freebsd-nq/lib/libthr/libthr.3

.\" Copyright (c) 2005 Robert N. M. Watson
.\" Copyright (c) 2014,2015,2021 The FreeBSD Foundation, Inc.
.\" All rights reserved.
.\"
.\" Part of this documentation was written by
.\" Konstantin Belousov <kib@FreeBSD.org> under sponsorship
.\" from the FreeBSD Foundation.
.\"
.\" 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.
.\"
.\" THIS SOFTWARE IS PROVIDED BY THE AUTHORS 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
.\" ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS 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.
.\"
.\" $FreeBSD$
.\"
.Dd October 1, 2021
.Dt LIBTHR 3
.Os
.Sh NAME
.Nm libthr
.Nd "1:1 POSIX threads library"
.Sh LIBRARY
.Lb libthr
.Sh SYNOPSIS
.In pthread.h
.Sh DESCRIPTION
The
.Nm
library provides a 1:1 implementation of the
.Xr pthread 3
library interfaces for application threading.
It
has been optimized for use by applications expecting system scope thread
semantics.
.Pp
The library is tightly integrated with the run-time link editor
.Xr ld-elf.so.1 1
and
.Lb libc ;
all three components must be built from the same source tree.
Mixing
.Li libc
and
.Nm
libraries from different versions of
.Fx
is not supported.
The run-time linker
.Xr ld-elf.so.1 1
has some code to ensure backward-compatibility with older versions of
.Nm .
.Pp
The man page documents the quirks and tunables of the
.Nm .
When linking with
.Li -lpthread ,
the run-time dependency
.Li libthr.so.3
is recorded in the produced object.
.Sh MUTEX ACQUISITION
A locked mutex (see
.Xr pthread_mutex_lock 3 )
is represented by a volatile variable of type
.Dv lwpid_t ,
which records the global system identifier of the thread
owning the lock.
.Nm
performs a contested mutex acquisition in three stages, each of which
is more resource-consuming than the previous.
The first two stages are only applied for a mutex of
.Dv PTHREAD_MUTEX_ADAPTIVE_NP
type and
.Dv PTHREAD_PRIO_NONE
protocol (see
.Xr pthread_mutexattr 3 ) .
.Pp
First, on SMP systems, a spin loop
is performed, where the library attempts to acquire the lock by
.Xr atomic 9
operations.
The loop count is controlled by the
.Ev LIBPTHREAD_SPINLOOPS
environment variable, with a default value of 2000.
.Pp
If the spin loop
was unable to acquire the mutex, a yield loop
is executed, performing the same
.Xr atomic 9
acquisition attempts as the spin loop,
but each attempt is followed by a yield of the CPU time
of the thread using the
.Xr sched_yield 2
syscall.
By default, the yield loop
is not executed.
This is controlled by the
.Ev LIBPTHREAD_YIELDLOOPS
environment variable.
.Pp
If both the spin and yield loops
failed to acquire the lock, the thread is taken off the CPU and
put to sleep in the kernel with the
.Xr _umtx_op 2
syscall.
The kernel wakes up a thread and hands the ownership of the lock to
the woken thread when the lock becomes available.
.Sh THREAD STACKS
Each thread is provided with a private user-mode stack area
used by the C runtime.
The size of the main (initial) thread stack is set by the kernel, and is
controlled by the
.Dv RLIMIT_STACK
process resource limit (see
.Xr getrlimit 2 ) .
.Pp
By default, the main thread's stack size is equal to the value of
.Dv RLIMIT_STACK
for the process.
If the
.Ev LIBPTHREAD_SPLITSTACK_MAIN
environment variable is present in the process environment
(its value does not matter),
the main thread's stack is reduced to 4MB on 64bit architectures, and to
2MB on 32bit architectures, when the threading library is initialized.
The rest of the address space area which has been reserved by the
kernel for the initial process stack is used for non-initial thread stacks
in this case.
The presence of the
.Ev LIBPTHREAD_BIGSTACK_MAIN
environment variable overrides
.Ev LIBPTHREAD_SPLITSTACK_MAIN ;
it is kept for backward-compatibility.
.Pp
The size of stacks for threads created by the process at run-time
with the
.Xr pthread_create 3
call is controlled by thread attributes: see
.Xr pthread_attr 3 ,
in particular, the
.Xr pthread_attr_setstacksize 3 ,
.Xr pthread_attr_setguardsize 3
and
.Xr pthread_attr_setstackaddr 3
functions.
If no attributes for the thread stack size are specified, the default
non-initial thread stack size is 2MB for 64bit architectures, and 1MB
for 32bit architectures.
.Sh RUN-TIME SETTINGS
The following environment variables are recognized by
.Nm
and adjust the operation of the library at run-time:
.Bl -tag -width "Ev LIBPTHREAD_SPLITSTACK_MAIN"
.It Ev LIBPTHREAD_BIGSTACK_MAIN
Disables the reduction of the initial thread stack enabled by
.Ev LIBPTHREAD_SPLITSTACK_MAIN .
.It Ev LIBPTHREAD_SPLITSTACK_MAIN
Causes a reduction of the initial thread stack, as described in the
section
.Sx THREAD STACKS .
This was the default behaviour of
.Nm
before
.Fx 11.0 .
.It Ev LIBPTHREAD_SPINLOOPS
The integer value of the variable overrides the default count of
iterations in the
.Li spin loop
of the mutex acquisition.
The default count is 2000, set by the
.Dv MUTEX_ADAPTIVE_SPINS
constant in the
.Nm
sources.
