binutils-gdb/gdbsupport/thread-pool.cc

252 lines
6.5 KiB
C++
Raw Normal View History

/* Thread pool
Copyright (C) 2019-2024 Free Software Foundation, Inc.
This file is part of GDB.
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>. */
#include "gdbsupport/thread-pool.h"
#if CXX_STD_THREAD
#include "gdbsupport/alt-stack.h"
#include "gdbsupport/block-signals.h"
#include <algorithm>
#include <system_error>
/* On the off chance that we have the pthread library on a Windows
host, but std::thread is not using it, avoid calling
pthread_setname_np on Windows. */
#ifndef _WIN32
#ifdef HAVE_PTHREAD_SETNAME_NP
#define USE_PTHREAD_SETNAME_NP
#endif
#endif
#ifdef USE_PTHREAD_SETNAME_NP
#include <pthread.h>
/* Handle platform discrepancies in pthread_setname_np: macOS uses a
single-argument form, while Linux uses a two-argument form. NetBSD
takes a printf-style format and an argument. This wrapper handles the
difference. */
ATTRIBUTE_UNUSED static void
do_set_thread_name (int (*set_name) (pthread_t, const char *, void *),
const char *name)
{
set_name (pthread_self (), "%s", const_cast<char *> (name));
}
ATTRIBUTE_UNUSED static void
do_set_thread_name (int (*set_name) (pthread_t, const char *),
const char *name)
{
set_name (pthread_self (), name);
}
/* The macOS man page says that pthread_setname_np returns "void", but
the headers actually declare it returning "int". */
ATTRIBUTE_UNUSED static void
do_set_thread_name (int (*set_name) (const char *), const char *name)
{
set_name (name);
}
static void
set_thread_name (const char *name)
{
do_set_thread_name (pthread_setname_np, name);
}
#elif defined (USE_WIN32API)
#include <windows.h>
typedef HRESULT WINAPI (SetThreadDescription_ftype) (HANDLE, PCWSTR);
static SetThreadDescription_ftype *dyn_SetThreadDescription;
static bool initialized;
static void
init_windows ()
{
initialized = true;
HMODULE hm = LoadLibrary (TEXT ("kernel32.dll"));
if (hm)
dyn_SetThreadDescription
= (SetThreadDescription_ftype *) GetProcAddress (hm,
"SetThreadDescription");
/* On some versions of Windows, this function is only available in
KernelBase.dll, not kernel32.dll. */
if (dyn_SetThreadDescription == nullptr)
{
hm = LoadLibrary (TEXT ("KernelBase.dll"));
if (hm)
dyn_SetThreadDescription
= (SetThreadDescription_ftype *) GetProcAddress (hm,
"SetThreadDescription");
}
}
static void
do_set_thread_name (const wchar_t *name)
{
if (!initialized)
init_windows ();
if (dyn_SetThreadDescription != nullptr)
dyn_SetThreadDescription (GetCurrentThread (), name);
}
#define set_thread_name(NAME) do_set_thread_name (L ## NAME)
#else /* USE_WIN32API */
static void
set_thread_name (const char *name)
{
}
#endif
#endif /* CXX_STD_THREAD */
namespace gdb
{
/* The thread pool detach()s its threads, so that the threads will not
prevent the process from exiting. However, it was discovered that
if any detached threads were still waiting on a condition variable,
then the condition variable's destructor would wait for the threads
to exit -- defeating the purpose.
Allocating the thread pool on the heap and simply "leaking" it
avoids this problem.
*/
thread_pool *thread_pool::g_thread_pool = new thread_pool ();
thread_pool::~thread_pool ()
{
/* Because this is a singleton, we don't need to clean up. The
threads are detached so that they won't prevent process exit.
