From: Jonathan Wakely <jwakely@redhat.com>
To: libstdc++@gcc.gnu.org, gcc-patches@gcc.gnu.org
Cc: Thomas Rodgers <rodgert@twrodgers.com>,
Nate Eldredge <nate@thatsmathematics.com>
Subject: Re: [PATCH 1/2] libstdc++: Atomic wait/notify ABI stabilization
Date: Sat, 9 Mar 2024 12:18:52 +0000 [thread overview]
Message-ID: <CACb0b4=x1YbFFW7arQAjM5EissFsUwTkXwYPS3CK4MRTykmM4Q@mail.gmail.com> (raw)
In-Reply-To: <20231116134736.1287539-1-jwakely@redhat.com>
[-- Attachment #1: Type: text/plain, Size: 36769 bytes --]
On Thu, 16 Nov 2023 at 13:49, Jonathan Wakely <jwakely@redhat.com> wrote:
> From: Thomas Rodgers <trodgers@redhat.com>
>
> These two patches were written by Tom earlier this year, before he left
> Red Hat. We should finish reviewing them for GCC 14 (and probably squash
> them into one?)
>
> Tom, you mentioned further work that changes the __platform_wait_t* to
> uintptr_t, is that ready, or likely to be ready very soon?
>
> Tested x86_64-linux, testing underway on powerpc-aix and sparc-solaris.
>
>
> -- >8 --
>
> This represents a major refactoring of the previous atomic::wait
> and atomic::notify implementation detail. The aim of this change
> is to simplify the implementation details and position the resulting
> implementation so that much of the current header-only detail
> can be moved into the shared library, while also accounting for
> anticipated changes to wait/notify functionality for C++26.
>
> The previous implementation implemented spin logic in terms of
> the types __default_spin_policy, __timed_backoff_spin_policy, and
> the free function __atomic_spin. These are replaced in favor of
> two new free functions; __spin_impl and __spin_until_impl. These
> currently inline free functions are expected to be moved into the
> libstdc++ shared library in a future commit.
>
> The previous implementation derived untimed and timed wait
> implementation detail from __detail::__waiter_pool_base. This
> is-a relationship is removed in the new version and the previous
> implementation detail is renamed to reflect this change. The
> static _S_for member has been renamed as well to indicate that it
> returns the __waiter_pool_impl entry in the static 'side table'
> for a given awaited address.
>
> This new implementation replaces all of the non-templated waiting
> detail of __waiter_base, __waiter_pool, __waiter, __enters_wait, and
> __bare_wait with the __wait_impl free function, and the supporting
> __wait_flags enum and __wait_args struct. This currenly inline free
> function is expected to be moved into the libstdc++ shared library
> in a future commit.
>
> This new implementation replaces all of the non-templated notifying
> detail of __waiter_base, __waiter_pool, and __waiter with the
> __notify_impl free function. This currently inline free function
> is expected to be moved into the libstdc++ shared library in a
> future commit.
>
> The __atomic_wait_address template function is updated to account
> for the above changes and to support the expected C++26 change to
> pass the most recent observed value to the caller supplied predicate.
>
> A new non-templated __atomic_wait_address_v free function is added
> that only works for atomic types that operate only on __platform_wait_t
> and requires the caller to supply a memory order. This is intended
> to be the simplest code path for such types.
>
> The __atomic_wait_address_v template function is now implemented in
> terms of new __atomic_wait_address template and continues to accept
> a user supplied "value function" to retrieve the current value of
> the atomic.
>
> The __atomic_notify_address template function is updated to account
> for the above changes.
>
> The template __platform_wait_until_impl is renamed to
> __wait_clock_t. The previous __platform_wait_until template is deleted
> and the functionality previously provided is moved t the new tempalate
> function __wait_until. A similar change is made to the
> __cond_wait_until_impl/__cond_wait_until implementation.
>
> This new implementation similarly replaces all of the non-templated
> waiting detail of __timed_waiter_pool, __timed_waiter, etc. with
> the new __wait_until_impl free function. This currently inline free
> function is expected to be moved into the libstdc++ shared library
> in a future commit.
>
> This implementation replaces all templated waiting functions that
> manage clock conversion as well as relative waiting (wait_for) with
> the new template functions __wait_until and __wait_for.
>
> Similarly the previous implementation detail for the various
> __atomic_wait_address_Xxx templates is adjusted to account for the
> implementation changes outlined above.
>
> All of the "bare wait" versions of __atomic_wait_Xxx have been removed
> and replaced with a defaulted boolean __bare_wait parameter on the
> new version of these templates.
>
> libstdc++-v3/ChangeLog:
>
> * include/bits/atomic_timed_wait.h:
> (__detail::__platform_wait_until_impl): Rename to
> __platform_wait_until.
> (__detail::__platform_wait_until): Remove previous
> definition.
