isthreaded == 0)) {
SIGFILLSET(sigset);
__sys_sigprocmask(SIG_SETMASK, &sigset, &_thr_initial->sigmask);
/*
* Tell the kernel to create a KSE for the initial thread
* and enable upcalls in it.
*/
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
_kse_initial->k_flags |= KF_STARTED;
#ifdef SYSTEM_SCOPE_ONLY
/*
* For bound thread, kernel reads mailbox pointer once,
* we'd set it here before calling kse_create
*/
_tcb_set(_kse_initial->k_kcb, _thr_initial->tcb);
KSE_SET_MBOX(_kse_initial, _thr_initial);
_kse_initial->k_kcb->kcb_kmbx.km_flags |= KMF_BOUND;
#else
_thr_initial->attr.flags &= ~PTHREAD_SCOPE_SYSTEM;
_kse_initial->k_kseg->kg_flags &= ~KGF_SINGLE_THREAD;
_kse_initial->k_kcb->kcb_kmbx.km_curthread = NULL;
#endif
/*
* Locking functions in libc are required when there are
* threads other than the initial thread.
*/
_thr_rtld_init();
__isthreaded = 1;
if (kse_create(&_kse_initial->k_kcb->kcb_kmbx, 0) != 0) {
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
_kse_initial->k_flags &= ~KF_STARTED;
__isthreaded = 0;
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
PANIC("kse_create() failed\n");
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
return (-1);
}
#ifndef SYSTEM_SCOPE_ONLY
/* Set current thread to initial thread */
_tcb_set(_kse_initial->k_kcb, _thr_initial->tcb);
KSE_SET_MBOX(_kse_initial, _thr_initial);
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
_thr_start_sig_daemon();
_thr_setmaxconcurrency();
#else
__sys_sigprocmask(SIG_SETMASK, &_thr_initial->sigmask, NULL);
#endif
}
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
return (0);
}
/*
* Lock wait and wakeup handlers for KSE locks. These are only used by
* KSEs, and should never be used by threads. KSE locks include the
* KSE group lock (used for locking the scheduling queue) and the
* kse_lock defined above.
*
* When a KSE lock attempt blocks, the entire KSE blocks allowing another
* KSE to run. For the most part, it doesn't make much sense to try and
* schedule another thread because you need to lock the scheduling queue
* in order to do that. And since the KSE lock is used to lock the scheduling
* queue, you would just end up blocking again.
*/
void
_kse_lock_wait(struct lock *lock, struct lockuser *lu)
{
struct kse *curkse = (struct kse *)_LCK_GET_PRIVATE(lu);
struct timespec ts;
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
int saved_flags;
2002-10-30 06:07:18 +00:00
if (curkse->k_kcb->kcb_kmbx.km_curthread != NULL)
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
PANIC("kse_lock_wait does not disable upcall.\n");
2002-10-30 06:07:18 +00:00
/*
* Enter a loop to wait until we get the lock.
2002-10-30 06:07:18 +00:00
*/
ts.tv_sec = 0;
ts.tv_nsec = 1000000; /* 1 sec */
while (!_LCK_GRANTED(lu)) {
/*
* Yield the kse and wait to be notified when the lock
* is granted.
*/
saved_flags = curkse->k_kcb->kcb_kmbx.km_flags;
curkse->k_kcb->kcb_kmbx.km_flags |= KMF_NOUPCALL |
KMF_NOCOMPLETED;
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
kse_release(&ts);
curkse->k_kcb->kcb_kmbx.km_flags = saved_flags;
}
}
void
_kse_lock_wakeup(struct lock *lock, struct lockuser *lu)
{
struct kse *curkse;
struct kse *kse;
struct kse_mailbox *mbx;
curkse = _get_curkse();
kse = (struct kse *)_LCK_GET_PRIVATE(lu);
if (kse == curkse)
PANIC("KSE trying to wake itself up in lock");
else {
mbx = &kse->k_kcb->kcb_kmbx;
_lock_grant(lock, lu);
2002-10-30 06:07:18 +00:00
/*
* Notify the owning kse that it has the lock.
* It is safe to pass invalid address to kse_wakeup
* even if the mailbox is not in kernel at all,
* and waking up a wrong kse is also harmless.
2002-10-30 06:07:18 +00:00
*/
kse_wakeup(mbx);
}
}
/*
* Thread wait and wakeup handlers for thread locks. These are only used
* by threads, never by KSEs. Thread locks include the per-thread lock
* (defined in its structure), and condition variable and mutex locks.
*/
void
_thr_lock_wait(struct lock *lock, struct lockuser *lu)
{
struct pthread *curthread = (struct pthread *)lu->lu_private;
do {
THR_LOCK_SWITCH(curthread);
THR_SET_STATE(curthread, PS_LOCKWAIT);
_thr_sched_switch_unlocked(curthread);
} while (!_LCK_GRANTED(lu));
}
2002-10-30 06:07:18 +00:00
void
_thr_lock_wakeup(struct lock *lock, struct lockuser *lu)
{
struct pthread *thread;
struct pthread *curthread;
struct kse_mailbox *kmbx;
curthread = _get_curthread();
thread = (struct pthread *)_LCK_GET_PRIVATE(lu);
THR_SCHED_LOCK(curthread, thread);
_lock_grant(lock, lu);
kmbx = _thr_setrunnable_unlocked(thread);
THR_SCHED_UNLOCK(curthread, thread);
if (kmbx != NULL)
kse_wakeup(kmbx);
}
kse_critical_t
_kse_critical_enter(void)
{
kse_critical_t crit;
crit = (kse_critical_t)_kcb_critical_enter();
return (crit);
}
void
_kse_critical_leave(kse_critical_t crit)
{
struct pthread *curthread;
_kcb_critical_leave((struct kse_thr_mailbox *)crit);
if ((crit != NULL) && ((curthread = _get_curthread()) != NULL))
THR_YIELD_CHECK(curthread);
}
int
_kse_in_critical(void)
{
return (_kcb_in_critical());
}
void
_thr_critical_enter(struct pthread *thread)
{
thread->critical_count++;
}
void
_thr_critical_leave(struct pthread *thread)
{
thread->critical_count--;
THR_YIELD_CHECK(thread);
}
void
_thr_sched_switch(struct pthread *curthread)
{
struct kse *curkse;
(void)_kse_critical_enter();
curkse = _get_curkse();
KSE_SCHED_LOCK(curkse, curkse->k_kseg);
_thr_sched_switch_unlocked(curthread);
}
/*
* XXX - We may need to take the scheduling lock before calling
* this, or perhaps take the lock within here before
* doing anything else.
*/
void
_thr_sched_switch_unlocked(struct pthread *curthread)
{
struct pthread *td;
struct pthread_sigframe psf;
struct kse *curkse;
int ret;
volatile int uts_once;
volatile int resume_once = 0;
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
ucontext_t uc;
/* We're in the scheduler, 5 by 5: */
curkse = _get_curkse();
curthread->need_switchout = 1; /* The thread yielded on its own. */
curthread->critical_yield = 0; /* No need to yield anymore. */
thr_accounting(curthread);
/* Thread can unlock the scheduler lock. */
curthread->lock_switch = 1;
/*
* The signal frame is allocated off the stack because
* a thread can be interrupted by other signals while
* it is running down pending signals.
*/
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
psf.psf_valid = 0;
curthread->curframe = &psf;
/*
* Enter the scheduler if any one of the following is true:
*
* o The current thread is dead; it's stack needs to be
* cleaned up and it can't be done while operating on
* it.
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
* o The current thread has signals pending, should
* let scheduler install signal trampoline for us.
* o There are no runnable threads.
* o The next thread to run won't unlock the scheduler
* lock. A side note: the current thread may be run
* instead of the next thread in the run queue, but
* we don't bother checking for that.
*/
if (curthread->attr.flags & PTHREAD_SCOPE_SYSTEM)
kse_sched_single(&curkse->k_kcb->kcb_kmbx);
else if ((curthread->state == PS_DEAD) ||
(((td = KSE_RUNQ_FIRST(curkse)) == NULL) &&
(curthread->state != PS_RUNNING)) ||
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
((td != NULL) && (td->lock_switch == 0))) {
curkse->k_switch = 1;
_thread_enter_uts(curthread->tcb, curkse->k_kcb);
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
}
else {
uts_once = 0;
THR_GETCONTEXT(&curthread->tcb->tcb_tmbx.tm_context);
if (uts_once == 0) {
uts_once = 1;
/* Switchout the current thread. */
kse_switchout_thread(curkse, curthread);
_tcb_set(curkse->k_kcb, NULL);
/* Choose another thread to run. */
td = KSE_RUNQ_FIRST(curkse);
KSE_RUNQ_REMOVE(curkse, td);
curkse->k_curthread = td;
/*
* Make sure the current thread's kse points to
* this kse.
*/
td->kse = curkse;
/*
* Reset the time slice if this thread is running
* for the first time or running again after using
* its full time slice allocation.
*/
if (td->slice_usec == -1)
td->slice_usec = 0;
/* Mark the thread active. */
td->active = 1;
/* Remove the frame reference. */
td->curframe = NULL;
/*
* Continue the thread at its current frame.
* Note: TCB is set in _thread_switch
*/
ret = _thread_switch(curkse->k_kcb, td->tcb, 0);
/* This point should not be reached. */
if (ret != 0)
PANIC("Bad return from _thread_switch");
PANIC("Thread has returned from _thread_switch");
}
}
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
if (psf.psf_valid) {
/*
* It is ugly we must increase critical count, because we
* have a frame saved, we must backout state in psf
* before we can process signals.
*/
curthread->critical_count++;
}
if (curthread->lock_switch != 0) {
/*
* Unlock the scheduling queue and leave the
* critical region.
*/
/* Don't trust this after a switch! */
curkse = _get_curkse();
curthread->lock_switch = 0;
KSE_SCHED_UNLOCK(curkse, curkse->k_kseg);
_kse_critical_leave(&curthread->tcb->tcb_tmbx);
}
/*
* This thread is being resumed; check for cancellations.
*/
if ((psf.psf_valid ||
(curthread->check_pending && !THR_IN_CRITICAL(curthread)))) {
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
resume_once = 0;
THR_GETCONTEXT(&uc);
if (resume_once == 0) {
resume_once = 1;
curthread->check_pending = 0;
thr_resume_check(curthread, &uc, &psf);
}
}
THR_ACTIVATE_LAST_LOCK(curthread);
}
/*
* This is the scheduler for a KSE which runs a scope system thread.
* The multi-thread KSE scheduler should also work for a single threaded
* KSE, but we use a separate scheduler so that it can be fine-tuned
* to be more efficient (and perhaps not need a separate stack for
* the KSE, allowing it to use the thread's stack).
*/
static void
kse_sched_single(struct kse_mailbox *kmbx)
{
struct kse *curkse;
struct pthread *curthread;
struct timespec ts;
sigset_t sigmask;
int i, sigseqno, level, first = 0;
curkse = (struct kse *)kmbx->km_udata;
curthread = curkse->k_curthread;
if ((curkse->k_flags & KF_INITIALIZED) == 0) {
/* Setup this KSEs specific data. */
_kcb_set(curkse->k_kcb);
_tcb_set(curkse->k_kcb, curthread->tcb);
curkse->k_flags |= KF_INITIALIZED;
first = 1;
curthread->active = 1;
/* Setup kernel signal masks for new thread. */
__sys_sigprocmask(SIG_SETMASK, &curthread->sigmask, NULL);
/*
* Enter critical region, this is meanless for bound thread,
* It is used to let other code work, those code want mailbox
* to be cleared.
*/
(void)_kse_critical_enter();
} else {
/*
* Bound thread always has tcb set, this prevent some
* code from blindly setting bound thread tcb to NULL,
* buggy code ?
*/
_tcb_set(curkse->k_kcb, curthread->tcb);
}
2002-10-30 06:07:18 +00:00
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
curthread->critical_yield = 0;
curthread->need_switchout = 0;
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
/*
* Lock the scheduling queue.
*
* There is no scheduling queue for single threaded KSEs,
* but we need a lock for protection regardless.
*/
if (curthread->lock_switch == 0)
KSE_SCHED_LOCK(curkse, curkse->k_kseg);
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
/*
* This has to do the job of kse_switchout_thread(), only
* for a single threaded KSE/KSEG.
*/
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
switch (curthread->state) {
case PS_DEAD:
curthread->check_pending = 0;
/* Unlock the scheduling queue and exit the KSE and thread. */
thr_cleanup(curkse, curthread);
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
KSE_SCHED_UNLOCK(curkse, curkse->k_kseg);
PANIC("bound thread shouldn't get here\n");
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
break;
case PS_SIGWAIT:
PANIC("bound thread does not have SIGWAIT state\n");
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
case PS_SLEEP_WAIT:
PANIC("bound thread does not have SLEEP_WAIT state\n");
case PS_SIGSUSPEND:
PANIC("bound thread does not have SIGSUSPEND state\n");
case PS_COND_WAIT:
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
break;
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
case PS_LOCKWAIT:
/*
* This state doesn't timeout.
*/
curthread->wakeup_time.tv_sec = -1;
curthread->wakeup_time.tv_nsec = -1;
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
level = curthread->locklevel - 1;
if (_LCK_GRANTED(&curthread->lockusers[level]))
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
THR_SET_STATE(curthread, PS_RUNNING);
break;
case PS_RUNNING:
if ((curthread->flags & THR_FLAGS_SUSPENDED) != 0) {
THR_SET_STATE(curthread, PS_SUSPENDED);
}
curthread->wakeup_time.tv_sec = -1;
curthread->wakeup_time.tv_nsec = -1;
break;
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
case PS_JOIN:
case PS_MUTEX_WAIT:
case PS_SUSPENDED:
case PS_DEADLOCK:
default:
/*
* These states don't timeout and don't need
* to be in the waiting queue.
*/
curthread->wakeup_time.tv_sec = -1;
curthread->wakeup_time.tv_nsec = -1;
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
break;
}
while (curthread->state != PS_RUNNING) {
sigseqno = curkse->k_sigseqno;
if (curthread->check_pending != 0) {
/*
* Install pending signals into the frame, possible
* cause mutex or condvar backout.
*/
curthread->check_pending = 0;
SIGFILLSET(sigmask);
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
/*
* Lock out kernel signal code when we are processing
* signals, and get a fresh copy of signal mask.
*/
__sys_sigprocmask(SIG_SETMASK, &sigmask,
&curthread->sigmask);
for (i = 1; i <= _SIG_MAXSIG; i++) {
if (SIGISMEMBER(curthread->sigmask, i))
continue;
if (SIGISMEMBER(curthread->sigpend, i))
(void)_thr_sig_add(curthread, i,
&curthread->siginfo[i-1]);
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
}
__sys_sigprocmask(SIG_SETMASK, &curthread->sigmask,
NULL);
/* The above code might make thread runnable */
if (curthread->state == PS_RUNNING)
break;
}
THR_DEACTIVATE_LAST_LOCK(curthread);
kse_wait(curkse, curthread, sigseqno);
THR_ACTIVATE_LAST_LOCK(curthread);
KSE_GET_TOD(curkse, &ts);
if (thr_timedout(curthread, &ts)) {
/* Indicate the thread timedout: */
curthread->timeout = 1;
/* Make the thread runnable. */
THR_SET_STATE(curthread, PS_RUNNING);
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
}
}
/* Remove the frame reference. */
curthread->curframe = NULL;
if (curthread->lock_switch == 0) {
/* Unlock the scheduling queue. */
KSE_SCHED_UNLOCK(curkse, curkse->k_kseg);
}
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
DBG_MSG("Continuing bound thread %p\n", curthread);
if (first) {
_kse_critical_leave(&curthread->tcb->tcb_tmbx);
pthread_exit(curthread->start_routine(curthread->arg));
}
}
#ifdef DEBUG_THREAD_KERN
static void
dump_queues(struct kse *curkse)
{
struct pthread *thread;
DBG_MSG("Threads in waiting queue:\n");
TAILQ_FOREACH(thread, &curkse->k_kseg->kg_schedq.sq_waitq, pqe) {
DBG_MSG(" thread %p, state %d, blocked %d\n",
thread, thread->state, thread->blocked);
}
}
#endif
/*
* This is the scheduler for a KSE which runs multiple threads.
