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 88598C54EE9 for ; Thu, 29 Sep 2022 02:15:44 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S234434AbiI2CPn (ORCPT ); Wed, 28 Sep 2022 22:15:43 -0400 Received: from lindbergh.monkeyblade.net ([23.128.96.19]:59050 "EHLO lindbergh.monkeyblade.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S234551AbiI2CPl (ORCPT ); Wed, 28 Sep 2022 22:15:41 -0400 Received: from dfw.source.kernel.org (dfw.source.kernel.org [139.178.84.217]) by lindbergh.monkeyblade.net (Postfix) with ESMTPS id 59A9D10952C; Wed, 28 Sep 2022 19:15:28 -0700 (PDT) Received: from smtp.kernel.org (relay.kernel.org [52.25.139.140]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by dfw.source.kernel.org (Postfix) with ESMTPS id 0917E61AD2; Thu, 29 Sep 2022 02:15:28 +0000 (UTC) Received: by smtp.kernel.org (Postfix) with ESMTPSA id 660A1C433D7; Thu, 29 Sep 2022 02:15:27 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=kernel.org; s=k20201202; t=1664417727; bh=k2wflFgQECzPufVWTIiUxC7W+iON7k0EvxVQlitKNZ4=; h=Date:From:To:Cc:Subject:References:In-Reply-To:From; b=Jmzhf9wzYUFIDKe6epFS2KMsQj/U1Avnfky0FL3FMC+CzfYZOEC2GTH36nAJ1tPQF 4f4FYTuLfiw1gYNnL6dLn712yt9ib3YNmBz5ktvgNYxniSvetmCqsx0equWnHUL3hk GIA6kJMlLQxGiWGqyk6wK1/zeAuVvtnCnJA0IKM+2Y6ASGc2AJQxQUHFygcAhmbq6D ExN0EQCEW3nSJmKVUP5HSZKQOShO/QUd0mJoFwf3EH/DM2XPCxu2kbm6QbG4Ijz3C+ 3bEsMwBp7t5o4gH3vgR+xcnrW70C0dEbz7MkTxmM9k8Kp2wORUk+bDHV9oxOtskoU4 e10GAgLls1aOQ== Date: Wed, 28 Sep 2022 19:15:26 -0700 From: "Darrick J. Wong" To: Dave Chinner Cc: linux-xfs@vger.kernel.org, linux-fsdevel@vger.kernel.org Subject: Re: [RFC PATCH 0/2] iomap/xfs: fix data corruption due to stale cached iomaps Message-ID: References: <20220921082959.1411675-1-david@fromorbit.com> <20220922225934.GU3600936@dread.disaster.area> <20220929021128.GF3600936@dread.disaster.area> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20220929021128.GF3600936@dread.disaster.area> Precedence: bulk List-ID: X-Mailing-List: linux-fsdevel@vger.kernel.org On Thu, Sep 29, 2022 at 12:11:28PM +1000, Dave Chinner wrote: > On Tue, Sep 27, 2022 at 10:16:42PM -0700, Darrick J. Wong wrote: > > On Fri, Sep 23, 2022 at 08:59:34AM +1000, Dave Chinner wrote: > > > On Wed, Sep 21, 2022 at 09:25:26PM -0700, Darrick J. Wong wrote: > > > > On Wed, Sep 21, 2022 at 06:29:57PM +1000, Dave Chinner wrote: > > > > > Hi folks, > > > > > > > > > > THese patches address the data corruption first described here: > > > > > > > > > > https://lore.kernel.org/linux-xfs/20220817093627.GZ3600936@dread.disaster.area/ > > > > > > > > > > This data corruption has been seen in high profile production > > > > > systems so there is some urgency to fix it. The underlying flaw is > > > > > essentially a zero-day iomap bug, so whatever fix we come up with > > > > > needs to be back portable to all supported stable kernels (i.e. > > > > > ~4.18 onwards). > > > > > > > > > > A combination of concurrent write()s, writeback IO completion, and > > > > > memory reclaim combine to expose the fact that the cached iomap that > > > > > is held across an iomap_begin/iomap_end iteration can become stale > > > > > without the iomap iterator actor being aware that the underlying > > > > > filesystem extent map has changed. > > > > > > > > > > Hence actions based on the iomap state (e.g. is unwritten or newly > > > > > allocated) may actually be incorrect as writeback actions may have > > > > > changed the state (unwritten to written, delalloc to unwritten or > > > > > written, etc). This affects partial block/page operations, where we > > > > > may need to read from disk or zero cached pages depending on the > > > > > actual extent state. Memory reclaim plays it's part here in that it > > > > > removes pages containing partial state from the page cache, exposing > > > > > future partial page/block operations to incorrect behaviour. > > > > > > > > > > Really, we should have known that this would be a problem - we have > > > > > exactly the same issue with cached iomaps for writeback, and the > > > > > ->map_blocks callback that occurs for every filesystem block we need > > > > > to write back is responsible for validating the cached iomap is > > > > > still valid. The data corruption on the write() side is a result of > > > > > not validating that the iomap is still valid before we initialise > > > > > new pages and prepare them for data to be copied in to them.... > > > > > > > > > > I'm