> (__detail::__cond_wait_until_impl): Rename to
> __cond_wait_until.
> (__detail::__cond_wait_until): Remove previous
> definition.
> (__detail::__spin_until_impl): New function.
> (__detail::__wait_until_impl): New function.
> (__detail::__wait_until): New function.
> (__detail::__wait_for): New function.
> (__detail::__timed_waiter_pool): Remove type.
> (__detail::__timed_backoff_spin_policy): Remove type.
> (__detail::__timed_waiter): Remove type.
> (__detail::__enters_timed_wait): Remove type alias.
> (__detail::__bare_timed_wait): Remove type alias.
> (__atomic_wait_address_until): Adjust to new implementation
> detail.
> (__atomic_wait_address_until_v): Likewise.
> (__atomic_wait_address_bare): Remove.
> (__atomic_wait_address_for): Adjust to new implementation
> detail.
> (__atomic_wait_address_for_v): Likewise.
> (__atomic_wait_address_for_bare): Remove.
> * include/bits/atomic_wait.h: Include bits/stl_pair.h.
> (__detail::__default_spin_policy): Remove type.
> (__detail::__atomic_spin): Remove function.
> (__detail::__waiter_pool_base): Rename to __waiter_pool_impl.
> Remove _M_notify. Rename _S_for to _S_impl_for.
> (__detail::__waiter_base): Remove type.
> (__detail::__waiter_pool): Remove type.
> (__detail::__waiter): Remove type.
> (__detail::__enters_wait): Remove type alias.
> (__detail::__bare_wait): Remove type alias.
> (__detail::__wait_flags): New enum.
> (__detail::__wait_args): New struct.
> (__detail::__wait_result_type): New type alias.
> (__detail::__spin_impl): New function.
> (__detail::__wait_impl): New function.
> (__atomic_wait_address): Adjust to new implementation detail.
> (__atomic_wait_address_v): Likewise.
> (__atomic_notify_address): Likewise.
> (__atomic_wait_address_bare): Delete.
> (__atomic_notify_address_bare): Likewise.
> * include/bits/semaphore_base.h: Adjust implementation to
> use new __atomic_wait_address_v contract.
> * include/std/barrier: Adjust implementation to use new
> __atomic_wait contract.
> * include/std/latch: Adjust implementation to use new
> __atomic_wait contract.
> * testsuite/29_atomics/atomic/wait_notify/100334.cc (main):
> Adjust to for __detail::__waiter_pool_base renaming.
> ---
> libstdc++-v3/include/bits/atomic_timed_wait.h | 549 ++++++++----------
> libstdc++-v3/include/bits/atomic_wait.h | 486 ++++++++--------
> libstdc++-v3/include/bits/semaphore_base.h | 53 +-
> libstdc++-v3/include/std/barrier | 6 +-
> libstdc++-v3/include/std/latch | 5 +-
> .../29_atomics/atomic/wait_notify/100334.cc | 4 +-
> 6 files changed, 514 insertions(+), 589 deletions(-)
>
> diff --git a/libstdc++-v3/include/bits/atomic_timed_wait.h
> b/libstdc++-v3/include/bits/atomic_timed_wait.h
> index ba21ab20a11..90427ef8b42 100644
> --- a/libstdc++-v3/include/bits/atomic_timed_wait.h
> +++ b/libstdc++-v3/include/bits/atomic_timed_wait.h
> @@ -74,62 +74,32 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
> #ifdef _GLIBCXX_HAVE_LINUX_FUTEX
> #define _GLIBCXX_HAVE_PLATFORM_TIMED_WAIT
> // returns true if wait ended before timeout
> - template<typename _Dur>
> - bool
> - __platform_wait_until_impl(const __platform_wait_t* __addr,
> - __platform_wait_t __old,
> - const chrono::time_point<__wait_clock_t,
> _Dur>&
> - __atime) noexcept
> - {
> - auto __s = chrono::time_point_cast<chrono::seconds>(__atime);
> - auto __ns = chrono::duration_cast<chrono::nanoseconds>(__atime -
> __s);
>
For much of the mail below I've removed many of the - lines in the diff,
because the changes are huge and often the removed - lines are completely
unrelated to the added + lines adjacent to them.
Just seeing the new code without unrelated code next to it is clearer.
