5504 lines · plain
1.. SPDX-License-Identifier: GPL-2.02.. _xfs_online_fsck_design:3 4..5 Mapping of heading styles within this document:6 Heading 1 uses "====" above and below7 Heading 2 uses "===="8 Heading 3 uses "----"9 Heading 4 uses "````"10 Heading 5 uses "^^^^"11 Heading 6 uses "~~~~"12 Heading 7 uses "...."13 14 Sections are manually numbered because apparently that's what everyone15 does in the kernel.16 17======================18XFS Online Fsck Design19======================20 21This document captures the design of the online filesystem check feature for22XFS.23The purpose of this document is threefold:24 25- To help kernel distributors understand exactly what the XFS online fsck26 feature is, and issues about which they should be aware.27 28- To help people reading the code to familiarize themselves with the relevant29 concepts and design points before they start digging into the code.30 31- To help developers maintaining the system by capturing the reasons32 supporting higher level decision making.33 34As the online fsck code is merged, the links in this document to topic branches35will be replaced with links to code.36 37This document is licensed under the terms of the GNU Public License, v2.38The primary author is Darrick J. Wong.39 40This design document is split into seven parts.41Part 1 defines what fsck tools are and the motivations for writing a new one.42Parts 2 and 3 present a high level overview of how online fsck process works43and how it is tested to ensure correct functionality.44Part 4 discusses the user interface and the intended usage modes of the new45program.46Parts 5 and 6 show off the high level components and how they fit together, and47then present case studies of how each repair function actually works.48Part 7 sums up what has been discussed so far and speculates about what else49might be built atop online fsck.50 51.. contents:: Table of Contents52 :local:53 541. What is a Filesystem Check?55==============================56 57A Unix filesystem has four main responsibilities:58 59- Provide a hierarchy of names through which application programs can associate60 arbitrary blobs of data for any length of time,61 62- Virtualize physical storage media across those names, and63 64- Retrieve the named data blobs at any time.65 66- Examine resource usage.67 68Metadata directly supporting these functions (e.g. files, directories, space69mappings) are sometimes called primary metadata.70Secondary metadata (e.g. reverse mapping and directory parent pointers) support71operations internal to the filesystem, such as internal consistency checking72and reorganization.73Summary metadata, as the name implies, condense information contained in74primary metadata for performance reasons.75 76The filesystem check (fsck) tool examines all the metadata in a filesystem77to look for errors.78In addition to looking for obvious metadata corruptions, fsck also79cross-references different types of metadata records with each other to look80for inconsistencies.81People do not like losing data, so most fsck tools also contains some ability82to correct any problems found.83As a word of caution -- the primary goal of most Linux fsck tools is to restore84the filesystem metadata to a consistent state, not to maximize the data85recovered.86That precedent will not be challenged here.87 88Filesystems of the 20th century generally lacked any redundancy in the ondisk89format, which means that fsck can only respond to errors by erasing files until90errors are no longer detected.91More recent filesystem designs contain enough redundancy in their metadata that92it is now possible to regenerate data structures when non-catastrophic errors93occur; this capability aids both strategies.94 95+--------------------------------------------------------------------------+96| **Note**: |97+--------------------------------------------------------------------------+98| System administrators avoid data loss by increasing the number of |99| separate storage systems through the creation of backups; and they avoid |100| downtime by increasing the redundancy of each storage system through the |101| creation of RAID arrays. |102| fsck tools address only the first problem. |103+--------------------------------------------------------------------------+104 105TLDR; Show Me the Code!106-----------------------107 108Code is posted to the kernel.org git trees as follows:109`kernel changes <https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-symlink>`_,110`userspace changes <https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfsprogs-dev.git/log/?h=scrub-media-scan-service>`_, and111`QA test changes <https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfstests-dev.git/log/?h=repair-dirs>`_.112Each kernel patchset adding an online repair function will use the same branch113name across the kernel, xfsprogs, and fstests git repos.114 115Existing Tools116--------------117 118The online fsck tool described here will be the third tool in the history of119XFS (on Linux) to check and repair filesystems.120Two programs precede it:121 122The first program, ``xfs_check``, was created as part of the XFS debugger123(``xfs_db``) and can only be used with unmounted filesystems.124It walks all metadata in the filesystem looking for inconsistencies in the125metadata, though it lacks any ability to repair what it finds.126Due to its high memory requirements and inability to repair things, this127program is now deprecated and will not be discussed further.128 129The second program, ``xfs_repair``, was created to be faster and more robust130than the first program.131Like its predecessor, it can only be used with unmounted filesystems.132It uses extent-based in-memory data structures to reduce memory consumption,133and tries to schedule readahead IO appropriately to reduce I/O waiting time134while it scans the metadata of the entire filesystem.135The most important feature of this tool is its ability to respond to136inconsistencies in file metadata and directory tree by erasing things as needed137to eliminate problems.138Space usage metadata are rebuilt from the observed file metadata.139 140Problem Statement141-----------------142 143The current XFS tools leave several problems unsolved:144 1451. **User programs** suddenly **lose access** to the filesystem when unexpected146 shutdowns occur as a result of silent corruptions in the metadata.147 These occur **unpredictably** and often without warning.148 1492. **Users** experience a **total loss of service** during the recovery period150 after an **unexpected shutdown** occurs.151 1523. **Users** experience a **total loss of service** if the filesystem is taken153 offline to **look for problems** proactively.154 1554. **Data owners** cannot **check the integrity** of their stored data without156 reading all of it.157 This may expose them to substantial billing costs when a linear media scan158 performed by the storage system administrator might suffice.159 1605. **System administrators** cannot **schedule** a maintenance window to deal161 with corruptions if they **lack the means** to assess filesystem health162 while the filesystem is online.163 1646. **Fleet monitoring tools** cannot **automate periodic checks** of filesystem165 health when doing so requires **manual intervention** and downtime.166 1677. **Users** can be tricked into **doing things they do not desire** when168 malicious actors **exploit quirks of Unicode** to place misleading names169 in directories.170 171Given this definition of the problems to be solved and the actors who would172benefit, the proposed solution is a third fsck tool that acts on a running173filesystem.174 175This new third program has three components: an in-kernel facility to check176metadata, an in-kernel facility to repair metadata, and a userspace driver177program to drive fsck activity on a live filesystem.178``xfs_scrub`` is the name of the driver program.179The rest of this document presents the goals and use cases of the new fsck180tool, describes its major design points in connection to those goals, and181discusses the similarities and differences with existing tools.182 183+--------------------------------------------------------------------------+184| **Note**: |185+--------------------------------------------------------------------------+186| Throughout this document, the existing offline fsck tool can also be |187| referred to by its current name "``xfs_repair``". |188| The userspace driver program for the new online fsck tool can be |189| referred to as "``xfs_scrub``". |190| The kernel portion of online fsck that validates metadata is called |191| "online scrub", and portion of the kernel that fixes metadata is called |192| "online repair". |193+--------------------------------------------------------------------------+194 195The naming hierarchy is broken up into objects known as directories and files196and the physical space is split into pieces known as allocation groups.197Sharding enables better performance on highly parallel systems and helps to198contain the damage when corruptions occur.199The division of the filesystem into principal objects (allocation groups and200inodes) means that there are ample opportunities to perform targeted checks and201repairs on a subset of the filesystem.202 203While this is going on, other parts continue processing IO requests.204Even if a piece of filesystem metadata can only be regenerated by scanning the205entire system, the scan can still be done in the background while other file206operations continue.207 208In summary, online fsck takes advantage of resource sharding and redundant209metadata to enable targeted checking and repair operations while the system210is running.211This capability will be coupled to automatic system management so that212autonomous self-healing of XFS maximizes service availability.213 2142. Theory of Operation215======================216 217Because it is necessary for online fsck to lock and scan live metadata objects,218online fsck consists of three separate code components.219The first is the userspace driver program ``xfs_scrub``, which is responsible220for identifying individual metadata items, scheduling work items for them,221reacting to the outcomes appropriately, and reporting results to the system222administrator.223The second and third are in the kernel, which implements functions to check224and repair each type of online fsck work item.225 226+------------------------------------------------------------------+227| **Note**: |228+------------------------------------------------------------------+229| For brevity, this document shortens the phrase "online fsck work |230| item" to "scrub item". |231+------------------------------------------------------------------+232 233Scrub item types are delineated in a manner consistent with the Unix design234philosophy, which is to say that each item should handle one aspect of a235metadata structure, and handle it well.236 237Scope238-----239 240In principle, online fsck should be able to check and to repair everything that241the offline fsck program can handle.242However, online fsck cannot be running 100% of the time, which means that243latent errors may creep in after a scrub completes.244If these errors cause the next mount to fail, offline fsck is the only245solution.246This limitation means that maintenance of the offline fsck tool will continue.247A second limitation of online fsck is that it must follow the same resource248sharing and lock acquisition rules as the regular filesystem.249This means that scrub cannot take *any* shortcuts to save time, because doing250so could lead to concurrency problems.251In other words, online fsck is not a complete replacement for offline fsck, and252a complete run of online fsck may take longer than online fsck.253However, both of these limitations are acceptable tradeoffs to satisfy the254different motivations of online fsck, which are to **minimize system downtime**255and to **increase predictability of operation**.256 257.. _scrubphases:258 259Phases of Work260--------------261 262The userspace driver program ``xfs_scrub`` splits the work of checking and263repairing an entire filesystem into seven phases.264Each phase concentrates on checking specific types of scrub items and depends265on the success of all previous phases.266The seven phases are as follows:267 2681. Collect geometry information about the mounted filesystem and computer,269 discover the online fsck capabilities of the kernel, and open the270 underlying storage devices.271 2722. Check allocation group metadata, all realtime volume metadata, and all quota273 files.274 Each metadata structure is scheduled as a separate scrub item.275 If corruption is found in the inode header or inode btree and ``xfs_scrub``276 is permitted to perform repairs, then those scrub items are repaired to277 prepare for phase 3.278 Repairs are implemented by using the information in the scrub item to279 resubmit the kernel scrub call with the repair flag enabled; this is280 discussed in the next section.281 Optimizations and all other repairs are deferred to phase 4.282 2833. Check all metadata of every file in the filesystem.284 Each metadata structure is also scheduled as a separate scrub item.285 If repairs are needed and ``xfs_scrub`` is permitted to perform repairs,286 and there were no problems detected during phase 2, then those scrub items287 are repaired immediately.288 Optimizations, deferred repairs, and unsuccessful repairs are deferred to289 phase 4.290 2914. All remaining repairs and scheduled optimizations are performed during this292 phase, if the caller permits them.293 Before starting repairs, the summary counters are checked and any necessary294 repairs are performed so that subsequent repairs will not fail the resource295 reservation step due to wildly incorrect summary counters.296 Unsuccessful repairs are requeued as long as forward progress on repairs is297 made somewhere in the filesystem.298 Free space in the filesystem is trimmed at the end of phase 4 if the299 filesystem is clean.300 3015. By the start of this phase, all primary and secondary filesystem metadata302 must be correct.303 Summary counters such as the free space counts and quota resource counts304 are checked and corrected.305 Directory entry names and extended attribute names are checked for306 suspicious entries such as control characters or confusing Unicode sequences307 appearing in names.308 3096. If the caller asks for a media scan, read all allocated and written data310 file extents in the filesystem.311 The ability to use hardware-assisted data file integrity checking is new312 to online fsck; neither of the previous tools have this capability.313 If media errors occur, they will be mapped to the owning files and reported.314 3157. Re-check the summary counters and presents the caller with a summary of316 space usage and file counts.317 318This allocation of responsibilities will be :ref:`revisited <scrubcheck>`319later in this document.320 321Steps for Each Scrub Item322-------------------------323 324The kernel scrub code uses a three-step strategy for checking and repairing325the one aspect of a metadata object represented by a scrub item:326 3271. The scrub item of interest is checked for corruptions; opportunities for328 optimization; and for values that are directly controlled by the system329 administrator but look suspicious.330 If the item is not corrupt or does not need optimization, resource are331 released and the positive scan results are returned to userspace.332 If the item is corrupt or could be optimized but the caller does not permit333 this, resources are released and the negative scan results are returned to334 userspace.335 Otherwise, the kernel moves on to the second step.336 3372. The repair function is called to rebuild the data structure.338 Repair functions generally choose rebuild a structure from other metadata339 rather than try to salvage the existing structure.340 If the repair fails, the scan results from the first step are returned to341 userspace.342 Otherwise, the kernel moves on to the third step.343 3443. In the third step, the kernel runs the same checks over the new metadata345 item to assess the efficacy of the repairs.346 The results of the reassessment are returned to userspace.347 348Classification of Metadata349--------------------------350 351Each type of metadata object (and therefore each type of scrub item) is352classified as follows:353 354Primary Metadata355````````````````356 357Metadata structures in this category should be most familiar to filesystem358users either because they are directly created by the user or they index359objects created by the user360Most filesystem objects fall into this class:361 362- Free space and reference count information363 364- Inode records and indexes365 366- Storage mapping information for file data367 368- Directories369 370- Extended attributes371 372- Symbolic links373 374- Quota limits375 376Scrub obeys the same rules as regular filesystem accesses for resource and lock377acquisition.378 379Primary metadata objects are the simplest for scrub to process.380The principal filesystem object (either an allocation group or an inode) that381owns the item being scrubbed is locked to guard against concurrent updates.382The check function examines every record associated with the type for obvious383errors and cross-references healthy records against other metadata to look for384inconsistencies.385Repairs for this class of scrub item are simple, since the repair function386starts by holding all the resources acquired in the previous step.387The repair function scans available metadata as needed to record all the388observations needed to complete the structure.389Next, it stages the observations in a new ondisk structure and commits it390atomically to complete the repair.391Finally, the storage from the old data structure are carefully reaped.392 393Because ``xfs_scrub`` locks a primary object for the duration of the repair,394this is effectively an offline repair operation performed on a subset of the395filesystem.396This minimizes the complexity of the repair code because it is not necessary to397handle concurrent updates from other threads, nor is it necessary to access398any other part of the filesystem.399As a result, indexed structures can be rebuilt very quickly, and programs400trying to access the damaged structure will be blocked until repairs complete.401The only infrastructure needed by the repair code are the staging area for402observations and a means to write new structures to disk.403Despite these limitations, the advantage that online repair holds is clear:404targeted work on individual shards of the filesystem avoids total loss of405service.406 407This mechanism is described in section 2.1 ("Off-Line Algorithm") of408V. Srinivasan and M. J. Carey, `"Performance of On-Line Index Construction409Algorithms" <https://minds.wisconsin.edu/bitstream/handle/1793/59524/TR1047.pdf>`_,410*Extending Database Technology*, pp. 293-309, 1992.411 412Most primary metadata repair functions stage their intermediate results in an413in-memory array prior to formatting the new ondisk structure, which is very414similar to the list-based algorithm discussed in section 2.3 ("List-Based415Algorithms") of Srinivasan.416However, any data structure builder that maintains a resource lock for the417duration of the repair is *always* an offline algorithm.418 419.. _secondary_metadata:420 421Secondary Metadata422``````````````````423 424Metadata structures in this category reflect records found in primary metadata,425but are only needed for online fsck or for reorganization of the filesystem.426 427Secondary metadata include:428 429- Reverse mapping information430 431- Directory parent pointers432 433This class of metadata is difficult for scrub to process because scrub attaches434to the secondary object but needs to check primary metadata, which runs counter435to the usual order of resource acquisition.436Frequently, this means that full filesystems scans are necessary to rebuild the437metadata.438Check functions can be limited in scope to reduce runtime.439Repairs, however, require a full scan of primary metadata, which can take a440long time to complete.441Under these conditions, ``xfs_scrub`` cannot lock resources for the entire442duration of the repair.443 444Instead, repair functions set up an in-memory staging structure to store445observations.446Depending on the requirements of the specific repair function, the staging447index will either have the same format as the ondisk structure or a design448specific to that repair function.449The next step is to release all locks and start the filesystem scan.450When the repair scanner needs to record an observation, the staging data are451locked long enough to apply the update.452While the filesystem scan is in progress, the repair function hooks the453filesystem so that it can apply pending filesystem updates to the staging454information.455Once the scan is done, the owning object is re-locked, the live data is used to456write a new ondisk structure, and the repairs are committed atomically.457The hooks are disabled and the staging staging area is freed.458Finally, the storage from the old data structure are carefully reaped.459 460Introducing concurrency helps online repair avoid various locking problems, but461comes at a high cost to code complexity.462Live filesystem code has to be hooked so that the repair function can observe463updates in progress.464The staging area has to become a fully functional parallel structure so that465updates can be merged from the hooks.466Finally, the hook, the filesystem scan, and the inode locking model must be467sufficiently well integrated that a hook event can decide if a given update468should be applied to the staging structure.469 470In theory, the scrub implementation could apply these same techniques for471primary metadata, but doing so would make it massively more complex and less472performant.473Programs attempting to access the damaged structures are not blocked from474operation, which may cause application failure or an unplanned filesystem475shutdown.476 477Inspiration for the secondary metadata repair strategy was drawn from section4782.4 of Srinivasan above, and sections 2 ("NSF: Inded Build Without Side-File")479and 3.1.1 ("Duplicate Key Insert Problem") in C. Mohan, `"Algorithms for480Creating Indexes for Very Large Tables Without Quiescing Updates"481<https://dl.acm.org/doi/10.1145/130283.130337>`_, 1992.482 483The sidecar index mentioned above bears some resemblance to the side file484method mentioned in Srinivasan and Mohan.485Their method consists of an index builder that extracts relevant record data