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 9F956C5479D for ; Wed, 11 Jan 2023 17:54:17 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S232923AbjAKRyQ (ORCPT ); Wed, 11 Jan 2023 12:54:16 -0500 Received: from lindbergh.monkeyblade.net ([23.128.96.19]:46126 "EHLO lindbergh.monkeyblade.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S232042AbjAKRyO (ORCPT ); Wed, 11 Jan 2023 12:54:14 -0500 Received: from dfw.source.kernel.org (dfw.source.kernel.org [139.178.84.217]) by lindbergh.monkeyblade.net (Postfix) with ESMTPS id 7F202A469; Wed, 11 Jan 2023 09:54:12 -0800 (PST) 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 0E76E61D98; Wed, 11 Jan 2023 17:54:12 +0000 (UTC) Received: by smtp.kernel.org (Postfix) with ESMTPSA id 62E63C433EF; Wed, 11 Jan 2023 17:54:11 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=kernel.org; s=k20201202; t=1673459651; bh=4bvncI8qdilTRyPNQlsk6bcTHlK2AIug7R45fmJ8gp0=; h=Date:From:To:Cc:Subject:References:In-Reply-To:From; b=YY75u5iab83oByo4YBuKdlWlQNHo/Wb+N1b4wKT5oXr1M9OvY7iyk7Hjky17t+yVr to6isQZfPP2oexWLEuUmwLwXe/98wDYUMsBYssXKin3I4H0jN1aQDLXObDbDT/2kvI iHPaiXJE3/Rkrc4uwqZN/Q91MgeI4ZkaPvdd2es5mcXSR2EGE0+0v0aICjVdvlZgrh WUEALHL+pm8KAxM9pD1+wtbnzBG6Pi02qsyYXlWRZdA8Nqf6PYhMdY0lZsxkbh8w8U RkGqpyGTkFR5jZbR0MnQAns+wzLa4ueAz9BMTaO0MP1mIE79gJSMM72nevtVYTCKdG KllfrX5PM3PWg== Date: Wed, 11 Jan 2023 09:54:10 -0800 From: "Darrick J. Wong" To: Amir Goldstein Cc: linux-xfs@vger.kernel.org, willy@infradead.org, chandan.babu@oracle.com, allison.henderson@oracle.com, linux-fsdevel@vger.kernel.org, hch@infradead.org, catherine.hoang@oracle.com, david@fromorbit.com Subject: Re: [PATCH 06/14] xfs: document how online fsck deals with eventual consistency Message-ID: References: <167243825144.682859.12802259329489258661.stgit@magnolia> <167243825245.682859.4827095718073568782.stgit@magnolia> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: Precedence: bulk List-ID: X-Mailing-List: linux-fsdevel@vger.kernel.org On Fri, Jan 06, 2023 at 05:33:00AM +0200, Amir Goldstein wrote: > On Thu, Jan 5, 2023 at 9:40 PM Darrick J. Wong wrote: > > > > On Thu, Jan 05, 2023 at 11:08:51AM +0200, Amir Goldstein wrote: > > > On Sat, Dec 31, 2022 at 12:32 AM Darrick J. Wong wrote: > > > > > > > > From: Darrick J. Wong > > > > > > > > Writes to an XFS filesystem employ an eventual consistency update model > > > > to break up complex multistep metadata updates into small chained > > > > transactions. This is generally good for performance and scalability > > > > because XFS doesn't need to prepare for enormous transactions, but it > > > > also means that online fsck must be careful not to attempt a fsck action > > > > unless it can be shown that there are no other threads processing a > > > > transaction chain. This part of the design documentation covers the > > > > thinking behind the consistency model and how scrub deals with it. > > > > > > > > Signed-off-by: Darrick J. Wong > > > > --- > > > > .../filesystems/xfs-online-fsck-design.rst | 303 ++++++++++++++++++++ > > > > 1 file changed, 303 insertions(+) > > > > > > > > > > > > diff --git a/Documentation/filesystems/xfs-online-fsck-design.rst b/Documentation/filesystems/xfs-online-fsck-design.rst > > > > index f45bf97fa9c4..419eb54ee200 100644 > > > > --- a/Documentation/filesystems/xfs-online-fsck-design.rst > > > > +++ b/Documentation/filesystems/xfs-online-fsck-design.rst > > > > @@ -1443,3 +1443,306 @@ This step is critical for enabling system administrator to monitor the status > > > > of the filesystem and the progress of any repairs. > > > > For developers, it is a useful means to judge the efficacy of error detection > > > > and correction in the online and offline checking tools. > > > > + > > > > +Eventual Consistency vs. Online Fsck > > > > +------------------------------------ > > > > + > > > > +Midway through the development of online scrubbing, the fsstress tests > > > > +uncovered a misinteraction between online fsck and compound transaction chains > > > > +created by other writer threads that resulted in false reports of metadata > > > > +inconsistency. > > > > +The root cause of these reports is the eventual consistency model introduced by > > > > +the expansion of deferred work items and compound transaction chains when > > > > +reverse mapping and reflink were introduced. > > > > + > > > > +Originally, transaction chains were added to XFS to avoid deadlocks when > > > > +unmapping space from files. > > > > +Deadlock avoidance rules require that AGs only be locked in increasing order, > > > > +which makes it impossible (say) to use a single transaction to free a space > > > > +extent in AG 7 and then try to free a now superfluous block mapping btree block > > > > +in AG 3. > > > > +To avoid these kinds of deadlocks, XFS creates Extent Freeing Intent (EFI) log > > > > +items to commit to freeing some space in one transaction while deferring the > > > > +actual metadata updates to a fresh transaction. > > > > +The transaction sequence looks like this: > > > > + > > > > +1. The first transaction contains a physical update to the file's block mapping > > > > + structures to remove the mapping from the btree blocks. > > > > + It then attaches to the in-memory transaction an action item to schedule > > > > + deferred freeing of space. > > > > + Concretely, each transaction maintains a list of ``struct > > > > + xfs_defer_pending`` objects, each of which maintains a list of ``struct > > > > + xfs_extent_free_item`` objects. > > > > + Returning to the example above, the action item tracks the freeing of both > > > > + the unmapped space from AG 7 and the block mapping btree (BMBT) block from > > > > + AG 3. > > > > + Deferred frees recorded in this manner are committed in the log by creating > > > > + an EFI log item from the ``struct xfs_extent_free_item`` object and > > > > + attaching the log item to the transaction. > > > > + When the log is persisted to disk, the EFI item is written into the ondisk > > > > + transaction record. > > > > + EFIs can list up to 16 extents to free, all sorted in AG order. > > > > + > > > > +2. The second transaction contains a physical update to the free space btrees > > > > + of AG 3 to release the former BMBT block and a second physical update to the > > > > + free space btrees of AG 7 to release the unmapped file space. > > > > + Observe that the the physical updates are resequenced in the correct order > > > > + when possible. > > > > + Attached to the transaction is a an extent free done (EFD) log item. > > > > + The EFD contains a pointer to the EFI logged in transaction #1 so that log > > > > + recovery can tell if the EFI needs to be replayed. > > > > + > > > > +If the system goes down after transaction #1 is written back to the filesystem > > > > +but before #2 is committed, a scan of the filesystem metadata would show > > > > +inconsistent filesystem metadata because there would not appear to be any owner > > > > +of the unmapped space. > > > > +Happily, log recovery corrects this inconsistency for us -- when recovery finds > > > > +an intent log item but does not find a corresponding intent done item, it will > > > > +reconstruct the incore state of the intent item and finish it. > > > > +In the example above, the log must replay both frees described in the recovered > > > > +EFI to complete the recovery phase. > > > > + > > > > +There are two subtleties to XFS' transaction chaining strategy to consider. > > > > +The first is that log items must be added to a transaction in the correct order > > > > +to prevent conflicts with principal objects that are not held by the > > > > +transaction. > > > > +In other words, all per-AG metadata updates for an unmapped block must be > > > > +completed before the last update to free the extent, and extents should not > > > > +be reallocated until that last update commits to the log. > > > > +The second subtlety comes from the fact that AG header buffers are (usually) > > > > +released between each transaction in a chain. > > > > +This means that other threads can observe an AG in an intermediate state, > > > > +but as long as the first subtlety is handled, this should not affect the > > > > +correctness of filesystem operations. > > > > +Unmounting the filesystem flushes all pending work to disk, which means that > > > > +offline fsck never sees the temporary inconsistencies caused by deferred work > > > > +item processing. > > > > +In this manner, XFS employs a form of eventual consistency to avoid deadlocks > > > > +and increase parallelism. > > > > + > > > > +During the design phase of the reverse mapping and reflink features, it was > > > > +decided that it was impractical to cram all the reverse mapping updates for a > > > > +single filesystem change into a single transaction because a single file > > > > +mapping operation can explode into many small updates: > > > > + > > > > +* The block mapping update itself > > > > +* A reverse mapping update for the block mapping update > > > > +* Fixing the freelist > > > > +* A reverse mapping update for the freelist fix > > > > + > > > > +* A shape change to the block mapping btree > > > > +* A reverse mapping update for the btree update > > > > +* Fixing the freelist (again) > > > > +* A reverse mapping update for the freelist fix > > > > + > > > > +* An update to the reference counting information > > > > +* A reverse mapping update for the refcount update > > > > +* Fixing the freelist (a third time) > > > > +* A reverse mapping update for the freelist fix > > > > + > > > > +* Freeing any space that was unmapped and not owned by any other file > > > > +* Fixing the freelist (a fourth time) > > > > +* A reverse mapping update for the freelist fix > > > > + > > > > +* Freeing the space used by the block mapping btree > > > > +* Fixing the freelist (a fifth time) > > > > +* A reverse mapping update for the freelist fix > > > > + > > > > +Free list fixups are not usually needed more than once per AG per transaction > > > > +chain, but it is theoretically possible if space is very tight. > > > > +For copy-on-write updates this is even worse, because this must be done once to > > > > +remove the space from a staging area and again to map it into the