freebsd-dev/lib/libc/sys/kse.2
2012-03-29 16:20:20 +00:00

680 lines
21 KiB
Groff

.\" Copyright (c) 2002 Packet Design, LLC.
.\" All rights reserved.
.\"
.\" Subject to the following obligations and disclaimer of warranty,
.\" use and redistribution of this software, in source or object code
.\" forms, with or without modifications are expressly permitted by
.\" Packet Design; provided, however, that:
.\"
.\" (i) Any and all reproductions of the source or object code
.\" must include the copyright notice above and the following
.\" disclaimer of warranties; and
.\" (ii) No rights are granted, in any manner or form, to use
.\" Packet Design trademarks, including the mark "PACKET DESIGN"
.\" on advertising, endorsements, or otherwise except as such
.\" appears in the above copyright notice or in the software.
.\"
.\" THIS SOFTWARE IS BEING PROVIDED BY PACKET DESIGN "AS IS", AND
.\" TO THE MAXIMUM EXTENT PERMITTED BY LAW, PACKET DESIGN MAKES NO
.\" REPRESENTATIONS OR WARRANTIES, EXPRESS OR IMPLIED, REGARDING
.\" THIS SOFTWARE, INCLUDING WITHOUT LIMITATION, ANY AND ALL IMPLIED
.\" WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE,
.\" OR NON-INFRINGEMENT. PACKET DESIGN DOES NOT WARRANT, GUARANTEE,
.\" OR MAKE ANY REPRESENTATIONS REGARDING THE USE OF, OR THE RESULTS
.\" OF THE USE OF THIS SOFTWARE IN TERMS OF ITS CORRECTNESS, ACCURACY,
.\" RELIABILITY OR OTHERWISE. IN NO EVENT SHALL PACKET DESIGN BE
.\" LIABLE FOR ANY DAMAGES RESULTING FROM OR ARISING OUT OF ANY USE
.\" OF THIS SOFTWARE, INCLUDING WITHOUT LIMITATION, ANY DIRECT,
.\" INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, PUNITIVE, OR CONSEQUENTIAL
.\" DAMAGES, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES, LOSS OF
.\" USE, DATA OR PROFITS, HOWEVER CAUSED AND UNDER 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 PACKET DESIGN IS ADVISED OF
.\" THE POSSIBILITY OF SUCH DAMAGE.
.\"
.\" $FreeBSD$
.\"
.Dd February 13, 2007
.Dt KSE 2
.Os
.Sh NAME
.Nm kse
.Nd "kernel support for user threads"
.Sh LIBRARY
.Lb libc
.Sh SYNOPSIS
.In sys/types.h
.In sys/kse.h
.Ft int
.Fn kse_create "struct kse_mailbox *mbx" "int sys-scope"
.Ft int
.Fn kse_exit void
.Ft int
.Fn kse_release "struct timespec *timeout"
.Ft int
.Fn kse_switchin "struct kse_thr_mailbox *tmbx" "int flags"
.Ft int
.Fn kse_thr_interrupt "struct kse_thr_mailbox *tmbx" "int cmd" "long data"
.Ft int
.Fn kse_wakeup "struct kse_mailbox *mbx"
.Sh DESCRIPTION
These system calls implement kernel support for multi-threaded processes.
.\"
.Ss Overview
.\"
Traditionally, user threading has been implemented in one of two ways:
either all threads are managed in user space and the kernel is unaware
of any threading (also known as
.Dq "N to 1" ) ,
or else separate processes sharing
a common memory space are created for each thread (also known as
.Dq "N to N" ) .
These approaches have advantages and disadvantages:
.Bl -column "- Cannot utilize multiple CPUs" "+ Can utilize multiple CPUs"
.It Sy "User threading Kernel threading"
.It "+ Lightweight - Heavyweight"
.It "+ User controls scheduling - Kernel controls scheduling"
.It "- Syscalls must be wrapped + No syscall wrapping required"
.It "- Cannot utilize multiple CPUs + Can utilize multiple CPUs"
.El
.Pp
The KSE system is a
hybrid approach that achieves the advantages of both the user and kernel
threading approaches.
The underlying philosophy of the KSE system is to give kernel support
for user threading without taking away any of the user threading library's
ability to make scheduling decisions.
A kernel-to-user upcall mechanism is used to pass control to the user
threading library whenever a scheduling decision needs to be made.