+ bool
> + __platform_wait_until(const __platform_wait_t* __addr,
> + __platform_wait_t __old,
> + const __wait_clock_t::time_point& __atime)
> noexcept
> + {
> + auto __s = chrono::time_point_cast<chrono::seconds>(__atime);
> + auto __ns = chrono::duration_cast<chrono::nanoseconds>(__atime -
> __s);
>
> - struct timespec __rt =
> + struct timespec __rt =
> {
> static_cast<std::time_t>(__s.time_since_epoch().count()),
> static_cast<long>(__ns.count())
> };
>
> + auto __e = syscall (SYS_futex, __addr,
> +
> static_cast<int>(__futex_wait_flags::__wait_bitset_private),
> + __old, &__rt, nullptr,
> +
> static_cast<int>(__futex_wait_flags::__bitset_match_any));
> + if (__e)
> + {
> + if (errno == ETIMEDOUT)
> return false;
> - }
> + if (errno != EINTR && errno != EAGAIN)
> + __throw_system_error(errno);
> + }
> + return true;
> }
> #else
> // define _GLIBCXX_HAVE_PLATFORM_TIMED_WAIT and implement
> __platform_wait_until()
> @@ -139,15 +109,10 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
>
> #ifdef _GLIBCXX_HAS_GTHREADS
> // Returns true if wait ended before timeout.
>
> + bool
> + __cond_wait_until(__condvar& __cv, mutex& __mx,
> + const __wait_clock_t::time_point& __atime)
> + {
> auto __s = chrono::time_point_cast<chrono::seconds>(__atime);
> auto __ns = chrono::duration_cast<chrono::nanoseconds>(__atime -
> __s);
>
> @@ -158,293 +123,261 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
> };
>
> #ifdef _GLIBCXX_USE_PTHREAD_COND_CLOCKWAIT
> - if constexpr (is_same_v<chrono::steady_clock, _Clock>)
> + if constexpr (is_same_v<chrono::steady_clock, __wait_clock_t>)
> __cv.wait_until(__mx, CLOCK_MONOTONIC, __ts);
> else
> #endif
> __cv.wait_until(__mx, __ts);
>
> + return __wait_clock_t::now() < __atime;
> }
> #endif // _GLIBCXX_HAS_GTHREADS
>
> + inline __wait_result_type
> + __spin_until_impl(const __platform_wait_t* __addr, __wait_args __args,
> + const __wait_clock_t::time_point& __deadline)
> {
>
> + auto __t0 = __wait_clock_t::now();
> + using namespace literals::chrono_literals;
> +
> + __platform_wait_t __val;
> + auto __now = __wait_clock_t::now();
> + for (; __now < __deadline; __now = __wait_clock_t::now())
> + {
> + auto __elapsed = __now - __t0;
> +#ifndef _GLIBCXX_NO_SLEEP
> + if (__elapsed > 128ms)
> + {
> + this_thread::sleep_for(64ms);
> + }
> + else if (__elapsed > 64us)
> + {
> + this_thread::sleep_for(__elapsed / 2);
> + }
> + else
> +#endif
> + if (__elapsed > 4us)
> + {
> + __thread_yield();
> + }
> + else
> + {
> + auto __res = __detail::__spin_impl(__addr, __args);
> + if (__res.first)
> + return __res;
> + }
> +
> + __atomic_load(__addr, &__val, __args._M_order);
> + if (__val != __args._M_old)
> + return make_pair(true, __val);
> + }
> + return make_pair(false, __val);
> + }
> +
> + inline __wait_result_type
> + __wait_until_impl(const __platform_wait_t* __addr, __wait_args __args,
> + const __wait_clock_t::time_point& __atime)
> + {
> +#ifdef _GLIBCXX_HAVE_PLATFORM_TIMED_WAIT
> + __waiter_pool_impl* __pool = nullptr;
> +#else
> + // if we don't have __platform_wait, we always need the side-table
> + __waiter_pool_impl* __pool =
> &__waiter_pool_impl::_S_impl_for(__addr);
> +#endif
> +
> + __platform_wait_t* __wait_addr;
> + if (__args & __wait_flags::__proxy_wait)
> {
> #ifdef _GLIBCXX_HAVE_PLATFORM_TIMED_WAIT
>
>
> + __pool = &__waiter_pool_impl::_S_impl_for(__addr);
> +#endif
> + __wait_addr = &__pool->_M_ver;
> + __atomic_load(__wait_addr, &__args._M_old, __args._M_order);
> }
>
> + else
> + __wait_addr = const_cast<__platform_wait_t*>(__addr);
>
> + if (__args & __wait_flags::__do_spin)
> + {
> + auto __res = __detail::__spin_until_impl(__wait_addr, __args,
> __atime);
> + if (__res.first)
> + return __res;
> + if (__args & __wait_flags::__spin_only)
> + return __res;
>
As observed by Nate Eldredge, this early return is always taken, even if
the caller didn't set __spin_only.
That's because args & __spin_only is args & 12 which is true if do_spin is
set (which it must be or we wouldn't have entered this block).
That means that any call to __wait_impl with __do_spin set in the flags
(which is in the default flags) will return before ever waiting, resulting
in the caller doing a busy loop.
+ }
>
> + if (!(__args & __wait_flags::__track_contention))
>
This condition seems backwards. We want to track contention (i.e. use
_M_enter_wait and _M_leave_wait) when the flag is set, but this does it
when it's not set.