file! > > > > + > > > > +To deal with this explosion in a calm manner, XFS expands its use of deferred > > > > +work items to cover most reverse mapping updates and all refcount updates. > > > > +This reduces the worst case size of transaction reservations by breaking the > > > > +work into a long chain of small updates, which increases the degree of eventual > > > > +consistency in the system. > > > > +Again, this generally isn't a problem because XFS orders its deferred work > > > > +items carefully to avoid resource reuse conflicts between unsuspecting threads. > > > > + > > > > +However, online fsck changes the rules -- remember that although physical > > > > +updates to per-AG structures are coordinated by locking the buffers for AG > > > > +headers, buffer locks are dropped between transactions. > > > > +Once scrub acquires resources and takes locks for a data structure, it must do > > > > +all the validation work without releasing the lock. > > > > +If the main lock for a space btree is an AG header buffer lock, scrub may have > > > > +interrupted another thread that is midway through finishing a chain. > > > > +For example, if a thread performing a copy-on-write has completed a reverse > > > > +mapping update but not the corresponding refcount update, the two AG btrees > > > > +will appear inconsistent to scrub and an observation of corruption will be > > > > +recorded. This observation will not be correct. > > > > +If a repair is attempted in this state, the results will be catastrophic! > > > > + > > > > +Several solutions to this problem were evaluated upon discovery of this flaw: > > > > + > > > > +1. Add a higher level lock to allocation groups and require writer threads to > > > > + acquire the higher level lock in AG order before making any changes. > > > > + This would be very difficult to implement in practice because it is > > > > + difficult to determine which locks need to be obtained, and in what order, > > > > + without simulating the entire operation. > > > > + Performing a dry run of a file operation to discover necessary locks would > > > > + make the filesystem very slow. > > > > + > > > > +2. Make the deferred work coordinator code aware of consecutive intent items > > > > + targeting the same AG and have it hold the AG header buffers locked across > > > > + the transaction roll between updates. > > > > + This would introduce a lot of complexity into the coordinator since it is > > > > + only loosely coupled with the actual deferred work items. > > > > + It would also fail to solve the problem because deferred work items can > > > > + generate new deferred subtasks, but all subtasks must be complete before > > > > + work can start on a new sibling task. > > > > + > > > > +3. Teach online fsck to walk all transactions waiting for whichever lock(s) > > > > + protect the data structure being scrubbed to look for pending operations. > > > > + The checking and repair operations must factor these pending operations into > > > > + the evaluations being performed. > > > > + This solution is a nonstarter because it is *extremely* invasive to the main > > > > + filesystem. > > > > + > > > > +4. Recognize that only online fsck has this requirement of total consistency > > > > + of AG metadata, and that online fsck should be relatively rare as compared > > > > + to filesystem change operations. > > > > + For each AG, maintain a count of intent items targetting that AG. > > > > + When online fsck wants to examine an AG, it should lock the AG header > > > > + buffers to quiesce all transaction chains that want to modify that AG, and > > > > + only proceed with the scrub if the count is zero. > > > > + In other words, scrub only proceeds if it can lock the AG header buffers and > > > > + there can't possibly be any intents in progress. > > > > + This may lead to fairness and starvation issues, but regular filesystem > > > > + updates take precedence over online fsck activity. > > > > + > > > > > > Is there any guarantee that some silly real life regular filesystem workload > > > won't starve online fsck forever? > > > IOW, is forward progress of online fsck guaranteed? > > > > Nope, forward progress isn't guaranteed. > > That sounds like a problem. So far it hasn't been. I prefer to sacrifice performance of the background fsck service for the sake of foreground tasks. The fsstress and fsx fstests haven't shown any particularly serious issues. I've also kicked off xfs_scrub on the same VM hosts that are running the fuzz test suite (~52 VMs per host) and scrub can still finish the filesystem in a couple of hours. Things get markedly worse on spinning rust with a lot of parallel unwritten extent conversions and allocations going on (aka the disk backup systems). Normally a backup from flash to rust takes about an hour; with scrub and backup contending for the head actuator, it'll go up to about 2-3 hours, but both tasks can make (verrrry slow) forward progress. That said -- the backup program spends a lot of iowait time waiting for file data blocks to read in or get written back, so the contention is on the storage hardware, not the filesystem