to486build the new structure as quickly as possible; and an auxiliary structure that487captures all updates that would be committed to the index by other threads were488the new index already online.489After the index building scan finishes, the updates recorded in the side file490are applied to the new index.491To avoid conflicts between the index builder and other writer threads, the492builder maintains a publicly visible cursor that tracks the progress of the493scan through the record space.494To avoid duplication of work between the side file and the index builder, side495file updates are elided when the record ID for the update is greater than the496cursor position within the record ID space.497 498To minimize changes to the rest of the codebase, XFS online repair keeps the499replacement index hidden until it's completely ready to go.500In other words, there is no attempt to expose the keyspace of the new index501while repair is running.502The complexity of such an approach would be very high and perhaps more503appropriate to building *new* indices.504 505**Future Work Question**: Can the full scan and live update code used to506facilitate a repair also be used to implement a comprehensive check?507 508*Answer*: In theory, yes. Check would be much stronger if each scrub function509employed these live scans to build a shadow copy of the metadata and then510compared the shadow records to the ondisk records.511However, doing that is a fair amount more work than what the checking functions512do now.513The live scans and hooks were developed much later.514That in turn increases the runtime of those scrub functions.515 516Summary Information517```````````````````518 519Metadata structures in this last category summarize the contents of primary520metadata records.521These are often used to speed up resource usage queries, and are many times522smaller than the primary metadata which they represent.523 524Examples of summary information include:525 526- Summary counts of free space and inodes527 528- File link counts from directories529 530- Quota resource usage counts531 532Check and repair require full filesystem scans, but resource and lock533acquisition follow the same paths as regular filesystem accesses.534 535The superblock summary counters have special requirements due to the underlying536implementation of the incore counters, and will be treated separately.537Check and repair of the other types of summary counters (quota resource counts538and file link counts) employ the same filesystem scanning and hooking539techniques as outlined above, but because the underlying data are sets of540integer counters, the staging data need not be a fully functional mirror of the541ondisk structure.542 543Inspiration for quota and file link count repair strategies were drawn from544sections 2.12 ("Online Index Operations") through 2.14 ("Incremental View545Maintenance") of G. Graefe, `"Concurrent Queries and Updates in Summary Views546and Their Indexes"547<http://www.odbms.org/wp-content/uploads/2014/06/Increment-locks.pdf>`_, 2011.548 549Since quotas are non-negative integer counts of resource usage, online550quotacheck can use the incremental view deltas described in section 2.14 to551track pending changes to the block and inode usage counts in each transaction,552and commit those changes to a dquot side file when the transaction commits.553Delta tracking is necessary for dquots because the index builder scans inodes,554whereas the data structure being rebuilt is an index of dquots.555Link count checking combines the view deltas and commit step into one because556it sets attributes of the objects being scanned instead of writing them to a557separate data structure.558Each online fsck function will be discussed as case studies later in this559document.560 561Risk Management562---------------563 564During the development of online fsck, several risk factors were identified565that may make the feature unsuitable for certain distributors and users.566Steps can be taken to mitigate or eliminate those risks, though at a cost to567functionality.568 569- **Decreased performance**: Adding metadata indices to the filesystem570 increases the time cost of persisting changes to disk, and the reverse space571 mapping and directory parent pointers are no exception.572 System administrators who require the maximum performance can disable the573 reverse mapping features at format time, though this choice dramatically574 reduces the ability of online fsck to find inconsistencies and repair them.575 576- **Incorrect repairs**: As with all software, there might be defects in the577 software that result in incorrect repairs being written to the filesystem.578 Systematic fuzz testing (detailed in the next section) is employed by the579 authors to find bugs early, but it might not catch everything.580 The kernel build system provides Kconfig options (``CONFIG_XFS_ONLINE_SCRUB``581 and ``CONFIG_XFS_ONLINE_REPAIR``) to enable distributors to choose not to582 accept this risk.583 The xfsprogs build system has a configure option (``--enable-scrub=no``) that584 disables building of the ``xfs_scrub`` binary, though this is not a risk585 mitigation if the kernel functionality remains enabled.586 587- **Inability to repair**: Sometimes, a filesystem is too badly damaged to be588 repairable.589 If the keyspaces of several metadata indices overlap in some manner but a590 coherent narrative cannot be formed from records collected, then the repair591 fails.592 To reduce the chance that a repair will fail with a dirty transaction and593 render the filesystem unusable, the online repair functions have been594 designed to stage and validate all new records before committing the new595 structure.596 597- **Misbehavior**: Online fsck requires many privileges -- raw IO to block598 devices, opening files by handle, ignoring Unix discretionary access control,599 and the ability to perform administrative changes.600 Running this automatically in the background scares people, so the systemd601 background service is configured to run with only the privileges required.602 Obviously, this cannot address certain problems like the kernel crashing or603 deadlocking, but it should be sufficient to prevent the scrub process from604 escaping and reconfiguring the system.605 The cron job does not have this protection.606 607- **Fuzz Kiddiez**: There are many people now who seem to think that running608 automated fuzz testing of ondisk artifacts to find mischievous behavior and609 spraying exploit code onto the public mailing list for instant zero-day610 disclosure is somehow of some social benefit.611 In the view of this author, the benefit is realized only when the fuzz612 operators help to **fix** the flaws, but this opinion apparently is not613 widely shared among security "researchers".614 The XFS maintainers' continuing ability to manage these events presents an615 ongoing risk to the stability of the development process.616 Automated testing should front-load some of the risk while the feature is617 considered EXPERIMENTAL.618 619Many of these risks are inherent to software programming.620Despite this, it is hoped that this new functionality will prove useful in621reducing unexpected downtime.622 6233. Testing Plan624===============625 626As stated before, fsck tools have three main goals:627 6281. Detect inconsistencies in the metadata;629 6302. Eliminate those inconsistencies; and631 6323. Minimize further loss of data.633 634Demonstrations of correct operation are necessary to build users' confidence635that the software behaves within expectations.636Unfortunately, it was not really feasible to perform regular exhaustive testing637of every aspect of a fsck tool until the introduction of low-cost virtual638machines with high-IOPS storage.639With ample hardware availability in mind, the testing strategy for the online640fsck project involves differential analysis against the existing fsck tools and641systematic testing of every attribute of every type of metadata object.642Testing can be split into four major categories, as discussed below.643 644Integrated Testing with fstests645-------------------------------646 647The primary goal of any free software QA effort is to make testing as648inexpensive and widespread as possible to maximize the scaling advantages of649community.650In other words, testing should maximize the breadth of filesystem configuration651scenarios and hardware setups.652This improves code quality by enabling the authors of online fsck to find and653fix bugs early, and helps developers of new features to find integration654issues earlier in their development effort.655 656The Linux filesystem community shares a common QA testing suite,657`fstests <https://git.kernel.org/pub/scm/fs/xfs/xfstests-dev.git/>`_, for658functional and regression testing.659Even before development work began on online fsck, fstests (when run on XFS)660would run both the ``xfs_check`` and ``xfs_repair -n`` commands on the test and661scratch filesystems between each test.662This provides a level of assurance that the kernel and the fsck tools stay in663alignment about what constitutes consistent metadata.664During development of the online checking code, fstests was modified to run665``xfs_scrub -n`` between each test to ensure that the new checking code666produces the same results as the two existing fsck tools.667 668To start development of online repair, fstests was modified to run669``xfs_repair`` to rebuild the filesystem's metadata indices between tests.670This ensures that offline repair does not crash, leave a corrupt filesystem671after it exists, or trigger complaints from the online check.672This also established a baseline for what can and cannot be repaired offline.673To complete the first phase of development of online repair, fstests was674modified to be able to run ``xfs_scrub`` in a "force rebuild" mode.675This enables a comparison of the effectiveness of online repair as compared to676the existing offline repair tools.677 678General Fuzz Testing of Metadata Blocks679---------------------------------------680 681XFS benefits greatly from having a very robust debugging tool, ``xfs_db``.682 683Before development of online fsck even began, a set of fstests were created684to test the rather common fault that entire metadata blocks get corrupted.685This required the creation of fstests library code that can create a filesystem686containing every possible type of metadata object.687Next, individual test cases were created to create a test filesystem, identify688a single block of a specific type of metadata object, trash it with the689existing ``blocktrash`` command in ``xfs_db``, and test the reaction of a690particular metadata validation strategy.691 692This earlier test suite enabled XFS developers to test the ability of the693in-kernel validation functions and the ability of the offline fsck tool to694detect and eliminate the inconsistent metadata.695This part of the test suite was extended to cover online fsck in exactly the696same manner.697 698In other words, for a given fstests filesystem configuration:699 700* For each metadata object existing on the filesystem:701 702 * Write garbage to it703 704 * Test the reactions of:705 706 1. The kernel verifiers to stop obviously bad metadata707 2. Offline repair (``xfs_repair``) to detect and fix708 3. Online repair (``xfs_scrub``) to detect and fix709 710Targeted Fuzz Testing of Metadata Records711-----------------------------------------712 713The testing plan for online fsck includes extending the existing fs testing714infrastructure to provide a much more powerful facility: targeted fuzz testing715of every metadata field of every metadata object in the filesystem.716``xfs_db`` can modify every field of every metadata structure in every717block in the filesystem to simulate the effects of memory corruption and718software bugs.719Given that fstests already contains the ability to create a filesystem720containing every metadata format known to the filesystem, ``xfs_db`` can be721used to perform exhaustive fuzz testing!722 723For a given fstests filesystem configuration:724 725* For each metadata object existing on the filesystem...726 727 * For each record inside that metadata object...728 729 * For each field inside that record...730 731 * For each conceivable type of transformation that can be applied to a bit field...732 733 1. Clear all bits734 2. Set all bits735 3. Toggle the most significant bit736 4. Toggle the middle bit737 5. Toggle the least significant bit738 6. Add a small quantity739 7. Subtract a small quantity740 8. Randomize the contents741 742 * ...test the reactions of:743 744 1. The kernel verifiers to stop obviously bad metadata745 2. Offline checking (``xfs_repair -n``)746 3. Offline repair (``xfs_repair``)747 4. Online checking (``xfs_scrub -n``)748 5. Online repair (``xfs_scrub``)749 6. Both repair tools (``xfs_scrub`` and then ``xfs_repair`` if online repair doesn't succeed)750 751This is quite the combinatoric explosion!752 753Fortunately, having this much test coverage makes it easy for XFS developers to754check the responses of XFS' fsck tools.755Since the introduction of the fuzz testing framework, these tests have been756used to discover incorrect repair code and missing functionality for entire757classes of metadata objects in ``xfs_repair``.758The enhanced testing was used to finalize the deprecation of ``xfs_check`` by759confirming that ``xfs_repair`` could detect at least as many corruptions as760the older tool.761 762These tests have been very valuable for ``xfs_scrub`` in the same ways -- they763allow the online fsck developers to compare online fsck against offline fsck,764and they enable XFS developers to find deficiencies in the code base.765 766Proposed patchsets include767`general fuzzer improvements768<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfstests-dev.git/log/?h=fuzzer-improvements>`_,769`fuzzing baselines770<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfstests-dev.git/log/?h=fuzz-baseline>`_,771and `improvements in fuzz testing comprehensiveness772<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfstests-dev.git/log/?h=more-fuzz-testing>`_.773 774Stress Testing775--------------776 777A unique requirement to online fsck is the ability to operate on a filesystem778concurrently with regular workloads.779Although it is of course impossible to run ``xfs_scrub`` with *zero* observable780impact on the running system, the online repair code should never introduce781inconsistencies into the filesystem metadata, and regular workloads should782never notice resource starvation.783To verify that these conditions are being met, fstests has been enhanced in784the following ways:785 786* For each scrub item type, create a test to exercise checking that item type787 while running ``fsstress``.788* For each scrub item type, create a test to exercise repairing that item type789 while running ``fsstress``.790* Race ``fsstress`` and ``xfs_scrub -n`` to ensure that checking the whole791 filesystem doesn't cause problems.792* Race ``fsstress`` and ``xfs_scrub`` in force-rebuild mode to ensure that793 force-repairing the whole filesystem doesn't cause problems.794* Race ``xfs_scrub`` in check and force-repair mode against ``fsstress`` while795 freezing and thawing the filesystem.796* Race ``xfs_scrub`` in check and force-repair mode against ``fsstress`` while797 remounting the filesystem read-only and read-write.798* The same, but running ``fsx`` instead of ``fsstress``. (Not done yet?)799 800Success is defined by the ability to run all of these tests without observing801any unexpected filesystem shutdowns due to corrupted metadata, kernel hang802check warnings, or any other sort of mischief.803 804Proposed patchsets include `general stress testing805<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfstests-dev.git/log/?h=race-scrub-and-mount-state-changes>`_806and the `evolution of existing per-function stress testing807<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfstests-dev.git/log/?h=refactor-scrub-stress>`_.808 8094. User Interface810=================811 812The primary user of online fsck is the system administrator, just like offline813repair.814Online fsck presents two modes of operation to administrators:815A foreground CLI process for online fsck on demand, and a background service816that performs autonomous checking and repair.817 818Checking on Demand819------------------820 821For administrators who want the absolute freshest information about the822metadata in a filesystem, ``xfs_scrub`` can be run as a foreground process on823a command line.824The program checks every piece of metadata in the filesystem while the825administrator waits for the results to be reported, just like the existing826``xfs_repair`` tool.827Both tools share a ``-n`` option to perform a read-only scan, and a ``-v``828option to increase the verbosity of the information reported.829 830A new feature of ``xfs_scrub`` is the ``-x`` option, which employs the error831correction capabilities of the hardware to check data file contents.832The media scan is not enabled by default because it may dramatically increase833program runtime and consume a lot of bandwidth on older storage hardware.834 835The output of a foreground invocation is captured in the system log.836 837The ``xfs_scrub_all`` program walks the list of mounted filesystems and838initiates ``xfs_scrub`` for each of them in parallel.839It serializes scans for any filesystems that resolve to the same top level840kernel block device to prevent resource overconsumption.841 842Background Service843------------------844 845To reduce the workload of system administrators, the ``xfs_scrub`` package846provides a suite of `systemd <https://systemd.io/>`_ timers and services that847run online fsck automatically on weekends by default.848The background service configures scrub to run with as little privilege as849possible, the lowest CPU and IO priority, and in a CPU-constrained single850threaded mode.851This can be tuned by the systemd administrator at any time to suit the latency852and throughput requirements of customer workloads.853 854The output of the background service is also captured in the system log.855If desired, reports of failures (either due to inconsistencies or mere runtime856errors) can be emailed automatically by setting the ``EMAIL_ADDR`` environment857variable in the following service files:858 859* ``xfs_scrub_fail@.service``860* ``xfs_scrub_media_fail@.service``861* ``xfs_scrub_all_fail.service``862 863The decision to enable the background scan is left to the system administrator.864This can be done by enabling either of the following services:865 866* ``xfs_scrub_all.timer`` on systemd systems867* ``xfs_scrub_all.cron`` on non-systemd systems868 869This automatic weekly scan is configured out of the box to perform an870additional media scan of all file data once per month.871This is less foolproof than, say, storing file data block checksums, but much872more performant if application software provides its own integrity checking,873redundancy can be provided elsewhere above the filesystem, or the storage874device's integrity guarantees are deemed sufficient.875 876The systemd unit file definitions have been subjected to a security audit877(as of systemd 249) to ensure that the xfs_scrub processes have as little878access to the rest of the system as possible.879This was performed via ``systemd-analyze security``, after which privileges880were restricted to the minimum required, sandboxing was set up to the maximal881extent possible with sandboxing and system call filtering; and access to the882filesystem tree was restricted to the minimum needed to start the program and883access the filesystem being scanned.884The service definition files restrict CPU usage to 80% of one CPU core, and885apply as nice of a priority to IO and CPU scheduling as possible.886This measure was taken to minimize delays in the rest of the filesystem.887No such hardening has been performed for the cron job.888 889Proposed patchset:890`Enabling the xfs_scrub background service891<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfsprogs-dev.git/log/?h=scrub-media-scan-service>`_.892 893Health Reporting894----------------895 896XFS caches a summary of each filesystem's health status in memory.897The information is updated whenever ``xfs_scrub`` is run, or whenever898inconsistencies are detected in the filesystem metadata during regular899operations.900System administrators should use the ``health`` command of ``xfs_spaceman`` to901download this information into a human-readable format.902If problems have been observed, the administrator can schedule a reduced903service window to run the online repair tool to correct the problem.904Failing that, the administrator can decide to schedule a maintenance window to905run the traditional offline repair tool to correct the problem.906 907**Future Work Question**: Should the health reporting integrate with the new908inotify fs error notification system?909Would it be helpful for sysadmins to have a daemon to listen for corruption910notifications and initiate a repair?911 912*Answer*: These questions remain unanswered, but should be a part of the913conversation with early adopters and potential downstream users of XFS.914 915Proposed patchsets include916`wiring up health reports to correction returns917<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=corruption-health-reports>`_918and919`preservation of sickness info during memory reclaim920<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=indirect-health-reporting>`_.921 9225. Kernel Algorithms and Data Structures923========================================924 925This section discusses the key algorithms and data structures of the kernel926code that provide the ability to check and repair metadata while the system927is running.928The first chapters in this section reveal the pieces that provide the929foundation for checking metadata.930The remainder of this section presents the mechanisms through which XFS931regenerates itself.932 933Self Describing Metadata934------------------------935 936Starting with XFS version 5 in 2012, XFS updated the format of nearly every937ondisk block header to record a magic number, a checksum, a universally938"unique" identifier (UUID), an owner code, the ondisk address of the block,939and a log sequence number.940When loading a block buffer from disk, the magic number, UUID, owner, and941ondisk address confirm that the retrieved block matches the specific owner of942the current filesystem, and that the information contained in the block is943supposed to be found at the ondisk address.944The first three components enable checking tools