{
>
>
> + // caller does not externally track contention
> +#ifdef _GLIBCXX_HAVE_PLATFORM_TIMED_WAIT
> + __pool = (__pool == nullptr) ?
> &__waiter_pool_impl::_S_impl_for(__addr)
> + : __pool;
>
This would be simpler as just:
if (!__pool)
__pool = ...;
> +#endif
> + __pool->_M_enter_wait();
> }
>
> + __wait_result_type __res;
> +#ifdef _GLIBCXX_HAVE_PLATFORM_TIMED_WAIT
> + if (__platform_wait_until(__wait_addr, __args._M_old, __atime))
> + __res = make_pair(true, __args._M_old);
> + else
> + __res = make_pair(false, __args._M_old);
> +#else
> + __platform_wait_t __val;
> + __atomic_load(__wait_addr, &__val, __args._M_order);
> + if (__val == __args._M_old)
> + {
> + lock_guard<mutex> __l{ __pool->_M_mtx };
> + __atomic_load(__wait_addr, &__val, __args._M_order);
> + if (__val == __args._M_old &&
> + __cond_wait_until(__pool->_M_cv, __pool->_M_mtx, __atime))
> + __res = make_pair(true, __val);
>
There's a missing 'else' here, so we don't set anything in __res in this
case.
I think it probably wants to be {false, __val} ?
+ }
> + else
> + __res = make_pair(false, __val);
> +#endif
> +
> + if (!(__args & __wait_flags::__track_contention))
> + // caller does not externally track contention
> + __pool->_M_leave_wait();
> + return __res;
> + }
> +
> + template<typename _Clock, typename _Dur>
> + __wait_result_type
> + __wait_until(const __platform_wait_t* __addr, __wait_args __args,
> + const chrono::time_point<_Clock, _Dur>& __atime)
> noexcept
> {
>
> + if constexpr (is_same_v<__wait_clock_t, _Clock>)
> + return __detail::__wait_until_impl(__addr, __args, __atime);
> + else
> {
>
> + auto __res = __detail::__wait_until_impl(__addr, __args,
> + __to_wait_clock(__atime));
> + if (!__res.first)
> + {
> + // We got a timeout when measured against __clock_t but
> + // we need to check against the caller-supplied clock
> + // to tell whether we should return a timeout.
> + if (_Clock::now() < __atime)
> + return make_pair(true, __res.second);
> + }
> + return __res;
> + }
> + }
>
>
> + template<typename _Rep, typename _Period>
> + __wait_result_type
> + __wait_for(const __platform_wait_t* __addr, __wait_args __args,
> + const chrono::duration<_Rep, _Period>& __rtime) noexcept
> + {
> + if (!__rtime.count())
> + // no rtime supplied, just spin a bit
> + return __detail::__wait_impl(__addr, __args |
> __wait_flags::__spin_only);
>
This should set __do_spin | __spin_only if the latter no longer implies the
former.
> + auto const __reltime =
> chrono::ceil<__wait_clock_t::duration>(__rtime);
> + auto const __atime = chrono::steady_clock::now() + __reltime;
> + return __detail::__wait_until(__addr, __args, __atime);
> + }
> } // namespace __detail
>
> // returns true if wait ended before timeout
> + template<typename _Tp,
> + typename _Pred, typename _ValFn,
> + typename _Clock, typename _Dur>
> + bool
> + __atomic_wait_address_until(const _Tp* __addr, _Pred&& __pred,
> + _ValFn&& __vfn,
> + const chrono::time_point<_Clock, _Dur>&
> __atime,
> + bool __bare_wait = false) noexcept
> + {
> + const auto __wait_addr =
> + reinterpret_cast<const __detail::__platform_wait_t*>(__addr);
> + __detail::__wait_args __args{ __addr, __bare_wait };
> + _Tp __val = __vfn();
> + while (!__pred(__val))
> + {
> + auto __res = __detail::__wait_until(__wait_addr, __args,
> __atime);
> + if (!__res.first)
> + // timed out
> + return __res.first; // C++26 will also return last observed
> __val
> + __val = __vfn();
> + }
> + return true; // C++26 will also return last observed __val
> + }
> +
> + template<typename _Clock, typename _Dur>
> + bool
> + __atomic_wait_address_until_v(const __detail::__platform_wait_t*
> __addr,
> + __detail::__platform_wait_t __old,
> + int __order,
> + const chrono::time_point<_Clock, _Dur>&
> __atime,
> + bool __bare_wait = false) noexcept
> + {
> + __detail::__wait_args __args{ __addr, __old, __order, __bare_wait };
> + auto __res = __detail::__wait_until(__addr, __args, __atime);
> + return __res.first; // C++26 will also return last observed __val
> + }
> +
> template<typename _Tp, typename _ValFn,
> typename _Clock, typename _Dur>
> bool
> __atomic_wait_address_until_v(const _Tp* __addr, _Tp&& __old,
> _ValFn&& __vfn,
> - const chrono::time_point<_Clock, _Dur>&
> - __atime) noexcept
> + const chrono::time_point<_Clock, _Dur>&
> __atime,
> + bool __bare_wait = false) noexcept
> {
> - __detail::__enters_timed_wait __w{__addr};
> - return __w._M_do_wait_until_v(__old, __vfn, __atime);
> + auto __pfn = [&](const _Tp& __val)
