LKML Archive on lore.kernel.org
help / color / mirror / Atom feed
* [PATCH 0/4] futex2: Add wait on multiple futexes syscall
@ 2021-08-05 19:04 André Almeida
  2021-08-05 19:04 ` [PATCH 1/4] futex: Prepare for futex_wait_multiple() André Almeida
                   ` (3 more replies)
  0 siblings, 4 replies; 11+ messages in thread
From: André Almeida @ 2021-08-05 19:04 UTC (permalink / raw)
  To: Thomas Gleixner, Ingo Molnar, Peter Zijlstra, Darren Hart,
	linux-kernel, Steven Rostedt, Sebastian Andrzej Siewior
  Cc: kernel, krisman, linux-api, libc-alpha, mtk.manpages,
	Davidlohr Bueso, André Almeida

Hi,

As opposed to previous futex2 patchsets, this one adds only one syscall:
futex_waitv(). This way we can focus on this operation that already have
a well defined use case and has been tested for months now.

The patchset reuses as much as possible of original futex code for the new
operation, so the first commit move some stuff to futex header to make
accessible for futex2.

Ideally, I would completely replace futex_wait_setup() with
futex_wait_multiple(): it is basic the same logic, but for n futexes,
so for existing operations it was a matter of calling it with nr_futexes=1.
This worked pretty well for futex_wait(): I tested with glibc tests,
tested with a complete distro running on top of it and perf benchs
presented no performance difference. However, it didn't work for
futex_wait_requeue_pi(), since the wait path for it is slightly different
of the normal wait, that would require some refactor to get it in a way to
be easily replaced. So I decided to not replace it at all.

 Use case

The use case of this syscall is to allow low level locking libraries to
wait for multiple locks at the same time. This is specially useful for
emulating Windows' WaitForMultipleObjects. A futex_waitv()-based solution
has been used for some time at Proton's Wine (a compatibility layer to
run Windows games on Linux). Compared to a solution that uses eventfd(),
futex was able to reduce CPU utilization for games, and even increase
frames per second for some games. This happens because eventfd doesn't
scale very well for a huge number of read, write and poll calls compared
to futex. Native game engines will benefit of this as well, given that
this wait pattern is common for games.

 Testing

Selftest is provided as part of this patchset. As stated above, I used
the futex_wait_multiple() in FUTEX_WAIT path and it worked fine in a
full distro. Throught Proton, I've tested futex_waitv() with modern games
that issue more than 40k futex calls per second.

André Almeida (4):
  futex: Prepare for futex_wait_multiple()
  futex2: Implement vectorized wait
  selftests: futex2: Add waitv test
  futex2: Documentation: Document futex_waitv() uAPI

 Documentation/userspace-api/futex2.rst        |  79 ++++++
 Documentation/userspace-api/index.rst         |   1 +
 arch/x86/entry/syscalls/syscall_32.tbl        |   1 +
 arch/x86/entry/syscalls/syscall_64.tbl        |   1 +
 include/linux/compat.h                        |   9 +
 include/linux/futex.h                         |  75 ++++++
 include/uapi/asm-generic/unistd.h             |   5 +-
 include/uapi/linux/futex.h                    |  17 ++
 init/Kconfig                                  |   7 +
 kernel/Makefile                               |   1 +
 kernel/futex.c                                | 224 ++++++++++++++----
 kernel/futex2.c                               | 198 ++++++++++++++++
 kernel/sys_ni.c                               |   4 +
 .../selftests/futex/functional/.gitignore     |   1 +
 .../selftests/futex/functional/Makefile       |   3 +-
 .../selftests/futex/functional/futex2_waitv.c | 154 ++++++++++++
 .../testing/selftests/futex/functional/run.sh |   3 +
 .../selftests/futex/include/futex2test.h      |  72 ++++++
 18 files changed, 812 insertions(+), 43 deletions(-)
 create mode 100644 Documentation/userspace-api/futex2.rst
 create mode 100644 kernel/futex2.c
 create mode 100644 tools/testing/selftests/futex/functional/futex2_waitv.c
 create mode 100644 tools/testing/selftests/futex/include/futex2test.h

-- 
2.32.0


^ permalink raw reply	[flat|nested] 11+ messages in thread

end of thread, other threads:[~2021-08-19  3:39 UTC | newest]

Thread overview: 11+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-08-05 19:04 [PATCH 0/4] futex2: Add wait on multiple futexes syscall André Almeida
2021-08-05 19:04 ` [PATCH 1/4] futex: Prepare for futex_wait_multiple() André Almeida
2021-08-18  8:22   ` Thomas Gleixner
2021-08-05 19:04 ` [PATCH 2/4] futex2: Implement vectorized wait André Almeida
2021-08-18 11:00   ` Thomas Gleixner
2021-08-18 16:20     ` Peter Zijlstra
2021-08-18 16:34       ` André Almeida
2021-08-18 19:45       ` Thomas Gleixner
2021-08-19  3:38         ` André Almeida
2021-08-05 19:04 ` [PATCH 3/4] selftests: futex2: Add waitv test André Almeida
2021-08-05 19:04 ` [PATCH 4/4] futex2: Documentation: Document futex_waitv() uAPI André Almeida

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 NNTP newsgroup(s).