to disregard alleged metadata945that doesn't belong to the filesystem, and the fourth component enables the946filesystem to detect lost writes.947 948Whenever a file system operation modifies a block, the change is submitted949to the log as part of a transaction.950The log then processes these transactions marking them done once they are951safely persisted to storage.952The logging code maintains the checksum and the log sequence number of the last953transactional update.954Checksums are useful for detecting torn writes and other discrepancies that can955be introduced between the computer and its storage devices.956Sequence number tracking enables log recovery to avoid applying out of date957log updates to the filesystem.958 959These two features improve overall runtime resiliency by providing a means for960the filesystem to detect obvious corruption when reading metadata blocks from961disk, but these buffer verifiers cannot provide any consistency checking962between metadata structures.963 964For more information, please see the documentation for965Documentation/filesystems/xfs/xfs-self-describing-metadata.rst966 967Reverse Mapping968---------------969 970The original design of XFS (circa 1993) is an improvement upon 1980s Unix971filesystem design.972In those days, storage density was expensive, CPU time was scarce, and973excessive seek time could kill performance.974For performance reasons, filesystem authors were reluctant to add redundancy to975the filesystem, even at the cost of data integrity.976Filesystems designers in the early 21st century choose different strategies to977increase internal redundancy -- either storing nearly identical copies of978metadata, or more space-efficient encoding techniques.979 980For XFS, a different redundancy strategy was chosen to modernize the design:981a secondary space usage index that maps allocated disk extents back to their982owners.983By adding a new index, the filesystem retains most of its ability to scale984well to heavily threaded workloads involving large datasets, since the primary985file metadata (the directory tree, the file block map, and the allocation986groups) remain unchanged.987Like any system that improves redundancy, the reverse-mapping feature increases988overhead costs for space mapping activities.989However, it has two critical advantages: first, the reverse index is key to990enabling online fsck and other requested functionality such as free space991defragmentation, better media failure reporting, and filesystem shrinking.992Second, the different ondisk storage format of the reverse mapping btree993defeats device-level deduplication because the filesystem requires real994redundancy.995 996+--------------------------------------------------------------------------+997| **Sidebar**: |998+--------------------------------------------------------------------------+999| A criticism of adding the secondary index is that it does nothing to |1000| improve the robustness of user data storage itself. |1001| This is a valid point, but adding a new index for file data block |1002| checksums increases write amplification by turning data overwrites into |1003| copy-writes, which age the filesystem prematurely. |1004| In keeping with thirty years of precedent, users who want file data |1005| integrity can supply as powerful a solution as they require. |1006| As for metadata, the complexity of adding a new secondary index of space |1007| usage is much less than adding volume management and storage device |1008| mirroring to XFS itself. |1009| Perfection of RAID and volume management are best left to existing |1010| layers in the kernel. |1011+--------------------------------------------------------------------------+1012 1013The information captured in a reverse space mapping record is as follows:1014 1015.. code-block:: c1016 1017 struct xfs_rmap_irec {1018 xfs_agblock_t rm_startblock; /* extent start block */1019 xfs_extlen_t rm_blockcount; /* extent length */1020 uint64_t rm_owner; /* extent owner */1021 uint64_t rm_offset; /* offset within the owner */1022 unsigned int rm_flags; /* state flags */1023 };1024 1025The first two fields capture the location and size of the physical space,1026in units of filesystem blocks.1027The owner field tells scrub which metadata structure or file inode have been1028assigned this space.1029For space allocated to files, the offset field tells scrub where the space was1030mapped within the file fork.1031Finally, the flags field provides extra information about the space usage --1032is this an attribute fork extent? A file mapping btree extent? Or an1033unwritten data extent?1034 1035Online filesystem checking judges the consistency of each primary metadata1036record by comparing its information against all other space indices.1037The reverse mapping index plays a key role in the consistency checking process1038because it contains a centralized alternate copy of all space allocation1039information.1040Program runtime and ease of resource acquisition are the only real limits to1041what online checking can consult.1042For example, a file data extent mapping can be checked against:1043 1044* The absence of an entry in the free space information.1045* The absence of an entry in the inode index.1046* The absence of an entry in the reference count data if the file is not1047 marked as having shared extents.1048* The correspondence of an entry in the reverse mapping information.1049 1050There are several observations to make about reverse mapping indices:1051 10521. Reverse mappings can provide a positive affirmation of correctness if any of1053 the above primary metadata are in doubt.1054 The checking code for most primary metadata follows a path similar to the1055 one outlined above.1056 10572. Proving the consistency of secondary metadata with the primary metadata is1058 difficult because that requires a full scan of all primary space metadata,1059 which is very time intensive.1060 For example, checking a reverse mapping record for a file extent mapping1061 btree block requires locking the file and searching the entire btree to1062 confirm the block.1063 Instead, scrub relies on rigorous cross-referencing during the primary space1064 mapping structure checks.1065 10663. Consistency scans must use non-blocking lock acquisition primitives if the1067 required locking order is not the same order used by regular filesystem1068 operations.1069 For example, if the filesystem normally takes a file ILOCK before taking1070 the AGF buffer lock but scrub wants to take a file ILOCK while holding1071 an AGF buffer lock, scrub cannot block on that second acquisition.1072 This means that forward progress during this part of a scan of the reverse1073 mapping data cannot be guaranteed if system load is heavy.1074 1075In summary, reverse mappings play a key role in reconstruction of primary1076metadata.1077The details of how these records are staged, written to disk, and committed1078into the filesystem are covered in subsequent sections.1079 1080Checking and Cross-Referencing1081------------------------------1082 1083The first step of checking a metadata structure is to examine every record1084contained within the structure and its relationship with the rest of the1085system.1086XFS contains multiple layers of checking to try to prevent inconsistent1087metadata from wreaking havoc on the system.1088Each of these layers contributes information that helps the kernel to make1089three decisions about the health of a metadata structure:1090 1091- Is a part of this structure obviously corrupt (``XFS_SCRUB_OFLAG_CORRUPT``) ?1092- Is this structure inconsistent with the rest of the system1093 (``XFS_SCRUB_OFLAG_XCORRUPT``) ?1094- Is there so much damage around the filesystem that cross-referencing is not1095 possible (``XFS_SCRUB_OFLAG_XFAIL``) ?1096- Can the structure be optimized to improve performance or reduce the size of1097 metadata (``XFS_SCRUB_OFLAG_PREEN``) ?1098- Does the structure contain data that is not inconsistent but deserves review1099 by the system administrator (``XFS_SCRUB_OFLAG_WARNING``) ?1100 1101The following sections describe how the metadata scrubbing process works.1102 1103Metadata Buffer Verification1104````````````````````````````1105 1106The lowest layer of metadata protection in XFS are the metadata verifiers built1107into the buffer cache.1108These functions perform inexpensive internal consistency checking of the block1109itself, and answer these questions:1110 1111- Does the block belong to this filesystem?1112 1113- Does the block belong to the structure that asked for the read?1114 This assumes that metadata blocks only have one owner, which is always true1115 in XFS.1116 1117- Is the type of data stored in the block within a reasonable range of what1118 scrub is expecting?1119 1120- Does the physical location of the block match the location it was read from?1121 1122- Does the block checksum match the data?1123 1124The scope of the protections here are very limited -- verifiers can only1125establish that the filesystem code is reasonably free of gross corruption bugs1126and that the storage system is reasonably competent at retrieval.1127Corruption problems observed at runtime cause the generation of health reports,1128failed system calls, and in the extreme case, filesystem shutdowns if the1129corrupt metadata force the cancellation of a dirty transaction.1130 1131Every online fsck scrubbing function is expected to read every ondisk metadata1132block of a structure in the course of checking the structure.1133Corruption problems observed during a check are immediately reported to1134userspace as corruption; during a cross-reference, they are reported as a1135failure to cross-reference once the full examination is complete.1136Reads satisfied by a buffer already in cache (and hence already verified)1137bypass these checks.1138 1139Internal Consistency Checks1140```````````````````````````1141 1142After the buffer cache, the next level of metadata protection is the internal1143record verification code built into the filesystem.1144These checks are split between the buffer verifiers, the in-filesystem users of1145the buffer cache, and the scrub code itself, depending on the amount of higher1146level context required.1147The scope of checking is still internal to the block.1148These higher level checking functions answer these questions:1149 1150- Does the type of data stored in the block match what scrub is expecting?1151 1152- Does the block belong to the owning structure that asked for the read?1153 1154- If the block contains records, do the records fit within the block?1155 1156- If the block tracks internal free space information, is it consistent with1157 the record areas?1158 1159- Are the records contained inside the block free of obvious corruptions?1160 1161Record checks in this category are more rigorous and more time-intensive.1162For example, block pointers and inumbers are checked to ensure that they point1163within the dynamically allocated parts of an allocation group and within1164the filesystem.1165Names are checked for invalid characters, and flags are checked for invalid1166combinations.1167Other record attributes are checked for sensible values.1168Btree records spanning an interval of the btree keyspace are checked for1169correct order and lack of mergeability (except for file fork mappings).1170For performance reasons, regular code may skip some of these checks unless1171debugging is enabled or a write is about to occur.1172Scrub functions, of course, must check all possible problems.1173 1174Validation of Userspace-Controlled Record Attributes1175````````````````````````````````````````````````````1176 1177Various pieces of filesystem metadata are directly controlled by userspace.1178Because of this nature, validation work cannot be more precise than checking1179that a value is within the possible range.1180These fields include:1181 1182- Superblock fields controlled by mount options1183- Filesystem labels1184- File timestamps1185- File permissions1186- File size1187- File flags1188- Names present in directory entries, extended attribute keys, and filesystem1189 labels1190- Extended attribute key namespaces1191- Extended attribute values1192- File data block contents1193- Quota limits1194- Quota timer expiration (if resource usage exceeds the soft limit)1195 1196Cross-Referencing Space Metadata1197````````````````````````````````1198 1199After internal block checks, the next higher level of checking is1200cross-referencing records between metadata structures.1201For regular runtime code, the cost of these checks is considered to be1202prohibitively expensive, but as scrub is dedicated to rooting out1203inconsistencies, it must pursue all avenues of inquiry.1204The exact set of cross-referencing is highly dependent on the context of the1205data structure being checked.1206 1207The XFS btree code has keyspace scanning functions that online fsck uses to1208cross reference one structure with another.1209Specifically, scrub can scan the key space of an index to determine if that1210keyspace is fully, sparsely, or not at all mapped to records.1211For the reverse mapping btree, it is possible to mask parts of the key for the1212purposes of performing a keyspace scan so that scrub can decide if the rmap1213btree contains records mapping a certain extent of physical space without the1214sparsenses of the rest of the rmap keyspace getting in the way.1215 1216Btree blocks undergo the following checks before cross-referencing:1217 1218- Does the type of data stored in the block match what scrub is expecting?1219 1220- Does the block belong to the owning structure that asked for the read?1221 1222- Do the records fit within the block?1223 1224- Are the records contained inside the block free of obvious corruptions?1225 1226- Are the name hashes in the correct order?1227 1228- Do node pointers within the btree point to valid block addresses for the type1229 of btree?1230 1231- Do child pointers point towards the leaves?1232 1233- Do sibling pointers point across the same level?1234 1235- For each node block record, does the record key accurate reflect the contents1236 of the child block?1237 1238Space allocation records are cross-referenced as follows:1239 12401. Any space mentioned by any metadata structure are cross-referenced as1241 follows:1242 1243 - Does the reverse mapping index list only the appropriate owner as the1244 owner of each block?1245 1246 - Are none of the blocks claimed as free space?1247 1248 - If these aren't file data blocks, are none of the blocks claimed as space1249 shared by different owners?1250 12512. Btree blocks are cross-referenced as follows:1252 1253 - Everything in class 1 above.1254 1255 - If there's a parent node block, do the keys listed for this block match the1256 keyspace of this block?1257 1258 - Do the sibling pointers point to valid blocks? Of the same level?1259 1260 - Do the child pointers point to valid blocks? Of the next level down?1261 12623. Free space btree records are cross-referenced as follows:1263 1264 - Everything in class 1 and 2 above.1265 1266 - Does the reverse mapping index list no owners of this space?1267 1268 - Is this space not claimed by the inode index for inodes?1269 1270 - Is it not mentioned by the reference count index?1271 1272 - Is there a matching record in the other free space btree?1273 12744. Inode btree records are cross-referenced as follows:1275 1276 - Everything in class 1 and 2 above.1277 1278 - Is there a matching record in free inode btree?1279 1280 - Do cleared bits in the holemask correspond with inode clusters?1281 1282 - Do set bits in the freemask correspond with inode records with zero link1283 count?1284 12855. Inode records are cross-referenced as follows:1286 1287 - Everything in class 1.1288 1289 - Do all the fields that summarize information about the file forks actually1290 match those forks?1291 1292 - Does each inode with zero link count correspond to a record in the free1293 inode btree?1294 12956. File fork space mapping records are cross-referenced as follows:1296 1297 - Everything in class 1 and 2 above.1298 1299 - Is this space not mentioned by the inode btrees?1300 1301 - If this is a CoW fork mapping, does it correspond to a CoW entry in the1302 reference count btree?1303 13047. Reference count records are cross-referenced as follows:1305 1306 - Everything in class 1 and 2 above.1307 1308 - Within the space subkeyspace of the rmap btree (that is to say, all1309 records mapped to a particular space extent and ignoring the owner info),1310 are there the same number of reverse mapping records for each block as the1311 reference count record claims?1312 1313Proposed patchsets are the series to find gaps in1314`refcount btree1315<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=scrub-detect-refcount-gaps>`_,1316`inode btree1317<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=scrub-detect-inobt-gaps>`_, and1318`rmap btree1319<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=scrub-detect-rmapbt-gaps>`_ records;1320to find1321`mergeable records1322<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=scrub-detect-mergeable-records>`_;1323and to1324`improve cross referencing with rmap1325<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=scrub-strengthen-rmap-checking>`_1326before starting a repair.1327 1328Checking Extended Attributes1329````````````````````````````1330 1331Extended attributes implement a key-value store that enable fragments of data1332to be attached to any file.1333Both the kernel and userspace can access the keys and values, subject to1334namespace and privilege restrictions.1335Most typically these fragments are metadata about the file -- origins, security1336contexts, user-supplied labels, indexing information, etc.1337 1338Names can be as long as 255 bytes and can exist in several different1339namespaces.1340Values can be as large as 64KB.1341A file's extended attributes are stored in blocks mapped by the attr fork.1342The mappings point to leaf blocks, remote value blocks, or dabtree blocks.1343Block 0 in the attribute fork is always the top of the structure, but otherwise1344each of the three types of blocks can be found at any offset in the attr fork.1345Leaf blocks contain attribute key records that point to the name and the value.1346Names are always stored elsewhere in the same leaf block.1347Values that are less than 3/4 the size of a filesystem block are also stored1348elsewhere in the same leaf block.1349Remote value blocks contain values that are too large to fit inside a leaf.1350If the leaf information exceeds a single filesystem block, a dabtree (also1351rooted at block 0) is created to map hashes of the attribute names to leaf1352blocks in the attr fork.1353 1354Checking an extended attribute structure is not so straightforward due to the1355lack of separation between attr blocks and index blocks.1356Scrub must read each block mapped by the attr fork and ignore the non-leaf1357blocks:1358 13591. Walk the dabtree in the attr fork (if present) to ensure that there are no1360 irregularities in the blocks or dabtree mappings that do not point to1361 attr leaf blocks.1362 13632. Walk the blocks of the attr fork looking for leaf blocks.1364 For each entry inside a leaf:1365 1366 a. Validate that the name does not contain invalid characters.1367 1368 b. Read the attr value.1369 This performs a named lookup of the attr name to ensure the correctness1370 of the dabtree.1371 If the value is stored in a remote block, this also validates the1372 integrity of the remote value block.1373 1374Checking and Cross-Referencing Directories1375``````````````````````````````````````````1376 1377The filesystem directory tree is a directed acylic graph structure, with files1378constituting the nodes, and directory entries (dirents) constituting the edges.1379Directories are a special type of file containing a set of mappings from a1380255-byte sequence (name) to an inumber.1381These are called directory entries, or dirents for short.1382Each directory file must have exactly one directory pointing to the file.1383A root directory points to itself.1384Directory entries point to files of any type.1385Each non-directory file may have multiple directories point to it.1386 1387In XFS, directories are implemented as a file containing up to three 32GB1388partitions.1389The first partition contains directory entry data blocks.1390Each data block contains variable-sized records associating a user-provided1391name with an inumber and, optionally, a file type.1392If the directory entry data grows beyond one block, the second partition (which1393exists as post-EOF extents) is populated with a block containing free space1394information and an index that maps hashes of the dirent names to directory data1395blocks in the first partition.1396This makes directory name lookups very fast.1397If this second partition grows beyond one block, the third partition is1398populated with a linear array of free space information for faster1399expansions.1400If the free space has been separated and the second partition grows again1401beyond one block, then a dabtree is used to map hashes of dirent names to1402directory data blocks.1403 1404Checking a directory is pretty straightforward:1405 14061. Walk the dabtree in the second partition (if present) to ensure that there1407 are no irregularities in the blocks or dabtree mappings that do not point to1408 dirent blocks.1409 14102. Walk the blocks of the first partition looking for directory entries.1411 Each dirent is checked as follows:1412 1413 a. Does the name contain no invalid characters?1414 1415 b. Does the inumber correspond to an actual, allocated inode?1416 1417 c. Does the child inode have a nonzero link count?1418 1419 d. If a file type is included in the dirent, does it match the type of the1420 inode?1421 1422 e. If the child is a subdirectory, does the child's dotdot pointer point1423 back to the parent?1424 1425 f. If the directory has a second partition, perform a named lookup of the1426 dirent name to ensure the correctness of the dabtree.1427 14283. Walk the free space list in the third partition (if present) to ensure that1429 the free spaces it describes are really unused.1430 1431Checking operations involving :ref:`parents <dirparent>` and1432:ref:`file link counts <nlinks>` are discussed in more detail in later1433sections.1434 1435Checking Directory/Attribute Btrees1436^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^1437 1438As stated in previous sections, the directory/attribute btree (dabtree) index1439maps user-provided names to improve lookup times by avoiding linear scans.1440Internally, it maps a 32-bit hash of the name to a block offset within the1441appropriate