> + { return !__detail::__atomic_compare(__old, __val); };
> + return __atomic_wait_address_until(__addr, __pfn,
> forward<_ValFn>(__vfn),
> + __atime, __bare_wait);
> }
>
> - template<typename _Tp, typename _Pred,
> - typename _Clock, typename _Dur>
> - bool
> - __atomic_wait_address_until(const _Tp* __addr, _Pred __pred,
> - const chrono::time_point<_Clock, _Dur>&
> - __atime)
> noexcept
> - {
> - __detail::__enters_timed_wait __w{__addr};
> - return __w._M_do_wait_until(__pred, __atime);
> - }
> + template<typename _Tp,
> + typename _Pred, typename _ValFn,
> + typename _Rep, typename _Period>
> + bool
> + __atomic_wait_address_for(const _Tp* __addr, _Pred&& __pred,
> + _ValFn&& __vfn,
> + const chrono::duration<_Rep, _Period>&
> __rtime,
> + bool __bare_wait = false) noexcept
> + {
> + const auto __wait_addr =
> + reinterpret_cast<const __detail::__platform_wait_t*>(__addr);
> + __detail::__wait_args __args{ __addr, __bare_wait };
> + _Tp __val = __vfn();
> + while (!__pred(__val))
> + {
> + auto __res = __detail::__wait_for(__wait_addr, __args, __rtime);
> + if (!__res.first)
> + // timed out
> + return __res.first; // C++26 will also return last observed
> __val
> + __val = __vfn();
> + }
> + return true; // C++26 will also return last observed __val
> + }
>
> - template<typename _Pred,
> - typename _Clock, typename _Dur>
> + template<typename _Rep, typename _Period>
> bool
> - __atomic_wait_address_until_bare(const __detail::__platform_wait_t*
> __addr,
> - _Pred __pred,
> - const chrono::time_point<_Clock, _Dur>&
> - __atime)
> noexcept
> - {
> - __detail::__bare_timed_wait __w{__addr};
> - return __w._M_do_wait_until(__pred, __atime);
> - }
> + __atomic_wait_address_for_v(const __detail::__platform_wait_t* __addr,
> + __detail::__platform_wait_t __old,
> + int __order,
> + const chrono::time_point<_Rep, _Period>&
> __rtime,
> + bool __bare_wait = false) noexcept
> + {
> + __detail::__wait_args __args{ __addr, __old, __order, __bare_wait };
> + auto __res = __detail::__wait_for(__addr, __args, __rtime);
> + return __res.first; // C++26 will also return last observed __Val
> + }
>
> template<typename _Tp, typename _ValFn,
> typename _Rep, typename _Period>
> bool
> __atomic_wait_address_for_v(const _Tp* __addr, _Tp&& __old, _ValFn&&
> __vfn,
> - const chrono::duration<_Rep, _Period>& __rtime)
> noexcept
> + const chrono::duration<_Rep, _Period>&
> __rtime,
> + bool __bare_wait = false) noexcept
> {
> - __detail::__enters_timed_wait __w{__addr};
> - return __w._M_do_wait_for_v(__old, __vfn, __rtime);
> - }
> -
> - template<typename _Tp, typename _Pred,
> - typename _Rep, typename _Period>
> - bool
> - __atomic_wait_address_for(const _Tp* __addr, _Pred __pred,
> - const chrono::duration<_Rep, _Period>& __rtime)
> noexcept
> - {
> -
> - __detail::__enters_timed_wait __w{__addr};
> - return __w._M_do_wait_for(__pred, __rtime);
> - }
> -
> - template<typename _Pred,
> - typename _Rep, typename _Period>
> - bool
> - __atomic_wait_address_for_bare(const __detail::__platform_wait_t*
> __addr,
> - _Pred __pred,
> - const chrono::duration<_Rep, _Period>& __rtime)
> noexcept
> - {
> - __detail::__bare_timed_wait __w{__addr};
> - return __w._M_do_wait_for(__pred, __rtime);
> + auto __pfn = [&](const _Tp& __val)
> + { return !__detail::__atomic_compare(__old, __val); };
> + return __atomic_wait_address_for(__addr, __pfn,
> forward<_ValFn>(__vfn),
> + __rtime, __bare_wait);
> }
> _GLIBCXX_END_NAMESPACE_VERSION
> } // namespace std
> diff --git a/libstdc++-v3/include/bits/atomic_wait.h
> b/libstdc++-v3/include/bits/atomic_wait.h
> index 506b7f0cc9d..63d132fc9a8 100644
> --- a/libstdc++-v3/include/bits/atomic_wait.h
> +++ b/libstdc++-v3/include/bits/atomic_wait.h
> @@ -47,7 +47,8 @@
> # include <bits/functexcept.h>
> #endif
>
> -# include <bits/std_mutex.h> // std::mutex, std::__condvar
> +#include <bits/stl_pair.h>
> +#include <bits/std_mutex.h> // std::mutex, std::__condvar
>
> namespace std _GLIBCXX_VISIBILITY(default)
> {
> @@ -131,7 +132,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
> __thread_yield() noexcept
> {
> #if defined _GLIBCXX_HAS_GTHREADS && defined _GLIBCXX_USE_SCHED_YIELD
> - __gthread_yield();
> + __gthread_yield();
> #endif
> }
>
> @@ -188,7 +157,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
> return __builtin_memcmp(&__a, &__b, sizeof(_Tp)) == 0;
> }
>
> - struct __waiter_pool_base
> + struct __waiter_pool_impl
>
Would it make sense to call this type "__wait_proxy" or something like that?