*/
static void
kse_sched_multi(struct kse_mailbox *kmbx)
{
struct kse *curkse;
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
struct pthread *curthread, *td_wait;
struct pthread_sigframe *curframe;
int ret;
curkse = (struct kse *)kmbx->km_udata;
THR_ASSERT(curkse->k_kcb->kcb_kmbx.km_curthread == NULL,
"Mailbox not null in kse_sched_multi");
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
/* Check for first time initialization: */
if ((curkse->k_flags & KF_INITIALIZED) == 0) {
/* Setup this KSEs specific data. */
_kcb_set(curkse->k_kcb);
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
/* Set this before grabbing the context. */
curkse->k_flags |= KF_INITIALIZED;
}
/*
* No current thread anymore, calling _get_curthread in UTS
* should dump core
*/
_tcb_set(curkse->k_kcb, NULL);
/* If this is an upcall; take the scheduler lock. */
if (curkse->k_switch == 0)
KSE_SCHED_LOCK(curkse, curkse->k_kseg);
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
curkse->k_switch = 0;
/*
* Now that the scheduler lock is held, get the current
* thread. The KSE's current thread cannot be safely
* examined without the lock because it could have returned
* as completed on another KSE. See kse_check_completed().
*/
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
curthread = curkse->k_curthread;
2002-10-30 06:07:18 +00:00
if (KSE_IS_IDLE(curkse)) {
KSE_CLEAR_IDLE(curkse);
curkse->k_kseg->kg_idle_kses--;
}
/*
* If the current thread was completed in another KSE, then
* it will be in the run queue. Don't mark it as being blocked.
*/
if ((curthread != NULL) &&
((curthread->flags & THR_FLAGS_IN_RUNQ) == 0) &&
(curthread->need_switchout == 0)) {
/*
* Assume the current thread is blocked; when the
* completed threads are checked and if the current
* thread is among the completed, the blocked flag
* will be cleared.
*/
curthread->blocked = 1;
}
/* Check for any unblocked threads in the kernel. */
kse_check_completed(curkse);
2002-10-30 06:07:18 +00:00
/*
* Check for threads that have timed-out.
*/
kse_check_waitq(curkse);
2002-10-30 06:07:18 +00:00
/*
* Switchout the current thread, if necessary, as the last step
* so that it is inserted into the run queue (if it's runnable)
* _after_ any other threads that were added to it above.
*/
if (curthread == NULL)
; /* Nothing to do here. */
else if ((curthread->need_switchout == 0) &&
(curthread->blocked == 0) && (THR_IN_CRITICAL(curthread))) {
/*
* Resume the thread and tell it to yield when
* it leaves the critical region.
*/
curthread->critical_yield = 1;
curthread->active = 1;
if ((curthread->flags & THR_FLAGS_IN_RUNQ) != 0)
KSE_RUNQ_REMOVE(curkse, curthread);
curkse->k_curthread = curthread;
curthread->kse = curkse;
DBG_MSG("Continuing thread %p in critical region\n",
curthread);
kse_wakeup_multi(curkse);
KSE_SCHED_UNLOCK(curkse, curkse->k_kseg);
ret = _thread_switch(curkse->k_kcb, curthread->tcb, 1);
if (ret != 0)
PANIC("Can't resume thread in critical region\n");
2002-10-30 06:07:18 +00:00
}
else if ((curthread->flags & THR_FLAGS_IN_RUNQ) == 0)
kse_switchout_thread(curkse, curthread);
curkse->k_curthread = NULL;
#ifdef DEBUG_THREAD_KERN
dump_queues(curkse);
#endif
/* Check if there are no threads ready to run: */
while (((curthread = KSE_RUNQ_FIRST(curkse)) == NULL) &&
(curkse->k_kseg->kg_threadcount != 0) &&
((curkse->k_flags & KF_TERMINATED) == 0)) {
/*
* Wait for a thread to become active or until there are
* no more threads.
*/
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
td_wait = KSE_WAITQ_FIRST(curkse);
kse_wait(curkse, td_wait, 0);
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
kse_check_completed(curkse);
kse_check_waitq(curkse);
}
/* Check for no more threads: */
if ((curkse->k_kseg->kg_threadcount == 0) ||
((curkse->k_flags & KF_TERMINATED) != 0)) {
/*
* Normally this shouldn't return, but it will if there
* are other KSEs running that create new threads that
* are assigned to this KSE[G]. For instance, if a scope
* system thread were to create a scope process thread
* and this kse[g] is the initial kse[g], then that newly
* created thread would be assigned to us (the initial
* kse[g]).
*/
kse_wakeup_multi(curkse);
KSE_SCHED_UNLOCK(curkse, curkse->k_kseg);
kse_fini(curkse);
/* never returns */
}
THR_ASSERT(curthread != NULL,
"Return from kse_wait/fini without thread.");
THR_ASSERT(curthread->state != PS_DEAD,
"Trying to resume dead thread!");
KSE_RUNQ_REMOVE(curkse, curthread);
Implement zero system call thread switching. Performance of thread switches should be on par with that under scheduler activations. o Timing is achieved through the use of a fixed interval timer (ITIMER_PROF) to count scheduling ticks instead of retrieving the time-of-day upon every thread switch and calculating elapsed real time. o Polling for I/O readiness is performed once for each scheduling tick instead of every thread switch. o The non-signal saving/restoring versions of setjmp/longjmp are used to save and restore thread contexts. This may allow the removal of _THREAD_SAFE macros from setjmp() and longjmp() - needs more investigation. Change signal handling so that signals are handled in the context of the thread that is receiving the signal. When signals are dispatched to a thread, a special signal handling frame is created on top of the target threads stack. The frame contains the threads saved state information and a new context in which the thread can run. The applications signal handler is invoked through a wrapper routine that knows how to restore the threads saved state and unwind to previous frames. Fix interruption of threads due to signals. Some states were being improperly interrupted while other states were not being interrupted. This should fix several PRs. Signal handlers, which are invoked as a result of a process signal (not by pthread_kill()), are now called with the code (or siginfo_t if SA_SIGINFO was set in sa_flags) and sigcontext_t as received from the process signal handler. Modify the search for a thread to which a signal is delivered. The search algorithm is now: o First thread found in sigwait() with signal in wait mask. o First thread found sigsuspend()'d on the signal. o Current thread if signal is unmasked. o First thread found with signal unmasked. Collapse machine dependent support into macros defined in pthread_private.h. These should probably eventually be moved into separate MD files. Change the range of settable priorities to be compliant with POSIX (0-31). The threads library uses higher priorities internally for real-time threads (not yet implemented) and threads executing signal handlers. Real-time threads and threads running signal handlers add 64 and 32, respectively, to a threads base priority. Some other small changes and cleanups. PR: 17757 18559 21943 Reviewed by: jasone
2000-10-13 22:12:32 +00:00
/*
* Make the selected thread the current thread.
*/
curkse->k_curthread = curthread;
/*
* Make sure the current thread's kse points to this kse.
*/
curthread->kse = curkse;
/*
* Reset the time slice if this thread is running for the first
* time or running again after using its full time slice allocation.
*/
if (curthread->slice_usec == -1)
curthread->slice_usec = 0;
/* Mark the thread active. */
curthread->active = 1;
/* Remove the frame reference. */
curframe = curthread->curframe;
curthread->curframe = NULL;
/*
* The thread's current signal frame will only be NULL if it
* is being resumed after being blocked in the kernel. In
* this case, and if the thread needs to run down pending
* signals or needs a cancellation check, we need to add a
* signal frame to the thread's context.
*/
#ifdef NOT_YET
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
if ((((curframe == NULL) && (curthread->check_pending != 0)) ||
(((curthread->cancelflags & THR_AT_CANCEL_POINT) == 0) &&
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
((curthread->cancelflags & PTHREAD_CANCEL_ASYNCHRONOUS) != 0))) &&
!THR_IN_CRITICAL(curthread))
signalcontext(&curthread->tcb->tcb_tmbx.tm_context, 0,
(__sighandler_t *)thr_resume_wrapper);
#else
if ((curframe == NULL) && (curthread->state == PS_RUNNING) &&
(curthread->check_pending != 0) && !THR_IN_CRITICAL(curthread)) {
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
curthread->check_pending = 0;
signalcontext(&curthread->tcb->tcb_tmbx.tm_context, 0,
(__sighandler_t *)thr_resume_wrapper);
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
}
#endif
kse_wakeup_multi(curkse);
/*
* Continue the thread at its current frame:
*/
if (curthread->lock_switch != 0) {
/*
* This thread came from a scheduler switch; it will
* unlock the scheduler lock and set the mailbox.
*/
ret = _thread_switch(curkse->k_kcb, curthread->tcb, 0);
} else {
/* This thread won't unlock the scheduler lock. */
KSE_SCHED_UNLOCK(curkse, curkse->k_kseg);
ret = _thread_switch(curkse->k_kcb, curthread->tcb, 1);
}
if (ret != 0)
PANIC("Thread has returned from _thread_switch");
/* This point should not be reached. */
PANIC("Thread has returned from _thread_switch");
}
static void
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
thr_resume_wrapper(int sig, siginfo_t *siginfo, ucontext_t *ucp)
Implement zero system call thread switching. Performance of thread switches should be on par with that under scheduler activations. o Timing is achieved through the use of a fixed interval timer (ITIMER_PROF) to count scheduling ticks instead of retrieving the time-of-day upon every thread switch and calculating elapsed real time. o Polling for I/O readiness is performed once for each scheduling tick instead of every thread switch. o The non-signal saving/restoring versions of setjmp/longjmp are used to save and restore thread contexts. This may allow the removal of _THREAD_SAFE macros from setjmp() and longjmp() - needs more investigation. Change signal handling so that signals are handled in the context of the thread that is receiving the signal. When signals are dispatched to a thread, a special signal handling frame is created on top of the target threads stack. The frame contains the threads saved state information and a new context in which the thread can run. The applications signal handler is invoked through a wrapper routine that knows how to restore the threads saved state and unwind to previous frames. Fix interruption of threads due to signals. Some states were being improperly interrupted while other states were not being interrupted. This should fix several PRs. Signal handlers, which are invoked as a result of a process signal (not by pthread_kill()), are now called with the code (or siginfo_t if SA_SIGINFO was set in sa_flags) and sigcontext_t as received from the process signal handler. Modify the search for a thread to which a signal is delivered. The search algorithm is now: o First thread found in sigwait() with signal in wait mask. o First thread found sigsuspend()'d on the signal. o Current thread if signal is unmasked. o First thread found with signal unmasked. Collapse machine dependent support into macros defined in pthread_private.h. These should probably eventually be moved into separate MD files. Change the range of settable priorities to be compliant with POSIX (0-31). The threads library uses higher priorities internally for real-time threads (not yet implemented) and threads executing signal handlers. Real-time threads and threads running signal handlers add 64 and 32, respectively, to a threads base priority. Some other small changes and cleanups. PR: 17757 18559 21943 Reviewed by: jasone
2000-10-13 22:12:32 +00:00
{
struct pthread *curthread = _get_curthread();
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
struct kse *curkse;
int ret, err_save = errno;
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
DBG_MSG(">>> sig wrapper\n");
if (curthread->lock_switch)
PANIC("thr_resume_wrapper, lock_switch != 0\n");
thr_resume_check(curthread, ucp, NULL);
errno = err_save;
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
_kse_critical_enter();
curkse = _get_curkse();
curthread->tcb->tcb_tmbx.tm_context = *ucp;
ret = _thread_switch(curkse->k_kcb, curthread->tcb, 1);
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
if (ret != 0)
PANIC("thr_resume_wrapper: thread has returned "
"from _thread_switch");
/* THR_SETCONTEXT(ucp); */ /* not work, why ? */
}
static void
thr_resume_check(struct pthread *curthread, ucontext_t *ucp,
struct pthread_sigframe *psf)
{
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
_thr_sig_rundown(curthread, ucp, psf);
#ifdef NOT_YET
if (((curthread->cancelflags & THR_AT_CANCEL_POINT) == 0) &&
((curthread->cancelflags & PTHREAD_CANCEL_ASYNCHRONOUS) != 0))
pthread_testcancel();
#endif
}
/*
* Clean up a thread. This must be called with the thread's KSE
* scheduling lock held. The thread must be a thread from the
* KSE's group.
*/
static void
thr_cleanup(struct kse *curkse, struct pthread *thread)
{
struct pthread *joiner;
struct kse_mailbox *kmbx = NULL;
int sys_scope;
if ((joiner = thread->joiner) != NULL) {
/* Joinee scheduler lock held; joiner won't leave. */
if (joiner->kseg == curkse->k_kseg) {
if (joiner->join_status.thread == thread) {
joiner->join_status.thread = NULL;
joiner->join_status.ret = thread->ret;
(void)_thr_setrunnable_unlocked(joiner);
}
} else {
KSE_SCHED_UNLOCK(curkse, curkse->k_kseg);
/* The joiner may have removed itself and exited. */
if (_thr_ref_add(thread, joiner, 0) == 0) {
KSE_SCHED_LOCK(curkse, joiner->kseg);
if (joiner->join_status.thread == thread) {
joiner->join_status.thread = NULL;
joiner->join_status.ret = thread->ret;
kmbx = _thr_setrunnable_unlocked(joiner);
}
KSE_SCHED_UNLOCK(curkse, joiner->kseg);
_thr_ref_delete(thread, joiner);
if (kmbx != NULL)
kse_wakeup(kmbx);
}
KSE_SCHED_LOCK(curkse, curkse->k_kseg);
}
thread->attr.flags |= PTHREAD_DETACHED;
}
if (!(sys_scope = (thread->attr.flags & PTHREAD_SCOPE_SYSTEM))) {
/*
* Remove the thread from the KSEG's list of threads.
*/
KSEG_THRQ_REMOVE(thread->kseg, thread);
/*
* Migrate the thread to the main KSE so that this
* KSE and KSEG can be cleaned when their last thread
* exits.
*/
thread->kseg = _kse_initial->k_kseg;
thread->kse = _kse_initial;
}
thread->flags |= THR_FLAGS_GC_SAFE;
/*
* We can't hold the thread list lock while holding the
* scheduler lock.
*/
KSE_SCHED_UNLOCK(curkse, curkse->k_kseg);
DBG_MSG("Adding thread %p to GC list\n", thread);
KSE_LOCK_ACQUIRE(curkse, &_thread_list_lock);
/* Use thread_list_lock */
active_threads--;
#ifdef SYSTEM_SCOPE_ONLY
if (active_threads == 0) {
#else
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
if (active_threads == 1) {
#endif
KSE_LOCK_RELEASE(curkse, &_thread_list_lock);
/* Possible use a signalcontext wrapper to call exit ? */
curkse->k_curthread = thread;
_tcb_set(curkse->k_kcb, thread->tcb);
exit(0);
}
THR_GCLIST_ADD(thread);
KSE_LOCK_RELEASE(curkse, &_thread_list_lock);
if (sys_scope) {
/*
* System scope thread is single thread group,
* when thread is exited, its kse and ksegrp should
* be recycled as well.
* kse upcall stack belongs to thread, clear it here.
