Re: [PATCH v4 00/15] Add futex2 syscalls

From: "André Almeida" <andrealmeid@collabora.com>
To: Zebediah Figura <z.figura12@gmail.com>,
	Thomas Gleixner <tglx@linutronix.de>,
	Ingo Molnar <mingo@redhat.com>,
	Peter Zijlstra <peterz@infradead.org>,
	Darren Hart <dvhart@infradead.org>,
	linux-kernel@vger.kernel.org,
	Steven Rostedt <rostedt@goodmis.org>,
	Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: kernel@collabora.com, krisman@collabora.com,
	pgriffais@valvesoftware.com, joel@joelfernandes.org,
	malteskarupke@fastmail.fm, linux-api@vger.kernel.org,
	fweimer@redhat.com, libc-alpha@sourceware.org,
	linux-kselftest@vger.kernel.org, shuah@kernel.org,
	acme@kernel.org, corbet@lwn.net, Peter Oskolkov <posk@posk.io>,
	Andrey Semashev <andrey.semashev@gmail.com>,
	Davidlohr Bueso <dave@stgolabs.net>
Subject: Re: [PATCH v4 00/15] Add futex2 syscalls
Date: Fri, 4 Jun 2021 14:04:21 -0300	[thread overview]
Message-ID: <d3d2a8f9-4a5e-fff5-c0f3-2563366b4c11@collabora.com> (raw)
In-Reply-To: <dab34fd2-b494-8686-bcd7-68beeba4f386@gmail.com>

Às 01:51 de 04/06/21, Zebediah Figura escreveu:
> On 6/3/21 2:59 PM, André Almeida wrote:
>>   ** The wait on multiple problem
>>
>>   The use case lies in the Wine implementation of the Windows NT
>> interface
>>   WaitMultipleObjects. This Windows API function allows a thread to sleep
>>   waiting on the first of a set of event sources (mutexes, timers,
>> signal,
>>   console input, etc) to signal.  Considering this is a primitive
>>   synchronization operation for Windows applications, being able to
>> quickly
>>   signal events on the producer side, and quickly go to sleep on the
>>   consumer side is essential for good performance of those running
>> over Wine.
>>
> 
> I know this is part of the cover letter, but I really do want to clarify
> that this isn't really accurate. The use case that this is referring to
> is not "the Wine implementation of WaitForMultipleObjects", it is an
> out-of-tree implementation of WaitForMultipleObjects that provides
> improved performance compared to the in-tree implementation.
> 
> This is especially salient because:
> 
> (1) this out-of-tree implementation is only in a small handful of cases
> any more performant than a different out-of-tree implementation which
> uses eventfd and poll() instead;
> 
> (2) these implementations will remain out-of-tree due to compatibility
> and robustness problems;
> 
> (3) I believe there is potential for an upstreamable implementation
> which does not rely on futex or futex2.

I'll let it more clear next time that this applies to Proton's Wine, and
not Wine.

Along with that, wait on multiple will be useful for other workloads,
such as the ones that uses Boost's mass locking algorithms and native
game engines for instance.