"Waiter pool impl" doesn't really tell me much, where is the waiter pool
that this is implementing?
There used to be __waiter_pool and __waiter_pool_base and __waiter, but now
there's just __waiter_pool_impl which isn't very descriptive in isolation.
> {
> // Don't use std::hardware_destructive_interference_size here
> because we
> // don't want the layout of library types to depend on compiler
> options.
> @@ -205,7 +174,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
> #ifndef _GLIBCXX_HAVE_PLATFORM_WAIT
> __condvar _M_cv;
> #endif
> - __waiter_pool_base() = default;
> + __waiter_pool_impl() = default;
>
> void
> _M_enter_wait() noexcept
>
> -
> - static __waiter_pool_base&
> - _S_for(const void* __addr) noexcept
> + static __waiter_pool_impl&
> + _S_impl_for(const void* __addr) noexcept
> {
> constexpr uintptr_t __ct = 16;
> - static __waiter_pool_base __w[__ct];
> + static __waiter_pool_impl __w[__ct];
> auto __key = (uintptr_t(__addr) >> 2) % __ct;
> return __w[__key];
> }
> };
>
> + enum class __wait_flags : uint32_t
> {
>
> + __abi_version = 0,
> + __proxy_wait = 1,
> + __track_contention = 2,
> + __do_spin = 4,
> + __spin_only = 8 | __do_spin, // implies __do_spin
>
This should be just 8 and not also imply __do_spin.
> + __abi_version_mask = 0xffff0000,
> };
>
> - template<typename _Tp>
> - struct __waiter_base
> + struct __wait_args
> + {
> + __platform_wait_t _M_old;
> + int _M_order = __ATOMIC_ACQUIRE;
> + __wait_flags _M_flags;
> +
> + template<typename _Tp>
> + explicit __wait_args(const _Tp* __addr,
> + bool __bare_wait = false) noexcept
> + : _M_flags{ _S_flags_for(__addr, __bare_wait) }
> + { }
> +
> + __wait_args(const __platform_wait_t* __addr, __platform_wait_t
> __old,
> + int __order, bool __bare_wait = false) noexcept
> + : _M_old{ __old }
> + , _M_order{ __order }
> + , _M_flags{ _S_flags_for(__addr, __bare_wait) }
> + { }
> +
> + __wait_args(const __wait_args&) noexcept = default;
> + __wait_args&
> + operator=(const __wait_args&) noexcept = default;
> +
> + bool
> + operator&(__wait_flags __flag) const noexcept
> {
>
>
> + using __t = underlying_type_t<__wait_flags>;
> + return static_cast<__t>(_M_flags)
> + & static_cast<__t>(__flag);
> + }
>
> + __wait_args
> + operator|(__wait_flags __flag) const noexcept
> + {
> + using __t = underlying_type_t<__wait_flags>;
> + __wait_args __res{ *this };
> + const auto __flags = static_cast<__t>(__res._M_flags)
> + | static_cast<__t>(__flag);
> + __res._M_flags = __wait_flags{ __flags };
> + return __res;
> + }
>
Seems like it would be clearer to define operator| for the actual enum
type, so that these member functions don't need to repeat all the casting
logic.
> + private:
> + static int
> + constexpr _S_default_flags() noexcept
> + {
> + using __t = underlying_type_t<__wait_flags>;
> + return static_cast<__t>(__wait_flags::__abi_version)
> + | static_cast<__t>(__wait_flags::__do_spin);
> + }
>
> + template<typename _Tp>
> + static int
> + constexpr _S_flags_for(const _Tp*, bool __bare_wait) noexcept
> {
>
> + auto __res = _S_default_flags();
> + if (!__bare_wait)
> + __res |= static_cast<int>(__wait_flags::__track_contention);
> + if constexpr (!__platform_wait_uses_type<_Tp>)
> + __res |= static_cast<int>(__wait_flags::__proxy_wait);
> + return __res;
> }
>
+ template<typename _Tp>
> + static int
> + _S_memory_order_for(const _Tp*, int __order) noexcept
> {
>
> + if constexpr (__platform_wait_uses_type<_Tp>)
> + return __order;
> + return __ATOMIC_ACQUIRE;
> + }
> + };
>
This function is never used. Should it be?
