Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1751798AbdHSMvU (ORCPT ); Sat, 19 Aug 2017 08:51:20 -0400 Received: from mail-oi0-f66.google.com ([209.85.218.66]:34490 "EHLO mail-oi0-f66.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751038AbdHSMvT (ORCPT ); Sat, 19 Aug 2017 08:51:19 -0400 MIME-Version: 1.0 In-Reply-To: <20170818234348.GE11771@tardis> References: <1502089981-21272-1-git-send-email-byungchul.park@lge.com> <1502089981-21272-10-git-send-email-byungchul.park@lge.com> <20170818234348.GE11771@tardis> From: Arnd Bergmann Date: Sat, 19 Aug 2017 14:51:17 +0200 X-Google-Sender-Auth: NVZRi-XQt9SNLFmbblrbpz1gPEk Message-ID: Subject: Re: [PATCH v8 09/14] lockdep: Apply crossrelease to completions To: Boqun Feng Cc: Byungchul Park , Peter Zijlstra , Ingo Molnar , Thomas Gleixner , Michel Lespinasse , kirill@shutemov.name, Linux Kernel Mailing List , Linux-MM , Andrew Morton , willy@infradead.org, Nicholas Piggin , kernel-team@lge.com Content-Type: text/plain; charset="UTF-8" Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 6249 Lines: 140 On Sat, Aug 19, 2017 at 1:43 AM, Boqun Feng wrote: > Hi Arnd, > > On Mon, Aug 14, 2017 at 10:50:24AM +0200, Arnd Bergmann wrote: >> On Mon, Aug 7, 2017 at 9:12 AM, Byungchul Park wrote: >> > Although wait_for_completion() and its family can cause deadlock, the >> > lock correctness validator could not be applied to them until now, >> > because things like complete() are usually called in a different context >> > from the waiting context, which violates lockdep's assumption. >> > >> > Thanks to CONFIG_LOCKDEP_CROSSRELEASE, we can now apply the lockdep >> > detector to those completion operations. Applied it. >> > >> > Signed-off-by: Byungchul Park >> >> This patch introduced a significant growth in kernel stack usage for a small >> set of functions. I see two new warnings for functions that get tipped over the >> 1024 or 2048 byte frame size limit in linux-next (with a few other patches >> applied): >> >> Before: >> >> drivers/md/dm-integrity.c: In function 'write_journal': >> drivers/md/dm-integrity.c:827:1: error: the frame size of 504 bytes is >> larger than xxx bytes [-Werror=frame-larger-than=] >> drivers/mmc/core/mmc_test.c: In function 'mmc_test_area_io_seq': >> drivers/mmc/core/mmc_test.c:1491:1: error: the frame size of 680 bytes >> is larger than 104 bytes [-Werror=frame-larger-than=] >> >> After: >> >> drivers/md/dm-integrity.c: In function 'write_journal': >> drivers/md/dm-integrity.c:827:1: error: the frame size of 1280 bytes >> is larger than 1024 bytes [-Werror=frame-larger-than=] >> drivers/mmc/core/mmc_test.c: In function 'mmc_test_area_io_seq': >> drivers/mmc/core/mmc_test.c:1491:1: error: the frame size of 1072 >> bytes is larger than 1024 bytes [-Werror=frame-larger-than=] >> >> I have not checked in detail why this happens, but I'm guessing that >> there is an overall increase in stack usage with >> CONFIG_LOCKDEP_COMPLETE in functions using completions, >> and I think it would be good to try to come up with a version that doesn't >> add as much. >> > > So I have been staring at this for a while, and below is what I found: > > (BTW, Arnd, may I know your compiler version? Mine is 7.1.1) That is what I used as well, on x86, arm32 and arm64. > In write_journal(), I can see the code generated like this on x86: > > io_comp.comp = COMPLETION_INITIALIZER_ONSTACK(io_comp.comp); > 2462: e8 00 00 00 00 callq 2467 > 2467: 48 8d 85 80 fd ff ff lea -0x280(%rbp),%rax > 246e: 48 c7 c6 00 00 00 00 mov $0x0,%rsi > 2475: 48 c7 c2 00 00 00 00 mov $0x0,%rdx > x->done = 0; > 247c: c7 85 90 fd ff ff 00 movl $0x0,-0x270(%rbp) > 2483: 00 00 00 > init_waitqueue_head(&x->wait); > 2486: 48 8d 78 18 lea 0x18(%rax),%rdi > 248a: e8 00 00 00 00 callq 248f > if (commit_start + commit_sections <= ic->journal_sections) { > 248f: 41 8b 87 a8 00 00 00 mov 0xa8(%r15),%eax > io_comp.comp = COMPLETION_INITIALIZER_ONSTACK(io_comp.comp); > 2496: 48 8d bd e8 f9 ff ff lea -0x618(%rbp),%rdi > 249d: 48 8d b5 90 fd ff ff lea -0x270(%rbp),%rsi > 24a4: b9 17 00 00 00 mov $0x17,%ecx > 24a9: f3 48 a5 rep movsq %ds:(%rsi),%es:(%rdi) > if (commit_start + commit_sections <= ic->journal_sections) { > 24ac: 41 39 c6 cmp %eax,%r14d > io_comp.comp = COMPLETION_INITIALIZER_ONSTACK(io_comp.comp); > 24af: 48 8d bd 90 fd ff ff lea -0x270(%rbp),%rdi > 24b6: 48 8d b5 e8 f9 ff ff lea -0x618(%rbp),%rsi > 24bd: b9 17 00 00 00 mov $0x17,%ecx > 24c2: f3 48 a5 rep movsq %ds:(%rsi),%es:(%rdi) > > Those two "rep movsq"s are very suspicious, because > COMPLETION_INITIALIZER_ONSTACK() should initialize the data in-place, > rather than move it to some temporary variable and copy it back. Right. I've seen this behavior before when using c99 compound literals, but I was surprised to see it here. I also submitted a patch for the one driver that turned up a new warning because of this behavior: https://www.spinics.net/lists/raid/msg58766.html In case of the mmc driver, the behavior was as expected, it was just a little too large and I sent the obvious workaround for it https://patchwork.kernel.org/patch/9902063/ > I tried to reduce the size of completion struct, and the "rep movsq" did > go away, however it seemed the compiler still allocated the memory for > the temporary variables on the stack, because whenever I > increased/decreased the size of completion, the stack size of > write_journal() got increased/decreased *7* times, but there are only > 3 journal_completion structures in write_journal(). So the *4* callsites > of COMPLETION_INITIALIZER_ONSTACK() looked very suspicous. > > So I come up with the following patch, trying to teach the compiler not > to do the unnecessary allocation, could you give it a try? > > Besides, I could also observe the stack size reduction of > write_journal() even for !LOCKDEP kernel. Ok. > ------------------- > Reported-by: Arnd Bergmann > Signed-off-by: Boqun Feng > --- > include/linux/completion.h | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/include/linux/completion.h b/include/linux/completion.h > index 791f053f28b7..cae5400022a3 100644 > --- a/include/linux/completion.h > +++ b/include/linux/completion.h > @@ -74,7 +74,7 @@ static inline void complete_release_commit(struct completion *x) {} > #endif > > #define COMPLETION_INITIALIZER_ONSTACK(work) \ > - ({ init_completion(&work); work; }) > + (*({ init_completion(&work); &work; })) > > /** > * DECLARE_COMPLETION - declare and initialize a completion structure Nice hack. Any idea why that's different to the compiler? I've applied that one to my test tree now, and reverted my own patch, will let you know if anything else shows up. I think we probably want to merge both patches to mainline. Arnd