not really happy with the solution I have for triggering > > > > > remapping of an iomap when the current one is considered stale. > > > > > Doing the right thing requires both iomap_iter() to handle stale > > > > > iomaps correctly (esp. the "map is invalid before the first actor > > > > > operation" case), and it requires the filesystem > > > > > iomap_begin/iomap_end operations to co-operate and be aware of stale > > > > > iomaps. > > > > > > > > > > There are a bunch of *nasty* issues around handling failed writes in > > > > > XFS taht this has exposed - a failed write() that races with a > > > > > mmap() based write to the same delalloc page will result in the mmap > > > > > writes being silently lost if we punch out the delalloc range we > > > > > allocated but didn't write to. g/344 and g/346 expose this bug > > > > > directly if we punch out delalloc regions allocated by now stale > > > > > mappings. > > > > > > > > Yuck. I'm pretty sure that callers (xfs_buffered_write_iomap_end) is > > > > supposed to call truncate_pagecache_range with the invalidatelock (fka > > > > MMAPLOCK) held. > > > > > > Yup, there's multiple problems with this code; apart from > > > recognising that it is obviously broken and definitely problematic, > > > I haven't dug into it further. > > > > ...and I've been so buried in attending meetings and livedebug sessions > > related to a 4.14 corruption that now I'm starved of time to fully think > > through all the implications of this one. :( > > > > > > > Then, because we can't punch out the delalloc we allocated region > > > > > safely when we have a stale iomap, we have to ensure when we remap > > > > > it the IOMAP_F_NEW flag is preserved so that the iomap code knows > > > > > that it is uninitialised space that is being written into so it will > > > > > zero sub page/sub block ranges correctly. > > > > > > > > Hm. IOMAP_F_NEW results in zeroing around, right? So if the first > > > > ->iomap_begin got a delalloc mapping, but by the time we got the folio > > > > locked someone else managed to writeback and evict the page, we'd no > > > > longer want that zeroing ... right? > > > > > > Yes, and that is one of the sources of the data corruption - zeroing > > > when we shouldn't. > > > > > > There are multiple vectors to having a stale iomap here: > > > > > > 1. we allocate the delalloc range, giving us IOMAP_DELALLOC and > > > IOMAP_F_NEW. Writeback runs, allocating the range as unwritten. > > > Even though the iomap is now stale, there is no data corruption > > > in this case because the range is unwritten and so we still need > > > zeroing. > > > > ...and I guess this at least happens more often now that writeback does > > delalloc -> unwritten -> write -> unwritten conversion? > > *nod* > > > > 2. Same as above, but IO completion converts the range to written. > > > Data corruption occurs in this case because IOMAP_F_NEW causes > > > incorrect page cache zeroing to occur on partial page writes. > > > > > > 3. We have an unwritten extent (prealloc, writeback in progress, > > > etc) so we have IOMAP_UNWRITTEN. These require zeroing, > > > regardless of whether IOMAP_F_NEW is set or not. Extent is > > > written behind our backs, unwritten conversion occurs, and now we > > > zero partial pages when we shouldn't. > > > > Yikes. > > > > > Other issues I've found: > > > > > > 4. page faults can run the buffered write path concurrently with > > > write() because they aren't serialised against each other. Hence > > > we can have overlapping concurrent iomap_iter() operations with > > > different zeroing requirements and it's anyone's guess as to > > > which will win the race to the page lock and do the initial > > > zeroing. This is a potential silent mmap() write data loss > > > vector. > > > > TBH I've long wondered why IOLOCK and MMAPLOCK both seemingly protected > > pagecache operations but the buffered io paths never seemed to take the > > MMAPLOCK, and if there was some subtle way things could go wrong. > > We can't take MMAPLOCK in the buffered IO path because the user > buffer could be a mmap()d range of the same file and we need to be > able to fault in those pages during copyin/copyout. Hence we can't > hold the MMAPLOCK across iomap_iter(), nor across > .iomap_begin/.iomap_end context pairs. Ahh, right, I forgot that case. >:O > taking the MMAPLOCK and dropping it again can be done in iomap_begin > or iomap_end, as long