Is the idea to force the use of acquire ordering when waiting on a proxy
instead of waiting on the atomic variable directly?
Is that done elsewhere? Can this function be removed, or should it be used
somewhere?
> +
> + using __wait_result_type = pair<bool, __platform_wait_t>;
> + inline __wait_result_type
> + __spin_impl(const __platform_wait_t* __addr, __wait_args __args)
> + {
> + __platform_wait_t __val;
> + for (auto __i = 0; __i < __atomic_spin_count; ++__i)
> + {
> + __atomic_load(__addr, &__val, __args._M_order);
> + if (__val != __args._M_old)
> + return make_pair(true, __val);
> + if (__i < __atomic_spin_count_relax)
> + __detail::__thread_relax();
> + else
> + __detail::__thread_yield();
> + }
> + return make_pair(false, __val);
>
Every use of make_pair can be replaced by just {x, y} to avoid making a
function call (which will compile faster too).
> + }
> +
> + inline __wait_result_type
> + __wait_impl(const __platform_wait_t* __addr, __wait_args __args)
> + {
> +#ifdef _GLIBCXX_HAVE_PLATFORM_WAIT
> + __waiter_pool_impl* __pool = nullptr;
> +#else
> + // if we don't have __platform_wait, we always need the side-table
> + __waiter_pool_impl* __pool =
> &__waiter_pool_impl::_S_impl_for(__addr);
> +#endif
> +
> + __platform_wait_t* __wait_addr;
>
This could be const __platform_wait_t*
> + if (__args & __wait_flags::__proxy_wait)
> + {
> +#ifdef _GLIBCXX_HAVE_PLATFORM_WAIT
> + __pool = &__waiter_pool_impl::_S_impl_for(__addr);
> +#endif
> + __wait_addr = &__pool->_M_ver;
> + __atomic_load(__wait_addr, &__args._M_old, __args._M_order);
> + }
> + else
> + __wait_addr = const_cast<__platform_wait_t*>(__addr);
> +
> + if (__args & __wait_flags::__do_spin)
> + {
> + auto __res = __detail::__spin_impl(__wait_addr, __args);
> + if (__res.first)
> + return __res;
> + if (__args & __wait_flags::__spin_only)
> + return __res;
>
As above, this early return is always taken, resulting in busy loops.
}
>
> + if (!(__args & __wait_flags::__track_contention))
> + {
> + // caller does not externally track contention
> +#ifdef _GLIBCXX_HAVE_PLATFORM_WAIT
> + __pool = (__pool == nullptr) ?
> &__waiter_pool_impl::_S_impl_for(__addr)
> + : __pool;
>
if (!pool)
pool = ...;
> +#endif
> + __pool->_M_enter_wait();
> + }
>
> + __wait_result_type __res;
> +#ifdef _GLIBCXX_HAVE_PLATFORM_WAIT
> + __platform_wait(__wait_addr, __args._M_old);
> + __res = make_pair(false, __args._M_old);
> +#else
> + __platform_wait_t __val;
> + __atomic_load(__wait_addr, &__val, __args._M_order);
> + if (__val == __args._M_old)
> + {
> + lock_guard<mutex> __l{ __pool->_M_mtx };
> + __atomic_load(__wait_addr, &__val, __args._M_order);
> + if (__val == __args._M_old)
> + __pool->_M_cv.wait(__pool->_M_mtx);
> + }
> + __res = make_pair(false, __val);
> +#endif
>
> - using __enters_wait = __waiter<std::true_type>;
> - using __bare_wait = __waiter<std::false_type>;
> + if (!(__args & __wait_flags::__track_contention))
>
Condition is backwards?
> + // caller does not externally track contention
> + __pool->_M_leave_wait();
>
This _M_leave_wait() could be done by an RAII type's destructor instead.
> + return __res;
> + }
> +
> + inline void
> + __notify_impl(const __platform_wait_t* __addr, [[maybe_unused]] bool
> __all,
> + __wait_args __args)
> + {
> +#ifdef _GLIBCXX_HAVE_PLATFORM_WAIT
> + __waiter_pool_impl* __pool = nullptr;
> +#else
> + // if we don't have __platform_notify, we always need the side-table
> + __waiter_pool_impl* __pool =
> &__waiter_pool_impl::_S_impl_for(__addr);
> +#endif
> +
> + if (!(__args & __wait_flags::__track_contention))
>
Condition is backwards?