file fork.1442 1443The internal structure of a dabtree closely resembles the btrees that record1444fixed-size metadata records -- each dabtree block contains a magic number, a1445checksum, sibling pointers, a UUID, a tree level, and a log sequence number.1446The format of leaf and node records are the same -- each entry points to the1447next level down in the hierarchy, with dabtree node records pointing to dabtree1448leaf blocks, and dabtree leaf records pointing to non-dabtree blocks elsewhere1449in the fork.1450 1451Checking and cross-referencing the dabtree is very similar to what is done for1452space btrees:1453 1454- Does the type of data stored in the block match what scrub is expecting?1455 1456- Does the block belong to the owning structure that asked for the read?1457 1458- Do the records fit within the block?1459 1460- Are the records contained inside the block free of obvious corruptions?1461 1462- Are the name hashes in the correct order?1463 1464- Do node pointers within the dabtree point to valid fork offsets for dabtree1465 blocks?1466 1467- Do leaf pointers within the dabtree point to valid fork offsets for directory1468 or attr leaf blocks?1469 1470- Do child pointers point towards the leaves?1471 1472- Do sibling pointers point across the same level?1473 1474- For each dabtree node record, does the record key accurate reflect the1475 contents of the child dabtree block?1476 1477- For each dabtree leaf record, does the record key accurate reflect the1478 contents of the directory or attr block?1479 1480Cross-Referencing Summary Counters1481``````````````````````````````````1482 1483XFS maintains three classes of summary counters: available resources, quota1484resource usage, and file link counts.1485 1486In theory, the amount of available resources (data blocks, inodes, realtime1487extents) can be found by walking the entire filesystem.1488This would make for very slow reporting, so a transactional filesystem can1489maintain summaries of this information in the superblock.1490Cross-referencing these values against the filesystem metadata should be a1491simple matter of walking the free space and inode metadata in each AG and the1492realtime bitmap, but there are complications that will be discussed in1493:ref:`more detail <fscounters>` later.1494 1495:ref:`Quota usage <quotacheck>` and :ref:`file link count <nlinks>`1496checking are sufficiently complicated to warrant separate sections.1497 1498Post-Repair Reverification1499``````````````````````````1500 1501After performing a repair, the checking code is run a second time to validate1502the new structure, and the results of the health assessment are recorded1503internally and returned to the calling process.1504This step is critical for enabling system administrator to monitor the status1505of the filesystem and the progress of any repairs.1506For developers, it is a useful means to judge the efficacy of error detection1507and correction in the online and offline checking tools.1508 1509Eventual Consistency vs. Online Fsck1510------------------------------------1511 1512Complex operations can make modifications to multiple per-AG data structures1513with a chain of transactions.1514These chains, once committed to the log, are restarted during log recovery if1515the system crashes while processing the chain.1516Because the AG header buffers are unlocked between transactions within a chain,1517online checking must coordinate with chained operations that are in progress to1518avoid incorrectly detecting inconsistencies due to pending chains.1519Furthermore, online repair must not run when operations are pending because1520the metadata are temporarily inconsistent with each other, and rebuilding is1521not possible.1522 1523Only online fsck has this requirement of total consistency of AG metadata, and1524should be relatively rare as compared to filesystem change operations.1525Online fsck coordinates with transaction chains as follows:1526 1527* For each AG, maintain a count of intent items targeting that AG.1528 The count should be bumped whenever a new item is added to the chain.1529 The count should be dropped when the filesystem has locked the AG header1530 buffers and finished the work.1531 1532* When online fsck wants to examine an AG, it should lock the AG header1533 buffers to quiesce all transaction chains that want to modify that AG.1534 If the count is zero, proceed with the checking operation.1535 If it is nonzero, cycle the buffer locks to allow the chain to make forward1536 progress.1537 1538This may lead to online fsck taking a long time to complete, but regular1539filesystem updates take precedence over background checking activity.1540Details about the discovery of this situation are presented in the1541:ref:`next section <chain_coordination>`, and details about the solution1542are presented :ref:`after that<intent_drains>`.1543 1544.. _chain_coordination:1545 1546Discovery of the Problem1547````````````````````````1548 1549Midway through the development of online scrubbing, the fsstress tests1550uncovered a misinteraction between online fsck and compound transaction chains1551created by other writer threads that resulted in false reports of metadata1552inconsistency.1553The root cause of these reports is the eventual consistency model introduced by1554the expansion of deferred work items and compound transaction chains when1555reverse mapping and reflink were introduced.1556 1557Originally, transaction chains were added to XFS to avoid deadlocks when1558unmapping space from files.1559Deadlock avoidance rules require that AGs only be locked in increasing order,1560which makes it impossible (say) to use a single transaction to free a space1561extent in AG 7 and then try to free a now superfluous block mapping btree block1562in AG 3.1563To avoid these kinds of deadlocks, XFS creates Extent Freeing Intent (EFI) log1564items to commit to freeing some space in one transaction while deferring the1565actual metadata updates to a fresh transaction.1566The transaction sequence looks like this:1567 15681. The first transaction contains a physical update to the file's block mapping1569 structures to remove the mapping from the btree blocks.1570 It then attaches to the in-memory transaction an action item to schedule1571 deferred freeing of space.1572 Concretely, each transaction maintains a list of ``struct1573 xfs_defer_pending`` objects, each of which maintains a list of ``struct1574 xfs_extent_free_item`` objects.1575 Returning to the example above, the action item tracks the freeing of both1576 the unmapped space from AG 7 and the block mapping btree (BMBT) block from1577 AG 3.1578 Deferred frees recorded in this manner are committed in the log by creating1579 an EFI log item from the ``struct xfs_extent_free_item`` object and1580 attaching the log item to the transaction.1581 When the log is persisted to disk, the EFI item is written into the ondisk1582 transaction record.1583 EFIs can list up to 16 extents to free, all sorted in AG order.1584 15852. The second transaction contains a physical update to the free space btrees1586 of AG 3 to release the former BMBT block and a second physical update to the1587 free space btrees of AG 7 to release the unmapped file space.1588 Observe that the physical updates are resequenced in the correct order1589 when possible.1590 Attached to the transaction is a an extent free done (EFD) log item.1591 The EFD contains a pointer to the EFI logged in transaction #1 so that log1592 recovery can tell if the EFI needs to be replayed.1593 1594If the system goes down after transaction #1 is written back to the filesystem1595but before #2 is committed, a scan of the filesystem metadata would show1596inconsistent filesystem metadata because there would not appear to be any owner1597of the unmapped space.1598Happily, log recovery corrects this inconsistency for us -- when recovery finds1599an intent log item but does not find a corresponding intent done item, it will1600reconstruct the incore state of the intent item and finish it.1601In the example above, the log must replay both frees described in the recovered1602EFI to complete the recovery phase.1603 1604There are subtleties to XFS' transaction chaining strategy to consider:1605 1606* Log items must be added to a transaction in the correct order to prevent1607 conflicts with principal objects that are not held by the transaction.1608 In other words, all per-AG metadata updates for an unmapped block must be1609 completed before the last update to free the extent, and extents should not1610 be reallocated until that last update commits to the log.1611 1612* AG header buffers are released between each transaction in a chain.1613 This means that other threads can observe an AG in an intermediate state,1614 but as long as the first subtlety is handled, this should not affect the1615 correctness of filesystem operations.1616 1617* Unmounting the filesystem flushes all pending work to disk, which means that1618 offline fsck never sees the temporary inconsistencies caused by deferred1619 work item processing.1620 1621In this manner, XFS employs a form of eventual consistency to avoid deadlocks1622and increase parallelism.1623 1624During the design phase of the reverse mapping and reflink features, it was1625decided that it was impractical to cram all the reverse mapping updates for a1626single filesystem change into a single transaction because a single file1627mapping operation can explode into many small updates:1628 1629* The block mapping update itself1630* A reverse mapping update for the block mapping update1631* Fixing the freelist1632* A reverse mapping update for the freelist fix1633 1634* A shape change to the block mapping btree1635* A reverse mapping update for the btree update1636* Fixing the freelist (again)1637* A reverse mapping update for the freelist fix1638 1639* An update to the reference counting information1640* A reverse mapping update for the refcount update1641* Fixing the freelist (a third time)1642* A reverse mapping update for the freelist fix1643 1644* Freeing any space that was unmapped and not owned by any other file1645* Fixing the freelist (a fourth time)1646* A reverse mapping update for the freelist fix1647 1648* Freeing the space used by the block mapping btree1649* Fixing the freelist (a fifth time)1650* A reverse mapping update for the freelist fix1651 1652Free list fixups are not usually needed more than once per AG per transaction1653chain, but it is theoretically possible if space is very tight.1654For copy-on-write updates this is even worse, because this must be done once to1655remove the space from a staging area and again to map it into the file!1656 1657To deal with this explosion in a calm manner, XFS expands its use of deferred1658work items to cover most reverse mapping updates and all refcount updates.1659This reduces the worst case size of transaction reservations by breaking the1660work into a long chain of small updates, which increases the degree of eventual1661consistency in the system.1662Again, this generally isn't a problem because XFS orders its deferred work1663items carefully to avoid resource reuse conflicts between unsuspecting threads.1664 1665However, online fsck changes the rules -- remember that although physical1666updates to per-AG structures are coordinated by locking the buffers for AG1667headers, buffer locks are dropped between transactions.1668Once scrub acquires resources and takes locks for a data structure, it must do1669all the validation work without releasing the lock.1670If the main lock for a space btree is an AG header buffer lock, scrub may have1671interrupted another thread that is midway through finishing a chain.1672For example, if a thread performing a copy-on-write has completed a reverse1673mapping update but not the corresponding refcount update, the two AG btrees1674will appear inconsistent to scrub and an observation of corruption will be1675recorded. This observation will not be correct.1676If a repair is attempted in this state, the results will be catastrophic!1677 1678Several other solutions to this problem were evaluated upon discovery of this1679flaw and rejected:1680 16811. Add a higher level lock to allocation groups and require writer threads to1682 acquire the higher level lock in AG order before making any changes.1683 This would be very difficult to implement in practice because it is1684 difficult to determine which locks need to be obtained, and in what order,1685 without simulating the entire operation.1686 Performing a dry run of a file operation to discover necessary locks would1687 make the filesystem very slow.1688 16892. Make the deferred work coordinator code aware of consecutive intent items1690 targeting the same AG and have it hold the AG header buffers locked across1691 the transaction roll between updates.1692 This would introduce a lot of complexity into the coordinator since it is1693 only loosely coupled with the actual deferred work items.1694 It would also fail to solve the problem because deferred work items can1695 generate new deferred subtasks, but all subtasks must be complete before1696 work can start on a new sibling task.1697 16983. Teach online fsck to walk all transactions waiting for whichever lock(s)1699 protect the data structure being scrubbed to look for pending operations.1700 The checking and repair operations must factor these pending operations into1701 the evaluations being performed.1702 This solution is a nonstarter because it is *extremely* invasive to the main1703 filesystem.1704 1705.. _intent_drains:1706 1707Intent Drains1708`````````````1709 1710Online fsck uses an atomic intent item counter and lock cycling to coordinate1711with transaction chains.1712There are two key properties to the drain mechanism.1713First, the counter is incremented when a deferred work item is *queued* to a1714transaction, and it is decremented after the associated intent done log item is1715*committed* to another transaction.1716The second property is that deferred work can be added to a transaction without1717holding an AG header lock, but per-AG work items cannot be marked done without1718locking that AG header buffer to log the physical updates and the intent done1719log item.1720The first property enables scrub to yield to running transaction chains, which1721is an explicit deprioritization of online fsck to benefit file operations.1722The second property of the drain is key to the correct coordination of scrub,1723since scrub will always be able to decide if a conflict is possible.1724 1725For regular filesystem code, the drain works as follows:1726 17271. Call the appropriate subsystem function to add a deferred work item to a1728 transaction.1729 17302. The function calls ``xfs_defer_drain_bump`` to increase the counter.1731 17323. When the deferred item manager wants to finish the deferred work item, it1733 calls ``->finish_item`` to complete it.1734 17354. The ``->finish_item`` implementation logs some changes and calls1736 ``xfs_defer_drain_drop`` to decrease the sloppy counter and wake up any threads1737 waiting on the drain.1738 17395. The subtransaction commits, which unlocks the resource associated with the1740 intent item.1741 1742For scrub, the drain works as follows:1743 17441. Lock the resource(s) associated with the metadata being scrubbed.1745 For example, a scan of the refcount btree would lock the AGI and AGF header1746 buffers.1747 17482. If the counter is zero (``xfs_defer_drain_busy`` returns false), there are no1749 chains in progress and the operation may proceed.1750 17513. Otherwise, release the resources grabbed in step 1.1752 17534. Wait for the intent counter to reach zero (``xfs_defer_drain_intents``), then go1754 back to step 1 unless a signal has been caught.1755 1756To avoid polling in step 4, the drain provides a waitqueue for scrub threads to1757be woken up whenever the intent count drops to zero.1758 1759The proposed patchset is the1760`scrub intent drain series1761<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=scrub-drain-intents>`_.1762 1763.. _jump_labels:1764 1765Static Keys (aka Jump Label Patching)1766`````````````````````````````````````1767 1768Online fsck for XFS separates the regular filesystem from the checking and1769repair code as much as possible.1770However, there are a few parts of online fsck (such as the intent drains, and1771later, live update hooks) where it is useful for the online fsck code to know1772what's going on in the rest of the filesystem.1773Since it is not expected that online fsck will be constantly running in the1774background, it is very important to minimize the runtime overhead imposed by1775these hooks when online fsck is compiled into the kernel but not actively1776running on behalf of userspace.1777Taking locks in the hot path of a writer thread to access a data structure only1778to find that no further action is necessary is expensive -- on the author's1779computer, this have an overhead of 40-50ns per access.1780Fortunately, the kernel supports dynamic code patching, which enables XFS to1781replace a static branch to hook code with ``nop`` sleds when online fsck isn't1782running.1783This sled has an overhead of however long it takes the instruction decoder to1784skip past the sled, which seems to be on the order of less than 1ns and1785does not access memory outside of instruction fetching.1786 1787When online fsck enables the static key, the sled is replaced with an1788unconditional branch to call the hook code.1789The switchover is quite expensive (~22000ns) but is paid entirely by the1790program that invoked online fsck, and can be amortized if multiple threads1791enter online fsck at the same time, or if multiple filesystems are being1792checked at the same time.1793Changing the branch direction requires taking the CPU hotplug lock, and since1794CPU initialization requires memory allocation, online fsck must be careful not1795to change a static key while holding any locks or resources that could be1796accessed in the memory reclaim paths.1797To minimize contention on the CPU hotplug lock, care should be taken not to1798enable or disable static keys unnecessarily.1799 1800Because static keys are intended to minimize hook overhead for regular1801filesystem operations when xfs_scrub is not running, the intended usage1802patterns are as follows:1803 1804- The hooked part of XFS should declare a static-scoped static key that1805 defaults to false.1806 The ``DEFINE_STATIC_KEY_FALSE`` macro takes care of this.1807 The static key itself should be declared as a ``static`` variable.1808 1809- When deciding to invoke code that's only used by scrub, the regular1810 filesystem should call the ``static_branch_unlikely`` predicate to avoid the1811 scrub-only hook code if the static key is not enabled.1812 1813- The regular filesystem should export helper functions that call1814 ``static_branch_inc`` to enable and ``static_branch_dec`` to disable the1815 static key.1816 Wrapper functions make it easy to compile out the relevant code if the kernel1817 distributor turns off online fsck at build time.1818 1819- Scrub functions wanting to turn on scrub-only XFS functionality should call1820 the ``xchk_fsgates_enable`` from the setup function to enable a specific1821 hook.1822 This must be done before obtaining any resources that are used by memory1823 reclaim.1824 Callers had better be sure they really need the functionality gated by the1825 static key; the ``TRY_HARDER`` flag is useful here.1826 1827Online scrub has resource acquisition helpers (e.g. ``xchk_perag_lock``) to1828handle locking AGI and AGF buffers for all scrubber functions.1829If it detects a conflict between scrub and the running transactions, it will1830try to wait for intents to complete.1831If the caller of the helper has not enabled the static key, the helper will1832return -EDEADLOCK, which should result in the scrub being restarted with the1833``TRY_HARDER`` flag set.1834The scrub setup function should detect that flag, enable the static key, and1835try the scrub again.1836Scrub teardown disables all static keys obtained by ``xchk_fsgates_enable``.1837 1838For more information, please see the kernel documentation of1839Documentation/staging/static-keys.rst.1840 1841.. _xfile:1842 1843Pageable Kernel Memory1844----------------------1845 1846Some online checking functions work by scanning the filesystem to build a1847shadow copy of an ondisk metadata structure in memory and comparing the two1848copies.1849For online repair to rebuild a metadata structure, it must compute the record1850set that will be stored in the new structure before it can persist that new1851structure to disk.1852Ideally, repairs complete with a single atomic commit that introduces1853a new data structure.1854To meet these goals, the kernel needs to collect a large amount of information1855in a place that doesn't require the correct operation of the filesystem.1856 1857Kernel memory isn't suitable because:1858 1859* Allocating a contiguous region of memory to create a C array is very1860 difficult, especially on 32-bit systems.1861 1862* Linked lists of records introduce double pointer overhead which is very high1863 and eliminate the possibility of indexed lookups.1864 1865* Kernel memory is pinned, which can drive the system into OOM conditions.1866 1867* The system might not have sufficient memory to stage all the information.1868 1869At any given time, online fsck does not need to keep the entire record set in1870memory, which means that individual records can be paged out if necessary.1871Continued development of online fsck demonstrated that the ability to perform1872indexed data storage would also be very useful.1873Fortunately, the Linux kernel already has a facility for byte-addressable and1874pageable storage: tmpfs.1875In-kernel graphics drivers (most notably i915) take advantage of tmpfs files1876to store intermediate data that doesn't need to be in memory at all times, so1877that usage precedent is already established.1878Hence, the ``xfile`` was born!1879 1880+--------------------------------------------------------------------------+1881| **Historical Sidebar**: |1882+--------------------------------------------------------------------------+1883| The first edition of online repair inserted records into a new btree as |1884| it found them, which failed because filesystem could shut down with a |1885| built data structure, which would be live after recovery finished. |1886| |1887| The second edition solved the half-rebuilt structure problem by storing |1888| everything in memory, but frequently ran the system out of memory. |1889| |1890| The third edition solved the OOM problem by using linked lists, but the |1891| memory overhead of