An arbitrarily number of user threads are multiplexed onto a fixed number of
virtual CPUs supplied by the kernel.
This can be thought of as an
.Dq "N to M"
threading scheme.
.Pp
Some general implications of this approach include:
.Bl -bullet
.It
The user process can run multiple threads simultaneously on multi-processor
machines.
The kernel grants the process virtual CPUs to schedule as it
wishes; these may run concurrently on real CPUs.
.It
All operations that block in the kernel become asynchronous, allowing
the user process to schedule another thread when any thread blocks.
.El
.\"
.Ss Definitions
.\"
KSE allows a user process to have multiple
.Sy threads
of execution in existence at the same time, some of which may be blocked
in the kernel while others may be executing or blocked in user space.
A
.Sy "kernel scheduling entity"
(KSE) is a
.Dq "virtual CPU"
granted to the process for the purpose of executing threads.
A thread that is currently executing is always associated with
exactly one KSE, whether executing in user space or in the kernel.
The KSE is said to be
.Sy assigned
to the thread.
KSEs (a user abstraction) are implemented on top
of kernel threads using an 'upcall' entity.
.Pp
The KSE becomes
.Sy unassigned ,
and the associated thread is suspended, when the KSE has an associated
.Sy mailbox ,
(see below) the thread has an associated
.Sy thread mailbox ,
(also see below) and any of the following occurs:
.Bl -bullet
.It
The thread invokes a system call that blocks.
.It
The thread makes any other demand of the kernel that cannot be immediately
satisfied, e.g., touches a page of memory that needs to be fetched from disk,
causing a page fault.
.It
Another thread that was previously blocked in the kernel completes its
work in the kernel (or is
.Sy interrupted )
and becomes ready to return to user space, and the current thread is returning
to user space.
.It
A signal is delivered to the process, and this KSE is chosen to deliver it.
.El
.Pp
In other words, as soon as there is a scheduling decision to be made,
the KSE becomes unassigned, because the kernel does not presume to know
how the process' other runnable threads should be scheduled.
Unassigned KSEs always return to user space as soon as possible via
the
.Sy upcall
mechanism (described below), allowing the user process to decide how
that KSE should be utilized next.
KSEs always complete as much work as possible in the kernel before
becoming unassigned.
.Pp
Individual KSEs within a process are effectively indistinguishable,
and any KSE in a process may be assigned by the kernel to any runnable
(in the kernel) thread associated with that process.
In practice, the kernel attempts to preserve the affinity between threads
and actual CPUs to optimize cache behavior, but this is invisible to the
user process.
(Affinity is not yet fully implemented.)
.Pp
Each KSE has a unique
.Sy "KSE mailbox"
supplied by the user process.
A mailbox consists of a control structure containing a pointer to an
.Sy "upcall function"
and a user stack.
The KSE invokes this function whenever it becomes unassigned.
The kernel updates this structure with information about threads that have
become runnable and signals that have been delivered before each upcall.
Upcalls may be temporarily blocked by the user thread scheduling code
during critical sections.
.Pp
Each user thread has a unique
.Sy "thread mailbox"
as well.
Threads are referred to using pointers to these mailboxes when communicating
between the kernel and the user thread scheduler.
Each KSE's mailbox contains a pointer to the mailbox of the user thread
that the KSE is currently executing.
This pointer is saved when the thread blocks in the kernel.
.Pp
Whenever a thread blocked in the kernel is ready to return to user space,
it is added to the process's list of
.Sy completed
threads.
This list is presented to the user code at the next upcall as a linked list
of thread mailboxes.
.Pp
There is a kernel-imposed limit on the number of threads in a process
that may be simultaneously blocked in the kernel (this number is not
currently visible to the user).
When this limit is reached, upcalls are blocked and no work is performed
for the process until one of the threads completes (or a signal is
received).
.\"
.Ss Managing KSEs
.\"
To become multi-threaded, a process must first invoke
.Fn kse_create .
The
.Fn kse_create
system call
creates a new KSE (except for the very first invocation; see below).
The KSE will be associated with the mailbox pointed to by
.Fa mbx .
If
.Fa sys_scope
is non-zero, then the new thread will be counted as a system scope
thread. Other things must be done as well to make a system scope thread
so this is not sufficient (yet).