*/
curkse->k_stack.ss_sp = 0;
curkse->k_stack.ss_size = 0;
kse_exit();
PANIC("kse_exit() failed for system scope thread");
}
KSE_SCHED_LOCK(curkse, curkse->k_kseg);
}
void
_thr_gc(struct pthread *curthread)
{
thread_gc(curthread);
kse_gc(curthread);
kseg_gc(curthread);
}
static void
thread_gc(struct pthread *curthread)
{
struct pthread *td, *td_next;
kse_critical_t crit;
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
TAILQ_HEAD(, pthread) worklist;
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
TAILQ_INIT(&worklist);
crit = _kse_critical_enter();
KSE_LOCK_ACQUIRE(curthread->kse, &_thread_list_lock);
/* Check the threads waiting for GC. */
for (td = TAILQ_FIRST(&_thread_gc_list); td != NULL; td = td_next) {
td_next = TAILQ_NEXT(td, gcle);
if ((td->flags & THR_FLAGS_GC_SAFE) == 0)
continue;
else if (((td->attr.flags & PTHREAD_SCOPE_SYSTEM) != 0) &&
((td->kse->k_kcb->kcb_kmbx.km_flags & KMF_DONE) == 0)) {
/*
* The thread and KSE are operating on the same
* stack. Wait for the KSE to exit before freeing
* the thread's stack as well as everything else.
*/
continue;
}
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
/*
* Remove the thread from the GC list. If the thread
* isn't yet detached, it will get added back to the
* GC list at a later time.
*/
THR_GCLIST_REMOVE(td);
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
DBG_MSG("Freeing thread %p stack\n", td);
/*
* We can free the thread stack since it's no longer
* in use.
*/
_thr_stack_free(&td->attr);
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
if (((td->attr.flags & PTHREAD_DETACHED) != 0) &&
(td->refcount == 0)) {
/*
* The thread has detached and is no longer
* referenced. It is safe to remove all
* remnants of the thread.
*/
THR_LIST_REMOVE(td);
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
TAILQ_INSERT_HEAD(&worklist, td, gcle);
}
}
KSE_LOCK_RELEASE(curthread->kse, &_thread_list_lock);
_kse_critical_leave(crit);
while ((td = TAILQ_FIRST(&worklist)) != NULL) {
TAILQ_REMOVE(&worklist, td, gcle);
/*
* XXX we don't free initial thread and its kse
* (if thread is a bound thread), because there might
* have some code referencing initial thread and kse.
*/
if (td == _thr_initial) {
DBG_MSG("Initial thread won't be freed\n");
continue;
}
if ((td->attr.flags & PTHREAD_SCOPE_SYSTEM) != 0) {
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
crit = _kse_critical_enter();
KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock);
kse_free_unlocked(td->kse);
kseg_free_unlocked(td->kseg);
KSE_LOCK_RELEASE(curthread->kse, &kse_lock);
_kse_critical_leave(crit);
}
DBG_MSG("Freeing thread %p\n", td);
_thr_free(curthread, td);
}
}
static void
kse_gc(struct pthread *curthread)
{
kse_critical_t crit;
TAILQ_HEAD(, kse) worklist;
struct kse *kse;
if (free_kse_count <= MAX_CACHED_KSES)
return;
TAILQ_INIT(&worklist);
crit = _kse_critical_enter();
KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock);
while (free_kse_count > MAX_CACHED_KSES) {
kse = TAILQ_FIRST(&free_kseq);
TAILQ_REMOVE(&free_kseq, kse, k_qe);
TAILQ_INSERT_HEAD(&worklist, kse, k_qe);
free_kse_count--;
}
KSE_LOCK_RELEASE(curthread->kse, &kse_lock);
_kse_critical_leave(crit);
while ((kse = TAILQ_FIRST(&worklist))) {
TAILQ_REMOVE(&worklist, kse, k_qe);
kse_destroy(kse);
}
}
static void
kseg_gc(struct pthread *curthread)
{
kse_critical_t crit;
TAILQ_HEAD(, kse_group) worklist;
struct kse_group *kseg;
if (free_kseg_count <= MAX_CACHED_KSEGS)
return;
crit = _kse_critical_enter();
KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock);
while (free_kseg_count > MAX_CACHED_KSEGS) {
kseg = TAILQ_FIRST(&free_kse_groupq);
TAILQ_REMOVE(&free_kse_groupq, kseg, kg_qe);
free_kseg_count--;
TAILQ_INSERT_HEAD(&worklist, kseg, kg_qe);
}
KSE_LOCK_RELEASE(curthread->kse, &kse_lock);
_kse_critical_leave(crit);
while ((kseg = TAILQ_FIRST(&worklist))) {
TAILQ_REMOVE(&worklist, kseg, kg_qe);
kseg_destroy(kseg);
}
}
/*
* Only new threads that are running or suspended may be scheduled.
*/
int
_thr_schedule_add(struct pthread *curthread, struct pthread *newthread)
{
kse_critical_t crit;
int ret;
/* Add the new thread. */
thr_link(newthread);
/*
* If this is the first time creating a thread, make sure
* the mailbox is set for the current thread.
*/
if ((newthread->attr.flags & PTHREAD_SCOPE_SYSTEM) != 0) {
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
/* We use the thread's stack as the KSE's stack. */
newthread->kse->k_kcb->kcb_kmbx.km_stack.ss_sp =
newthread->attr.stackaddr_attr;
newthread->kse->k_kcb->kcb_kmbx.km_stack.ss_size =
newthread->attr.stacksize_attr;
/*
* No need to lock the scheduling queue since the
* KSE/KSEG pair have not yet been started.
*/
KSEG_THRQ_ADD(newthread->kseg, newthread);
/* this thread never gives up kse */
newthread->active = 1;
newthread->kse->k_curthread = newthread;
newthread->kse->k_kcb->kcb_kmbx.km_flags = KMF_BOUND;
newthread->kse->k_kcb->kcb_kmbx.km_func =
(kse_func_t *)kse_sched_single;
newthread->kse->k_kcb->kcb_kmbx.km_quantum = 0;
KSE_SET_MBOX(newthread->kse, newthread);
/*
* This thread needs a new KSE and KSEG.
*/
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
newthread->kse->k_flags &= ~KF_INITIALIZED;
newthread->kse->k_flags |= KF_STARTED;
/* Fire up! */
ret = kse_create(&newthread->kse->k_kcb->kcb_kmbx, 1);
if (ret != 0)
ret = errno;
}
else {
/*
* Lock the KSE and add the new thread to its list of
* assigned threads. If the new thread is runnable, also
* add it to the KSE's run queue.
*/
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
crit = _kse_critical_enter();
KSE_SCHED_LOCK(curthread->kse, newthread->kseg);
KSEG_THRQ_ADD(newthread->kseg, newthread);
if (newthread->state == PS_RUNNING)
THR_RUNQ_INSERT_TAIL(newthread);
if ((newthread->kse->k_flags & KF_STARTED) == 0) {
/*
* This KSE hasn't been started yet. Start it
* outside of holding the lock.
*/
newthread->kse->k_flags |= KF_STARTED;
newthread->kse->k_kcb->kcb_kmbx.km_func =
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
(kse_func_t *)kse_sched_multi;
newthread->kse->k_kcb->kcb_kmbx.km_flags = 0;
kse_create(&newthread->kse->k_kcb->kcb_kmbx, 0);
} else if ((newthread->state == PS_RUNNING) &&
KSE_IS_IDLE(newthread->kse)) {
/*
* The thread is being scheduled on another KSEG.
*/
kse_wakeup_one(newthread);
In the words of the author: o The polling mechanism for I/O readiness was changed from select() to poll(). In additon, a wrapped version of poll() is now provided. o The wrapped select routine now converts each fd_set to a poll array so that the thread scheduler doesn't have to perform a bitwise search for selected fds each time file descriptors are polled for I/O readiness. o The thread scheduler was modified to use a new queue (_workq) for threads that need work. Threads waiting for I/O readiness and spinblocks are added to the work queue in addition to the waiting queue. This reduces the time spent forming/searching the array of file descriptors being polled. o The waiting queue (_waitingq) is now maintained in order of thread wakeup time. This allows the thread scheduler to find the nearest wakeup time by looking at the first thread in the queue instead of searching the entire queue. o Removed file descriptor locking for select/poll routines. An application should not rely on the threads library for providing this locking; if necessary, the application should use mutexes to protect selecting/polling of file descriptors. o Retrieve and use the kernel clock rate/resolution at startup instead of hardcoding the clock resolution to 10 msec (tested with kernel running at 1000 HZ). o All queues have been changed to use queue.h macros. These include the queues of all threads, dead threads, and threads waiting for file descriptor locks. o Added reinitialization of the GC mutex and condition variable after a fork. Also prevented reallocation of the ready queue after a fork. o Prevented the wrapped close routine from closing the thread kernel pipes. o Initialized file descriptor table for stdio entries at thread init. o Provided additional flags to indicate to what queues threads belong. o Moved TAILQ initialization for statically allocated mutex and condition variables to after the spinlock. o Added dispatching of signals to pthread_kill. Removing the dispatching of signals from thread activation broke sigsuspend when pthread_kill was used to send a signal to a thread. o Temporarily set the state of a thread to PS_SUSPENDED when it is first created and placed in the list of threads so that it will not be accidentally scheduled before becoming a member of one of the scheduling queues. o Change the signal handler to queue signals to the thread kernel pipe if the scheduling queues are protected. When scheduling queues are unprotected, signals are then dequeued and handled. o Ensured that all installed signal handlers block the scheduling signal and that the scheduling signal handler blocks all other signals. This ensures that the signal handler is only interruptible for and by non-scheduling signals. An atomic lock is used to decide which instance of the signal handler will handle pending signals. o Removed _lock_thread_list and _unlock_thread_list as they are no longer used to protect the thread list. o Added missing RCS IDs to modified files. o Added checks for appropriate queue membership and activity when adding, removing, and searching the scheduling queues. These checks add very little overhead and are enabled when compiled with _PTHREADS_INVARIANTS defined. Suggested and implemented by Tor Egge with some modification by me. o Close a race condition in uthread_close. (Tor Egge) o Protect the scheduling queues while modifying them in pthread_cond_signal and _thread_fd_unlock. (Tor Egge) o Ensure that when a thread gets a mutex, the mutex is on that threads list of owned mutexes. (Tor Egge) o Set the kernel-in-scheduler flag in _thread_kern_sched_state and _thread_kern_sched_state_unlock to prevent a scheduling signal from calling the scheduler again. (Tor Egge) o Don't use TAILQ_FOREACH macro while searching the waiting queue for threads in a sigwait state, because a change of state destroys the TAILQ link. It is actually safe to do so, though, because once a sigwaiting thread is found, the loop ends and the function returns. (Tor Egge) o When dispatching signals to threads, make the thread inherit the signal deferral flag of the currently running thread. (Tor Egge) Submitted by: Daniel Eischen <eischen@vigrid.com> and Tor Egge <Tor.Egge@fast.no>
1999-06-20 08:28:48 +00:00
}
KSE_SCHED_UNLOCK(curthread->kse, newthread->kseg);
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
_kse_critical_leave(crit);
ret = 0;
}
if (ret != 0)
thr_unlink(newthread);
return (ret);
}
In the words of the author: o The polling mechanism for I/O readiness was changed from select() to poll(). In additon, a wrapped version of poll() is now provided. o The wrapped select routine now converts each fd_set to a poll array so that the thread scheduler doesn't have to perform a bitwise search for selected fds each time file descriptors are polled for I/O readiness. o The thread scheduler was modified to use a new queue (_workq) for threads that need work. Threads waiting for I/O readiness and spinblocks are added to the work queue in addition to the waiting queue. This reduces the time spent forming/searching the array of file descriptors being polled. o The waiting queue (_waitingq) is now maintained in order of thread wakeup time. This allows the thread scheduler to find the nearest wakeup time by looking at the first thread in the queue instead of searching the entire queue. o Removed file descriptor locking for select/poll routines. An application should not rely on the threads library for providing this locking; if necessary, the application should use mutexes to protect selecting/polling of file descriptors. o Retrieve and use the kernel clock rate/resolution at startup instead of hardcoding the clock resolution to 10 msec (tested with kernel running at 1000 HZ). o All queues have been changed to use queue.h macros. These include the queues of all threads, dead threads, and threads waiting for file descriptor locks. o Added reinitialization of the GC mutex and condition variable after a fork. Also prevented reallocation of the ready queue after a fork. o Prevented the wrapped close routine from closing the thread kernel pipes. o Initialized file descriptor table for stdio entries at thread init. o Provided additional flags to indicate to what queues threads belong. o Moved TAILQ initialization for statically allocated mutex and condition variables to after the spinlock. o Added dispatching of signals to pthread_kill. Removing the dispatching of signals from thread activation broke sigsuspend when pthread_kill was used to send a signal to a thread. o Temporarily set the state of a thread to PS_SUSPENDED when it is first created and placed in the list of threads so that it will not be accidentally scheduled before becoming a member of one of the scheduling queues. o Change the signal handler to queue signals to the thread kernel pipe if the scheduling queues are protected. When scheduling queues are unprotected, signals are then dequeued and handled. o Ensured that all installed signal handlers block the scheduling signal and that the scheduling signal handler blocks all other signals. This ensures that the signal handler is only interruptible for and by non-scheduling signals. An atomic lock is used to decide which instance of the signal handler will handle pending signals. o Removed _lock_thread_list and _unlock_thread_list as they are no longer used to protect the thread list. o Added missing RCS IDs to modified files. o Added checks for appropriate queue membership and activity when adding, removing, and searching the scheduling queues. These checks add very little overhead and are enabled when compiled with _PTHREADS_INVARIANTS defined. Suggested and implemented by Tor Egge with some modification by me. o Close a race condition in uthread_close. (Tor Egge) o Protect the scheduling queues while modifying them in pthread_cond_signal and _thread_fd_unlock. (Tor Egge) o Ensure that when a thread gets a mutex, the mutex is on that threads list of owned mutexes. (Tor Egge) o Set the kernel-in-scheduler flag in _thread_kern_sched_state and _thread_kern_sched_state_unlock to prevent a scheduling signal from calling the scheduler again. (Tor Egge) o Don't use TAILQ_FOREACH macro while searching the waiting queue for threads in a sigwait state, because a change of state destroys the TAILQ link. It is actually safe to do so, though, because once a sigwaiting thread is found, the loop ends and the function returns. (Tor Egge) o When dispatching signals to threads, make the thread inherit the signal deferral flag of the currently running thread. (Tor Egge) Submitted by: Daniel Eischen <eischen@vigrid.com> and Tor Egge <Tor.Egge@fast.no>
1999-06-20 08:28:48 +00:00
void
kse_waitq_insert(struct pthread *thread)
{
struct pthread *td;
if (thread->wakeup_time.tv_sec == -1)
TAILQ_INSERT_TAIL(&thread->kse->k_schedq->sq_waitq, thread,
pqe);
else {
td = TAILQ_FIRST(&thread->kse->k_schedq->sq_waitq);
while ((td != NULL) && (td->wakeup_time.tv_sec != -1) &&
((td->wakeup_time.tv_sec < thread->wakeup_time.tv_sec) ||
((td->wakeup_time.tv_sec == thread->wakeup_time.tv_sec) &&
(td->wakeup_time.tv_nsec <= thread->wakeup_time.tv_nsec))))
td = TAILQ_NEXT(td, pqe);
if (td == NULL)
TAILQ_INSERT_TAIL(&thread->kse->k_schedq->sq_waitq,
thread, pqe);
else
TAILQ_INSERT_BEFORE(td, thread, pqe);
}
thread->flags |= THR_FLAGS_IN_WAITQ;
}
/*
* This must be called with the scheduling lock held.