locks. > > The kernel checks for fatal > > signals every time it backs off a scrub so at least we don't end up with > > unkillable processes. At one point I added a timeout field to the ioctl > > interface so that the kernel could time out an operation if it took too > > long to acquire the necessary resources. So far, the "race fsstress and > > xfs_scrub" tests have not shown scrub failing to make any forward > > progress. > > > > That said, I have /not/ yet had a chance to try it out any of these > > massive 1000-core systems with an according workload. > > > > Don't know if fsstress is the best way to check the worst case scenario. > > Can you think of a workload, say several threads creating and deleting > temp files, with deferred parent pointer items preventing the queue from > ever draining? The worst workload would be one that is entirely metadata based -- a giant directory tree full of empty files with all information being stored as extended attributes. > Considering that a "full journal" scenario is always going to be a possible > worst case incident, how bad would it be to block new transactions > instead of the possibility of starving scrub consistency checks forever? First of all, scrub has already allocated a transaction by the time it gets to the intent drain step. There's no good way to block new transactions once we've reached this stage, nor should there be. Blocking transactions stalls xfs garbage collection and memory reclaim. > Wouldn't the consistency checks be much faster than freeing journal > space would be in a "full journal" situation? I haven't investigated this in depth, but yes, scrub should be faster than forcing the log and checkpointing the log to move the log tail forward to empty out the journal. > I don't know if there is a "mission statement" for online fsck, but > I think it would say "minimal user interference" not "no user interference". Yes. The section about eventual consistency states that "...regular filesystem updates take precedence over online fsck activity". > It sounds like the interference we are trying to avoid is light years away > from the downtime of offline fsck, so online fsck would still be a huge win. > online fsck that never ends OTOH... maybe less so. Well you /can/ just kill the xfs_scrub processes if they are taking too much time. One of the nastier papercuts of the background scrub is that the fs cannot be unmounted while it's running, and systemd doesn't have a good mechanism for "kill this service before stopping this mount". Or maybe it does and I haven't yet found it? (The cronjob variant definitely suffers from that...) > > > Good luck with landing online fsck before the 2024 NYE deluge ;) > > > > Thank *you* for reading this chapter of the design document!! :) > > > > Oh I read them all at the summer submission, but it took me so long > that I forgot to follow up.. Yeah, that seems to be a common problem with large new features. :/ > My other question was regarding memory usage control. > I have horrid memories from e2fsck unpredictable memory usage > and unpredictable runtime due to swapping. > > xfs_repair -m was a huge improvement compared to e2fsck. > I don't remember reading about memory usage limits for online repair, > so I was concerned about unpredictable memory usage and swapping. > Can you say something to ease those concerns? Both e2fsck and xfs_repair have to be capable of repairing the entire filesystem all at once, which means that they allocate many many of incore objects from which all of the ondisk space metadata (ag btrees in the case of xfs, bitmaps for e2fsck) is regenerated. Since the fs is offline, it's considered advantageous to perform *one* scan and rebuild everything all at once, even if the memory cost is high. xfs_scrub scans and repairs each metadata object individually, which means that it only needs to allocate as much (kernel/xfile) memory as needed to scan a single btree/inode record/quota record/bitmap. For scans the memory requirements are usually minimal since it creates a bunch of btree cursors and cross-references records. For repairs, the memory requirements are on the order of the size of the new data structure that will be written out. We scan the fs to build the new recordset in memory, compute the size of the new btree, allocate some blocks, and format the records into the blocks before committing the btree root. For summary data (e.g. link counts, dquots) we build a shadow copy in memory, so the memory requirements are on the order of the number of files in the fs and the number of uid/gid/projid in the filesystem, respectively. Most of the intermediate structures are stuffed into a tmpfs file, which means they can be paged out to disk. If there's really no memory available, scrub can abort all the way out to userspace provided it hasn't committed anything to disk yet. IOWs, online fsck generally only requires enough memory to build a new copy of whichever objects it happens to be scanning at any given moment. The background service runs single-threaded to avoid consuming a lot of CPU or memory. --D > Thanks, > Amir.