.It Ev LIBPTHREAD_YIELDLOOPS
A non-zero integer value enables the yield loop
in the process of the mutex acquisition.
The value is the count of loop operations.
.It Ev LIBPTHREAD_QUEUE_FIFO
The integer value of the variable specifies how often blocked
threads are inserted at the head of the sleep queue, instead of its tail.
Bigger values reduce the frequency of the FIFO discipline.
The value must be between 0 and 255.
.Pp
.El
The following
.Dv sysctl
MIBs affect the operation of the library:
.Bl -tag -width "Dv debug.umtx.robust_faults_verbose"
.It Dv kern.ipc.umtx_vnode_persistent
By default, a shared lock backed by a mapped file in memory is
automatically destroyed on the last unmap of the corresponding file's page,
which is allowed by POSIX.
Setting the sysctl to 1 makes such a shared lock object persist until
the vnode is recycled by the Virtual File System.
Note that in case file is not opened and not mapped, the kernel might
recycle it at any moment, making this sysctl less useful than it sounds.
.It Dv kern.ipc.umtx_max_robust
The maximal number of robust mutexes allowed for one thread.
The kernel will not unlock more mutexes than specified, see
.Xr _umtx_op
for more details.
The default value is large enough for most useful applications.
.It Dv debug.umtx.robust_faults_verbose
A non zero value makes kernel emit some diagnostic when the robust
mutexes unlock was prematurely aborted after detecting some inconsistency,
as a measure to prevent memory corruption.
.El
.Pp
The
.Dv RLIMIT_UMTXP
limit (see
.Xr getrlimit 2 )
defines how many shared locks a given user may create simultaneously.
.Sh INTERACTION WITH RUN-TIME LINKER
On load,
.Nm
installs interposing handlers into the hooks exported by
.Li libc .
The interposers provide real locking implementation instead of the
stubs for single-threaded processes in
.Li libc ,
cancellation support and some modifications to the signal operations.
.Pp
.Nm
cannot be unloaded; the
.Xr dlclose 3
function does not perform any action when called with a handle for
.Nm .
One of the reasons is that the internal interposing of
.Li libc
functions cannot be undone.
.Sh SIGNALS
The implementation interposes the user-installed
.Xr signal 3
handlers.
This interposing is done to postpone signal delivery to threads which
entered (libthr-internal) critical sections, where the calling
of the user-provided signal handler is unsafe.
An example of such a situation is owning the internal library lock.
When a signal is delivered while the signal handler cannot be safely
called, the call is postponed and performed until after the exit from
the critical section.
This should be taken into account when interpreting
.Xr ktrace 1
logs.
.Sh PROCESS-SHARED SYNCHRONIZATION OBJECTS
In the
.Li libthr
implementation,
user-visible types for all synchronization objects (e.g. pthread_mutex_t)
are pointers to internal structures, allocated either by the corresponding
.Fn pthread_<objtype>_init
method call, or implicitly on first use when a static initializer
was specified.
The initial implementation of process-private locking object used this
model with internal allocation, and the addition of process-shared objects
was done in a way that did not break the application binary interface.
.Pp
For process-private objects, the internal structure is allocated using
either
.Xr malloc 3
or, for
.Xr pthread_mutex_init 3 ,
an internal memory allocator implemented in
.Nm .
The internal allocator for mutexes is used to avoid bootstrap issues
with many
.Xr malloc 3
implementations which need working mutexes to function.
The same allocator is used for thread-specific data, see
.Xr pthread_setspecific 3 ,
for the same reason.
.Pp
For process-shared objects, the internal structure is created by first
allocating a shared memory segment using
.Xr _umtx_op 2
operation
.Dv UMTX_OP_SHM ,
and then mapping it into process address space with
.Xr mmap 2
with the
.Dv MAP_SHARED
flag.
The POSIX standard requires that:
.Bd -literal
only the process-shared synchronization object itself can be used for
performing synchronization.  It need not be referenced at the address
used to initialize it (that is, another mapping of the same object can
be used).
.Ed
.Pp
With the
.Fx
implementation, process-shared objects require initialization
in each process that use them.
In particular, if you map the shared memory containing the user portion of
a process-shared object already initialized in different process, locking
functions do not work on it.
.Pp
Another broken case is a forked child creating the object in memory shared
with the parent, which cannot be used from parent.
Note that processes should not use non-async-signal safe functions after
.Xr fork 2
anyway.
.Sh SEE ALSO
.Xr ktrace 1 ,
.Xr ld-elf.so.1 1 ,
.Xr getrlimit 2 ,
.Xr errno 2 ,
.Xr thr_exit 2 ,
.Xr thr_kill 2 ,
.Xr thr_kill2 2 ,
.Xr thr_new 2 ,
.Xr thr_self 2 ,
.Xr thr_set_name 2 ,
.Xr _umtx_op 2 ,
.Xr dlclose 3 ,
.Xr dlopen 3 ,
.Xr getenv 3 ,
.Xr pthread_attr 3 ,
.Xr pthread_attr_setstacksize 3 ,
.Xr pthread_create 3 ,
.Xr signal 3 ,
.Xr atomic 9
.Sh HISTORY
The
.Nm
library first appeared in
.Fx 5.2 .
.Sh AUTHORS
.An -nosplit
The
.Nm
library
was originally created by
.An Jeff Roberson Aq Mt jeff@FreeBSD.org ,
and enhanced by
.An Jonathan Mini Aq Mt mini@FreeBSD.org
and
.An Mike Makonnen Aq Mt mtm@FreeBSD.org .
It has been substantially rewritten and optimized by
.An David Xu Aq Mt davidxu@FreeBSD.org .