*/
static void
kse_check_completed(struct kse *kse)
{
struct pthread *thread;
struct kse_thr_mailbox *completed;
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
int sig;
if ((completed = kse->k_kcb->kcb_kmbx.km_completed) != NULL) {
kse->k_kcb->kcb_kmbx.km_completed = NULL;
while (completed != NULL) {
thread = completed->tm_udata;
DBG_MSG("Found completed thread %p, name %s\n",
thread,
(thread->name == NULL) ? "none" : thread->name);
thread->blocked = 0;
if (thread != kse->k_curthread) {
thr_accounting(thread);
if ((thread->flags & THR_FLAGS_SUSPENDED) != 0)
THR_SET_STATE(thread, PS_SUSPENDED);
else
KSE_RUNQ_INSERT_TAIL(kse, thread);
if ((thread->kse != kse) &&
(thread->kse->k_curthread == thread)) {
/*
* Remove this thread from its
* previous KSE so that it (the KSE)
* doesn't think it is still active.
*/
thread->kse->k_curthread = NULL;
thread->active = 0;
}
}
if ((sig = thread->tcb->tcb_tmbx.tm_syncsig.si_signo)
!= 0) {
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
if (SIGISMEMBER(thread->sigmask, sig))
SIGADDSET(thread->sigpend, sig);
else
(void)_thr_sig_add(thread, sig,
&thread->tcb->tcb_tmbx.tm_syncsig);
thread->tcb->tcb_tmbx.tm_syncsig.si_signo = 0;
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
}
completed = completed->tm_next;
}
}
}
In the words of the author: o The polling mechanism for I/O readiness was changed from select() to poll(). In additon, a wrapped version of poll() is now provided. o The wrapped select routine now converts each fd_set to a poll array so that the thread scheduler doesn't have to perform a bitwise search for selected fds each time file descriptors are polled for I/O readiness. o The thread scheduler was modified to use a new queue (_workq) for threads that need work. Threads waiting for I/O readiness and spinblocks are added to the work queue in addition to the waiting queue. This reduces the time spent forming/searching the array of file descriptors being polled. o The waiting queue (_waitingq) is now maintained in order of thread wakeup time. This allows the thread scheduler to find the nearest wakeup time by looking at the first thread in the queue instead of searching the entire queue. o Removed file descriptor locking for select/poll routines. An application should not rely on the threads library for providing this locking; if necessary, the application should use mutexes to protect selecting/polling of file descriptors. o Retrieve and use the kernel clock rate/resolution at startup instead of hardcoding the clock resolution to 10 msec (tested with kernel running at 1000 HZ). o All queues have been changed to use queue.h macros. These include the queues of all threads, dead threads, and threads waiting for file descriptor locks. o Added reinitialization of the GC mutex and condition variable after a fork. Also prevented reallocation of the ready queue after a fork. o Prevented the wrapped close routine from closing the thread kernel pipes. o Initialized file descriptor table for stdio entries at thread init. o Provided additional flags to indicate to what queues threads belong. o Moved TAILQ initialization for statically allocated mutex and condition variables to after the spinlock. o Added dispatching of signals to pthread_kill. Removing the dispatching of signals from thread activation broke sigsuspend when pthread_kill was used to send a signal to a thread. o Temporarily set the state of a thread to PS_SUSPENDED when it is first created and placed in the list of threads so that it will not be accidentally scheduled before becoming a member of one of the scheduling queues. o Change the signal handler to queue signals to the thread kernel pipe if the scheduling queues are protected. When scheduling queues are unprotected, signals are then dequeued and handled. o Ensured that all installed signal handlers block the scheduling signal and that the scheduling signal handler blocks all other signals. This ensures that the signal handler is only interruptible for and by non-scheduling signals. An atomic lock is used to decide which instance of the signal handler will handle pending signals. o Removed _lock_thread_list and _unlock_thread_list as they are no longer used to protect the thread list. o Added missing RCS IDs to modified files. o Added checks for appropriate queue membership and activity when adding, removing, and searching the scheduling queues. These checks add very little overhead and are enabled when compiled with _PTHREADS_INVARIANTS defined. Suggested and implemented by Tor Egge with some modification by me. o Close a race condition in uthread_close. (Tor Egge) o Protect the scheduling queues while modifying them in pthread_cond_signal and _thread_fd_unlock. (Tor Egge) o Ensure that when a thread gets a mutex, the mutex is on that threads list of owned mutexes. (Tor Egge) o Set the kernel-in-scheduler flag in _thread_kern_sched_state and _thread_kern_sched_state_unlock to prevent a scheduling signal from calling the scheduler again. (Tor Egge) o Don't use TAILQ_FOREACH macro while searching the waiting queue for threads in a sigwait state, because a change of state destroys the TAILQ link. It is actually safe to do so, though, because once a sigwaiting thread is found, the loop ends and the function returns. (Tor Egge) o When dispatching signals to threads, make the thread inherit the signal deferral flag of the currently running thread. (Tor Egge) Submitted by: Daniel Eischen <eischen@vigrid.com> and Tor Egge <Tor.Egge@fast.no>
1999-06-20 08:28:48 +00:00
/*
* This must be called with the scheduling lock held.
*/
static void
kse_check_waitq(struct kse *kse)
{
struct pthread *pthread;
struct timespec ts;
[ The author's description... ] o Runnable threads are now maintained in priority queues. The implementation requires two things: 1.) The priority queues must be protected during insertion and removal of threads. Since the kernel scheduler must modify the priority queues, a spinlock for protection cannot be used. The functions _thread_kern_sched_defer() and _thread_kern_sched_undefer() were added to {un}defer kernel scheduler activation. 2.) A thread (active) priority change can be performed only when the thread is removed from the priority queue. The implementation uses a threads active priority when inserting it into the queue. A by-product is that thread switches are much faster. A separate queue is used for waiting and/or blocked threads, and it is searched at most 2 times in the kernel scheduler when there are active threads. It should be possible to reduce this to once by combining polling of threads waiting on I/O with the loop that looks for timed out threads and the minimum timeout value. o Functions to defer kernel scheduler activation were added. These are _thread_kern_sched_defer() and _thread_kern_sched_undefer() and may be called recursively. These routines do not block the scheduling signal, but latch its occurrence. The signal handler will not call the kernel scheduler when the running thread has deferred scheduling, but it will be called when running thread undefers scheduling. o Added support for _POSIX_THREAD_PRIORITY_SCHEDULING. All the POSIX routines required by this should now be implemented. One note, SCHED_OTHER, SCHED_FIFO, and SCHED_RR are required to be defined by including pthread.h. These defines are currently in sched.h. I modified pthread.h to include sched.h but don't know if this is the proper thing to do. o Added support for priority protection and inheritence mutexes. This allows definition of _POSIX_THREAD_PRIO_PROTECT and _POSIX_THREAD_PRIO_INHERIT. o Added additional error checks required by POSIX for mutexes and condition variables. o Provided a wrapper for sigpending which is marked as a hidden syscall. o Added a non-portable function as a debugging aid to allow an application to monitor thread context switches. An application can install a routine that gets called everytime a thread (explicitly created by the application) gets context switched. The routine gets passed the pthread IDs of the threads that are being switched in and out. Submitted by: Dan Eischen <eischen@vigrid.com> Changes by me: o Added a PS_SPINBLOCK state to deal with the priority inversion problem most often (I think) seen by threads calling malloc/free/realloc. o Dispatch signals to the running thread directly rather than at a context switch to avoid the situation where the switch never occurs.
1999-03-23 05:07:56 +00:00
KSE_GET_TOD(kse, &ts);
In the words of the author: o The polling mechanism for I/O readiness was changed from select() to poll(). In additon, a wrapped version of poll() is now provided. o The wrapped select routine now converts each fd_set to a poll array so that the thread scheduler doesn't have to perform a bitwise search for selected fds each time file descriptors are polled for I/O readiness. o The thread scheduler was modified to use a new queue (_workq) for threads that need work. Threads waiting for I/O readiness and spinblocks are added to the work queue in addition to the waiting queue. This reduces the time spent forming/searching the array of file descriptors being polled. o The waiting queue (_waitingq) is now maintained in order of thread wakeup time. This allows the thread scheduler to find the nearest wakeup time by looking at the first thread in the queue instead of searching the entire queue. o Removed file descriptor locking for select/poll routines. An application should not rely on the threads library for providing this locking; if necessary, the application should use mutexes to protect selecting/polling of file descriptors. o Retrieve and use the kernel clock rate/resolution at startup instead of hardcoding the clock resolution to 10 msec (tested with kernel running at 1000 HZ). o All queues have been changed to use queue.h macros. These include the queues of all threads, dead threads, and threads waiting for file descriptor locks. o Added reinitialization of the GC mutex and condition variable after a fork. Also prevented reallocation of the ready queue after a fork. o Prevented the wrapped close routine from closing the thread kernel pipes. o Initialized file descriptor table for stdio entries at thread init. o Provided additional flags to indicate to what queues threads belong. o Moved TAILQ initialization for statically allocated mutex and condition variables to after the spinlock. o Added dispatching of signals to pthread_kill. Removing the dispatching of signals from thread activation broke sigsuspend when pthread_kill was used to send a signal to a thread. o Temporarily set the state of a thread to PS_SUSPENDED when it is first created and placed in the list of threads so that it will not be accidentally scheduled before becoming a member of one of the scheduling queues. o Change the signal handler to queue signals to the thread kernel pipe if the scheduling queues are protected. When scheduling queues are unprotected, signals are then dequeued and handled. o Ensured that all installed signal handlers block the scheduling signal and that the scheduling signal handler blocks all other signals. This ensures that the signal handler is only interruptible for and by non-scheduling signals. An atomic lock is used to decide which instance of the signal handler will handle pending signals. o Removed _lock_thread_list and _unlock_thread_list as they are no longer used to protect the thread list. o Added missing RCS IDs to modified files. o Added checks for appropriate queue membership and activity when adding, removing, and searching the scheduling queues. These checks add very little overhead and are enabled when compiled with _PTHREADS_INVARIANTS defined. Suggested and implemented by Tor Egge with some modification by me. o Close a race condition in uthread_close. (Tor Egge) o Protect the scheduling queues while modifying them in pthread_cond_signal and _thread_fd_unlock. (Tor Egge) o Ensure that when a thread gets a mutex, the mutex is on that threads list of owned mutexes. (Tor Egge) o Set the kernel-in-scheduler flag in _thread_kern_sched_state and _thread_kern_sched_state_unlock to prevent a scheduling signal from calling the scheduler again. (Tor Egge) o Don't use TAILQ_FOREACH macro while searching the waiting queue for threads in a sigwait state, because a change of state destroys the TAILQ link. It is actually safe to do so, though, because once a sigwaiting thread is found, the loop ends and the function returns. (Tor Egge) o When dispatching signals to threads, make the thread inherit the signal deferral flag of the currently running thread. (Tor Egge) Submitted by: Daniel Eischen <eischen@vigrid.com> and Tor Egge <Tor.Egge@fast.no>
1999-06-20 08:28:48 +00:00
/*
* Wake up threads that have timedout. This has to be
* done before adding the current thread to the run queue
* so that a CPU intensive thread doesn't get preference
* over waiting threads.
*/
while (((pthread = KSE_WAITQ_FIRST(kse)) != NULL) &&
thr_timedout(pthread, &ts)) {
/* Remove the thread from the wait queue: */
KSE_WAITQ_REMOVE(kse, pthread);
DBG_MSG("Found timedout thread %p in waitq\n", pthread);
/* Indicate the thread timedout: */
pthread->timeout = 1;
/* Add the thread to the priority queue: */
if ((pthread->flags & THR_FLAGS_SUSPENDED) != 0)
THR_SET_STATE(pthread, PS_SUSPENDED);
else {
THR_SET_STATE(pthread, PS_RUNNING);
KSE_RUNQ_INSERT_TAIL(kse, pthread);
}
}
}
static int
thr_timedout(struct pthread *thread, struct timespec *curtime)
{
if (thread->wakeup_time.tv_sec < 0)
return (0);
else if (thread->wakeup_time.tv_sec > curtime->tv_sec)
return (0);
else if ((thread->wakeup_time.tv_sec == curtime->tv_sec) &&
(thread->wakeup_time.tv_nsec > curtime->tv_nsec))
return (0);
else
return (1);
}
/*
* This must be called with the scheduling lock held.
*
* Each thread has a time slice, a wakeup time (used when it wants
* to wait for a specified amount of time), a run state, and an
* active flag.
*
* When a thread gets run by the scheduler, the active flag is
* set to non-zero (1). When a thread performs an explicit yield
* or schedules a state change, it enters the scheduler and the
* active flag is cleared. When the active flag is still seen
* set in the scheduler, that means that the thread is blocked in
* the kernel (because it is cleared before entering the scheduler
* in all other instances).
*
* The wakeup time is only set for those states that can timeout.
* It is set to (-1, -1) for all other instances.
*
* The thread's run state, aside from being useful when debugging,
* is used to place the thread in an appropriate queue. There
* are 2 basic queues:
*
* o run queue - queue ordered by priority for all threads
* that are runnable
* o waiting queue - queue sorted by wakeup time for all threads
* that are not otherwise runnable (not blocked
* in kernel, not waiting for locks)
*
* The thread's time slice is used for round-robin scheduling
* (the default scheduling policy). While a SCHED_RR thread
* is runnable it's time slice accumulates. When it reaches
* the time slice interval, it gets reset and added to the end
* of the queue of threads at its priority. When a thread no
* longer becomes runnable (blocks in kernel, waits, etc), its
* time slice is reset.
*
* The job of kse_switchout_thread() is to handle all of the above.
*/
static void
kse_switchout_thread(struct kse *kse, struct pthread *thread)
{
int level;
int i;
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
int restart;
siginfo_t siginfo;
/*
* Place the currently running thread into the
* appropriate queue(s).
*/
DBG_MSG("Switching out thread %p, state %d\n", thread, thread->state);
THR_DEACTIVATE_LAST_LOCK(thread);
if (thread->blocked != 0) {
thread->active = 0;
thread->need_switchout = 0;
/* This thread must have blocked in the kernel. */
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
/*
* Check for pending signals for this thread to
* see if we need to interrupt it in the kernel.
*/
if (thread->check_pending != 0) {
for (i = 1; i <= _SIG_MAXSIG; ++i) {
if (SIGISMEMBER(thread->sigpend, i) &&
!SIGISMEMBER(thread->sigmask, i)) {
restart = _thread_sigact[1 - 1].sa_flags & SA_RESTART;
kse_thr_interrupt(&thread->tcb->tcb_tmbx,
restart ? KSE_INTR_RESTART : KSE_INTR_INTERRUPT, 0);
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
break;
}
}
}
}
else {
switch (thread->state) {
case PS_DEAD:
/*
* The scheduler is operating on a different
* stack. It is safe to do garbage collecting
* here.
*/
thread->active = 0;
thread->need_switchout = 0;
thread->lock_switch = 0;
thr_cleanup(kse, thread);
return;
break;
case PS_RUNNING:
if ((thread->flags & THR_FLAGS_SUSPENDED) != 0)
THR_SET_STATE(thread, PS_SUSPENDED);
break;
case PS_COND_WAIT:
case PS_SLEEP_WAIT:
/* Insert into the waiting queue: */
KSE_WAITQ_INSERT(kse, thread);
break;
case PS_LOCKWAIT:
/*
* This state doesn't timeout.
*/
thread->wakeup_time.tv_sec = -1;
thread->wakeup_time.tv_nsec = -1;
level = thread->locklevel - 1;
if (!_LCK_GRANTED(&thread->lockusers[level]))
KSE_WAITQ_INSERT(kse, thread);
else
THR_SET_STATE(thread, PS_RUNNING);
break;
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
case PS_SIGWAIT:
KSE_WAITQ_INSERT(kse, thread);
break;
case PS_JOIN:
case PS_MUTEX_WAIT:
case PS_SIGSUSPEND:
case PS_SUSPENDED:
case PS_DEADLOCK:
default:
/*
* These states don't timeout.