System scope variables are not covered
in detail in this manual page yet, but briefly, they never perform
upcalls and do not return to the user thread scheduler.
Once launched they run autonomously.
The pthreads library knows how to make system
scope threads and users are encouraged to use the library interface.
.Pp
Each process initially has a single KSE executing a single user thread.
Since the KSE does not have an associated mailbox, it must remain assigned
to the thread and does not perform any upcalls.
(It is by definition a system scope thread).
The result is the traditional, unthreaded mode of operation.
Therefore, as a special case, the first call to
.Fn kse_create
by this initial thread with
.Fa sys_scope
equal to zero does not create a new KSE; instead, it simply associates the
current KSE with the supplied KSE mailbox, and no immediate upcall results.
However, an upcall will be triggered the next time the thread blocks and
the required conditions are met.
.Pp
The kernel does not allow more KSEs to exist in a process than the
number of physical CPUs in the system (this number is available as the
.Xr sysctl 3
variable
.Va hw.ncpu ) .
Having more KSEs than CPUs would not add any value to the user process,
as the additional KSEs would just compete with each other for access to
the real CPUs.
Since the extra KSEs would always be side-lined, the result
to the application would be exactly the same as having fewer KSEs.
There may however be arbitrarily many user threads, and it is up to the
user thread scheduler to handle mapping the application's user threads
onto the available KSEs.
.Pp
The
.Fn kse_exit
system call
causes the KSE assigned to the currently running thread to be destroyed.
If this KSE is the last one in the process, there must be no remaining
threads associated with that process blocked in the kernel.
This system call does not return unless there is an error.
Calling
.Fn kse_exit
from the last thread is the same as calling
.Fn exit .
.Pp
The
.Fn kse_release
system call
is used to
.Dq park
the KSE assigned to the currently running thread when it is not needed,
e.g., when there are more available KSEs than runnable user threads.
The thread converts to an upcall but does not get scheduled until
there is a new reason to do so, e.g., a previously
blocked thread becomes runnable, or the timeout expires.
If successful,
.Fn kse_release
does not return to the caller.
.Pp
The
.Fn kse_switchin
system call can be used by the UTS, when it has selected a new thread,
to switch to the context of that thread.
The use of
.Fn kse_switchin
is machine dependent.
Some platforms do not need a system call to switch to a new context,
while others require its use in particular cases.
.Pp
The
.Fn kse_wakeup
system call
is the opposite of
.Fn kse_release .
It causes the (parked) KSE associated with the mailbox pointed to by
.Fa mbx
to be woken up, causing it to upcall.
If the KSE has already woken up for another reason, this system call has no
effect.
The
.Fa mbx
argument
may be
.Dv NULL
to specify
.Dq "any KSE in the current process" .
.Pp
The
.Fn kse_thr_interrupt
system call
is used to interrupt a currently blocked thread.
The thread must either be blocked in the kernel or assigned to a KSE
(i.e., executing).
The thread is then marked as interrupted.
As soon as the thread invokes an interruptible system call (or immediately
for threads already blocked in one), the thread will be made runnable again,
even though the kernel operation may not have completed.
The effect on the interrupted system call is the same as if it had been
interrupted by a signal; typically this means an error is returned with
.Va errno
set to
.Er EINTR .
.\"
.Ss Signals
.\"
The current implementation creates a special signal thread.
Kernel threads (KSEs) in a process mask all signals, and only the signal
thread waits for signals to be delivered to the process, the signal thread
is responsible
for dispatching signals to user threads.
.Pp
A downside of this is that if a multiplexed thread
calls the
.Fn execve
syscall, its signal mask and pending signals may not be
available in the kernel.
They are stored
in userland and the kernel does not know where to get them, however
.Tn POSIX
requires them to be restored and passed them to new process.
Just setting the mask for the thread before calling
.Fn execve
is only a
close approximation to the problem as it does not re-deliver back to the kernel
any pending signals that the old process may have blocked, and it allows a
window in which new signals may be delivered to the process between the setting
of the mask and the
.Fn execve .
.Pp
For now this problem has been solved by adding a special combined
.Fn kse_thr_interrupt Ns / Ns Fn execve
mode to the
.Fn kse_thr_interrupt
syscall.