the list pointers was extreme. |1892+--------------------------------------------------------------------------+1893 1894xfile Access Models1895```````````````````1896 1897A survey of the intended uses of xfiles suggested these use cases:1898 18991. Arrays of fixed-sized records (space management btrees, directory and1900 extended attribute entries)1901 19022. Sparse arrays of fixed-sized records (quotas and link counts)1903 19043. Large binary objects (BLOBs) of variable sizes (directory and extended1905 attribute names and values)1906 19074. Staging btrees in memory (reverse mapping btrees)1908 19095. Arbitrary contents (realtime space management)1910 1911To support the first four use cases, high level data structures wrap the xfile1912to share functionality between online fsck functions.1913The rest of this section discusses the interfaces that the xfile presents to1914four of those five higher level data structures.1915The fifth use case is discussed in the :ref:`realtime summary <rtsummary>` case1916study.1917 1918XFS is very record-based, which suggests that the ability to load and store1919complete records is important.1920To support these cases, a pair of ``xfile_load`` and ``xfile_store``1921functions are provided to read and persist objects into an xfile that treat any1922error as an out of memory error. For online repair, squashing error conditions1923in this manner is an acceptable behavior because the only reaction is to abort1924the operation back to userspace.1925 1926However, no discussion of file access idioms is complete without answering the1927question, "But what about mmap?"1928It is convenient to access storage directly with pointers, just like userspace1929code does with regular memory.1930Online fsck must not drive the system into OOM conditions, which means that1931xfiles must be responsive to memory reclamation.1932tmpfs can only push a pagecache folio to the swap cache if the folio is neither1933pinned nor locked, which means the xfile must not pin too many folios.1934 1935Short term direct access to xfile contents is done by locking the pagecache1936folio and mapping it into kernel address space. Object load and store uses this1937mechanism. Folio locks are not supposed to be held for long periods of time, so1938long term direct access to xfile contents is done by bumping the folio refcount,1939mapping it into kernel address space, and dropping the folio lock.1940These long term users *must* be responsive to memory reclaim by hooking into1941the shrinker infrastructure to know when to release folios.1942 1943The ``xfile_get_folio`` and ``xfile_put_folio`` functions are provided to1944retrieve the (locked) folio that backs part of an xfile and to release it.1945The only code to use these folio lease functions are the xfarray1946:ref:`sorting<xfarray_sort>` algorithms and the :ref:`in-memory1947btrees<xfbtree>`.1948 1949xfile Access Coordination1950`````````````````````````1951 1952For security reasons, xfiles must be owned privately by the kernel.1953They are marked ``S_PRIVATE`` to prevent interference from the security system,1954must never be mapped into process file descriptor tables, and their pages must1955never be mapped into userspace processes.1956 1957To avoid locking recursion issues with the VFS, all accesses to the shmfs file1958are performed by manipulating the page cache directly.1959xfile writers call the ``->write_begin`` and ``->write_end`` functions of the1960xfile's address space to grab writable pages, copy the caller's buffer into the1961page, and release the pages.1962xfile readers call ``shmem_read_mapping_page_gfp`` to grab pages directly1963before copying the contents into the caller's buffer.1964In other words, xfiles ignore the VFS read and write code paths to avoid1965having to create a dummy ``struct kiocb`` and to avoid taking inode and1966freeze locks.1967tmpfs cannot be frozen, and xfiles must not be exposed to userspace.1968 1969If an xfile is shared between threads to stage repairs, the caller must provide1970its own locks to coordinate access.1971For example, if a scrub function stores scan results in an xfile and needs1972other threads to provide updates to the scanned data, the scrub function must1973provide a lock for all threads to share.1974 1975.. _xfarray:1976 1977Arrays of Fixed-Sized Records1978`````````````````````````````1979 1980In XFS, each type of indexed space metadata (free space, inodes, reference1981counts, file fork space, and reverse mappings) consists of a set of fixed-size1982records indexed with a classic B+ tree.1983Directories have a set of fixed-size dirent records that point to the names,1984and extended attributes have a set of fixed-size attribute keys that point to1985names and values.1986Quota counters and file link counters index records with numbers.1987During a repair, scrub needs to stage new records during the gathering step and1988retrieve them during the btree building step.1989 1990Although this requirement can be satisfied by calling the read and write1991methods of the xfile directly, it is simpler for callers for there to be a1992higher level abstraction to take care of computing array offsets, to provide1993iterator functions, and to deal with sparse records and sorting.1994The ``xfarray`` abstraction presents a linear array for fixed-size records atop1995the byte-accessible xfile.1996 1997.. _xfarray_access_patterns:1998 1999Array Access Patterns2000^^^^^^^^^^^^^^^^^^^^^2001 2002Array access patterns in online fsck tend to fall into three categories.2003Iteration of records is assumed to be necessary for all cases and will be2004covered in the next section.2005 2006The first type of caller handles records that are indexed by position.2007Gaps may exist between records, and a record may be updated multiple times2008during the collection step.2009In other words, these callers want a sparse linearly addressed table file.2010The typical use case are quota records or file link count records.2011Access to array elements is performed programmatically via ``xfarray_load`` and2012``xfarray_store`` functions, which wrap the similarly-named xfile functions to2013provide loading and storing of array elements at arbitrary array indices.2014Gaps are defined to be null records, and null records are defined to be a2015sequence of all zero bytes.2016Null records are detected by calling ``xfarray_element_is_null``.2017They are created either by calling ``xfarray_unset`` to null out an existing2018record or by never storing anything to an array index.2019 2020The second type of caller handles records that are not indexed by position2021and do not require multiple updates to a record.2022The typical use case here is rebuilding space btrees and key/value btrees.2023These callers can add records to the array without caring about array indices2024via the ``xfarray_append`` function, which stores a record at the end of the2025array.2026For callers that require records to be presentable in a specific order (e.g.2027rebuilding btree data), the ``xfarray_sort`` function can arrange the sorted2028records; this function will be covered later.2029 2030The third type of caller is a bag, which is useful for counting records.2031The typical use case here is constructing space extent reference counts from2032reverse mapping information.2033Records can be put in the bag in any order, they can be removed from the bag2034at any time, and uniqueness of records is left to callers.2035The ``xfarray_store_anywhere`` function is used to insert a record in any2036null record slot in the bag; and the ``xfarray_unset`` function removes a2037record from the bag.2038 2039The proposed patchset is the2040`big in-memory array2041<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=big-array>`_.2042 2043Iterating Array Elements2044^^^^^^^^^^^^^^^^^^^^^^^^2045 2046Most users of the xfarray require the ability to iterate the records stored in2047the array.2048Callers can probe every possible array index with the following:2049 2050.. code-block:: c2051 2052 xfarray_idx_t i;2053 foreach_xfarray_idx(array, i) {2054 xfarray_load(array, i, &rec);2055 2056 /* do something with rec */2057 }2058 2059All users of this idiom must be prepared to handle null records or must already2060know that there aren't any.2061 2062For xfarray users that want to iterate a sparse array, the ``xfarray_iter``2063function ignores indices in the xfarray that have never been written to by2064calling ``xfile_seek_data`` (which internally uses ``SEEK_DATA``) to skip areas2065of the array that are not populated with memory pages.2066Once it finds a page, it will skip the zeroed areas of the page.2067 2068.. code-block:: c2069 2070 xfarray_idx_t i = XFARRAY_CURSOR_INIT;2071 while ((ret = xfarray_iter(array, &i, &rec)) == 1) {2072 /* do something with rec */2073 }2074 2075.. _xfarray_sort:2076 2077Sorting Array Elements2078^^^^^^^^^^^^^^^^^^^^^^2079 2080During the fourth demonstration of online repair, a community reviewer remarked2081that for performance reasons, online repair ought to load batches of records2082into btree record blocks instead of inserting records into a new btree one at a2083time.2084The btree insertion code in XFS is responsible for maintaining correct ordering2085of the records, so naturally the xfarray must also support sorting the record2086set prior to bulk loading.2087 2088Case Study: Sorting xfarrays2089~~~~~~~~~~~~~~~~~~~~~~~~~~~~2090 2091The sorting algorithm used in the xfarray is actually a combination of adaptive2092quicksort and a heapsort subalgorithm in the spirit of2093`Sedgewick <https://algs4.cs.princeton.edu/23quicksort/>`_ and2094`pdqsort <https://github.com/orlp/pdqsort>`_, with customizations for the Linux2095kernel.2096To sort records in a reasonably short amount of time, ``xfarray`` takes2097advantage of the binary subpartitioning offered by quicksort, but it also uses2098heapsort to hedge against performance collapse if the chosen quicksort pivots2099are poor.2100Both algorithms are (in general) O(n * lg(n)), but there is a wide performance2101gulf between the two implementations.2102 2103The Linux kernel already contains a reasonably fast implementation of heapsort.2104It only operates on regular C arrays, which limits the scope of its usefulness.2105There are two key places where the xfarray uses it:2106 2107* Sorting any record subset backed by a single xfile page.2108 2109* Loading a small number of xfarray records from potentially disparate parts2110 of the xfarray into a memory buffer, and sorting the buffer.2111 2112In other words, ``xfarray`` uses heapsort to constrain the nested recursion of2113quicksort, thereby mitigating quicksort's worst runtime behavior.2114 2115Choosing a quicksort pivot is a tricky business.2116A good pivot splits the set to sort in half, leading to the divide and conquer2117behavior that is crucial to O(n * lg(n)) performance.2118A poor pivot barely splits the subset at all, leading to O(n\ :sup:`2`)2119runtime.2120The xfarray sort routine tries to avoid picking a bad pivot by sampling nine2121records into a memory buffer and using the kernel heapsort to identify the2122median of the nine.2123 2124Most modern quicksort implementations employ Tukey's "ninther" to select a2125pivot from a classic C array.2126Typical ninther implementations pick three unique triads of records, sort each2127of the triads, and then sort the middle value of each triad to determine the2128ninther value.2129As stated previously, however, xfile accesses are not entirely cheap.2130It turned out to be much more performant to read the nine elements into a2131memory buffer, run the kernel's in-memory heapsort on the buffer, and choose2132the 4th element of that buffer as the pivot.2133Tukey's ninthers are described in J. W. Tukey, `The ninther, a technique for2134low-effort robust (resistant) location in large samples`, in *Contributions to2135Survey Sampling and Applied Statistics*, edited by H. David, (Academic Press,21361978), pp. 251–257.2137 2138The partitioning of quicksort is fairly textbook -- rearrange the record2139subset around the pivot, then set up the current and next stack frames to2140sort with the larger and the smaller halves of the pivot, respectively.2141This keeps the stack space requirements to log2(record count).2142 2143As a final performance optimization, the hi and lo scanning phase of quicksort2144keeps examined xfile pages mapped in the kernel for as long as possible to2145reduce map/unmap cycles.2146Surprisingly, this reduces overall sort runtime by nearly half again after2147accounting for the application of heapsort directly onto xfile pages.2148 2149.. _xfblob:2150 2151Blob Storage2152````````````2153 2154Extended attributes and directories add an additional requirement for staging2155records: arbitrary byte sequences of finite length.2156Each directory entry record needs to store entry name,2157and each extended attribute needs to store both the attribute name and value.2158The names, keys, and values can consume a large amount of memory, so the2159``xfblob`` abstraction was created to simplify management of these blobs2160atop an xfile.2161 2162Blob arrays provide ``xfblob_load`` and ``xfblob_store`` functions to retrieve2163and persist objects.2164The store function returns a magic cookie for every object that it persists.2165Later, callers provide this cookie to the ``xblob_load`` to recall the object.2166The ``xfblob_free`` function frees a specific blob, and the ``xfblob_truncate``2167function frees them all because compaction is not needed.2168 2169The details of repairing directories and extended attributes will be discussed2170in a subsequent section about atomic file content exchanges.2171However, it should be noted that these repair functions only use blob storage2172to cache a small number of entries before adding them to a temporary ondisk2173file, which is why compaction is not required.2174 2175The proposed patchset is at the start of the2176`extended attribute repair2177<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-xattrs>`_ series.2178 2179.. _xfbtree:2180 2181In-Memory B+Trees2182`````````````````2183 2184The chapter about :ref:`secondary metadata<secondary_metadata>` mentioned that2185checking and repairing of secondary metadata commonly requires coordination2186between a live metadata scan of the filesystem and writer threads that are2187updating that metadata.2188Keeping the scan data up to date requires requires the ability to propagate2189metadata updates from the filesystem into the data being collected by the scan.2190This *can* be done by appending concurrent updates into a separate log file and2191applying them before writing the new metadata to disk, but this leads to2192unbounded memory consumption if the rest of the system is very busy.2193Another option is to skip the side-log and commit live updates from the2194filesystem directly into the scan data, which trades more overhead for a lower2195maximum memory requirement.2196In both cases, the data structure holding the scan results must support indexed2197access to perform well.2198 2199Given that indexed lookups of scan data is required for both strategies, online2200fsck employs the second strategy of committing live updates directly into2201scan data.2202Because xfarrays are not indexed and do not enforce record ordering, they2203are not suitable for this task.2204Conveniently, however, XFS has a library to create and maintain ordered reverse2205mapping records: the existing rmap btree code!2206If only there was a means to create one in memory.2207 2208Recall that the :ref:`xfile <xfile>` abstraction represents memory pages as a2209regular file, which means that the kernel can create byte or block addressable2210virtual address spaces at will.2211The XFS buffer cache specializes in abstracting IO to block-oriented address2212spaces, which means that adaptation of the buffer cache to interface with2213xfiles enables reuse of the entire btree library.2214Btrees built atop an xfile are collectively known as ``xfbtrees``.2215The next few sections describe how they actually work.2216 2217The proposed patchset is the2218`in-memory btree2219<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=in-memory-btrees>`_2220series.2221 2222Using xfiles as a Buffer Cache Target2223^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^2224 2225Two modifications are necessary to support xfiles as a buffer cache target.2226The first is to make it possible for the ``struct xfs_buftarg`` structure to2227host the ``struct xfs_buf`` rhashtable, because normally those are held by a2228per-AG structure.2229The second change is to modify the buffer ``ioapply`` function to "read" cached2230pages from the xfile and "write" cached pages back to the xfile.2231Multiple access to individual buffers is controlled by the ``xfs_buf`` lock,2232since the xfile does not provide any locking on its own.2233With this adaptation in place, users of the xfile-backed buffer cache use2234exactly the same APIs as users of the disk-backed buffer cache.2235The separation between xfile and buffer cache implies higher memory usage since2236they do not share pages, but this property could some day enable transactional2237updates to an in-memory btree.2238Today, however, it simply eliminates the need for new code.2239 2240Space Management with an xfbtree2241^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^2242 2243Space management for an xfile is very simple -- each btree block is one memory2244page in size.2245These blocks use the same header format as an on-disk btree, but the in-memory2246block verifiers ignore the checksums, assuming that xfile memory is no more2247corruption-prone than regular DRAM.2248Reusing existing code here is more important than absolute memory efficiency.2249 2250The very first block of an xfile backing an xfbtree contains a header block.2251The header describes the owner, height, and the block number of the root2252xfbtree block.2253 2254To allocate a btree block, use ``xfile_seek_data`` to find a gap in the file.2255If there are no gaps, create one by extending the length of the xfile.2256Preallocate space for the block with ``xfile_prealloc``, and hand back the2257location.2258To free an xfbtree block, use ``xfile_discard`` (which internally uses2259``FALLOC_FL_PUNCH_HOLE``) to remove the memory page from the xfile.2260 2261Populating an xfbtree2262^^^^^^^^^^^^^^^^^^^^^2263 2264An online fsck function that wants to create an xfbtree should proceed as2265follows:2266 22671. Call ``xfile_create`` to create an xfile.2268 22692. Call ``xfs_alloc_memory_buftarg`` to create a buffer cache target structure2270 pointing to the xfile.2271 22723. Pass the buffer cache target, buffer ops, and other information to2273 ``xfbtree_init`` to initialize the passed in ``struct xfbtree`` and write an2274 initial root block to the xfile.2275 Each btree type should define a wrapper that passes necessary arguments to2276 the creation function.2277 For example, rmap btrees define ``xfs_rmapbt_mem_create`` to take care of2278 all the necessary details for callers.2279 22804. Pass the xfbtree object to the btree cursor creation function for the2281 btree type.2282 Following the example above, ``xfs_rmapbt_mem_cursor`` takes care of this2283 for callers.2284 22855. Pass the btree cursor to the regular btree functions to make queries against2286 and to update the in-memory btree.2287 For example, a btree cursor for an rmap xfbtree can be passed to the2288 ``xfs_rmap_*`` functions just like any other btree cursor.2289 See the :ref:`next section<xfbtree_commit>` for information on dealing with2290 xfbtree updates that are logged to a transaction.2291 22926. When finished, delete the btree cursor, destroy the xfbtree object, free the2293 buffer target, and the destroy the xfile to release all resources.2294 2295.. _xfbtree_commit:2296 2297Committing Logged xfbtree Buffers2298^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^2299 2300Although it is a clever hack to reuse the rmap btree code to handle the staging2301structure, the ephemeral nature of the in-memory btree block storage presents2302some challenges of its own.2303The XFS transaction manager must not commit buffer log items for buffers backed2304by an xfile because the log format does not understand updates for devices2305other than the data device.2306An ephemeral xfbtree probably will not exist by the time the AIL checkpoints2307log transactions back into the filesystem, and certainly won't exist during2308log recovery.2309For these reasons, any code updating an xfbtree in transaction context must2310remove the buffer log items from the transaction and write the updates into the2311backing xfile before committing or cancelling the transaction.2312 2313The ``xfbtree_trans_commit`` and ``xfbtree_trans_cancel`` functions implement2314this functionality as follows:2315 23161. Find each buffer log item whose buffer targets the xfile.2317 23182. Record the dirty/ordered status of the log item.2319 23203. Detach the log item from the buffer.2321 23224. Queue the buffer to a special delwri list.2323 23245. Clear the transaction dirty flag if the only dirty log items were the ones2325 that were detached in step 3.2326 23276. Submit the delwri list to commit the changes to the xfile, if the updates2328 are being committed.2329 2330After removing xfile logged buffers from the transaction in this manner, the2331transaction can be committed or cancelled.2332 2333Bulk Loading of Ondisk B+Trees2334------------------------------2335 2336As mentioned previously, early iterations of online repair built new btree2337structures by creating a new btree and adding observations individually.2338Loading a btree one record at a time had a slight advantage of not requiring2339the incore records to be sorted prior to commit, but was very slow and leaked2340blocks if the system went down during a repair.2341Loading records one at a time also meant that repair could not control the2342loading factor of the blocks in the new btree.2343 2344Fortunately, the venerable ``xfs_repair`` tool had a more efficient means for2345rebuilding a btree index from a collection of records -- bulk btree loading.2346This was implemented rather inefficiently code-wise, since ``xfs_repair``2347had separate copy-pasted implementations for each btree type.2348 2349To prepare for online fsck, each of the four bulk loaders were studied, notes2350were taken, and the four were refactored into a single generic btree bulk2351loading mechanism.2352Those notes in turn have been refreshed and are presented below.2353 2354Geometry Computation2355````````````````````2356 2357The zeroth step of bulk loading is to assemble the entire record set that will2358be stored in the new btree, and