*/
thread->wakeup_time.tv_sec = -1;
thread->wakeup_time.tv_nsec = -1;
/* Insert into the waiting queue: */
KSE_WAITQ_INSERT(kse, thread);
break;
}
thr_accounting(thread);
if (thread->state == PS_RUNNING) {
if (thread->slice_usec == -1) {
/*
* The thread exceeded its time quantum or
* it yielded the CPU; place it at the tail
* of the queue for its priority.
*/
KSE_RUNQ_INSERT_TAIL(kse, thread);
} else {
/*
* The thread hasn't exceeded its interval
* Place it at the head of the queue for its
* priority.
*/
KSE_RUNQ_INSERT_HEAD(kse, thread);
}
}
}
thread->active = 0;
thread->need_switchout = 0;
if (thread->check_pending != 0) {
/* Install pending signals into the frame. */
thread->check_pending = 0;
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
KSE_LOCK_ACQUIRE(kse, &_thread_signal_lock);
for (i = 1; i <= _SIG_MAXSIG; i++) {
if (SIGISMEMBER(thread->sigmask, i))
continue;
if (SIGISMEMBER(thread->sigpend, i))
(void)_thr_sig_add(thread, i,
&thread->siginfo[i-1]);
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
else if (SIGISMEMBER(_thr_proc_sigpending, i) &&
_thr_getprocsig_unlocked(i, &siginfo)) {
(void)_thr_sig_add(thread, i, &siginfo);
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
}
}
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
KSE_LOCK_RELEASE(kse, &_thread_signal_lock);
}
}
/*
* This function waits for the smallest timeout value of any waiting
* thread, or until it receives a message from another KSE.
*
* This must be called with the scheduling lock held.
*/
static void
kse_wait(struct kse *kse, struct pthread *td_wait, int sigseqno)
{
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
struct timespec ts, ts_sleep;
int saved_flags;
In the words of the author: o The polling mechanism for I/O readiness was changed from select() to poll(). In additon, a wrapped version of poll() is now provided. o The wrapped select routine now converts each fd_set to a poll array so that the thread scheduler doesn't have to perform a bitwise search for selected fds each time file descriptors are polled for I/O readiness. o The thread scheduler was modified to use a new queue (_workq) for threads that need work. Threads waiting for I/O readiness and spinblocks are added to the work queue in addition to the waiting queue. This reduces the time spent forming/searching the array of file descriptors being polled. o The waiting queue (_waitingq) is now maintained in order of thread wakeup time. This allows the thread scheduler to find the nearest wakeup time by looking at the first thread in the queue instead of searching the entire queue. o Removed file descriptor locking for select/poll routines. An application should not rely on the threads library for providing this locking; if necessary, the application should use mutexes to protect selecting/polling of file descriptors. o Retrieve and use the kernel clock rate/resolution at startup instead of hardcoding the clock resolution to 10 msec (tested with kernel running at 1000 HZ). o All queues have been changed to use queue.h macros. These include the queues of all threads, dead threads, and threads waiting for file descriptor locks. o Added reinitialization of the GC mutex and condition variable after a fork. Also prevented reallocation of the ready queue after a fork. o Prevented the wrapped close routine from closing the thread kernel pipes. o Initialized file descriptor table for stdio entries at thread init. o Provided additional flags to indicate to what queues threads belong. o Moved TAILQ initialization for statically allocated mutex and condition variables to after the spinlock. o Added dispatching of signals to pthread_kill. Removing the dispatching of signals from thread activation broke sigsuspend when pthread_kill was used to send a signal to a thread. o Temporarily set the state of a thread to PS_SUSPENDED when it is first created and placed in the list of threads so that it will not be accidentally scheduled before becoming a member of one of the scheduling queues. o Change the signal handler to queue signals to the thread kernel pipe if the scheduling queues are protected. When scheduling queues are unprotected, signals are then dequeued and handled. o Ensured that all installed signal handlers block the scheduling signal and that the scheduling signal handler blocks all other signals. This ensures that the signal handler is only interruptible for and by non-scheduling signals. An atomic lock is used to decide which instance of the signal handler will handle pending signals. o Removed _lock_thread_list and _unlock_thread_list as they are no longer used to protect the thread list. o Added missing RCS IDs to modified files. o Added checks for appropriate queue membership and activity when adding, removing, and searching the scheduling queues. These checks add very little overhead and are enabled when compiled with _PTHREADS_INVARIANTS defined. Suggested and implemented by Tor Egge with some modification by me. o Close a race condition in uthread_close. (Tor Egge) o Protect the scheduling queues while modifying them in pthread_cond_signal and _thread_fd_unlock. (Tor Egge) o Ensure that when a thread gets a mutex, the mutex is on that threads list of owned mutexes. (Tor Egge) o Set the kernel-in-scheduler flag in _thread_kern_sched_state and _thread_kern_sched_state_unlock to prevent a scheduling signal from calling the scheduler again. (Tor Egge) o Don't use TAILQ_FOREACH macro while searching the waiting queue for threads in a sigwait state, because a change of state destroys the TAILQ link. It is actually safe to do so, though, because once a sigwaiting thread is found, the loop ends and the function returns. (Tor Egge) o When dispatching signals to threads, make the thread inherit the signal deferral flag of the currently running thread. (Tor Egge) Submitted by: Daniel Eischen <eischen@vigrid.com> and Tor Egge <Tor.Egge@fast.no>
1999-06-20 08:28:48 +00:00
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
KSE_GET_TOD(kse, &ts);
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
if ((td_wait == NULL) || (td_wait->wakeup_time.tv_sec < 0)) {
/* Limit sleep to no more than 1 minute. */
ts_sleep.tv_sec = 60;
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
ts_sleep.tv_nsec = 0;
} else {
TIMESPEC_SUB(&ts_sleep, &td_wait->wakeup_time, &ts);
if (ts_sleep.tv_sec > 60) {
ts_sleep.tv_sec = 60;
ts_sleep.tv_nsec = 0;
}
}
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
/* Don't sleep for negative times. */
if ((ts_sleep.tv_sec >= 0) && (ts_sleep.tv_nsec >= 0)) {
KSE_SET_IDLE(kse);
kse->k_kseg->kg_idle_kses++;
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
KSE_SCHED_UNLOCK(kse, kse->k_kseg);
if ((kse->k_kseg->kg_flags & KGF_SINGLE_THREAD) &&
(kse->k_sigseqno != sigseqno))
; /* don't sleep */
else {
saved_flags = kse->k_kcb->kcb_kmbx.km_flags;
kse->k_kcb->kcb_kmbx.km_flags |= KMF_NOUPCALL;
kse_release(&ts_sleep);
kse->k_kcb->kcb_kmbx.km_flags = saved_flags;
}
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
KSE_SCHED_LOCK(kse, kse->k_kseg);
if (KSE_IS_IDLE(kse)) {
KSE_CLEAR_IDLE(kse);
kse->k_kseg->kg_idle_kses--;
}
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
}
}
/*
* Avoid calling this kse_exit() so as not to confuse it with the
* system call of the same name.
*/
static void
kse_fini(struct kse *kse)
{
/* struct kse_group *free_kseg = NULL; */
struct timespec ts;
In the words of the author: o The polling mechanism for I/O readiness was changed from select() to poll(). In additon, a wrapped version of poll() is now provided. o The wrapped select routine now converts each fd_set to a poll array so that the thread scheduler doesn't have to perform a bitwise search for selected fds each time file descriptors are polled for I/O readiness. o The thread scheduler was modified to use a new queue (_workq) for threads that need work. Threads waiting for I/O readiness and spinblocks are added to the work queue in addition to the waiting queue. This reduces the time spent forming/searching the array of file descriptors being polled. o The waiting queue (_waitingq) is now maintained in order of thread wakeup time. This allows the thread scheduler to find the nearest wakeup time by looking at the first thread in the queue instead of searching the entire queue. o Removed file descriptor locking for select/poll routines. An application should not rely on the threads library for providing this locking; if necessary, the application should use mutexes to protect selecting/polling of file descriptors. o Retrieve and use the kernel clock rate/resolution at startup instead of hardcoding the clock resolution to 10 msec (tested with kernel running at 1000 HZ). o All queues have been changed to use queue.h macros. These include the queues of all threads, dead threads, and threads waiting for file descriptor locks. o Added reinitialization of the GC mutex and condition variable after a fork. Also prevented reallocation of the ready queue after a fork. o Prevented the wrapped close routine from closing the thread kernel pipes. o Initialized file descriptor table for stdio entries at thread init. o Provided additional flags to indicate to what queues threads belong. o Moved TAILQ initialization for statically allocated mutex and condition variables to after the spinlock. o Added dispatching of signals to pthread_kill. Removing the dispatching of signals from thread activation broke sigsuspend when pthread_kill was used to send a signal to a thread. o Temporarily set the state of a thread to PS_SUSPENDED when it is first created and placed in the list of threads so that it will not be accidentally scheduled before becoming a member of one of the scheduling queues. o Change the signal handler to queue signals to the thread kernel pipe if the scheduling queues are protected. When scheduling queues are unprotected, signals are then dequeued and handled. o Ensured that all installed signal handlers block the scheduling signal and that the scheduling signal handler blocks all other signals. This ensures that the signal handler is only interruptible for and by non-scheduling signals. An atomic lock is used to decide which instance of the signal handler will handle pending signals. o Removed _lock_thread_list and _unlock_thread_list as they are no longer used to protect the thread list. o Added missing RCS IDs to modified files. o Added checks for appropriate queue membership and activity when adding, removing, and searching the scheduling queues. These checks add very little overhead and are enabled when compiled with _PTHREADS_INVARIANTS defined. Suggested and implemented by Tor Egge with some modification by me. o Close a race condition in uthread_close. (Tor Egge) o Protect the scheduling queues while modifying them in pthread_cond_signal and _thread_fd_unlock. (Tor Egge) o Ensure that when a thread gets a mutex, the mutex is on that threads list of owned mutexes. (Tor Egge) o Set the kernel-in-scheduler flag in _thread_kern_sched_state and _thread_kern_sched_state_unlock to prevent a scheduling signal from calling the scheduler again. (Tor Egge) o Don't use TAILQ_FOREACH macro while searching the waiting queue for threads in a sigwait state, because a change of state destroys the TAILQ link. It is actually safe to do so, though, because once a sigwaiting thread is found, the loop ends and the function returns. (Tor Egge) o When dispatching signals to threads, make the thread inherit the signal deferral flag of the currently running thread. (Tor Egge) Submitted by: Daniel Eischen <eischen@vigrid.com> and Tor Egge <Tor.Egge@fast.no>
1999-06-20 08:28:48 +00:00
/*
* Check to see if this is one of the main kses.
In the words of the author: o The polling mechanism for I/O readiness was changed from select() to poll(). In additon, a wrapped version of poll() is now provided. o The wrapped select routine now converts each fd_set to a poll array so that the thread scheduler doesn't have to perform a bitwise search for selected fds each time file descriptors are polled for I/O readiness. o The thread scheduler was modified to use a new queue (_workq) for threads that need work. Threads waiting for I/O readiness and spinblocks are added to the work queue in addition to the waiting queue. This reduces the time spent forming/searching the array of file descriptors being polled. o The waiting queue (_waitingq) is now maintained in order of thread wakeup time. This allows the thread scheduler to find the nearest wakeup time by looking at the first thread in the queue instead of searching the entire queue. o Removed file descriptor locking for select/poll routines. An application should not rely on the threads library for providing this locking; if necessary, the application should use mutexes to protect selecting/polling of file descriptors. o Retrieve and use the kernel clock rate/resolution at startup instead of hardcoding the clock resolution to 10 msec (tested with kernel running at 1000 HZ). o All queues have been changed to use queue.h macros. These include the queues of all threads, dead threads, and threads waiting for file descriptor locks. o Added reinitialization of the GC mutex and condition variable after a fork. Also prevented reallocation of the ready queue after a fork. o Prevented the wrapped close routine from closing the thread kernel pipes. o Initialized file descriptor table for stdio entries at thread init. o Provided additional flags to indicate to what queues threads belong. o Moved TAILQ initialization for statically allocated mutex and condition variables to after the spinlock. o Added dispatching of signals to pthread_kill. Removing the dispatching of signals from thread activation broke sigsuspend when pthread_kill was used to send a signal to a thread. o Temporarily set the state of a thread to PS_SUSPENDED when it is first created and placed in the list of threads so that it will not be accidentally scheduled before becoming a member of one of the scheduling queues. o Change the signal handler to queue signals to the thread kernel pipe if the scheduling queues are protected. When scheduling queues are unprotected, signals are then dequeued and handled. o Ensured that all installed signal handlers block the scheduling signal and that the scheduling signal handler blocks all other signals. This ensures that the signal handler is only interruptible for and by non-scheduling signals. An atomic lock is used to decide which instance of the signal handler will handle pending signals. o Removed _lock_thread_list and _unlock_thread_list as they are no longer used to protect the thread list. o Added missing RCS IDs to modified files. o Added checks for appropriate queue membership and activity when adding, removing, and searching the scheduling queues. These checks add very little overhead and are enabled when compiled with _PTHREADS_INVARIANTS defined. Suggested and implemented by Tor Egge with some modification by me. o Close a race condition in uthread_close. (Tor Egge) o Protect the scheduling queues while modifying them in pthread_cond_signal and _thread_fd_unlock. (Tor Egge) o Ensure that when a thread gets a mutex, the mutex is on that threads list of owned mutexes. (Tor Egge) o Set the kernel-in-scheduler flag in _thread_kern_sched_state and _thread_kern_sched_state_unlock to prevent a scheduling signal from calling the scheduler again. (Tor Egge) o Don't use TAILQ_FOREACH macro while searching the waiting queue for threads in a sigwait state, because a change of state destroys the TAILQ link. It is actually safe to do so, though, because once a sigwaiting thread is found, the loop ends and the function returns. (Tor Egge) o When dispatching signals to threads, make the thread inherit the signal deferral flag of the currently running thread. (Tor Egge) Submitted by: Daniel Eischen <eischen@vigrid.com> and Tor Egge <Tor.Egge@fast.no>
1999-06-20 08:28:48 +00:00
*/
if (kse->k_kseg != _kse_initial->k_kseg) {
PANIC("shouldn't get here");
/* This is for supporting thread groups. */
#ifdef NOT_YET
/* Remove this KSE from the KSEG's list of KSEs. */
KSE_SCHED_LOCK(kse, kse->k_kseg);
TAILQ_REMOVE(&kse->k_kseg->kg_kseq, kse, k_kgqe);
kse->k_kseg->kg_ksecount--;
if (TAILQ_EMPTY(&kse->k_kseg->kg_kseq))
free_kseg = kse->k_kseg;
KSE_SCHED_UNLOCK(kse, kse->k_kseg);
/*
* Add this KSE to the list of free KSEs along with
* the KSEG if is now orphaned.
*/
KSE_LOCK_ACQUIRE(kse, &kse_lock);
if (free_kseg != NULL)
kseg_free_unlocked(free_kseg);
kse_free_unlocked(kse);
KSE_LOCK_RELEASE(kse, &kse_lock);
kse_exit();
/* Never returns. */
PANIC("kse_exit()");
#endif
} else {
/*
* We allow program to kill kse in initial group (by
* lowering the concurrency).
