From mboxrd@z Thu Jan 1 00:00:00 1970 From: Nicholas Piggin Subject: Re: [RFC PATCH 4/7] x86: use exit_lazy_tlb rather than membarrier_mm_sync_core_before_usermode Date: Tue, 21 Jul 2020 20:04:27 +1000 Message-ID: <1595324577.x3bf55tpgu.astroid@bobo.none> References: <1594868476.6k5kvx8684.astroid@bobo.none> <20200716085032.GO10769@hirez.programming.kicks-ass.net> <1594892300.mxnq3b9a77.astroid@bobo.none> <20200716110038.GA119549@hirez.programming.kicks-ass.net> <1594906688.ikv6r4gznx.astroid@bobo.none> <1314561373.18530.1594993363050.JavaMail.zimbra@efficios.com> <1595213677.kxru89dqy2.astroid@bobo.none> <2055788870.20749.1595263590675.JavaMail.zimbra@efficios.com> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: quoted-printable Return-path: Received: from lindbergh.monkeyblade.net ([23.128.96.19]:41418 "EHLO lindbergh.monkeyblade.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1727942AbgGUKEh (ORCPT ); Tue, 21 Jul 2020 06:04:37 -0400 In-Reply-To: <2055788870.20749.1595263590675.JavaMail.zimbra@efficios.com> Sender: linux-arch-owner@vger.kernel.org List-ID: To: Mathieu Desnoyers Cc: Anton Blanchard , Arnd Bergmann , Jens Axboe , linux-arch , linux-kernel , linux-mm , linuxppc-dev , Andy Lutomirski , Andy Lutomirski , Peter Zijlstra , x86 Excerpts from Mathieu Desnoyers's message of July 21, 2020 2:46 am: > ----- On Jul 19, 2020, at 11:03 PM, Nicholas Piggin npiggin@gmail.com wro= te: >=20 >> Excerpts from Mathieu Desnoyers's message of July 17, 2020 11:42 pm: >>> ----- On Jul 16, 2020, at 7:26 PM, Nicholas Piggin npiggin@gmail.com wr= ote: >>> [...] >>>>=20 >>>> membarrier does replace barrier instructions on remote CPUs, which do >>>> order accesses performed by the kernel on the user address space. So >>>> membarrier should too I guess. >>>>=20 >>>> Normal process context accesses like read(2) will do so because they >>>> don't get filtered out from IPIs, but kernel threads using the mm may >>>> not. >>>=20 >>> But it should not be an issue, because membarrier's ordering is only wi= th >>> respect >>> to submit and completion of io_uring requests, which are performed thro= ugh >>> system calls from the context of user-space threads, which are called f= rom the >>> right mm. >>=20 >> Is that true? Can io completions be written into an address space via a >> kernel thread? I don't know the io_uring code well but it looks like >> that's asynchonously using the user mm context. >=20 > Indeed, the io completion appears to be signaled asynchronously between k= ernel > and user-space. Yep, many other places do similar with use_mm. [snip] > So as far as membarrier memory ordering dependencies are concerned, it re= lies > on the store-release/load-acquire dependency chain in the completion queu= e to > order against anything that was done prior to the completed requests. >=20 > What is in-flight while the requests are being serviced provides no memor= y > ordering guarantee whatsoever. Yeah you're probably right in this case I think. Quite likely most kernel=20 tasks that asynchronously write to user memory would at least have some=20 kind of producer-consumer barriers. But is that restriction of all async modifications documented and enforced anywhere? >> How about other memory accesses via kthread_use_mm? Presumably there is >> still ordering requirement there for membarrier, >=20 > Please provide an example case with memory accesses via kthread_use_mm wh= ere > ordering matters to support your concern. I think the concern Andy raised with io_uring was less a specific=20 problem he saw and more a general concern that we have these memory=20 accesses which are not synchronized with membarrier. >> so I really think >> it's a fragile interface with no real way for the user to know how >> kernel threads may use its mm for any particular reason, so membarrier >> should synchronize all possible kernel users as well. >=20 > I strongly doubt so, but perhaps something should be clarified in the doc= umentation > if you have that feeling. I'd rather go the other way and say if you have reasoning or numbers for=20 why PF_KTHREAD is an important optimisation above rq->curr =3D=3D rq->idle then we could think about keeping this subtlety with appropriate=20 documentation added, otherwise we can just kill it and remove all doubt. That being said, the x86 sync core gap that I imagined could be fixed=20 by changing to rq->curr =3D=3D rq->idle test does not actually exist becaus= e the global membarrier does not have a sync core option. So fixing the exit_lazy_tlb points that this series does *should* fix that. So PF_KTHREAD may be less problematic than I thought from implementation point of view, only semantics. Thanks, Nick