The
.Fn kse_thr_interrupt
syscall has a sub command
.Dv KSE_INTR_EXECVE ,
that allows it to accept a
.Vt kse_execv_args
structure, and allowing it to adjust the signals and then atomically
convert into an
.Fn execve
call.
Additional pending signals and the correct signal mask can be passed
to the kernel in this way.
The thread library overrides the
.Fn execve
syscall
and translates it into
.Fn kse_intr_interrupt
call, allowing a multiplexed thread
to restore pending signals and the correct signal mask before doing the
.Fn exec .
This solution to the problem may change.
.\"
.Ss KSE Mailboxes
.\"
Each KSE has a unique mailbox for user-kernel communication defined in
.In sys/kse.h .
Some of the fields there are:
.Pp
.Va km_version
describes the version of this structure and must be equal to
.Dv KSE_VER_0 .
.Va km_udata
is an opaque pointer ignored by the kernel.
.Pp
.Va km_func
points to the KSE's upcall function;
it will be invoked using
.Va km_stack ,
which must remain valid for the lifetime of the KSE.
.Pp
.Va km_curthread
always points to the thread that is currently assigned to this KSE if any,
or
.Dv NULL
otherwise.
This field is modified by both the kernel and the user process as follows.
.Pp
When
.Va km_curthread
is not
.Dv NULL ,
it is assumed to be pointing at the mailbox for the currently executing
thread, and the KSE may be unassigned, e.g., if the thread blocks in the
kernel.
The kernel will then save the contents of
.Va km_curthread
with the blocked thread, set
.Va km_curthread
to
.Dv NULL ,
and upcall to invoke
.Fn km_func .
.Pp
When
.Va km_curthread
is
.Dv NULL ,
the kernel will never perform any upcalls with this KSE; in other words,
the KSE remains assigned to the thread even if it blocks.
.Va km_curthread
must be
.Dv NULL
while the KSE is executing critical user thread scheduler
code that would be disrupted by an intervening upcall;
in particular, while
.Fn km_func
itself is executing.
.Pp
Before invoking
.Fn km_func
in any upcall, the kernel always sets
.Va km_curthread
to
.Dv NULL .
Once the user thread scheduler has chosen a new thread to run,
it should point
.Va km_curthread
at the thread's mailbox, re-enabling upcalls, and then resume the thread.
.Em Note :
modification of
.Va km_curthread
by the user thread scheduler must be atomic
with the loading of the context of the new thread, to avoid
the situation where the thread context area
may be modified by a blocking async operation, while there
is still valid information to be read out of it.
.Pp
.Va km_completed
points to a linked list of user threads that have completed their work
in the kernel since the last upcall.
The user thread scheduler should put these threads back into its
own runnable queue.
Each thread in a process that completes a kernel operation
(synchronous or asynchronous) that results in an upcall is guaranteed to be
linked into exactly one KSE's
.Va km_completed
list; which KSE in the group, however, is indeterminate.
Furthermore, the completion will be reported in only one upcall.
.Pp
.Va km_sigscaught
contains the list of signals caught by this process since the previous
upcall to any KSE in the process.
As long as there exists one or more KSEs with an associated mailbox in
the user process, signals are delivered this way rather than the
traditional way.
(This has not been implemented and may change.)
.Pp
.Va km_timeofday
is set by the kernel to the current system time before performing
each upcall.
.Pp
.Va km_flags
may contain any of the following bits OR'ed together:
.Bl -tag -width indent
.It Dv KMF_NOUPCALL
Block upcalls from happening.
The thread is in some critical section.
.It Dv KMF_NOCOMPLETED , KMF_DONE , KMF_BOUND
This thread should be considered to be permanently bound to
its KSE, and treated much like a non-threaded process would be.
It is a
.Dq "long term"
version of
.Dv KMF_NOUPCALL
in some ways.
.It Dv KMF_WAITSIGEVENT
Implement characteristics needed for the signal delivery thread.
.El
.\"
.Ss Thread Mailboxes
.\"
Each user thread must have associated with it a unique
.Vt "struct kse_thr_mailbox"
as defined in
.In sys/kse.h .
It includes the following fields.
.Pp
.Va tm_udata
is an opaque pointer ignored by the kernel.
.Pp
.Va tm_context
stores the context for the thread when the thread is blocked in user space.
This field is also updated by the kernel before a completed thread is returned
to the user thread scheduler via
.Va km_completed .