*/
if ((kse != _kse_initial) &&
((kse->k_flags & KF_TERMINATED) != 0)) {
KSE_SCHED_LOCK(kse, kse->k_kseg);
TAILQ_REMOVE(&kse->k_kseg->kg_kseq, kse, k_kgqe);
kse->k_kseg->kg_ksecount--;
KSE_SCHED_UNLOCK(kse, kse->k_kseg);
KSE_LOCK_ACQUIRE(kse, &kse_lock);
kse_free_unlocked(kse);
KSE_LOCK_RELEASE(kse, &kse_lock);
kse_exit();
/* Never returns. */
PANIC("kse_exit() failed for initial kseg");
}
KSE_SCHED_LOCK(kse, kse->k_kseg);
KSE_SET_IDLE(kse);
kse->k_kseg->kg_idle_kses++;
KSE_SCHED_UNLOCK(kse, kse->k_kseg);
ts.tv_sec = 120;
ts.tv_nsec = 0;
kse->k_kcb->kcb_kmbx.km_flags = 0;
kse_release(&ts);
/* Never reach */
}
}
void
_thr_set_timeout(const struct timespec *timeout)
{
struct pthread *curthread = _get_curthread();
struct timespec ts;
/* Reset the timeout flag for the running thread: */
curthread->timeout = 0;
/* Check if the thread is to wait forever: */
if (timeout == NULL) {
/*
* Set the wakeup time to something that can be recognised as
* different to an actual time of day:
*/
curthread->wakeup_time.tv_sec = -1;
curthread->wakeup_time.tv_nsec = -1;
}
/* Check if no waiting is required: */
else if ((timeout->tv_sec == 0) && (timeout->tv_nsec == 0)) {
/* Set the wake up time to 'immediately': */
curthread->wakeup_time.tv_sec = 0;
curthread->wakeup_time.tv_nsec = 0;
} else {
/* Calculate the time for the current thread to wakeup: */
KSE_GET_TOD(curthread->kse, &ts);
TIMESPEC_ADD(&curthread->wakeup_time, &ts, timeout);
}
}
[ The author's description... ] o Runnable threads are now maintained in priority queues. The implementation requires two things: 1.) The priority queues must be protected during insertion and removal of threads. Since the kernel scheduler must modify the priority queues, a spinlock for protection cannot be used. The functions _thread_kern_sched_defer() and _thread_kern_sched_undefer() were added to {un}defer kernel scheduler activation. 2.) A thread (active) priority change can be performed only when the thread is removed from the priority queue. The implementation uses a threads active priority when inserting it into the queue. A by-product is that thread switches are much faster. A separate queue is used for waiting and/or blocked threads, and it is searched at most 2 times in the kernel scheduler when there are active threads. It should be possible to reduce this to once by combining polling of threads waiting on I/O with the loop that looks for timed out threads and the minimum timeout value. o Functions to defer kernel scheduler activation were added. These are _thread_kern_sched_defer() and _thread_kern_sched_undefer() and may be called recursively. These routines do not block the scheduling signal, but latch its occurrence. The signal handler will not call the kernel scheduler when the running thread has deferred scheduling, but it will be called when running thread undefers scheduling. o Added support for _POSIX_THREAD_PRIORITY_SCHEDULING. All the POSIX routines required by this should now be implemented. One note, SCHED_OTHER, SCHED_FIFO, and SCHED_RR are required to be defined by including pthread.h. These defines are currently in sched.h. I modified pthread.h to include sched.h but don't know if this is the proper thing to do. o Added support for priority protection and inheritence mutexes. This allows definition of _POSIX_THREAD_PRIO_PROTECT and _POSIX_THREAD_PRIO_INHERIT. o Added additional error checks required by POSIX for mutexes and condition variables. o Provided a wrapper for sigpending which is marked as a hidden syscall. o Added a non-portable function as a debugging aid to allow an application to monitor thread context switches. An application can install a routine that gets called everytime a thread (explicitly created by the application) gets context switched. The routine gets passed the pthread IDs of the threads that are being switched in and out. Submitted by: Dan Eischen <eischen@vigrid.com> Changes by me: o Added a PS_SPINBLOCK state to deal with the priority inversion problem most often (I think) seen by threads calling malloc/free/realloc. o Dispatch signals to the running thread directly rather than at a context switch to avoid the situation where the switch never occurs.
1999-03-23 05:07:56 +00:00
void
_thr_panic_exit(char *file, int line, char *msg)
[ The author's description... ] o Runnable threads are now maintained in priority queues. The implementation requires two things: 1.) The priority queues must be protected during insertion and removal of threads. Since the kernel scheduler must modify the priority queues, a spinlock for protection cannot be used. The functions _thread_kern_sched_defer() and _thread_kern_sched_undefer() were added to {un}defer kernel scheduler activation. 2.) A thread (active) priority change can be performed only when the thread is removed from the priority queue. The implementation uses a threads active priority when inserting it into the queue. A by-product is that thread switches are much faster. A separate queue is used for waiting and/or blocked threads, and it is searched at most 2 times in the kernel scheduler when there are active threads. It should be possible to reduce this to once by combining polling of threads waiting on I/O with the loop that looks for timed out threads and the minimum timeout value. o Functions to defer kernel scheduler activation were added. These are _thread_kern_sched_defer() and _thread_kern_sched_undefer() and may be called recursively. These routines do not block the scheduling signal, but latch its occurrence. The signal handler will not call the kernel scheduler when the running thread has deferred scheduling, but it will be called when running thread undefers scheduling. o Added support for _POSIX_THREAD_PRIORITY_SCHEDULING. All the POSIX routines required by this should now be implemented. One note, SCHED_OTHER, SCHED_FIFO, and SCHED_RR are required to be defined by including pthread.h. These defines are currently in sched.h. I modified pthread.h to include sched.h but don't know if this is the proper thing to do. o Added support for priority protection and inheritence mutexes. This allows definition of _POSIX_THREAD_PRIO_PROTECT and _POSIX_THREAD_PRIO_INHERIT. o Added additional error checks required by POSIX for mutexes and condition variables. o Provided a wrapper for sigpending which is marked as a hidden syscall. o Added a non-portable function as a debugging aid to allow an application to monitor thread context switches. An application can install a routine that gets called everytime a thread (explicitly created by the application) gets context switched. The routine gets passed the pthread IDs of the threads that are being switched in and out. Submitted by: Dan Eischen <eischen@vigrid.com> Changes by me: o Added a PS_SPINBLOCK state to deal with the priority inversion problem most often (I think) seen by threads calling malloc/free/realloc. o Dispatch signals to the running thread directly rather than at a context switch to avoid the situation where the switch never occurs.
1999-03-23 05:07:56 +00:00
{
char buf[256];
snprintf(buf, sizeof(buf), "(%s:%d) %s\n", file, line, msg);
__sys_write(2, buf, strlen(buf));
abort();
[ The author's description... ] o Runnable threads are now maintained in priority queues. The implementation requires two things: 1.) The priority queues must be protected during insertion and removal of threads. Since the kernel scheduler must modify the priority queues, a spinlock for protection cannot be used. The functions _thread_kern_sched_defer() and _thread_kern_sched_undefer() were added to {un}defer kernel scheduler activation. 2.) A thread (active) priority change can be performed only when the thread is removed from the priority queue. The implementation uses a threads active priority when inserting it into the queue. A by-product is that thread switches are much faster. A separate queue is used for waiting and/or blocked threads, and it is searched at most 2 times in the kernel scheduler when there are active threads. It should be possible to reduce this to once by combining polling of threads waiting on I/O with the loop that looks for timed out threads and the minimum timeout value. o Functions to defer kernel scheduler activation were added. These are _thread_kern_sched_defer() and _thread_kern_sched_undefer() and may be called recursively. These routines do not block the scheduling signal, but latch its occurrence. The signal handler will not call the kernel scheduler when the running thread has deferred scheduling, but it will be called when running thread undefers scheduling. o Added support for _POSIX_THREAD_PRIORITY_SCHEDULING. All the POSIX routines required by this should now be implemented. One note, SCHED_OTHER, SCHED_FIFO, and SCHED_RR are required to be defined by including pthread.h. These defines are currently in sched.h. I modified pthread.h to include sched.h but don't know if this is the proper thing to do. o Added support for priority protection and inheritence mutexes. This allows definition of _POSIX_THREAD_PRIO_PROTECT and _POSIX_THREAD_PRIO_INHERIT. o Added additional error checks required by POSIX for mutexes and condition variables. o Provided a wrapper for sigpending which is marked as a hidden syscall. o Added a non-portable function as a debugging aid to allow an application to monitor thread context switches. An application can install a routine that gets called everytime a thread (explicitly created by the application) gets context switched. The routine gets passed the pthread IDs of the threads that are being switched in and out. Submitted by: Dan Eischen <eischen@vigrid.com> Changes by me: o Added a PS_SPINBLOCK state to deal with the priority inversion problem most often (I think) seen by threads calling malloc/free/realloc. o Dispatch signals to the running thread directly rather than at a context switch to avoid the situation where the switch never occurs.
1999-03-23 05:07:56 +00:00
}
void
_thr_setrunnable(struct pthread *curthread, struct pthread *thread)
[ The author's description... ] o Runnable threads are now maintained in priority queues. The implementation requires two things: 1.) The priority queues must be protected during insertion and removal of threads. Since the kernel scheduler must modify the priority queues, a spinlock for protection cannot be used. The functions _thread_kern_sched_defer() and _thread_kern_sched_undefer() were added to {un}defer kernel scheduler activation. 2.) A thread (active) priority change can be performed only when the thread is removed from the priority queue. The implementation uses a threads active priority when inserting it into the queue. A by-product is that thread switches are much faster. A separate queue is used for waiting and/or blocked threads, and it is searched at most 2 times in the kernel scheduler when there are active threads. It should be possible to reduce this to once by combining polling of threads waiting on I/O with the loop that looks for timed out threads and the minimum timeout value. o Functions to defer kernel scheduler activation were added. These are _thread_kern_sched_defer() and _thread_kern_sched_undefer() and may be called recursively. These routines do not block the scheduling signal, but latch its occurrence. The signal handler will not call the kernel scheduler when the running thread has deferred scheduling, but it will be called when running thread undefers scheduling. o Added support for _POSIX_THREAD_PRIORITY_SCHEDULING. All the POSIX routines required by this should now be implemented. One note, SCHED_OTHER, SCHED_FIFO, and SCHED_RR are required to be defined by including pthread.h. These defines are currently in sched.h. I modified pthread.h to include sched.h but don't know if this is the proper thing to do. o Added support for priority protection and inheritence mutexes. This allows definition of _POSIX_THREAD_PRIO_PROTECT and _POSIX_THREAD_PRIO_INHERIT. o Added additional error checks required by POSIX for mutexes and condition variables. o Provided a wrapper for sigpending which is marked as a hidden syscall. o Added a non-portable function as a debugging aid to allow an application to monitor thread context switches. An application can install a routine that gets called everytime a thread (explicitly created by the application) gets context switched. The routine gets passed the pthread IDs of the threads that are being switched in and out. Submitted by: Dan Eischen <eischen@vigrid.com> Changes by me: o Added a PS_SPINBLOCK state to deal with the priority inversion problem most often (I think) seen by threads calling malloc/free/realloc. o Dispatch signals to the running thread directly rather than at a context switch to avoid the situation where the switch never occurs.
1999-03-23 05:07:56 +00:00
{
kse_critical_t crit;
struct kse_mailbox *kmbx;
crit = _kse_critical_enter();
KSE_SCHED_LOCK(curthread->kse, thread->kseg);
kmbx = _thr_setrunnable_unlocked(thread);
KSE_SCHED_UNLOCK(curthread->kse, thread->kseg);
_kse_critical_leave(crit);
if (kmbx != NULL)
kse_wakeup(kmbx);
}
struct kse_mailbox *
_thr_setrunnable_unlocked(struct pthread *thread)
{
struct kse_mailbox *kmbx = NULL;
if ((thread->kseg->kg_flags & KGF_SINGLE_THREAD) != 0) {
/* No silly queues for these threads. */
if ((thread->flags & THR_FLAGS_SUSPENDED) != 0)
THR_SET_STATE(thread, PS_SUSPENDED);
else {
THR_SET_STATE(thread, PS_RUNNING);
kmbx = kse_wakeup_one(thread);
}
} else if (thread->state != PS_RUNNING) {
if ((thread->flags & THR_FLAGS_IN_WAITQ) != 0)
KSE_WAITQ_REMOVE(thread->kse, thread);
if ((thread->flags & THR_FLAGS_SUSPENDED) != 0)
THR_SET_STATE(thread, PS_SUSPENDED);
else {
THR_SET_STATE(thread, PS_RUNNING);
if ((thread->blocked == 0) && (thread->active == 0) &&
(thread->flags & THR_FLAGS_IN_RUNQ) == 0)
THR_RUNQ_INSERT_TAIL(thread);
/*
* XXX - Threads are not yet assigned to specific
* KSEs; they are assigned to the KSEG. So
* the fact that a thread's KSE is waiting
* doesn't necessarily mean that it will be
* the KSE that runs the thread after the
* lock is granted. But we don't know if the
* other KSEs within the same KSEG are also
* in a waiting state or not so we err on the
* side of caution and wakeup the thread's
* last known KSE. We ensure that the
* threads KSE doesn't change while it's
* scheduling lock is held so it is safe to
* reference it (the KSE). If the KSE wakes
* up and doesn't find any more work it will
* again go back to waiting so no harm is
* done.
*/
kmbx = kse_wakeup_one(thread);
}
}
return (kmbx);
}
static struct kse_mailbox *
kse_wakeup_one(struct pthread *thread)
{
struct kse *ke;
if (KSE_IS_IDLE(thread->kse)) {
KSE_CLEAR_IDLE(thread->kse);
thread->kseg->kg_idle_kses--;
return (&thread->kse->k_kcb->kcb_kmbx);
} else {
TAILQ_FOREACH(ke, &thread->kseg->kg_kseq, k_kgqe) {
if (KSE_IS_IDLE(ke)) {
KSE_CLEAR_IDLE(ke);
ke->k_kseg->kg_idle_kses--;
return (&ke->k_kcb->kcb_kmbx);
}
}
}
return (NULL);
}
static void
kse_wakeup_multi(struct kse *curkse)
{
struct kse *ke;
int tmp;
if ((tmp = KSE_RUNQ_THREADS(curkse)) && curkse->k_kseg->kg_idle_kses) {
TAILQ_FOREACH(ke, &curkse->k_kseg->kg_kseq, k_kgqe) {
if (KSE_IS_IDLE(ke)) {
KSE_CLEAR_IDLE(ke);
ke->k_kseg->kg_idle_kses--;
KSE_WAKEUP(ke);
if (--tmp == 0)
break;
}
}
}
}
/*
* Allocate a new KSEG.
*
* We allow the current thread to be NULL in the case that this
* is the first time a KSEG is being created (library initialization).
* In this case, we don't need to (and can't) take any locks.