sort the records.2359Next, call ``xfs_btree_bload_compute_geometry`` to compute the shape of the2360btree from the record set, the type of btree, and any load factor preferences.2361This information is required for resource reservation.2362 2363First, the geometry computation computes the minimum and maximum records that2364will fit in a leaf block from the size of a btree block and the size of the2365block header.2366Roughly speaking, the maximum number of records is::2367 2368 maxrecs = (block_size - header_size) / record_size2369 2370The XFS design specifies that btree blocks should be merged when possible,2371which means the minimum number of records is half of maxrecs::2372 2373 minrecs = maxrecs / 22374 2375The next variable to determine is the desired loading factor.2376This must be at least minrecs and no more than maxrecs.2377Choosing minrecs is undesirable because it wastes half the block.2378Choosing maxrecs is also undesirable because adding a single record to each2379newly rebuilt leaf block will cause a tree split, which causes a noticeable2380drop in performance immediately afterwards.2381The default loading factor was chosen to be 75% of maxrecs, which provides a2382reasonably compact structure without any immediate split penalties::2383 2384 default_load_factor = (maxrecs + minrecs) / 22385 2386If space is tight, the loading factor will be set to maxrecs to try to avoid2387running out of space::2388 2389 leaf_load_factor = enough space ? default_load_factor : maxrecs2390 2391Load factor is computed for btree node blocks using the combined size of the2392btree key and pointer as the record size::2393 2394 maxrecs = (block_size - header_size) / (key_size + ptr_size)2395 minrecs = maxrecs / 22396 node_load_factor = enough space ? default_load_factor : maxrecs2397 2398Once that's done, the number of leaf blocks required to store the record set2399can be computed as::2400 2401 leaf_blocks = ceil(record_count / leaf_load_factor)2402 2403The number of node blocks needed to point to the next level down in the tree2404is computed as::2405 2406 n_blocks = (n == 0 ? leaf_blocks : node_blocks[n])2407 node_blocks[n + 1] = ceil(n_blocks / node_load_factor)2408 2409The entire computation is performed recursively until the current level only2410needs one block.2411The resulting geometry is as follows:2412 2413- For AG-rooted btrees, this level is the root level, so the height of the new2414 tree is ``level + 1`` and the space needed is the summation of the number of2415 blocks on each level.2416 2417- For inode-rooted btrees where the records in the top level do not fit in the2418 inode fork area, the height is ``level + 2``, the space needed is the2419 summation of the number of blocks on each level, and the inode fork points to2420 the root block.2421 2422- For inode-rooted btrees where the records in the top level can be stored in2423 the inode fork area, then the root block can be stored in the inode, the2424 height is ``level + 1``, and the space needed is one less than the summation2425 of the number of blocks on each level.2426 This only becomes relevant when non-bmap btrees gain the ability to root in2427 an inode, which is a future patchset and only included here for completeness.2428 2429.. _newbt:2430 2431Reserving New B+Tree Blocks2432```````````````````````````2433 2434Once repair knows the number of blocks needed for the new btree, it allocates2435those blocks using the free space information.2436Each reserved extent is tracked separately by the btree builder state data.2437To improve crash resilience, the reservation code also logs an Extent Freeing2438Intent (EFI) item in the same transaction as each space allocation and attaches2439its in-memory ``struct xfs_extent_free_item`` object to the space reservation.2440If the system goes down, log recovery will use the unfinished EFIs to free the2441unused space, the free space, leaving the filesystem unchanged.2442 2443Each time the btree builder claims a block for the btree from a reserved2444extent, it updates the in-memory reservation to reflect the claimed space.2445Block reservation tries to allocate as much contiguous space as possible to2446reduce the number of EFIs in play.2447 2448While repair is writing these new btree blocks, the EFIs created for the space2449reservations pin the tail of the ondisk log.2450It's possible that other parts of the system will remain busy and push the head2451of the log towards the pinned tail.2452To avoid livelocking the filesystem, the EFIs must not pin the tail of the log2453for too long.2454To alleviate this problem, the dynamic relogging capability of the deferred ops2455mechanism is reused here to commit a transaction at the log head containing an2456EFD for the old EFI and new EFI at the head.2457This enables the log to release the old EFI to keep the log moving forwards.2458 2459EFIs have a role to play during the commit and reaping phases; please see the2460next section and the section about :ref:`reaping<reaping>` for more details.2461 2462Proposed patchsets are the2463`bitmap rework2464<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-bitmap-rework>`_2465and the2466`preparation for bulk loading btrees2467<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-prep-for-bulk-loading>`_.2468 2469 2470Writing the New Tree2471````````````````````2472 2473This part is pretty simple -- the btree builder (``xfs_btree_bulkload``) claims2474a block from the reserved list, writes the new btree block header, fills the2475rest of the block with records, and adds the new leaf block to a list of2476written blocks::2477 2478 ┌────┐2479 │leaf│2480 │RRR │2481 └────┘2482 2483Sibling pointers are set every time a new block is added to the level::2484 2485 ┌────┐ ┌────┐ ┌────┐ ┌────┐2486 │leaf│→│leaf│→│leaf│→│leaf│2487 │RRR │←│RRR │←│RRR │←│RRR │2488 └────┘ └────┘ └────┘ └────┘2489 2490When it finishes writing the record leaf blocks, it moves on to the node2491blocks2492To fill a node block, it walks each block in the next level down in the tree2493to compute the relevant keys and write them into the parent node::2494 2495 ┌────┐ ┌────┐2496 │node│──────→│node│2497 │PP │←──────│PP │2498 └────┘ └────┘2499 ↙ ↘ ↙ ↘2500 ┌────┐ ┌────┐ ┌────┐ ┌────┐2501 │leaf│→│leaf│→│leaf│→│leaf│2502 │RRR │←│RRR │←│RRR │←│RRR │2503 └────┘ └────┘ └────┘ └────┘2504 2505When it reaches the root level, it is ready to commit the new btree!::2506 2507 ┌─────────┐2508 │ root │2509 │ PP │2510 └─────────┘2511 ↙ ↘2512 ┌────┐ ┌────┐2513 │node│──────→│node│2514 │PP │←──────│PP │2515 └────┘ └────┘2516 ↙ ↘ ↙ ↘2517 ┌────┐ ┌────┐ ┌────┐ ┌────┐2518 │leaf│→│leaf│→│leaf│→│leaf│2519 │RRR │←│RRR │←│RRR │←│RRR │2520 └────┘ └────┘ └────┘ └────┘2521 2522The first step to commit the new btree is to persist the btree blocks to disk2523synchronously.2524This is a little complicated because a new btree block could have been freed2525in the recent past, so the builder must use ``xfs_buf_delwri_queue_here`` to2526remove the (stale) buffer from the AIL list before it can write the new blocks2527to disk.2528Blocks are queued for IO using a delwri list and written in one large batch2529with ``xfs_buf_delwri_submit``.2530 2531Once the new blocks have been persisted to disk, control returns to the2532individual repair function that called the bulk loader.2533The repair function must log the location of the new root in a transaction,2534clean up the space reservations that were made for the new btree, and reap the2535old metadata blocks:2536 25371. Commit the location of the new btree root.2538 25392. For each incore reservation:2540 2541 a. Log Extent Freeing Done (EFD) items for all the space that was consumed2542 by the btree builder. The new EFDs must point to the EFIs attached to2543 the reservation to prevent log recovery from freeing the new blocks.2544 2545 b. For unclaimed portions of incore reservations, create a regular deferred2546 extent free work item to be free the unused space later in the2547 transaction chain.2548 2549 c. The EFDs and EFIs logged in steps 2a and 2b must not overrun the2550 reservation of the committing transaction.2551 If the btree loading code suspects this might be about to happen, it must2552 call ``xrep_defer_finish`` to clear out the deferred work and obtain a2553 fresh transaction.2554 25553. Clear out the deferred work a second time to finish the commit and clean2556 the repair transaction.2557 2558The transaction rolling in steps 2c and 3 represent a weakness in the repair2559algorithm, because a log flush and a crash before the end of the reap step can2560result in space leaking.2561Online repair functions minimize the chances of this occurring by using very2562large transactions, which each can accommodate many thousands of block freeing2563instructions.2564Repair moves on to reaping the old blocks, which will be presented in a2565subsequent :ref:`section<reaping>` after a few case studies of bulk loading.2566 2567Case Study: Rebuilding the Inode Index2568^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^2569 2570The high level process to rebuild the inode index btree is:2571 25721. Walk the reverse mapping records to generate ``struct xfs_inobt_rec``2573 records from the inode chunk information and a bitmap of the old inode btree2574 blocks.2575 25762. Append the records to an xfarray in inode order.2577 25783. Use the ``xfs_btree_bload_compute_geometry`` function to compute the number2579 of blocks needed for the inode btree.2580 If the free space inode btree is enabled, call it again to estimate the2581 geometry of the finobt.2582 25834. Allocate the number of blocks computed in the previous step.2584 25855. Use ``xfs_btree_bload`` to write the xfarray records to btree blocks and2586 generate the internal node blocks.2587 If the free space inode btree is enabled, call it again to load the finobt.2588 25896. Commit the location of the new btree root block(s) to the AGI.2590 25917. Reap the old btree blocks using the bitmap created in step 1.2592 2593Details are as follows.2594 2595The inode btree maps inumbers to the ondisk location of the associated2596inode records, which means that the inode btrees can be rebuilt from the2597reverse mapping information.2598Reverse mapping records with an owner of ``XFS_RMAP_OWN_INOBT`` marks the2599location of the old inode btree blocks.2600Each reverse mapping record with an owner of ``XFS_RMAP_OWN_INODES`` marks the2601location of at least one inode cluster buffer.2602A cluster is the smallest number of ondisk inodes that can be allocated or2603freed in a single transaction; it is never smaller than 1 fs block or 4 inodes.2604 2605For the space represented by each inode cluster, ensure that there are no2606records in the free space btrees nor any records in the reference count btree.2607If there are, the space metadata inconsistencies are reason enough to abort the2608operation.2609Otherwise, read each cluster buffer to check that its contents appear to be2610ondisk inodes and to decide if the file is allocated2611(``xfs_dinode.i_mode != 0``) or free (``xfs_dinode.i_mode == 0``).2612Accumulate the results of successive inode cluster buffer reads until there is2613enough information to fill a single inode chunk record, which is 64 consecutive2614numbers in the inumber keyspace.2615If the chunk is sparse, the chunk record may include holes.2616 2617Once the repair function accumulates one chunk's worth of data, it calls2618``xfarray_append`` to add the inode btree record to the xfarray.2619This xfarray is walked twice during the btree creation step -- once to populate2620the inode btree with all inode chunk records, and a second time to populate the2621free inode btree with records for chunks that have free non-sparse inodes.2622The number of records for the inode btree is the number of xfarray records,2623but the record count for the free inode btree has to be computed as inode chunk2624records are stored in the xfarray.2625 2626The proposed patchset is the2627`AG btree repair2628<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-ag-btrees>`_2629series.2630 2631Case Study: Rebuilding the Space Reference Counts2632^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^2633 2634Reverse mapping records are used to rebuild the reference count information.2635Reference counts are required for correct operation of copy on write for shared2636file data.2637Imagine the reverse mapping entries as rectangles representing extents of2638physical blocks, and that the rectangles can be laid down to allow them to2639overlap each other.2640From the diagram below, it is apparent that a reference count record must start2641or end wherever the height of the stack changes.2642In other words, the record emission stimulus is level-triggered::2643 2644 █ ███2645 ██ █████ ████ ███ ██████2646 ██ ████ ███████████ ████ █████████2647 ████████████████████████████████ ███████████2648 ^ ^ ^^ ^^ ^ ^^ ^^^ ^^^^ ^ ^^ ^ ^ ^2649 2 1 23 21 3 43 234 2123 1 01 2 3 02650 2651The ondisk reference count btree does not store the refcount == 0 cases because2652the free space btree already records which blocks are free.2653Extents being used to stage copy-on-write operations should be the only records2654with refcount == 1.2655Single-owner file blocks aren't recorded in either the free space or the2656reference count btrees.2657 2658The high level process to rebuild the reference count btree is:2659 26601. Walk the reverse mapping records to generate ``struct xfs_refcount_irec``2661 records for any space having more than one reverse mapping and add them to2662 the xfarray.2663 Any records owned by ``XFS_RMAP_OWN_COW`` are also added to the xfarray2664 because these are extents allocated to stage a copy on write operation and2665 are tracked in the refcount btree.2666 2667 Use any records owned by ``XFS_RMAP_OWN_REFC`` to create a bitmap of old2668 refcount btree blocks.2669 26702. Sort the records in physical extent order, putting the CoW staging extents2671 at the end of the xfarray.2672 This matches the sorting order of records in the refcount btree.2673 26743. Use the ``xfs_btree_bload_compute_geometry`` function to compute the number2675 of blocks needed for the new tree.2676 26774. Allocate the number of blocks computed in the previous step.2678 26795. Use ``xfs_btree_bload`` to write the xfarray records to btree blocks and2680 generate the internal node blocks.2681 26826. Commit the location of new btree root block to the AGF.2683 26847. Reap the old btree blocks using the bitmap created in step 1.2685 2686Details are as follows; the same algorithm is used by ``xfs_repair`` to2687generate refcount information from reverse mapping records.2688 2689- Until the reverse mapping btree runs out of records:2690 2691 - Retrieve the next record from the btree and put it in a bag.2692 2693 - Collect all records with the same starting block from the btree and put2694 them in the bag.2695 2696 - While the bag isn't empty:2697 2698 - Among the mappings in the bag, compute the lowest block number where the2699 reference count changes.2700 This position will be either the starting block number of the next2701 unprocessed reverse mapping or the next block after the shortest mapping2702 in the bag.2703 2704 - Remove all mappings from the bag that end at this position.2705 2706 - Collect all reverse mappings that start at this position from the btree2707 and put them in the bag.2708 2709 - If the size of the bag changed and is greater than one, create a new2710 refcount record associating the block number range that we just walked to2711 the size of the bag.2712 2713The bag-like structure in this case is a type 2 xfarray as discussed in the2714:ref:`xfarray access patterns<xfarray_access_patterns>` section.2715Reverse mappings are added to the bag using ``xfarray_store_anywhere`` and2716removed via ``xfarray_unset``.2717Bag members are examined through ``xfarray_iter`` loops.2718 2719The proposed patchset is the2720`AG btree repair2721<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-ag-btrees>`_2722series.2723 2724Case Study: Rebuilding File Fork Mapping Indices2725^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^2726 2727The high level process to rebuild a data/attr fork mapping btree is:2728 27291. Walk the reverse mapping records to generate ``struct xfs_bmbt_rec``2730 records from the reverse mapping records for that inode and fork.2731 Append these records to an xfarray.2732 Compute the bitmap of the old bmap btree blocks from the ``BMBT_BLOCK``2733 records.2734 27352. Use the ``xfs_btree_bload_compute_geometry`` function to compute the number2736 of blocks needed for the new tree.2737 27383. Sort the records in file offset order.2739 27404. If the extent records would fit in the inode fork immediate area, commit the2741 records to that immediate area and skip to step 8.2742 27435. Allocate the number of blocks computed in the previous step.2744 27456. Use ``xfs_btree_bload`` to write the xfarray records to btree blocks and2746 generate the internal node blocks.2747 27487. Commit the new btree root block to the inode fork immediate area.2749 27508. Reap the old btree blocks using the bitmap created in step 1.2751 2752There are some complications here:2753First, it's possible to move the fork offset to adjust the sizes of the2754immediate areas if the data and attr forks are not both in BMBT format.2755Second, if there are sufficiently few fork mappings, it may be possible to use2756EXTENTS format instead of BMBT, which may require a conversion.2757Third, the incore extent map must be reloaded carefully to avoid disturbing2758any delayed allocation extents.2759 2760The proposed patchset is the2761`file mapping repair2762<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-file-mappings>`_2763series.2764 2765.. _reaping:2766 2767Reaping Old Metadata Blocks2768---------------------------2769 2770Whenever online fsck builds a new data structure to replace one that is2771suspect, there is a question of how to find and dispose of the blocks that2772belonged to the old structure.2773The laziest method of course is not to deal with them at all, but this slowly2774leads to service degradations as space leaks out of the filesystem.2775Hopefully, someone will schedule a rebuild of the free space information to2776plug all those leaks.2777Offline repair rebuilds all space metadata after recording the usage of2778the files and directories that it decides not to clear, hence it can build new2779structures in the discovered free space and avoid the question of reaping.2780 2781As part of a repair, online fsck relies heavily on the reverse mapping records2782to find space that is owned by the corresponding rmap owner yet truly free.2783Cross referencing rmap records with other rmap records is necessary because2784there may be other data structures that also think they own some of those2785blocks (e.g. crosslinked trees).2786Permitting the block allocator to hand them out again will not push the system2787towards consistency.2788 2789For space metadata, the process of finding extents to dispose of generally2790follows this format:2791 27921. Create a bitmap of space used by data structures that must be preserved.2793 The space reservations used to create the new metadata can be used here if2794 the same rmap owner code is used to denote all of the objects being rebuilt.2795 27962. Survey the reverse mapping data to create a bitmap of space owned by the2797 same ``XFS_RMAP_OWN_*`` number for the metadata that is being preserved.2798 27993. Use the bitmap disunion operator to subtract (1) from (2).2800 The remaining set bits represent candidate extents that could be freed.2801 The process moves on to step 4 below.2802 2803Repairs for file-based metadata such as extended attributes, directories,2804symbolic links, quota files and realtime bitmaps are performed by building a2805new structure attached to a temporary file and exchanging all mappings in the2806file forks.2807Afterward, the mappings in the old file fork are the candidate blocks for2808disposal.2809 2810The process for disposing of old extents is as follows:2811 28124. For each candidate extent, count the number of reverse mapping records for2813 the first block in that extent that do not have the same rmap owner for the2814 data structure being repaired.2815 2816 - If zero, the block has a single owner and can be freed.2817 2818 - If not, the block is part of a crosslinked structure and must not be2819 freed.2820 28215. Starting with the next block in the extent, figure out how many more blocks2822 have the same zero/nonzero other owner status as that first block.2823 28246. If the region is crosslinked, delete the reverse mapping entry for the2825 structure being repaired and move on to the next region.2826 28277. If the region is to be freed, mark any corresponding buffers in the buffer2828 cache as stale to prevent log writeback.2829 28308. Free the region and move on.2831 2832However, there is one complication to this procedure.2833Transactions are of finite size, so the reaping process must be careful to roll2834the transactions to avoid overruns.2835Overruns come from two sources:2836 2837a. EFIs logged on behalf of space that is no longer occupied2838 2839b. Log items for buffer invalidations2840 2841This is also a window in which a crash during the reaping process can leak2842blocks.2843As stated earlier, online repair functions use very large transactions to2844minimize the chances of this occurring.2845 2846The proposed patchset is the2847`preparation for bulk loading btrees2848<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-prep-for-bulk-loading>`_2849series.2850 2851Case Study: Reaping After a Regular Btree Repair2852````````````````````````````````````````````````2853 2854Old reference count and inode btrees are the easiest to reap because they have2855rmap records with special owner codes: ``XFS_RMAP_OWN_REFC`` for the refcount2856btree, and ``XFS_RMAP_OWN_INOBT`` for the inode and free inode btrees.2857Creating a list of extents to reap the old btree blocks is quite simple,2858conceptually:2859 28601. Lock the relevant AGI/AGF header buffers to prevent allocation and frees.2861 28622. For each reverse mapping record with an rmap owner corresponding to the2863 metadata structure being rebuilt, set the corresponding range in a bitmap.2864 28653. Walk the current data structures that have the same rmap owner.2866 For each block visited, clear that range in the above bitmap.2867 28684. Each set bit in the bitmap represents a block that could be a block from the2869 old data structures and hence is a candidate for reaping.2870 In other words, ``(rmap_records_owned_by & ~blocks_reachable_by_walk)``2871 are the blocks that might be freeable.2872 2873If it is possible to maintain the AGF