.Pp
.Va tm_next
links the
.Va km_completed
threads together when returned by the kernel with an upcall.
The end of the list is marked with a
.Dv NULL
pointer.
.Pp
.Va tm_uticks
and
.Va tm_sticks
are time counters for user mode and kernel mode execution, respectively.
These counters count ticks of the statistics clock (see
.Xr clocks 7 ) .
While any thread is actively executing in the kernel, the corresponding
.Va tm_sticks
counter is incremented.
While any KSE is executing in user space and that KSE's
.Va km_curthread
pointer is not equal to
.Dv NULL ,
the corresponding
.Va tm_uticks
counter is incremented.
.Pp
.Va tm_flags
may contain any of the following bits OR'ed together:
.Bl -tag -width indent
.It Dv TMF_NOUPCALL
Similar to
.Dv KMF_NOUPCALL .
This flag inhibits upcalling for critical sections.
Some architectures require this to be in one place and some in the other.
.El
.Sh RETURN VALUES
The
.Fn kse_create ,
.Fn kse_wakeup ,
and
.Fn kse_thr_interrupt
system calls
return zero if successful.
The
.Fn kse_exit
and
.Fn kse_release
system calls
do not return if successful.
.Pp
All of these system calls return a non-zero error code in case of an error.
.Sh ERRORS
The
.Fn kse_create
system call
will fail if:
.Bl -tag -width Er
.It Bq Er ENXIO
There are already as many KSEs in the process as hardware processors.
.It Bq Er EAGAIN
The user is not the super user, and the soft resource limit corresponding
to the
.Fa resource
argument
.Dv RLIMIT_NPROC
would be exceeded (see
.Xr getrlimit 2 ) .
.It Bq Er EFAULT
The
.Fa mbx
argument
points to an address which is not a valid part of the process address space.
.El
.Pp
The
.Fn kse_exit
system call
will fail if:
.Bl -tag -width Er
.It Bq Er EDEADLK
The current KSE is the last in its process and there are still one or more
threads associated with the process blocked in the kernel.
.It Bq Er ESRCH
The current KSE has no associated mailbox, i.e., the process is operating
in traditional, unthreaded mode (in this case use
.Xr _exit 2
to exit the process).
.El
.Pp
The
.Fn kse_release
system call
will fail if:
.Bl -tag -width Er
.It Bq Er ESRCH
The current KSE has no associated mailbox, i.e., the process is operating is
traditional, unthreaded mode.
.El
.Pp
The
.Fn kse_wakeup
system call
will fail if:
.Bl -tag -width Er
.It Bq Er ESRCH
The
.Fa mbx
argument
is not
.Dv NULL
and the mailbox pointed to by
.Fa mbx
is not associated with any KSE in the process.
.It Bq Er ESRCH
The
.Fa mbx
argument
is
.Dv NULL
and the current KSE has no associated mailbox, i.e., the process is operating
in traditional, unthreaded mode.
.El
.Pp
The
.Fn kse_thr_interrupt
system call
will fail if:
.Bl -tag -width Er
.It Bq Er ESRCH
The thread corresponding to
.Fa tmbx
is neither currently assigned to any KSE in the process nor blocked in the
kernel.
.El
.Sh SEE ALSO
.Xr rfork 2 ,
.Xr pthread 3 ,
.Xr ucontext 3
.Rs
.%A "Thomas E. Anderson"
.%A "Brian N. Bershad"
.%A "Edward D. Lazowska"
.%A "Henry M. Levy"
.%J "ACM Transactions on Computer Systems"
.%N Issue 1
.%V Volume 10
.%D February 1992
.%I ACM Press
.%P pp. 53-79
.%T "Scheduler activations: effective kernel support for the user-level management of parallelism"
.Re
.Sh HISTORY
The KSE system calls first appeared in
.Fx 5.0 .
.Sh AUTHORS
KSE was originally implemented by
.An -nosplit
.An "Julian Elischer" Aq julian@FreeBSD.org ,
with additional contributions by
.An "Jonathan Mini" Aq mini@FreeBSD.org ,
.An "Daniel Eischen" Aq deischen@FreeBSD.org ,
and
.An "David Xu" Aq davidxu@FreeBSD.org .
.Pp
This manual page was written by
.An "Archie Cobbs" Aq archie@FreeBSD.org .
.Sh BUGS
The KSE code is
.Ud