From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by smtp.lore.kernel.org (Postfix) with ESMTP id C9E0FC433F5 for ; Thu, 10 Mar 2022 18:08:29 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S245334AbiCJSJ1 convert rfc822-to-8bit (ORCPT ); Thu, 10 Mar 2022 13:09:27 -0500 Received: from lindbergh.monkeyblade.net ([23.128.96.19]:59018 "EHLO lindbergh.monkeyblade.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S245328AbiCJSJ0 (ORCPT ); Thu, 10 Mar 2022 13:09:26 -0500 Received: from elephants.elehost.com (elephants.elehost.com [216.66.27.132]) by lindbergh.monkeyblade.net (Postfix) with ESMTPS id 531FC4B855 for ; Thu, 10 Mar 2022 10:08:24 -0800 (PST) Received: from Mazikeen (cpe788df74d2cc1-cm788df74d2cc0.cpe.net.cable.rogers.com [72.138.27.250] (may be forged)) (authenticated bits=0) by elephants.elehost.com (8.16.1/8.16.1) with ESMTPSA id 22AI8MOD030772 (version=TLSv1.2 cipher=ECDHE-RSA-AES256-GCM-SHA384 bits=256 verify=NO); Thu, 10 Mar 2022 13:08:22 -0500 (EST) (envelope-from rsbecker@nexbridge.com) Reply-To: From: To: "'Neeraj Singh'" , "=?utf-8?Q?'=C3=86var_Arnfj=C3=B6r=C3=B0_Bjarmason'?=" Cc: "'Neeraj K. Singh via GitGitGadget'" , "'Git List'" , "'Johannes Schindelin'" , "'Jeff King'" , "'Jeff Hostetler'" , "'Christoph Hellwig'" , "'Bagas Sanjaya'" , "'Elijah Newren'" , "'Neeraj K. Singh'" , "'Patrick Steinhardt'" , "'Junio C Hamano'" , "'Eric Wong'" References: <211116.8635nwr055.gmgdl@evledraar.gmail.com> <211117.86ee7f8cm4.gmgdl@evledraar.gmail.com> <211201.864k7sbdjt.gmgdl@evledraar.gmail.com> <220310.86lexilo3d.gmgdl@evledraar.gmail.com> <220310.86r179ki38.gmgdl@evledraar.gmail.com> In-Reply-To: Subject: RE: [PATCH v9 0/9] Implement a batched fsync option for core.fsyncObjectFiles Date: Thu, 10 Mar 2022 13:08:16 -0500 Organization: Nexbridge Inc. Message-ID: <00ae01d834a9$d443a530$7ccaef90$@nexbridge.com> MIME-Version: 1.0 Content-Type: text/plain; charset="utf-8" Content-Transfer-Encoding: 8BIT X-Mailer: Microsoft Outlook 16.0 Thread-Index: AQGgIVyljHjFm8+Xd7vU2vU1n+z2ZQKRvucnAxQpxkcB5Wnx4wJZi8f1AnM48H0BjRpn+AFX0GFeAXM2sKsCoVUILQIqGAxUrH0ydOA= Content-Language: en-ca Precedence: bulk List-ID: X-Mailing-List: git@vger.kernel.org On March 10, 2022 12:53 PM, Neeraj Singh wrote: >On Thu, Mar 10, 2022 at 6:17 AM Ævar Arnfjörð Bjarmason >wrote: >> >> >> On Wed, Mar 09 2022, Neeraj Singh wrote: >> >> > On Wed, Mar 9, 2022 at 3:10 PM Ævar Arnfjörð Bjarmason > wrote: >> >> >> >> Replying to an old-ish E-Mail of mine with some more thought that >> >> came to mind after[1] (another recently resurrected fsync() thread). >> >> >> >> I wonder if there's another twist on the plan outlined in [2] that >> >> would be both portable & efficient, i.e. the "slow" POSIX way to >> >> write files A..Z is to open/write/close/fsync each one, so we'll >> >> trigger a HW flush N times. >> >> >> >> And as we've discussed, doing it just on Z will implicitly flush >> >> A..Y on common OS's in the wild, which we're taking advantage of here. >> >> >> >> But aside from the rename() dance in[2], what do those OS's do if >> >> you write A..Z, fsync() the "fd" for Z, and then fsync A..Y (or, >> >> presumably equivalently, in reverse order: Y..A). >> >> >> >> I'd think they'd be smart enough to know that they already >> >> implicitly flushed that data since Z was flushend, and make those >> >> fsync()'s a rather cheap noop. >> >> >> >> But I don't know, hence the question. >> >> >> >> If that's true then perhaps it's a path towards having our cake and >> >> eating it too in some cases? >> >> >> >> I.e. an FS that would flush A..Y if we flush Z would do so quickly >> >> and reliably, whereas a FS that doesn't have such an optimization >> >> might be just as slow for all of A..Y, but at least it'll be safe. >> >> >> >> 1. >> >> https://lore.kernel.org/git/220309.867d93lztw.gmgdl@evledraar.gmail >> >> .com/ 2. >> >> https://lore.kernel.org/git/e1747ce00af7ab3170a69955b07d995d5321d6f >> >> 3.1637020263.git.gitgitgadget@gmail.com/ >> > >> > The important angle here is that we need some way to indicate to the >> > OS what A..Y is before we fsync on Z. I.e. the OS will cache any >> > writes in memory until some sync-ish operation is done on *that >> > specific file*. Syncing just 'Z' with no sync operations on A..Y >> > doesn't indicate that A..Y would get written out. Apparently the >> > bad old ext3 behavior was similar to what you're proposing where a >> > sync on 'Z' would imply something about independent files. >> >> It's certainly starting to sound like I'm misunderstanding this whole >> thing, but just to clarify again I'm talking about the sort of loops >> mentioned upthread in my [1]. I.e. you have (to copy from that E-Mail): >> >> bulk_checkin_start_make_cookie(): >> n = 10 >> for i in 1..n: >> write_nth(i, fsync: 0); >> bulk_checkin_end_commit_cookie(); >> >> I.e. we have a "cookie" file in a given dir (where, in this example, >> we'd also write files A..Z). I.e. we write: >> >> cookie >> {A..Z} >> cookie >> >> And then only fsync() on the "cookie" at the end, which "flushes" the >> A..Z updates on some FS's (again, all per my possibly-incorrect >> understanding). >> >> Which is why I proposed that in many/all cases we could do this, i.e. >> just the same without the "cookie" file (which AFAICT isn't needed >> per-se, but was just added to make the API a bit simpler in not >> needing to modify the relevant loops): >> >> all_fsync = bulk_checkin_mode() ? 