*/
struct kse_group *
_kseg_alloc(struct pthread *curthread)
{
struct kse_group *kseg = NULL;
kse_critical_t crit;
if ((curthread != NULL) && (free_kseg_count > 0)) {
/* Use the kse lock for the kseg queue. */
crit = _kse_critical_enter();
KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock);
if ((kseg = TAILQ_FIRST(&free_kse_groupq)) != NULL) {
TAILQ_REMOVE(&free_kse_groupq, kseg, kg_qe);
free_kseg_count--;
active_kseg_count++;
TAILQ_INSERT_TAIL(&active_kse_groupq, kseg, kg_qe);
}
KSE_LOCK_RELEASE(curthread->kse, &kse_lock);
_kse_critical_leave(crit);
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
if (kseg)
kseg_reinit(kseg);
}
[ The author's description... ] o Runnable threads are now maintained in priority queues. The implementation requires two things: 1.) The priority queues must be protected during insertion and removal of threads. Since the kernel scheduler must modify the priority queues, a spinlock for protection cannot be used. The functions _thread_kern_sched_defer() and _thread_kern_sched_undefer() were added to {un}defer kernel scheduler activation. 2.) A thread (active) priority change can be performed only when the thread is removed from the priority queue. The implementation uses a threads active priority when inserting it into the queue. A by-product is that thread switches are much faster. A separate queue is used for waiting and/or blocked threads, and it is searched at most 2 times in the kernel scheduler when there are active threads. It should be possible to reduce this to once by combining polling of threads waiting on I/O with the loop that looks for timed out threads and the minimum timeout value. o Functions to defer kernel scheduler activation were added. These are _thread_kern_sched_defer() and _thread_kern_sched_undefer() and may be called recursively. These routines do not block the scheduling signal, but latch its occurrence. The signal handler will not call the kernel scheduler when the running thread has deferred scheduling, but it will be called when running thread undefers scheduling. o Added support for _POSIX_THREAD_PRIORITY_SCHEDULING. All the POSIX routines required by this should now be implemented. One note, SCHED_OTHER, SCHED_FIFO, and SCHED_RR are required to be defined by including pthread.h. These defines are currently in sched.h. I modified pthread.h to include sched.h but don't know if this is the proper thing to do. o Added support for priority protection and inheritence mutexes. This allows definition of _POSIX_THREAD_PRIO_PROTECT and _POSIX_THREAD_PRIO_INHERIT. o Added additional error checks required by POSIX for mutexes and condition variables. o Provided a wrapper for sigpending which is marked as a hidden syscall. o Added a non-portable function as a debugging aid to allow an application to monitor thread context switches. An application can install a routine that gets called everytime a thread (explicitly created by the application) gets context switched. The routine gets passed the pthread IDs of the threads that are being switched in and out. Submitted by: Dan Eischen <eischen@vigrid.com> Changes by me: o Added a PS_SPINBLOCK state to deal with the priority inversion problem most often (I think) seen by threads calling malloc/free/realloc. o Dispatch signals to the running thread directly rather than at a context switch to avoid the situation where the switch never occurs.
1999-03-23 05:07:56 +00:00
/*
* If requested, attempt to allocate a new KSE group only if the
* KSE allocation was successful and a KSE group wasn't found in
* the free list.
[ The author's description... ] o Runnable threads are now maintained in priority queues. The implementation requires two things: 1.) The priority queues must be protected during insertion and removal of threads. Since the kernel scheduler must modify the priority queues, a spinlock for protection cannot be used. The functions _thread_kern_sched_defer() and _thread_kern_sched_undefer() were added to {un}defer kernel scheduler activation. 2.) A thread (active) priority change can be performed only when the thread is removed from the priority queue. The implementation uses a threads active priority when inserting it into the queue. A by-product is that thread switches are much faster. A separate queue is used for waiting and/or blocked threads, and it is searched at most 2 times in the kernel scheduler when there are active threads. It should be possible to reduce this to once by combining polling of threads waiting on I/O with the loop that looks for timed out threads and the minimum timeout value. o Functions to defer kernel scheduler activation were added. These are _thread_kern_sched_defer() and _thread_kern_sched_undefer() and may be called recursively. These routines do not block the scheduling signal, but latch its occurrence. The signal handler will not call the kernel scheduler when the running thread has deferred scheduling, but it will be called when running thread undefers scheduling. o Added support for _POSIX_THREAD_PRIORITY_SCHEDULING. All the POSIX routines required by this should now be implemented. One note, SCHED_OTHER, SCHED_FIFO, and SCHED_RR are required to be defined by including pthread.h. These defines are currently in sched.h. I modified pthread.h to include sched.h but don't know if this is the proper thing to do. o Added support for priority protection and inheritence mutexes. This allows definition of _POSIX_THREAD_PRIO_PROTECT and _POSIX_THREAD_PRIO_INHERIT. o Added additional error checks required by POSIX for mutexes and condition variables. o Provided a wrapper for sigpending which is marked as a hidden syscall. o Added a non-portable function as a debugging aid to allow an application to monitor thread context switches. An application can install a routine that gets called everytime a thread (explicitly created by the application) gets context switched. The routine gets passed the pthread IDs of the threads that are being switched in and out. Submitted by: Dan Eischen <eischen@vigrid.com> Changes by me: o Added a PS_SPINBLOCK state to deal with the priority inversion problem most often (I think) seen by threads calling malloc/free/realloc. o Dispatch signals to the running thread directly rather than at a context switch to avoid the situation where the switch never occurs.
1999-03-23 05:07:56 +00:00
*/
if ((kseg == NULL) &&
((kseg = (struct kse_group *)malloc(sizeof(*kseg))) != NULL)) {
if (_pq_alloc(&kseg->kg_schedq.sq_runq,
THR_MIN_PRIORITY, THR_LAST_PRIORITY) != 0) {
free(kseg);
kseg = NULL;
} else {
kseg_init(kseg);
/* Add the KSEG to the list of active KSEGs. */
if (curthread != NULL) {
crit = _kse_critical_enter();
KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock);
active_kseg_count++;
TAILQ_INSERT_TAIL(&active_kse_groupq,
kseg, kg_qe);
KSE_LOCK_RELEASE(curthread->kse, &kse_lock);
_kse_critical_leave(crit);
} else {
active_kseg_count++;
TAILQ_INSERT_TAIL(&active_kse_groupq,
kseg, kg_qe);
}
}
}
return (kseg);
}
static void
kseg_init(struct kse_group *kseg)
{
kseg_reinit(kseg);
_lock_init(&kseg->kg_lock, LCK_ADAPTIVE, _kse_lock_wait,
_kse_lock_wakeup);
}
static void
kseg_reinit(struct kse_group *kseg)
{
TAILQ_INIT(&kseg->kg_kseq);
TAILQ_INIT(&kseg->kg_threadq);
TAILQ_INIT(&kseg->kg_schedq.sq_waitq);
kseg->kg_threadcount = 0;
kseg->kg_ksecount = 0;
kseg->kg_idle_kses = 0;
kseg->kg_flags = 0;
}
/*
* This must be called with the kse lock held and when there are
* no more threads that reference it.
*/
static void
kseg_free_unlocked(struct kse_group *kseg)
{
TAILQ_REMOVE(&active_kse_groupq, kseg, kg_qe);
TAILQ_INSERT_HEAD(&free_kse_groupq, kseg, kg_qe);
free_kseg_count++;
active_kseg_count--;
}
void
_kseg_free(struct kse_group *kseg)
{
struct kse *curkse;
kse_critical_t crit;
crit = _kse_critical_enter();
curkse = _get_curkse();
KSE_LOCK_ACQUIRE(curkse, &kse_lock);
kseg_free_unlocked(kseg);
KSE_LOCK_RELEASE(curkse, &kse_lock);
_kse_critical_leave(crit);
}
static void
kseg_destroy(struct kse_group *kseg)
{
_lock_destroy(&kseg->kg_lock);
_pq_free(&kseg->kg_schedq.sq_runq);
free(kseg);
}
/*
* Allocate a new KSE.
*
* We allow the current thread to be NULL in the case that this
* is the first time a KSE is being created (library initialization).
* In this case, we don't need to (and can't) take any locks.
*/
struct kse *
_kse_alloc(struct pthread *curthread, int sys_scope)
{
struct kse *kse = NULL;
char *stack;
kse_critical_t crit;
int i;
if ((curthread != NULL) && (free_kse_count > 0)) {
crit = _kse_critical_enter();
KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock);
/* Search for a finished KSE. */
kse = TAILQ_FIRST(&free_kseq);
while ((kse != NULL) &&
((kse->k_kcb->kcb_kmbx.km_flags & KMF_DONE) == 0)) {
kse = TAILQ_NEXT(kse, k_qe);
}
if (kse != NULL) {
DBG_MSG("found an unused kse.\n");
TAILQ_REMOVE(&free_kseq, kse, k_qe);
free_kse_count--;
TAILQ_INSERT_TAIL(&active_kseq, kse, k_qe);
active_kse_count++;
}
KSE_LOCK_RELEASE(curthread->kse, &kse_lock);
_kse_critical_leave(crit);
if (kse != NULL)
kse_reinit(kse, sys_scope);
In the words of the author: o The polling mechanism for I/O readiness was changed from select() to poll(). In additon, a wrapped version of poll() is now provided. o The wrapped select routine now converts each fd_set to a poll array so that the thread scheduler doesn't have to perform a bitwise search for selected fds each time file descriptors are polled for I/O readiness. o The thread scheduler was modified to use a new queue (_workq) for threads that need work. Threads waiting for I/O readiness and spinblocks are added to the work queue in addition to the waiting queue. This reduces the time spent forming/searching the array of file descriptors being polled. o The waiting queue (_waitingq) is now maintained in order of thread wakeup time. This allows the thread scheduler to find the nearest wakeup time by looking at the first thread in the queue instead of searching the entire queue. o Removed file descriptor locking for select/poll routines. An application should not rely on the threads library for providing this locking; if necessary, the application should use mutexes to protect selecting/polling of file descriptors. o Retrieve and use the kernel clock rate/resolution at startup instead of hardcoding the clock resolution to 10 msec (tested with kernel running at 1000 HZ). o All queues have been changed to use queue.h macros. These include the queues of all threads, dead threads, and threads waiting for file descriptor locks. o Added reinitialization of the GC mutex and condition variable after a fork. Also prevented reallocation of the ready queue after a fork. o Prevented the wrapped close routine from closing the thread kernel pipes. o Initialized file descriptor table for stdio entries at thread init. o Provided additional flags to indicate to what queues threads belong. o Moved TAILQ initialization for statically allocated mutex and condition variables to after the spinlock. o Added dispatching of signals to pthread_kill. Removing the dispatching of signals from thread activation broke sigsuspend when pthread_kill was used to send a signal to a thread. o Temporarily set the state of a thread to PS_SUSPENDED when it is first created and placed in the list of threads so that it will not be accidentally scheduled before becoming a member of one of the scheduling queues. o Change the signal handler to queue signals to the thread kernel pipe if the scheduling queues are protected. When scheduling queues are unprotected, signals are then dequeued and handled. o Ensured that all installed signal handlers block the scheduling signal and that the scheduling signal handler blocks all other signals. This ensures that the signal handler is only interruptible for and by non-scheduling signals. An atomic lock is used to decide which instance of the signal handler will handle pending signals. o Removed _lock_thread_list and _unlock_thread_list as they are no longer used to protect the thread list. o Added missing RCS IDs to modified files. o Added checks for appropriate queue membership and activity when adding, removing, and searching the scheduling queues. These checks add very little overhead and are enabled when compiled with _PTHREADS_INVARIANTS defined. Suggested and implemented by Tor Egge with some modification by me. o Close a race condition in uthread_close. (Tor Egge) o Protect the scheduling queues while modifying them in pthread_cond_signal and _thread_fd_unlock. (Tor Egge) o Ensure that when a thread gets a mutex, the mutex is on that threads list of owned mutexes. (Tor Egge) o Set the kernel-in-scheduler flag in _thread_kern_sched_state and _thread_kern_sched_state_unlock to prevent a scheduling signal from calling the scheduler again. (Tor Egge) o Don't use TAILQ_FOREACH macro while searching the waiting queue for threads in a sigwait state, because a change of state destroys the TAILQ link. It is actually safe to do so, though, because once a sigwaiting thread is found, the loop ends and the function returns. (Tor Egge) o When dispatching signals to threads, make the thread inherit the signal deferral flag of the currently running thread. (Tor Egge) Submitted by: Daniel Eischen <eischen@vigrid.com> and Tor Egge <Tor.Egge@fast.no>
1999-06-20 08:28:48 +00:00
}
if ((kse == NULL) &&
((kse = (struct kse *)malloc(sizeof(*kse))) != NULL)) {
if (sys_scope != 0)
stack = NULL;
else if ((stack = malloc(KSE_STACKSIZE)) == NULL) {
free(kse);
return (NULL);
}
bzero(kse, sizeof(*kse));
/* Initialize KCB without the lock. */
if ((kse->k_kcb = _kcb_ctor(kse)) == NULL) {
if (stack != NULL)
free(stack);
free(kse);
return (NULL);
}
/* Initialize the lockusers. */
for (i = 0; i < MAX_KSE_LOCKLEVEL; i++) {
_lockuser_init(&kse->k_lockusers[i], (void *)kse);
_LCK_SET_PRIVATE2(&kse->k_lockusers[i], NULL);
}
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
/* _lock_init(kse->k_lock, ...) */
if (curthread != NULL) {
crit = _kse_critical_enter();
KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock);
}
kse->k_flags = 0;
TAILQ_INSERT_TAIL(&active_kseq, kse, k_qe);
active_kse_count++;
if (curthread != NULL) {
KSE_LOCK_RELEASE(curthread->kse, &kse_lock);
_kse_critical_leave(crit);
}
/*
* Create the KSE context.
* Scope system threads (one thread per KSE) are not required
* to have a stack for an unneeded kse upcall.
Implement zero system call thread switching. Performance of thread switches should be on par with that under scheduler activations. o Timing is achieved through the use of a fixed interval timer (ITIMER_PROF) to count scheduling ticks instead of retrieving the time-of-day upon every thread switch and calculating elapsed real time. o Polling for I/O readiness is performed once for each scheduling tick instead of every thread switch. o The non-signal saving/restoring versions of setjmp/longjmp are used to save and restore thread contexts. This may allow the removal of _THREAD_SAFE macros from setjmp() and longjmp() - needs more investigation. Change signal handling so that signals are handled in the context of the thread that is receiving the signal. When signals are dispatched to a thread, a special signal handling frame is created on top of the target threads stack. The frame contains the threads saved state information and a new context in which the thread can run. The applications signal handler is invoked through a wrapper routine that knows how to restore the threads saved state and unwind to previous frames. Fix interruption of threads due to signals. Some states were being improperly interrupted while other states were not being interrupted. This should fix several PRs. Signal handlers, which are invoked as a result of a process signal (not by pthread_kill()), are now called with the code (or siginfo_t if SA_SIGINFO was set in sa_flags) and sigcontext_t as received from the process signal handler. Modify the search for a thread to which a signal is delivered. The search algorithm is now: o First thread found in sigwait() with signal in wait mask. o First thread found sigsuspend()'d on the signal. o Current thread if signal is unmasked. o First thread found with signal unmasked. Collapse machine dependent support into macros defined in pthread_private.h. These should probably eventually be moved into separate MD files. Change the range of settable priorities to be compliant with POSIX (0-31). The threads library uses higher priorities internally for real-time threads (not yet implemented) and threads executing signal handlers. Real-time threads and threads running signal handlers add 64 and 32, respectively, to a threads base priority. Some other small changes and cleanups. PR: 17757 18559 21943 Reviewed by: jasone
2000-10-13 22:12:32 +00:00
*/
if (!sys_scope) {
kse->k_kcb->kcb_kmbx.km_func = (kse_func_t *)kse_sched_multi;
kse->k_stack.ss_sp = stack;
kse->k_stack.ss_size = KSE_STACKSIZE;
} else {
kse->k_kcb->kcb_kmbx.km_func = (kse_func_t *)kse_sched_single;
kse->k_stack.ss_sp = NULL;
kse->k_stack.ss_size = 0;
}
kse->k_kcb->kcb_kmbx.km_udata = (void *)kse;
kse->k_kcb->kcb_kmbx.km_quantum = 20000;
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
/*
* We need to keep a copy of the stack in case it
* doesn't get used; a KSE running a scope system
* thread will use that thread's stack.