lock throughout the repair (which is the2874common case), then step 2 can be performed at the same time as the reverse2875mapping record walk that creates the records for the new btree.2876 2877Case Study: Rebuilding the Free Space Indices2878`````````````````````````````````````````````2879 2880The high level process to rebuild the free space indices is:2881 28821. Walk the reverse mapping records to generate ``struct xfs_alloc_rec_incore``2883 records from the gaps in the reverse mapping btree.2884 28852. Append the records to an xfarray.2886 28873. Use the ``xfs_btree_bload_compute_geometry`` function to compute the number2888 of blocks needed for each new tree.2889 28904. Allocate the number of blocks computed in the previous step from the free2891 space information collected.2892 28935. Use ``xfs_btree_bload`` to write the xfarray records to btree blocks and2894 generate the internal node blocks for the free space by length index.2895 Call it again for the free space by block number index.2896 28976. Commit the locations of the new btree root blocks to the AGF.2898 28997. Reap the old btree blocks by looking for space that is not recorded by the2900 reverse mapping btree, the new free space btrees, or the AGFL.2901 2902Repairing the free space btrees has three key complications over a regular2903btree repair:2904 2905First, free space is not explicitly tracked in the reverse mapping records.2906Hence, the new free space records must be inferred from gaps in the physical2907space component of the keyspace of the reverse mapping btree.2908 2909Second, free space repairs cannot use the common btree reservation code because2910new blocks are reserved out of the free space btrees.2911This is impossible when repairing the free space btrees themselves.2912However, repair holds the AGF buffer lock for the duration of the free space2913index reconstruction, so it can use the collected free space information to2914supply the blocks for the new free space btrees.2915It is not necessary to back each reserved extent with an EFI because the new2916free space btrees are constructed in what the ondisk filesystem thinks is2917unowned space.2918However, if reserving blocks for the new btrees from the collected free space2919information changes the number of free space records, repair must re-estimate2920the new free space btree geometry with the new record count until the2921reservation is sufficient.2922As part of committing the new btrees, repair must ensure that reverse mappings2923are created for the reserved blocks and that unused reserved blocks are2924inserted into the free space btrees.2925Deferrred rmap and freeing operations are used to ensure that this transition2926is atomic, similar to the other btree repair functions.2927 2928Third, finding the blocks to reap after the repair is not overly2929straightforward.2930Blocks for the free space btrees and the reverse mapping btrees are supplied by2931the AGFL.2932Blocks put onto the AGFL have reverse mapping records with the owner2933``XFS_RMAP_OWN_AG``.2934This ownership is retained when blocks move from the AGFL into the free space2935btrees or the reverse mapping btrees.2936When repair walks reverse mapping records to synthesize free space records, it2937creates a bitmap (``ag_owner_bitmap``) of all the space claimed by2938``XFS_RMAP_OWN_AG`` records.2939The repair context maintains a second bitmap corresponding to the rmap btree2940blocks and the AGFL blocks (``rmap_agfl_bitmap``).2941When the walk is complete, the bitmap disunion operation ``(ag_owner_bitmap &2942~rmap_agfl_bitmap)`` computes the extents that are used by the old free space2943btrees.2944These blocks can then be reaped using the methods outlined above.2945 2946The proposed patchset is the2947`AG btree repair2948<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-ag-btrees>`_2949series.2950 2951.. _rmap_reap:2952 2953Case Study: Reaping After Repairing Reverse Mapping Btrees2954``````````````````````````````````````````````````````````2955 2956Old reverse mapping btrees are less difficult to reap after a repair.2957As mentioned in the previous section, blocks on the AGFL, the two free space2958btree blocks, and the reverse mapping btree blocks all have reverse mapping2959records with ``XFS_RMAP_OWN_AG`` as the owner.2960The full process of gathering reverse mapping records and building a new btree2961are described in the case study of2962:ref:`live rebuilds of rmap data <rmap_repair>`, but a crucial point from that2963discussion is that the new rmap btree will not contain any records for the old2964rmap btree, nor will the old btree blocks be tracked in the free space btrees.2965The list of candidate reaping blocks is computed by setting the bits2966corresponding to the gaps in the new rmap btree records, and then clearing the2967bits corresponding to extents in the free space btrees and the current AGFL2968blocks.2969The result ``(new_rmapbt_gaps & ~(agfl | bnobt_records))`` are reaped using the2970methods outlined above.2971 2972The rest of the process of rebuildng the reverse mapping btree is discussed2973in a separate :ref:`case study<rmap_repair>`.2974 2975The proposed patchset is the2976`AG btree repair2977<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-ag-btrees>`_2978series.2979 2980Case Study: Rebuilding the AGFL2981```````````````````````````````2982 2983The allocation group free block list (AGFL) is repaired as follows:2984 29851. Create a bitmap for all the space that the reverse mapping data claims is2986 owned by ``XFS_RMAP_OWN_AG``.2987 29882. Subtract the space used by the two free space btrees and the rmap btree.2989 29903. Subtract any space that the reverse mapping data claims is owned by any2991 other owner, to avoid re-adding crosslinked blocks to the AGFL.2992 29934. Once the AGFL is full, reap any blocks leftover.2994 29955. The next operation to fix the freelist will right-size the list.2996 2997See `fs/xfs/scrub/agheader_repair.c <https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/fs/xfs/scrub/agheader_repair.c>`_ for more details.2998 2999Inode Record Repairs3000--------------------3001 3002Inode records must be handled carefully, because they have both ondisk records3003("dinodes") and an in-memory ("cached") representation.3004There is a very high potential for cache coherency issues if online fsck is not3005careful to access the ondisk metadata *only* when the ondisk metadata is so3006badly damaged that the filesystem cannot load the in-memory representation.3007When online fsck wants to open a damaged file for scrubbing, it must use3008specialized resource acquisition functions that return either the in-memory3009representation *or* a lock on whichever object is necessary to prevent any3010update to the ondisk location.3011 3012The only repairs that should be made to the ondisk inode buffers are whatever3013is necessary to get the in-core structure loaded.3014This means fixing whatever is caught by the inode cluster buffer and inode fork3015verifiers, and retrying the ``iget`` operation.3016If the second ``iget`` fails, the repair has failed.3017 3018Once the in-memory representation is loaded, repair can lock the inode and can3019subject it to comprehensive checks, repairs, and optimizations.3020Most inode attributes are easy to check and constrain, or are user-controlled3021arbitrary bit patterns; these are both easy to fix.3022Dealing with the data and attr fork extent counts and the file block counts is3023more complicated, because computing the correct value requires traversing the3024forks, or if that fails, leaving the fields invalid and waiting for the fork3025fsck functions to run.3026 3027The proposed patchset is the3028`inode3029<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-inodes>`_3030repair series.3031 3032Quota Record Repairs3033--------------------3034 3035Similar to inodes, quota records ("dquots") also have both ondisk records and3036an in-memory representation, and hence are subject to the same cache coherency3037issues.3038Somewhat confusingly, both are known as dquots in the XFS codebase.3039 3040The only repairs that should be made to the ondisk quota record buffers are3041whatever is necessary to get the in-core structure loaded.3042Once the in-memory representation is loaded, the only attributes needing3043checking are obviously bad limits and timer values.3044 3045Quota usage counters are checked, repaired, and discussed separately in the3046section about :ref:`live quotacheck <quotacheck>`.3047 3048The proposed patchset is the3049`quota3050<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-quota>`_3051repair series.3052 3053.. _fscounters:3054 3055Freezing to Fix Summary Counters3056--------------------------------3057 3058Filesystem summary counters track availability of filesystem resources such3059as free blocks, free inodes, and allocated inodes.3060This information could be compiled by walking the free space and inode indexes,3061but this is a slow process, so XFS maintains a copy in the ondisk superblock3062that should reflect the ondisk metadata, at least when the filesystem has been3063unmounted cleanly.3064For performance reasons, XFS also maintains incore copies of those counters,3065which are key to enabling resource reservations for active transactions.3066Writer threads reserve the worst-case quantities of resources from the3067incore counter and give back whatever they don't use at commit time.3068It is therefore only necessary to serialize on the superblock when the3069superblock is being committed to disk.3070 3071The lazy superblock counter feature introduced in XFS v5 took this even further3072by training log recovery to recompute the summary counters from the AG headers,3073which eliminated the need for most transactions even to touch the superblock.3074The only time XFS commits the summary counters is at filesystem unmount.3075To reduce contention even further, the incore counter is implemented as a3076percpu counter, which means that each CPU is allocated a batch of blocks from a3077global incore counter and can satisfy small allocations from the local batch.3078 3079The high-performance nature of the summary counters makes it difficult for3080online fsck to check them, since there is no way to quiesce a percpu counter3081while the system is running.3082Although online fsck can read the filesystem metadata to compute the correct3083values of the summary counters, there's no way to hold the value of a percpu3084counter stable, so it's quite possible that the counter will be out of date by3085the time the walk is complete.3086Earlier versions of online scrub would return to userspace with an incomplete3087scan flag, but this is not a satisfying outcome for a system administrator.3088For repairs, the in-memory counters must be stabilized while walking the3089filesystem metadata to get an accurate reading and install it in the percpu3090counter.3091 3092To satisfy this requirement, online fsck must prevent other programs in the3093system from initiating new writes to the filesystem, it must disable background3094garbage collection threads, and it must wait for existing writer programs to3095exit the kernel.3096Once that has been established, scrub can walk the AG free space indexes, the3097inode btrees, and the realtime bitmap to compute the correct value of all3098four summary counters.3099This is very similar to a filesystem freeze, though not all of the pieces are3100necessary:3101 3102- The final freeze state is set one higher than ``SB_FREEZE_COMPLETE`` to3103 prevent other threads from thawing the filesystem, or other scrub threads3104 from initiating another fscounters freeze.3105 3106- It does not quiesce the log.3107 3108With this code in place, it is now possible to pause the filesystem for just3109long enough to check and correct the summary counters.3110 3111+--------------------------------------------------------------------------+3112| **Historical Sidebar**: |3113+--------------------------------------------------------------------------+3114| The initial implementation used the actual VFS filesystem freeze |3115| mechanism to quiesce filesystem activity. |3116| With the filesystem frozen, it is possible to resolve the counter values |3117| with exact precision, but there are many problems with calling the VFS |3118| methods directly: |3119| |3120| - Other programs can unfreeze the filesystem without our knowledge. |3121| This leads to incorrect scan results and incorrect repairs. |3122| |3123| - Adding an extra lock to prevent others from thawing the filesystem |3124| required the addition of a ``->freeze_super`` function to wrap |3125| ``freeze_fs()``. |3126| This in turn caused other subtle problems because it turns out that |3127| the VFS ``freeze_super`` and ``thaw_super`` functions can drop the |3128| last reference to the VFS superblock, and any subsequent access |3129| becomes a UAF bug! |3130| This can happen if the filesystem is unmounted while the underlying |3131| block device has frozen the filesystem. |3132| This problem could be solved by grabbing extra references to the |3133| superblock, but it felt suboptimal given the other inadequacies of |3134| this approach. |3135| |3136| - The log need not be quiesced to check the summary counters, but a VFS |3137| freeze initiates one anyway. |3138| This adds unnecessary runtime to live fscounter fsck operations. |3139| |3140| - Quiescing the log means that XFS flushes the (possibly incorrect) |3141| counters to disk as part of cleaning the log. |3142| |3143| - A bug in the VFS meant that freeze could complete even when |3144| sync_filesystem fails to flush the filesystem and returns an error. |3145| This bug was fixed in Linux 5.17. |3146+--------------------------------------------------------------------------+3147 3148The proposed patchset is the3149`summary counter cleanup3150<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-fscounters>`_3151series.3152 3153Full Filesystem Scans3154---------------------3155 3156Certain types of metadata can only be checked by walking every file in the3157entire filesystem to record observations and comparing the observations against3158what's recorded on disk.3159Like every other type of online repair, repairs are made by writing those3160observations to disk in a replacement structure and committing it atomically.3161However, it is not practical to shut down the entire filesystem to examine3162hundreds of billions of files because the downtime would be excessive.3163Therefore, online fsck must build the infrastructure to manage a live scan of3164all the files in the filesystem.3165There are two questions that need to be solved to perform a live walk:3166 3167- How does scrub manage the scan while it is collecting data?3168 3169- How does the scan keep abreast of changes being made to the system by other3170 threads?3171 3172.. _iscan:3173 3174Coordinated Inode Scans3175```````````````````````3176 3177In the original Unix filesystems of the 1970s, each directory entry contained3178an index number (*inumber*) which was used as an index into on ondisk array3179(*itable*) of fixed-size records (*inodes*) describing a file's attributes and3180its data block mapping.3181This system is described by J. Lions, `"inode (5659)"3182<http://www.lemis.com/grog/Documentation/Lions/>`_ in *Lions' Commentary on3183UNIX, 6th Edition*, (Dept. of Computer Science, the University of New South3184Wales, November 1977), pp. 18-2; and later by D. Ritchie and K. Thompson,3185`"Implementation of the File System"3186<https://archive.org/details/bstj57-6-1905/page/n8/mode/1up>`_, from *The UNIX3187Time-Sharing System*, (The Bell System Technical Journal, July 1978), pp.31881913-4.3189 3190XFS retains most of this design, except now inumbers are search keys over all3191the space in the data section filesystem.3192They form a continuous keyspace that can be expressed as a 64-bit integer,3193though the inodes themselves are sparsely distributed within the keyspace.3194Scans proceed in a linear fashion across the inumber keyspace, starting from3195``0x0`` and ending at ``0xFFFFFFFFFFFFFFFF``.3196Naturally, a scan through a keyspace requires a scan cursor object to track the3197scan progress.3198Because this keyspace is sparse, this cursor contains two parts.3199The first part of this scan cursor object tracks the inode that will be3200examined next; call this the examination cursor.3201Somewhat less obviously, the scan cursor object must also track which parts of3202the keyspace have already been visited, which is critical for deciding if a3203concurrent filesystem update needs to be incorporated into the scan data.3204Call this the visited inode cursor.3205 3206Advancing the scan cursor is a multi-step process encapsulated in3207``xchk_iscan_iter``:3208 32091. Lock the AGI buffer of the AG containing the inode pointed to by the visited3210 inode cursor.3211 This guarantee that inodes in this AG cannot be allocated or freed while3212 advancing the cursor.3213 32142. Use the per-AG inode btree to look up the next inumber after the one that3215 was just visited, since it may not be keyspace adjacent.3216 32173. If there are no more inodes left in this AG:3218 3219 a. Move the examination cursor to the point of the inumber keyspace that3220 corresponds to the start of the next AG.3221 3222 b. Adjust the visited inode cursor to indicate that it has "visited" the3223 last possible inode in the current AG's inode keyspace.3224 XFS inumbers are segmented, so the cursor needs to be marked as having3225 visited the entire keyspace up to just before the start of the next AG's3226 inode keyspace.3227 3228 c. Unlock the AGI and return to step 1 if there are unexamined AGs in the3229 filesystem.3230 3231 d. If there are no more AGs to examine, set both cursors to the end of the3232 inumber keyspace.3233 The scan is now complete.3234 32354. Otherwise, there is at least one more inode to scan in this AG:3236 3237 a. Move the examination cursor ahead to the next inode marked as allocated3238 by the inode btree.3239 3240 b. Adjust the visited inode cursor to point to the inode just prior to where3241 the examination cursor is now.3242 Because the scanner holds the AGI buffer lock, no inodes could have been3243 created in the part of the inode keyspace that the visited inode cursor3244 just advanced.3245 32465. Get the incore inode for the inumber of the examination cursor.3247 By maintaining the AGI buffer lock until this point, the scanner knows that3248 it was safe to advance the examination cursor across the entire keyspace,3249 and that it has stabilized this next inode so that it cannot disappear from3250 the filesystem until the scan releases the incore inode.3251 32526. Drop the AGI lock and return the incore inode to the caller.3253 3254Online fsck functions scan all files in the filesystem as follows:3255 32561. Start a scan by calling ``xchk_iscan_start``.3257 32582. Advance the scan cursor (``xchk_iscan_iter``) to get the next inode.3259 If one is provided:3260 3261 a. Lock the inode to prevent updates during the scan.3262 3263 b. Scan the inode.3264 3265 c. While still holding the inode lock, adjust the visited inode cursor3266 (``xchk_iscan_mark_visited``) to point to this inode.3267 3268 d. Unlock and release the inode.3269 32708. Call ``xchk_iscan_teardown`` to complete the scan.3271 3272There are subtleties with the inode cache that complicate grabbing the incore3273inode for the caller.3274Obviously, it is an absolute requirement that the inode metadata be consistent3275enough to load it into the inode cache.3276Second, if the incore inode is stuck in some intermediate state, the scan3277coordinator must release the AGI and push the main filesystem to get the inode3278back into a loadable state.3279 3280The proposed patches are the3281`inode scanner3282<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=scrub-iscan>`_3283series.3284The first user of the new functionality is the3285`online quotacheck3286<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-quotacheck>`_3287series.3288 3289Inode Management3290````````````````3291 3292In regular filesystem code, references to allocated XFS incore inodes are3293always obtained (``xfs_iget``) outside of transaction context because the3294creation of the incore context for an existing file does not require metadata3295updates.3296However, it is important to note that references to incore inodes obtained as3297part of file creation must be performed in transaction context because the3298filesystem must ensure the atomicity of the ondisk inode btree index updates3299and the initialization of the actual ondisk inode.3300 3301References to incore inodes are always released (``xfs_irele``) outside of3302transaction context because there are a handful of activities that might3303require ondisk updates:3304 3305- The VFS may decide to kick off writeback as part of a ``DONTCACHE`` inode3306 release.3307 3308- Speculative preallocations need to be unreserved.3309 3310- An unlinked file may have lost its last reference, in which case the entire3311 file must be inactivated, which involves releasing all of its resources in3312 the ondisk metadata and freeing the inode.3313 3314These activities are collectively called inode inactivation.3315Inactivation has two parts -- the VFS part, which initiates writeback on all3316dirty file pages, and the XFS part, which cleans up XFS-specific information3317and frees the inode if it was unlinked.3318If the inode is unlinked (or unconnected after a file handle operation), the3319kernel drops the inode into the inactivation machinery immediately.3320 3321During normal operation, resource acquisition for an update follows this order3322to avoid deadlocks:3323 33241. Inode reference (``iget``).3325 33262. Filesystem freeze protection, if repairing (``mnt_want_write_file``).3327 33283. Inode ``IOLOCK`` (VFS ``i_rwsem``) lock to control file IO.3329 33304. Inode ``MMAPLOCK`` (page cache ``invalidate_lock``) lock for operations that3331 can update page cache mappings.3332 33335. Log feature enablement.3334 33356. Transaction log space grant.3336 33377. Space on the data and realtime devices for the transaction.3338 33398. Incore dquot references, if a file is being repaired.3340 Note that they are not locked, merely acquired.3341 33429. Inode ``ILOCK`` for file metadata updates.3343 334410. AG header buffer locks / Realtime metadata inode ILOCK.3345 334611. Realtime metadata buffer locks, if applicable.3347 334812. Extent mapping btree blocks, if applicable.3349 3350Resources are often released in the reverse order, though this is not required.3351However, online fsck differs from regular XFS operations because it may examine3352an object that normally is acquired