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>
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.
signal can arrive before the thread is woken from it's wait4. In this
case, don't return an EINTR, just set the thread state to running and
the wait4 wrapper will loop and get the exit status of the process.
line number every time a file descriptor is locked.
This looks like a big change but it isn't. It should reduce the size
of libc_r and make it run slightly faster.
for the process, not a separate set for each thread). By default, the
process now only has signal handlers installed for SIGVTALRM, SIGINFO
and SIGCHLD. The thread kernel signal handler is installed for other
signals on demand. This means that SIG_IGN and SIG_DFL processing is now
left to the kernel, not the thread kernel.
Change the signal dispatch to no longer use a signal thread, and
call the signal handler using the stack of the thread that has the
signal pending.
Change the atomic lock method to use test-and-set asm code with
a yield if blocked. This introduces separate locks for each type
of object instead of blocking signals to prevent a context
switch. It was this blocking of signals that caused the performance
degradation the people have noted.
This is a *big* change!
Here are the diffs for libc_r to get it one step closer to P1003.1c
These make most of the thread/mutex/condvar structures opaque to the
user. There are three functions which have been renamed with _np
suffixes because they are extensions to P1003.1c (I did them for JAVA,
which needs to suspend/resume threads and also start threads suspended).
I've created a new header (pthread_np.h) for the non-POSIX stuff.
The egrep tags stuff in /usr/src/lib/libc_r/Makefile that I uncommented
doesn't work. I think its best to delete it. I don't think libc_r needs
tags anyway, 'cause most of the source is in libc which does have tags.
also:
Here's the first batch of man pages for the thread functions.
The diff to /usr/src/lib/libc_r/Makefile removes some stuff that was
inherited from /usr/src/lib/libc/Makefile that should only be done with
libc.
also:
I should have sent this diff with the pthread(3) man page.
It allows people to type
make -DWANT_LIBC_R world
to get libc_r built with the rest of the world. I put this in the
pthread(3) man page. The default is still not to build libc_r.
also:
The diff attached adds a pthread(3) man page to /usr/src/share/man/man3.
The idea is that without libc_r installed, this man page will give people
enough info to know that they have to build libc_r.