And, cleaning up here would be actively harmful in at least one
case -- see the comment by the definition of g_thread_pool. */
}
void
thread_pool::set_thread_count (size_t num_threads)
{
#if CXX_STD_THREAD
std::lock_guard<std::mutex> guard (m_tasks_mutex);
gdb: Enable early init of thread pool size This commit enables the early initialization commands (92e4e97a9f5) to modify the number of threads used by gdb's thread pool. The motivation here is to prevent gdb from spawning a detrimental number of threads on many-core systems under environments with restrictive ulimits. With gdb before this commit, the thread pool takes the following sizes: 1. Thread pool size is initialized to 0. 2. After the maintenance commands are defined, the thread pool size is set to the number of system cores (if it has not already been set). 3. Using early initialization commands, the thread pool size can be changed using "maint set worker-threads". 4. After the first prompt, the thread pool size can be changed as in the previous step. Therefore after step 2. gdb has potentially launched hundreds of threads on a many-core system. After this change, step 2 and 3 are reversed so there is an opportunity to set the required number of threads without needing to default to the number of system cores first. There does exist a configure option (added in 261b07488b9) to disable multithreading, but this does not allow for an already deployed gdb to be configured. Additionally, the default number of worker threads is clamped at eight to control the number of worker threads spawned on many-core systems. This value was chosen as testing recorded on bugzilla issue 29959 indicates that parallel efficiency drops past this point. GDB built with GCC 13. No test suite regressions detected. Compilers: GCC, ACfL, Intel, Intel LLVM, NVHPC; Platforms: x86_64, aarch64. The scenario that interests me the most involves preventing GDB from spawning any worker threads at all. This was tested by counting the number of clones observed by strace: strace -e clone,clone3 gdb/gdb -q \ --early-init-eval-command="maint set worker-threads 0" \ -ex q ./gdb/gdb |& grep --count clone The new test relies on "gdb: install CLI uiout while processing early init files" developed by Andrew Burgess. This patch will need pushing prior to this change. The clamping was tested on machines with both 16 cores and a single core. "maint show worker-threads" correctly reported eight and one respectively. Approved-By: Tom Tromey <tom@tromey.com>
2023-12-04 22:23:17 +08:00
m_sized_at_least_once = true;
/* If the new size is larger, start some new threads. */
if (m_thread_count < num_threads)
{
/* Ensure that signals used by gdb are blocked in the new
threads. */
block_signals blocker;
for (size_t i = m_thread_count; i < num_threads; ++i)
{
try
{
std::thread thread (&thread_pool::thread_function, this);
thread.detach ();
}
catch (const std::system_error &)
{
/* libstdc++ may not implement std::thread, and will
throw an exception on use. It seems fine to ignore
this, and any other sort of startup failure here. */
num_threads = i;
break;
}
}
}
/* If the new size is smaller, terminate some existing threads. */
if (num_threads < m_thread_count)
{
for (size_t i = num_threads; i < m_thread_count; ++i)
m_tasks.emplace ();
m_tasks_cv.notify_all ();
}
m_thread_count = num_threads;
#else
/* No threads available, simply ignore the request. */
#endif /* CXX_STD_THREAD */
}
#if CXX_STD_THREAD
void
thread_pool::do_post_task (std::packaged_task<void ()> &&func)
{
gdb: Enable early init of thread pool size This commit enables the early initialization commands (92e4e97a9f5) to modify the number of threads used by gdb's thread pool. The motivation here is to prevent gdb from spawning a detrimental number of threads on many-core systems under environments with restrictive ulimits. With gdb before this commit, the thread pool takes the following sizes: 1. Thread pool size is initialized to 0. 2. After the maintenance commands are defined, the thread pool size is set to the number of system cores (if it has not already been set). 3. Using early initialization commands, the thread pool size can be changed using "maint set worker-threads". 4. After the first prompt, the thread pool size can be changed as in the previous step. Therefore after step 2. gdb has potentially launched hundreds of threads on a many-core system. After this change, step 2 and 3 are reversed so there is an opportunity to set the required number of threads without needing to default to the number of system cores first. There does exist a configure option (added in 261b07488b9) to disable multithreading, but this does not allow for an already deployed gdb to be configured. Additionally, the default number of worker threads is clamped at eight to control the number of worker threads spawned on many-core systems. This value was chosen as testing recorded on bugzilla issue 29959 indicates that parallel efficiency drops past this point. GDB built with GCC 13. No test suite regressions detected. Compilers: GCC, ACfL, Intel, Intel LLVM, NVHPC; Platforms: x86_64, aarch64. The scenario that interests me the most involves preventing GDB from spawning any worker threads at all. This was tested by counting the number of clones observed by strace: strace -e clone,clone3 gdb/gdb -q \ --early-init-eval-command="maint set worker-threads 0" \ -ex q ./gdb/gdb |& grep --count clone The new test relies on "gdb: install CLI uiout while processing early init files" developed by Andrew Burgess. This patch will need pushing prior to this change. The clamping was tested on machines with both 16 cores and a single core. "maint show worker-threads" correctly reported eight and one respectively. Approved-By: Tom Tromey <tom@tromey.com>
2023-12-04 22:23:17 +08:00
/* This assert is here to check that no tasks are posted to the pool between
its initialization and sizing. */
gdb_assert (m_sized_at_least_once);
std::packaged_task<void ()> t (std::move (func));
if (m_thread_count != 0)
{
std::lock_guard<std::mutex> guard (m_tasks_mutex);
m_tasks.emplace (std::move (t));
m_tasks_cv.notify_one ();
}
else
{
/* Just execute it now. */
t ();
}
}
void
thread_pool::thread_function ()
{
/* This must be done here, because on macOS one can only set the
name of the current thread. */
set_thread_name ("gdb worker");
/* Ensure that SIGSEGV is delivered to an alternate signal
stack. */
gdb::alternate_signal_stack signal_stack;
while (true)
{
std::optional<task_t> t;
{
/* We want to hold the lock while examining the task list, but
not while invoking the task function. */
std::unique_lock<std::mutex> guard (m_tasks_mutex);
while (m_tasks.empty ())
m_tasks_cv.wait (guard);
t = std::move (m_tasks.front());
m_tasks.pop ();
}
if (!t.has_value ())
break;
(*t) ();
}
}
#endif /* CXX_STD_THREAD */
} /* namespace gdb */