0 : fsync_turned_on_in_general(); >> end_fsync = bulk_checkin_mode() ? 1 : all_fsync; >> n = 10; >> for i in 1..n: >> write_nth(i, fsync: (i == n) ? end_fsync : all_fsync); >> >> I.e. we don't pay the cost of the fsync() as we're in the loop, but >> just for the last file, which "flushes" the rest. >> >> So far all of that's a paraphrasing of existing exchanges, but what I >> was wondering now in[2] is if we add this to this last example above: >> >> for i in 1..n-1: >> fsync_nth(i) >> >> Wouldn't those same OS's that are being clever about deferring the >> syncing of A..Z as a "batch" be clever enough to turn that >> (re-)syncing into a NOOP? >> >> Of course in this case we'd need to keep the fd's open and be clever >> about E[MN]FILE (i.e. "Too many open..."), or do an fsync() every Nth >> for some reasonable Nth, e.g. somewhere in the 2^10..2^12 range. >> >> But *if* this works it seems to me to be something we might be able to >> enable when "core.fsyncObjectFiles" is configured on those systems. >> >> I.e. the implicit assumption with that configuration was that if we >> sync N loose objects and then update and fsync the ref that the FS >> would queue up the ref update after the syncing of the loose objects. >> >> This new "cookie" (or my suggested "fsync last of N") is basically >> making the same assumption, just with the slight twist that some >> OSs/FSs are known to behave like that on a per-subdir basis, no? >> >> > Here's an interesting paper I recently came across that proposes the >> > interface we'd really want, 'syncv': >> > >https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.924.1168&rep=rep1 >&type=pdf. >> >> 1. >> https://lore.kernel.org/git/211201.864k7sbdjt.gmgdl@evledraar.gmail.co >> m/ 2. >> https://lore.kernel.org/git/220310.86lexilo3d.gmgdl@evledraar.gmail.co >> m/ > >On the actual FS implementations in the three common OSes I'm familiar with >(macOS, Windows, Linux), each file has its own independent data caching in OS >memory. Fsyncing one of them doesn't necessarily imply writing out the OS cache >for any other file. Except, apparently, on ext3 in data=ordered mode, but that FS >is no longer common. On Linux, we use sync_file_range to get the OS to write the >in-memory cache to the storage hardware, which is what makes the data >'available' to fsync. > >Now, we could consider an implementation where we call sync_file_range >without the wait flags (i.e. without SYNC_FILE_RANGE_WAIT_BEFORE and >SYNC_FILE_RANGE_WAIT_AFTER). Then we could later fsync every file (or batch >of files), which might be more efficient if the OS coalesces the disk cache flushes. I >expect that this method is less likely to give us the desired performance on >common linux FSes, however. > >The macOS and Windows APIs are defined a bit differently from Linux. >In both those OSes, >we're actually calling fsync-equivalent APIs that are defined to write back all the >relevant data and metadata, just without the storage cache flush. > >So to summarize: >1. We need to do write(2) to get the data out of Git and into the OS filesystem >cache. >2. We need some API (macOS-fsync, Windows-NtFlushBuffersFileEx, >Linux-sync_file_range) > to transfer the data per-file to the storage controller, but without flushing the >storage controller. >3. We need some api (macOS-F_FULLFSYNC, Windows-NtFlushBuffersFile, Linux- >fsync) > to push the storage controller cache to durable media. This only needs to be >done once > at the end to push out the data made available in step (2). While this might not be a surprise, on some platforms fsync is a thread-blocking operation. When the OS has kernel threads, fsync can potentially cause multiple processes (if implemented that way) to block, particularly where an fd is shared across threads (and thus processes), which may end up causing a deadlock. We might need to keep an eye out for this type of situation in the future and at least try to test for it. I cannot actually see a situation where this would occur in git, but that does not mean it is impossible. Food for thought. --Randall