> + {
> +#ifdef _GLIBCXX_HAVE_PLATFORM_WAIT
> + __pool = &__waiter_pool_impl::_S_impl_for(__addr);
> +#endif
> + if (!__pool->_M_waiting())
> + return;
> + }
> +
> + __platform_wait_t* __wait_addr;
> + if (__args & __wait_flags::__proxy_wait)
> + {
> +#ifdef _GLIBCXX_HAVE_PLATFORM_WAIT
> + __pool = (__pool == nullptr) ?
> &__waiter_pool_impl::_S_impl_for(__addr)
> + : __pool;
>
if (!pool)
pool = ...;
> +#endif
> + __wait_addr = &__pool->_M_ver;
> + __atomic_fetch_add(__wait_addr, 1, __ATOMIC_RELAXED);
> + __all = true;
> + }
>
For a non-proxied wait, __wait_addr is uninitialized here. So the futex
wait uses a random address.
There needs to be `else __wait_addr = __addr;`
+
> +#ifdef _GLIBCXX_HAVE_PLATFORM_WAIT
> + __platform_notify(__wait_addr, __all);
> +#else
> + lock_guard<mutex> __l{ __pool->_M_mtx };
> + __pool->_M_cv.notify_all();
> +#endif
> + }
> } // namespace __detail
>
> + template<typename _Tp,
> + typename _Pred, typename _ValFn>
> + void
> + __atomic_wait_address(const _Tp* __addr,
> + _Pred&& __pred, _ValFn&& __vfn,
> + bool __bare_wait = false) noexcept
> + {
> + const auto __wait_addr =
> + reinterpret_cast<const __detail::__platform_wait_t*>(__addr);
> + __detail::__wait_args __args{ __addr, __bare_wait };
> + _Tp __val = __vfn();
> + while (!__pred(__val))
> + {
>
As observed by Nate, the old value is never set in the __args, so the wait
always uses 0 (the value set in the __wait_args constructor).
It needs:
if constexpr (__platform_wait_uses_type<_Tp>)
__args._M_old = __builtin_bit_cast(__platform_wait_t, __val);
For proxied waits _M_old isn't used, pool->_M_ver is used instead.
> + __detail::__wait_impl(__wait_addr, __args);
> + __val = __vfn();
> + }
> + // C++26 will return __val
> + }
> +
> + inline void
> + __atomic_wait_address_v(const __detail::__platform_wait_t* __addr,
> + __detail::__platform_wait_t __old,
> + int __order)
> + {
> + __detail::__wait_args __args{ __addr, __old, __order };
> + // C++26 will not ignore the return value here
> + __detail::__wait_impl(__addr, __args);
> + }
> +
> template<typename _Tp, typename _ValFn>
> void
> __atomic_wait_address_v(const _Tp* __addr, _Tp __old,
> _ValFn __vfn) noexcept
> {
> diff --git a/libstdc++-v3/include/std/latch
> b/libstdc++-v3/include/std/latch
> index 27cd80d7100..525647c1701 100644
> --- a/libstdc++-v3/include/std/latch
> +++ b/libstdc++-v3/include/std/latch
> @@ -74,8 +74,9 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
> _GLIBCXX_ALWAYS_INLINE void
> wait() const noexcept
> {
> - auto const __pred = [this] { return this->try_wait(); };
> - std::__atomic_wait_address(&_M_a, __pred);
> + auto const __vfn = [this] { return this->try_wait(); };
> + auto const __pred = [this](bool __b) { return __b; };
>
This second lambda doesn't need to capture 'this'.
> + std::__atomic_wait_address(&_M_a, __pred, __vfn);
> }
>
> _GLIBCXX_ALWAYS_INLINE void
>
>
next prev parent reply other threads:[~2024-03-09 12:19 UTC|newest]
Thread overview: 10+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-11-16 13:45 Jonathan Wakely
2023-11-16 13:45 ` [PATCH 2/2] libstdc++: Pass __wait_args to internal API by const pointer Jonathan Wakely
2023-11-16 20:46 ` [PATCH 1/2] libstdc++: Atomic wait/notify ABI stabilization Jonathan Wakely
2024-03-09 12:18 ` Jonathan Wakely [this message]
2024-03-09 12:27 ` Jonathan Wakely
2023-12-11 8:16 Nate Eldredge
2023-12-11 19:24 ` Nate Eldredge
2023-12-11 20:18 ` Jonathan Wakely
2023-12-14 22:23 ` Thomas Rodgers
2023-12-14 23:30 ` Nate Eldredge
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to='CACb0b4=x1YbFFW7arQAjM5EissFsUwTkXwYPS3CK4MRTykmM4Q@mail.gmail.com' \
--to=jwakely@redhat.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=libstdc++@gcc.gnu.org \
--cc=nate@thatsmathematics.com \
--cc=rodgert@twrodgers.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).