*/
kse->k_kcb->kcb_kmbx.km_stack = kse->k_stack;
}
return (kse);
[ The author's description... ] o Runnable threads are now maintained in priority queues. The implementation requires two things: 1.) The priority queues must be protected during insertion and removal of threads. Since the kernel scheduler must modify the priority queues, a spinlock for protection cannot be used. The functions _thread_kern_sched_defer() and _thread_kern_sched_undefer() were added to {un}defer kernel scheduler activation. 2.) A thread (active) priority change can be performed only when the thread is removed from the priority queue. The implementation uses a threads active priority when inserting it into the queue. A by-product is that thread switches are much faster. A separate queue is used for waiting and/or blocked threads, and it is searched at most 2 times in the kernel scheduler when there are active threads. It should be possible to reduce this to once by combining polling of threads waiting on I/O with the loop that looks for timed out threads and the minimum timeout value. o Functions to defer kernel scheduler activation were added. These are _thread_kern_sched_defer() and _thread_kern_sched_undefer() and may be called recursively. These routines do not block the scheduling signal, but latch its occurrence. The signal handler will not call the kernel scheduler when the running thread has deferred scheduling, but it will be called when running thread undefers scheduling. o Added support for _POSIX_THREAD_PRIORITY_SCHEDULING. All the POSIX routines required by this should now be implemented. One note, SCHED_OTHER, SCHED_FIFO, and SCHED_RR are required to be defined by including pthread.h. These defines are currently in sched.h. I modified pthread.h to include sched.h but don't know if this is the proper thing to do. o Added support for priority protection and inheritence mutexes. This allows definition of _POSIX_THREAD_PRIO_PROTECT and _POSIX_THREAD_PRIO_INHERIT. o Added additional error checks required by POSIX for mutexes and condition variables. o Provided a wrapper for sigpending which is marked as a hidden syscall. o Added a non-portable function as a debugging aid to allow an application to monitor thread context switches. An application can install a routine that gets called everytime a thread (explicitly created by the application) gets context switched. The routine gets passed the pthread IDs of the threads that are being switched in and out. Submitted by: Dan Eischen <eischen@vigrid.com> Changes by me: o Added a PS_SPINBLOCK state to deal with the priority inversion problem most often (I think) seen by threads calling malloc/free/realloc. o Dispatch signals to the running thread directly rather than at a context switch to avoid the situation where the switch never occurs.
1999-03-23 05:07:56 +00:00
}
static void
kse_reinit(struct kse *kse, int sys_scope)
{
if (!sys_scope) {
kse->k_kcb->kcb_kmbx.km_func = (kse_func_t *)kse_sched_multi;
if (kse->k_stack.ss_sp == NULL) {
/* XXX check allocation failure */
kse->k_stack.ss_sp = (char *) malloc(KSE_STACKSIZE);
kse->k_stack.ss_size = KSE_STACKSIZE;
}
kse->k_kcb->kcb_kmbx.km_quantum = 20000;
} else {
kse->k_kcb->kcb_kmbx.km_func = (kse_func_t *)kse_sched_single;
if (kse->k_stack.ss_sp)
free(kse->k_stack.ss_sp);
kse->k_stack.ss_sp = NULL;
kse->k_stack.ss_size = 0;
kse->k_kcb->kcb_kmbx.km_quantum = 0;
}
kse->k_kcb->kcb_kmbx.km_stack = kse->k_stack;
kse->k_kcb->kcb_kmbx.km_udata = (void *)kse;
kse->k_kcb->kcb_kmbx.km_curthread = NULL;
kse->k_kcb->kcb_kmbx.km_flags = 0;
kse->k_curthread = NULL;
kse->k_kseg = 0;
kse->k_schedq = 0;
kse->k_locklevel = 0;
kse->k_flags = 0;
kse->k_idle = 0;
kse->k_error = 0;
kse->k_cpu = 0;
kse->k_done = 0;
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
kse->k_switch = 0;
kse->k_sigseqno = 0;
}
void
kse_free_unlocked(struct kse *kse)
[ The author's description... ] o Runnable threads are now maintained in priority queues. The implementation requires two things: 1.) The priority queues must be protected during insertion and removal of threads. Since the kernel scheduler must modify the priority queues, a spinlock for protection cannot be used. The functions _thread_kern_sched_defer() and _thread_kern_sched_undefer() were added to {un}defer kernel scheduler activation. 2.) A thread (active) priority change can be performed only when the thread is removed from the priority queue. The implementation uses a threads active priority when inserting it into the queue. A by-product is that thread switches are much faster. A separate queue is used for waiting and/or blocked threads, and it is searched at most 2 times in the kernel scheduler when there are active threads. It should be possible to reduce this to once by combining polling of threads waiting on I/O with the loop that looks for timed out threads and the minimum timeout value. o Functions to defer kernel scheduler activation were added. These are _thread_kern_sched_defer() and _thread_kern_sched_undefer() and may be called recursively. These routines do not block the scheduling signal, but latch its occurrence. The signal handler will not call the kernel scheduler when the running thread has deferred scheduling, but it will be called when running thread undefers scheduling. o Added support for _POSIX_THREAD_PRIORITY_SCHEDULING. All the POSIX routines required by this should now be implemented. One note, SCHED_OTHER, SCHED_FIFO, and SCHED_RR are required to be defined by including pthread.h. These defines are currently in sched.h. I modified pthread.h to include sched.h but don't know if this is the proper thing to do. o Added support for priority protection and inheritence mutexes. This allows definition of _POSIX_THREAD_PRIO_PROTECT and _POSIX_THREAD_PRIO_INHERIT. o Added additional error checks required by POSIX for mutexes and condition variables. o Provided a wrapper for sigpending which is marked as a hidden syscall. o Added a non-portable function as a debugging aid to allow an application to monitor thread context switches. An application can install a routine that gets called everytime a thread (explicitly created by the application) gets context switched. The routine gets passed the pthread IDs of the threads that are being switched in and out. Submitted by: Dan Eischen <eischen@vigrid.com> Changes by me: o Added a PS_SPINBLOCK state to deal with the priority inversion problem most often (I think) seen by threads calling malloc/free/realloc. o Dispatch signals to the running thread directly rather than at a context switch to avoid the situation where the switch never occurs.
1999-03-23 05:07:56 +00:00
{
TAILQ_REMOVE(&active_kseq, kse, k_qe);
active_kse_count--;
kse->k_kseg = NULL;
kse->k_kcb->kcb_kmbx.km_quantum = 20000;
kse->k_flags = 0;
TAILQ_INSERT_HEAD(&free_kseq, kse, k_qe);
free_kse_count++;
}
void
_kse_free(struct pthread *curthread, struct kse *kse)
{
kse_critical_t crit;
if (curthread == NULL)
kse_free_unlocked(kse);
else {
crit = _kse_critical_enter();
KSE_LOCK_ACQUIRE(curthread->kse, &kse_lock);
kse_free_unlocked(kse);
KSE_LOCK_RELEASE(curthread->kse, &kse_lock);
_kse_critical_leave(crit);
}
[ The author's description... ] o Runnable threads are now maintained in priority queues. The implementation requires two things: 1.) The priority queues must be protected during insertion and removal of threads. Since the kernel scheduler must modify the priority queues, a spinlock for protection cannot be used. The functions _thread_kern_sched_defer() and _thread_kern_sched_undefer() were added to {un}defer kernel scheduler activation. 2.) A thread (active) priority change can be performed only when the thread is removed from the priority queue. The implementation uses a threads active priority when inserting it into the queue. A by-product is that thread switches are much faster. A separate queue is used for waiting and/or blocked threads, and it is searched at most 2 times in the kernel scheduler when there are active threads. It should be possible to reduce this to once by combining polling of threads waiting on I/O with the loop that looks for timed out threads and the minimum timeout value. o Functions to defer kernel scheduler activation were added. These are _thread_kern_sched_defer() and _thread_kern_sched_undefer() and may be called recursively. These routines do not block the scheduling signal, but latch its occurrence. The signal handler will not call the kernel scheduler when the running thread has deferred scheduling, but it will be called when running thread undefers scheduling. o Added support for _POSIX_THREAD_PRIORITY_SCHEDULING. All the POSIX routines required by this should now be implemented. One note, SCHED_OTHER, SCHED_FIFO, and SCHED_RR are required to be defined by including pthread.h. These defines are currently in sched.h. I modified pthread.h to include sched.h but don't know if this is the proper thing to do. o Added support for priority protection and inheritence mutexes. This allows definition of _POSIX_THREAD_PRIO_PROTECT and _POSIX_THREAD_PRIO_INHERIT. o Added additional error checks required by POSIX for mutexes and condition variables. o Provided a wrapper for sigpending which is marked as a hidden syscall. o Added a non-portable function as a debugging aid to allow an application to monitor thread context switches. An application can install a routine that gets called everytime a thread (explicitly created by the application) gets context switched. The routine gets passed the pthread IDs of the threads that are being switched in and out. Submitted by: Dan Eischen <eischen@vigrid.com> Changes by me: o Added a PS_SPINBLOCK state to deal with the priority inversion problem most often (I think) seen by threads calling malloc/free/realloc. o Dispatch signals to the running thread directly rather than at a context switch to avoid the situation where the switch never occurs.
1999-03-23 05:07:56 +00:00
}
static void
kse_destroy(struct kse *kse)
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
{
int i;
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
if (kse->k_stack.ss_sp != NULL)
free(kse->k_stack.ss_sp);
_kcb_dtor(kse->k_kcb);
for (i = 0; i < MAX_KSE_LOCKLEVEL; ++i)
_lockuser_destroy(&kse->k_lockusers[i]);
_lock_destroy(&kse->k_lock);
free(kse);
}
struct pthread *
_thr_alloc(struct pthread *curthread)
{
kse_critical_t crit;
struct pthread *thread = NULL;
if (curthread != NULL) {
if (GC_NEEDED())
_thr_gc(curthread);
if (free_thread_count > 0) {
crit = _kse_critical_enter();
KSE_LOCK_ACQUIRE(curthread->kse, &thread_lock);
if ((thread = TAILQ_FIRST(&free_threadq)) != NULL) {
TAILQ_REMOVE(&free_threadq, thread, tle);
free_thread_count--;
}
KSE_LOCK_RELEASE(curthread->kse, &thread_lock);
_kse_critical_leave(crit);
}
}
if ((thread == NULL) &&
((thread = malloc(sizeof(struct pthread))) != NULL)) {
bzero(thread, sizeof(struct pthread));
if ((thread->tcb = _tcb_ctor(thread)) == NULL) {
free(thread);
thread = NULL;
}
}
return (thread);
}
void
_thr_free(struct pthread *curthread, struct pthread *thread)
{
kse_critical_t crit;
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
int i;
DBG_MSG("Freeing thread %p\n", thread);
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
if ((curthread == NULL) || (free_thread_count >= MAX_CACHED_THREADS)) {
for (i = 0; i < MAX_THR_LOCKLEVEL; i++) {
_lockuser_destroy(&thread->lockusers[i]);
}
_lock_destroy(&thread->lock);
_tcb_dtor(thread->tcb);
free(thread);
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
}
else {
/* Reinitialize any important fields here. */
thread->lock_switch = 0;
sigemptyset(&thread->sigpend);
thread->check_pending = 0;
/* Add the thread to the free thread list. */
crit = _kse_critical_enter();
Add an i386-specifc hack to always set %gs. There still seems to be instances where the kernel doesn't properly save and/or restore it. Use noupcall and nocompleted flags in the KSE mailbox. These require kernel changes to work which will be committed sometime later. Things still work without the changes. Remove the general kse entry function and use two different functions -- one for scope system threads and one for scope process threads. The scope system function is not yet enabled and we use the same function for all threads at the moment. Keep a copy of the KSE stack for the case that a KSE runs a scope system thread and uses the same stack as the thread (no upcalls are generated, so a separate stack isn't needed). This isn't enabled yet. Use a separate field for the KSE waiting flag. It isn't correct to use the mailbox flags field. The following fixes were provided by David Xu: o Initialize condition variable locks with thread versions of the low-level locking functions instead of the kse versions. o Enable threading before creating the first thread instead of after. o Don't enter critical regions when trying to malloc/free or call functions that malloc/free. o Take the scheduling lock when inheriting thread attributes. o Check the attribute's stack pointer instead of the attributes stack size for null when allocating a thread's stack. o Add a kseg reinit function so we don't have to destroy and then recreate the same lock. o Check the return value of kse_create() and return an appropriate error if it fails. o Don't forget to destroy a thread's locks when freeing it. o Examine the correct flags word for checking to see if a thread is in a synchronization queue. Things should now work on an SMP kernel.
2003-04-21 04:02:56 +00:00
KSE_LOCK_ACQUIRE(curthread->kse, &thread_lock);
TAILQ_INSERT_TAIL(&free_threadq, thread, tle);
free_thread_count++;
KSE_LOCK_RELEASE(curthread->kse, &thread_lock);
_kse_critical_leave(crit);
}
}
/*
* Add an active thread:
*
* o Assign the thread a unique id (which GDB uses to track
* threads.
* o Add the thread to the list of all threads and increment
* number of active threads.
*/
static void
thr_link(struct pthread *thread)
{
kse_critical_t crit;
struct kse *curkse;
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
struct pthread *curthread;
crit = _kse_critical_enter();
curkse = _get_curkse();
o Use a daemon thread to monitor signal events in kernel, if pending signals were changed in kernel, it will retrieve the pending set and try to find a thread to dispatch the signal. The dispatching process can be rolled back if the signal is no longer in kernel. o Create two functions _thr_signal_init() and _thr_signal_deinit(), all signal action settings are retrieved from kernel when threading mode is turned on, after a fork(), child process will reset them to user settings by calling _thr_signal_deinit(). when threading mode is not turned on, all signal operations are direct past to kernel. o When a thread generated a synchoronous signals and its context returned from completed list, UTS will retrieve the signal from its mailbox and try to deliver the signal to thread. o Context signal mask is now only used when delivering signals, thread's current signal mask is always the one in pthread structure. o Remove have_signals field in pthread structure, replace it with psf_valid in pthread_signal_frame. when psf_valid is true, in context switch time, thread will backout itself from some mutex/condition internal queues, then begin to process signals. when a thread is not at blocked state and running, check_pending indicates there are signals for the thread, after preempted and then resumed time, UTS will try to deliver signals to the thread. o At signal delivering time, not only pending signals in thread will be scanned, process's pending signals will be scanned too. o Change sigwait code a bit, remove field sigwait in pthread_wait_data, replace it with oldsigmask in pthread structure, when a thread calls sigwait(), its current signal mask is backuped to oldsigmask, and waitset is copied to its signal mask and when the thread gets a signal in the waitset range, its current signal mask is restored from oldsigmask, these are done in atomic fashion. o Two additional POSIX APIs are implemented, sigwaitinfo() and sigtimedwait(). o Signal code locking is better than previous, there is fewer race conditions. o Temporary disable most of code in _kse_single_thread as it is not safe after fork().
2003-06-28 09:55:02 +00:00
curthread = _get_curthread();
thread->sigmask = curthread->sigmask;
KSE_LOCK_ACQUIRE(curkse, &_thread_list_lock);
/*
* Initialize the unique id (which GDB uses to track
* threads), add the thread to the list of all threads,
* and
*/
thread->uniqueid = next_uniqueid++;
THR_LIST_ADD(thread);
active_threads++;
KSE_LOCK_RELEASE(curkse, &_thread_list_lock);
_kse_critical_leave(crit);
}
/*
* Remove an active thread.
*/
static void
thr_unlink(struct pthread *thread)
{
kse_critical_t crit;
struct kse *curkse;
crit = _kse_critical_enter();
curkse = _get_curkse();
KSE_LOCK_ACQUIRE(curkse, &_thread_list_lock);
THR_LIST_REMOVE(thread);
active_threads--;
KSE_LOCK_RELEASE(curkse, &_thread_list_lock);
_kse_critical_leave(crit);
}
void
_thr_hash_add(struct pthread *thread)
{
struct thread_hash_head *head;
head = &thr_hashtable[THREAD_HASH(thread)];
LIST_INSERT_HEAD(head, thread, hle);
}
void
_thr_hash_remove(struct pthread *thread)
{
LIST_REMOVE(thread, hle);
}
struct pthread *
_thr_hash_find(struct pthread *thread)
{
struct pthread *td;
struct thread_hash_head *head;
head = &thr_hashtable[THREAD_HASH(thread)];
LIST_FOREACH(td, head, hle) {
if (td == thread)
return (thread);
}
return (NULL);
}