as those methods aren't called from the page > fault path.... > > > > 5. anything that can modify the extent layout without holding the > > > i_rwsem exclusive can race with iomap iterating the extent list. > > > Holding the i_rwsem shared and modifying the extent list (e.g. > > > direct IO writes) can result in iomaps changing in the middle of, > > > say, buffered reads (e.g. hole->unwritten->written). > > > > Yep. I wonder, can this result in other incorrect write behavior that > > you and I haven't thought of yet? > > Entirely possible - this code is complex and there are lots of very > subtle interactions and we've already found several bonus broken > bits as a result. Hence I wouldn't be surprised if we've missed > other subtle issues and/or not fully grokked the implications of the > broken bits we've found... > > [....] > > > > > What happens if iomap_writepage_map errors out (say because ->map_blocks > > > > returns an error) without adding the folio to any ioend? > > > > > > Without reading further: > > > > > > 1. if we want to retry the write, we folio_redirty_for_writepage(), > > > unlock it and return with no error. Essentially we just skip over > > > it. > > > > If the fs isn't shut down, I guess we could redirty the page, though I > > guess the problem is that the page is now stuck in dirty state until > > xfs_scrub fixes the problem. If it fixes the problem. > > > > I think for bufferhead users it's nastier because we might have a > > situation where pagedirty is unset but BH_Dirty is still set. It > > certainly is a problem on 4.14. > > > > > 2. If we want to fail the write, we should call set_mapping_error() > > > to record the failure for the next syscall to report and, maybe, set > > > the error flag/clear the uptodate flag on the folio depending on > > > whether we want the data to remain valid in memory or not. > > > > That seems to be happening. Sort of. > > > > I think there's also a UAF in iomap_writepage_map -- if the folio is > > unlocked and we cleared (or never set) PageWriteback, isn't it possible > > that by the time we get to the mapping_set_error, the folio could have > > been torn out of the page cache and reused somewhere else? > > We still have a reference to the folio at this point from the lookup > in write_cache_pages(). Hence the folio can't be freed while we are > running iomap_writepage_map(). > > However, we have unlocked the folio, and we don't hold either the IO > lock or the invalidate lock and so the folio could get punched out > of the page cache.... > > > In which case, we're at best walking off a NULL mapping and crashing the > > system, and at worst setting an IO error on the wrong mapping? > > Yes, I think so - we could walk off a NULL mapping here, > but because write_cache_pages() still holds a page reference, the > page won't get freed from under us so we won't ever see the wrong > mapping being set here. > > I think we could fix that simply by using inode->i_mapping instead > of folio->mapping... Oh. Yes. I'll get on that tomorrow. > > > > I think in > > > > that case we'll follow the (error && !count) case, in which we unlock > > > > the folio and exit without calling folio_redirty_for_writepage, right? > > > > The error will get recorded in the mapping for the next fsync, I think, > > > > but I also wonder if we *should* redirty because the mapping failed, not > > > > the attempt at persistence. > > > > > > *nod* > > > > > > I think the question that needs to be answered here is this: in what > > > case is an error being returned from ->map_blocks a recoverable > > > error that a redirty + future writeback retry will succeed? > > > > > > AFAICT, all cases from XFS this is a fatal error (e.g. corruption of > > > the BMBT), so the failure will persist across all attempts to retry > > > the write? > > > > > > Perhaps online repair will change this (i.e. in the background > > > repair fixes the BMBT corruption and so the next attempt to write > > > the data will succeed) so I can see that we *might* need to redirty > > > the page in this case, but.... > > > > ...but I don't know that we can practically wait for repairs to happen > > because the page is now stuck in dirty state indefinitely. > > *nod* > > So do we treat it as fatal for now, and revisit it later when online > repair might be able to do something better here? Sounds good to me. --D > Cheers, > > Dave. > -- > Dave Chinner > david@fromorbit.com