in a later stage of the locking order, and3353then decide to cross-reference the object with an object that is acquired3354earlier in the order.3355The next few sections detail the specific ways in which online fsck takes care3356to avoid deadlocks.3357 3358iget and irele During a Scrub3359^^^^^^^^^^^^^^^^^^^^^^^^^^^^^3360 3361An inode scan performed on behalf of a scrub operation runs in transaction3362context, and possibly with resources already locked and bound to it.3363This isn't much of a problem for ``iget`` since it can operate in the context3364of an existing transaction, as long as all of the bound resources are acquired3365before the inode reference in the regular filesystem.3366 3367When the VFS ``iput`` function is given a linked inode with no other3368references, it normally puts the inode on an LRU list in the hope that it can3369save time if another process re-opens the file before the system runs out3370of memory and frees it.3371Filesystem callers can short-circuit the LRU process by setting a ``DONTCACHE``3372flag on the inode to cause the kernel to try to drop the inode into the3373inactivation machinery immediately.3374 3375In the past, inactivation was always done from the process that dropped the3376inode, which was a problem for scrub because scrub may already hold a3377transaction, and XFS does not support nesting transactions.3378On the other hand, if there is no scrub transaction, it is desirable to drop3379otherwise unused inodes immediately to avoid polluting caches.3380To capture these nuances, the online fsck code has a separate ``xchk_irele``3381function to set or clear the ``DONTCACHE`` flag to get the required release3382behavior.3383 3384Proposed patchsets include fixing3385`scrub iget usage3386<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=scrub-iget-fixes>`_ and3387`dir iget usage3388<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=scrub-dir-iget-fixes>`_.3389 3390.. _ilocking:3391 3392Locking Inodes3393^^^^^^^^^^^^^^3394 3395In regular filesystem code, the VFS and XFS will acquire multiple IOLOCK locks3396in a well-known order: parent → child when updating the directory tree, and3397in numerical order of the addresses of their ``struct inode`` object otherwise.3398For regular files, the MMAPLOCK can be acquired after the IOLOCK to stop page3399faults.3400If two MMAPLOCKs must be acquired, they are acquired in numerical order of3401the addresses of their ``struct address_space`` objects.3402Due to the structure of existing filesystem code, IOLOCKs and MMAPLOCKs must be3403acquired before transactions are allocated.3404If two ILOCKs must be acquired, they are acquired in inumber order.3405 3406Inode lock acquisition must be done carefully during a coordinated inode scan.3407Online fsck cannot abide these conventions, because for a directory tree3408scanner, the scrub process holds the IOLOCK of the file being scanned and it3409needs to take the IOLOCK of the file at the other end of the directory link.3410If the directory tree is corrupt because it contains a cycle, ``xfs_scrub``3411cannot use the regular inode locking functions and avoid becoming trapped in an3412ABBA deadlock.3413 3414Solving both of these problems is straightforward -- any time online fsck3415needs to take a second lock of the same class, it uses trylock to avoid an ABBA3416deadlock.3417If the trylock fails, scrub drops all inode locks and use trylock loops to3418(re)acquire all necessary resources.3419Trylock loops enable scrub to check for pending fatal signals, which is how3420scrub avoids deadlocking the filesystem or becoming an unresponsive process.3421However, trylock loops means that online fsck must be prepared to measure the3422resource being scrubbed before and after the lock cycle to detect changes and3423react accordingly.3424 3425.. _dirparent:3426 3427Case Study: Finding a Directory Parent3428^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^3429 3430Consider the directory parent pointer repair code as an example.3431Online fsck must verify that the dotdot dirent of a directory points up to a3432parent directory, and that the parent directory contains exactly one dirent3433pointing down to the child directory.3434Fully validating this relationship (and repairing it if possible) requires a3435walk of every directory on the filesystem while holding the child locked, and3436while updates to the directory tree are being made.3437The coordinated inode scan provides a way to walk the filesystem without the3438possibility of missing an inode.3439The child directory is kept locked to prevent updates to the dotdot dirent, but3440if the scanner fails to lock a parent, it can drop and relock both the child3441and the prospective parent.3442If the dotdot entry changes while the directory is unlocked, then a move or3443rename operation must have changed the child's parentage, and the scan can3444exit early.3445 3446The proposed patchset is the3447`directory repair3448<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-dirs>`_3449series.3450 3451.. _fshooks:3452 3453Filesystem Hooks3454`````````````````3455 3456The second piece of support that online fsck functions need during a full3457filesystem scan is the ability to stay informed about updates being made by3458other threads in the filesystem, since comparisons against the past are useless3459in a dynamic environment.3460Two pieces of Linux kernel infrastructure enable online fsck to monitor regular3461filesystem operations: filesystem hooks and :ref:`static keys<jump_labels>`.3462 3463Filesystem hooks convey information about an ongoing filesystem operation to3464a downstream consumer.3465In this case, the downstream consumer is always an online fsck function.3466Because multiple fsck functions can run in parallel, online fsck uses the Linux3467notifier call chain facility to dispatch updates to any number of interested3468fsck processes.3469Call chains are a dynamic list, which means that they can be configured at3470run time.3471Because these hooks are private to the XFS module, the information passed along3472contains exactly what the checking function needs to update its observations.3473 3474The current implementation of XFS hooks uses SRCU notifier chains to reduce the3475impact to highly threaded workloads.3476Regular blocking notifier chains use a rwsem and seem to have a much lower3477overhead for single-threaded applications.3478However, it may turn out that the combination of blocking chains and static3479keys are a more performant combination; more study is needed here.3480 3481The following pieces are necessary to hook a certain point in the filesystem:3482 3483- A ``struct xfs_hooks`` object must be embedded in a convenient place such as3484 a well-known incore filesystem object.3485 3486- Each hook must define an action code and a structure containing more context3487 about the action.3488 3489- Hook providers should provide appropriate wrapper functions and structs3490 around the ``xfs_hooks`` and ``xfs_hook`` objects to take advantage of type3491 checking to ensure correct usage.3492 3493- A callsite in the regular filesystem code must be chosen to call3494 ``xfs_hooks_call`` with the action code and data structure.3495 This place should be adjacent to (and not earlier than) the place where3496 the filesystem update is committed to the transaction.3497 In general, when the filesystem calls a hook chain, it should be able to3498 handle sleeping and should not be vulnerable to memory reclaim or locking3499 recursion.3500 However, the exact requirements are very dependent on the context of the hook3501 caller and the callee.3502 3503- The online fsck function should define a structure to hold scan data, a lock3504 to coordinate access to the scan data, and a ``struct xfs_hook`` object.3505 The scanner function and the regular filesystem code must acquire resources3506 in the same order; see the next section for details.3507 3508- The online fsck code must contain a C function to catch the hook action code3509 and data structure.3510 If the object being updated has already been visited by the scan, then the3511 hook information must be applied to the scan data.3512 3513- Prior to unlocking inodes to start the scan, online fsck must call3514 ``xfs_hooks_setup`` to initialize the ``struct xfs_hook``, and3515 ``xfs_hooks_add`` to enable the hook.3516 3517- Online fsck must call ``xfs_hooks_del`` to disable the hook once the scan is3518 complete.3519 3520The number of hooks should be kept to a minimum to reduce complexity.3521Static keys are used to reduce the overhead of filesystem hooks to nearly3522zero when online fsck is not running.3523 3524.. _liveupdate:3525 3526Live Updates During a Scan3527``````````````````````````3528 3529The code paths of the online fsck scanning code and the :ref:`hooked<fshooks>`3530filesystem code look like this::3531 3532 other program3533 ↓3534 inode lock ←────────────────────┐3535 ↓ │3536 AG header lock │3537 ↓ │3538 filesystem function │3539 ↓ │3540 notifier call chain │ same3541 ↓ ├─── inode3542 scrub hook function │ lock3543 ↓ │3544 scan data mutex ←──┐ same │3545 ↓ ├─── scan │3546 update scan data │ lock │3547 ↑ │ │3548 scan data mutex ←──┘ │3549 ↑ │3550 inode lock ←────────────────────┘3551 ↑3552 scrub function3553 ↑3554 inode scanner3555 ↑3556 xfs_scrub3557 3558These rules must be followed to ensure correct interactions between the3559checking code and the code making an update to the filesystem:3560 3561- Prior to invoking the notifier call chain, the filesystem function being3562 hooked must acquire the same lock that the scrub scanning function acquires3563 to scan the inode.3564 3565- The scanning function and the scrub hook function must coordinate access to3566 the scan data by acquiring a lock on the scan data.3567 3568- Scrub hook function must not add the live update information to the scan3569 observations unless the inode being updated has already been scanned.3570 The scan coordinator has a helper predicate (``xchk_iscan_want_live_update``)3571 for this.3572 3573- Scrub hook functions must not change the caller's state, including the3574 transaction that it is running.3575 They must not acquire any resources that might conflict with the filesystem3576 function being hooked.3577 3578- The hook function can abort the inode scan to avoid breaking the other rules.3579 3580The inode scan APIs are pretty simple:3581 3582- ``xchk_iscan_start`` starts a scan3583 3584- ``xchk_iscan_iter`` grabs a reference to the next inode in the scan or3585 returns zero if there is nothing left to scan3586 3587- ``xchk_iscan_want_live_update`` to decide if an inode has already been3588 visited in the scan.3589 This is critical for hook functions to decide if they need to update the3590 in-memory scan information.3591 3592- ``xchk_iscan_mark_visited`` to mark an inode as having been visited in the3593 scan3594 3595- ``xchk_iscan_teardown`` to finish the scan3596 3597This functionality is also a part of the3598`inode scanner3599<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=scrub-iscan>`_3600series.3601 3602.. _quotacheck:3603 3604Case Study: Quota Counter Checking3605^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^3606 3607It is useful to compare the mount time quotacheck code to the online repair3608quotacheck code.3609Mount time quotacheck does not have to contend with concurrent operations, so3610it does the following:3611 36121. Make sure the ondisk dquots are in good enough shape that all the incore3613 dquots will actually load, and zero the resource usage counters in the3614 ondisk buffer.3615 36162. Walk every inode in the filesystem.3617 Add each file's resource usage to the incore dquot.3618 36193. Walk each incore dquot.3620 If the incore dquot is not being flushed, add the ondisk buffer backing the3621 incore dquot to a delayed write (delwri) list.3622 36234. Write the buffer list to disk.3624 3625Like most online fsck functions, online quotacheck can't write to regular3626filesystem objects until the newly collected metadata reflect all filesystem3627state.3628Therefore, online quotacheck records file resource usage to a shadow dquot3629index implemented with a sparse ``xfarray``, and only writes to the real dquots3630once the scan is complete.3631Handling transactional updates is tricky because quota resource usage updates3632are handled in phases to minimize contention on dquots:3633 36341. The inodes involved are joined and locked to a transaction.3635 36362. For each dquot attached to the file:3637 3638 a. The dquot is locked.3639 3640 b. A quota reservation is added to the dquot's resource usage.3641 The reservation is recorded in the transaction.3642 3643 c. The dquot is unlocked.3644 36453. Changes in actual quota usage are tracked in the transaction.3646 36474. At transaction commit time, each dquot is examined again:3648 3649 a. The dquot is locked again.3650 3651 b. Quota usage changes are logged and unused reservation is given back to3652 the dquot.3653 3654 c. The dquot is unlocked.3655 3656For online quotacheck, hooks are placed in steps 2 and 4.3657The step 2 hook creates a shadow version of the transaction dquot context3658(``dqtrx``) that operates in a similar manner to the regular code.3659The step 4 hook commits the shadow ``dqtrx`` changes to the shadow dquots.3660Notice that both hooks are called with the inode locked, which is how the3661live update coordinates with the inode scanner.3662 3663The quotacheck scan looks like this:3664 36651. Set up a coordinated inode scan.3666 36672. For each inode returned by the inode scan iterator:3668 3669 a. Grab and lock the inode.3670 3671 b. Determine that inode's resource usage (data blocks, inode counts,3672 realtime blocks) and add that to the shadow dquots for the user, group,3673 and project ids associated with the inode.3674 3675 c. Unlock and release the inode.3676 36773. For each dquot in the system:3678 3679 a. Grab and lock the dquot.3680 3681 b. Check the dquot against the shadow dquots created by the scan and updated3682 by the live hooks.3683 3684Live updates are key to being able to walk every quota record without3685needing to hold any locks for a long duration.3686If repairs are desired, the real and shadow dquots are locked and their3687resource counts are set to the values in the shadow dquot.3688 3689The proposed patchset is the3690`online quotacheck3691<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-quotacheck>`_3692series.3693 3694.. _nlinks:3695 3696Case Study: File Link Count Checking3697^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^3698 3699File link count checking also uses live update hooks.3700The coordinated inode scanner is used to visit all directories on the3701filesystem, and per-file link count records are stored in a sparse ``xfarray``3702indexed by inumber.3703During the scanning phase, each entry in a directory generates observation3704data as follows:3705 37061. If the entry is a dotdot (``'..'``) entry of the root directory, the3707 directory's parent link count is bumped because the root directory's dotdot3708 entry is self referential.3709 37102. If the entry is a dotdot entry of a subdirectory, the parent's backref3711 count is bumped.3712 37133. If the entry is neither a dot nor a dotdot entry, the target file's parent3714 count is bumped.3715 37164. If the target is a subdirectory, the parent's child link count is bumped.3717 3718A crucial point to understand about how the link count inode scanner interacts3719with the live update hooks is that the scan cursor tracks which *parent*3720directories have been scanned.3721In other words, the live updates ignore any update about ``A → B`` when A has3722not been scanned, even if B has been scanned.3723Furthermore, a subdirectory A with a dotdot entry pointing back to B is3724accounted as a backref counter in the shadow data for A, since child dotdot3725entries affect the parent's link count.3726Live update hooks are carefully placed in all parts of the filesystem that3727create, change, or remove directory entries, since those operations involve3728bumplink and droplink.3729 3730For any file, the correct link count is the number of parents plus the number3731of child subdirectories.3732Non-directories never have children of any kind.3733The backref information is used to detect inconsistencies in the number of3734links pointing to child subdirectories and the number of dotdot entries3735pointing back.3736 3737After the scan completes, the link count of each file can be checked by locking3738both the inode and the shadow data, and comparing the link counts.3739A second coordinated inode scan cursor is used for comparisons.3740Live updates are key to being able to walk every inode without needing to hold3741any locks between inodes.3742If repairs are desired, the inode's link count is set to the value in the3743shadow information.3744If no parents are found, the file must be :ref:`reparented <orphanage>` to the3745orphanage to prevent the file from being lost forever.3746 3747The proposed patchset is the3748`file link count repair3749<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=scrub-nlinks>`_3750series.3751 3752.. _rmap_repair:3753 3754Case Study: Rebuilding Reverse Mapping Records3755^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^3756 3757Most repair functions follow the same pattern: lock filesystem resources,3758walk the surviving ondisk metadata looking for replacement metadata records,3759and use an :ref:`in-memory array <xfarray>` to store the gathered observations.3760The primary advantage of this approach is the simplicity and modularity of the3761repair code -- code and data are entirely contained within the scrub module,3762do not require hooks in the main filesystem, and are usually the most efficient3763in memory use.3764A secondary advantage of this repair approach is atomicity -- once the kernel3765decides a structure is corrupt, no other threads can access the metadata until3766the kernel finishes repairing and revalidating the metadata.3767 3768For repairs going on within a shard of the filesystem, these advantages3769outweigh the delays inherent in locking the shard while repairing parts of the3770shard.3771Unfortunately, repairs to the reverse mapping btree cannot use the "standard"3772btree repair strategy because it must scan every space mapping of every fork of3773every file in the filesystem, and the filesystem cannot stop.3774Therefore, rmap repair foregoes atomicity between scrub and repair.3775It combines a :ref:`coordinated inode scanner <iscan>`, :ref:`live update hooks3776<liveupdate>`, and an :ref:`in-memory rmap btree <xfbtree>` to complete the3777scan for reverse mapping records.3778 37791. Set up an xfbtree to stage rmap records.3780 37812. While holding the locks on the AGI and AGF buffers acquired during the3782 scrub, generate reverse mappings for all AG metadata: inodes, btrees, CoW3783 staging extents, and the internal log.3784 37853. Set up an inode scanner.3786 37874. Hook into rmap updates for the AG being repaired so that the live scan data3788 can receive updates to the rmap btree from the rest of the filesystem during3789 the file scan.3790 37915. For each space mapping found in either fork of each file scanned,3792 decide if the mapping matches the AG of interest.3793 If so:3794 3795 a. Create a btree cursor for the in-memory btree.3796 3797 b. Use the rmap code to add the record to the in-memory btree.3798 3799 c. Use the :ref:`special commit function <xfbtree_commit>` to write the3800 xfbtree changes to the xfile.3801 38026. For each live update received via the hook, decide if the owner has already3803 been scanned.3804 If so, apply the live update into the scan data:3805 3806 a. Create a btree cursor for the in-memory btree.3807 3808 b. Replay the operation into the in-memory btree.3809 3810 c. Use the :ref:`special commit function <xfbtree_commit>` to write the3811 xfbtree changes to the xfile.3812 This is performed with an empty transaction to avoid changing the3813 caller's state.3814 38157. When the inode scan finishes, create a new scrub transaction and relock the3816 two AG headers.3817 38188. Compute the new btree geometry using the number of rmap records in the3819 shadow btree, like all other btree rebuilding functions.3820 38219. Allocate the number of blocks computed in the previous step.3822 382310. Perform the usual btree bulk loading and commit to install the new rmap3824 btree.3825 382611. Reap the old rmap btree blocks as discussed in the case study about how3827 to :ref:`reap after rmap btree repair <rmap_reap>`.3828 382912. Free the xfbtree now that it not needed.3830 3831The proposed patchset is the3832`rmap repair3833<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-rmap-btree>`_3834series.3835 3836Staging Repairs with Temporary Files on Disk3837--------------------------------------------3838 3839XFS stores a substantial amount of metadata in file forks: directories,3840extended attributes, symbolic link targets, free space bitmaps and summary3841information for the realtime volume, and quota records.3842File forks map 64-bit logical file fork space extents to physical storage space3843extents, similar to how a memory management unit maps 64-bit virtual addresses3844to physical memory addresses.3845Therefore, file-based tree structures (such as directories and extended3846attributes) use blocks mapped in the file fork offset address space that point3847to other blocks mapped within that same address space, and file-based linear3848structures (such as bitmaps and quota records) compute array element offsets in3849the file fork offset address space.3850 3851Because file forks can consume as much space as the entire filesystem, repairs3852cannot be staged in memory, even when a paging scheme is available.3853Therefore, online repair of file-based metadata createas a temporary file in3854the XFS filesystem, writes a new structure at the correct offsets into the3855temporary file, and atomically exchanges all file fork mappings (and hence the3856fork contents) to commit the repair.3857Once the repair is complete, the old fork can be reaped as necessary; if the3858system goes down during the reap, the iunlink code will delete the blocks3859during log recovery.3860 3861**Note**: All space usage and inode indices in the filesystem *must* be3862consistent to use a temporary file safely!3863This dependency is the reason why online repair can only use pageable kernel3864memory to stage ondisk space usage information.3865 3866Exchanging metadata file mappings with a temporary file requires the owner3867field of the block headers to match the file being repaired and not the3868temporary file.3869The directory, extended attribute, and symbolic link functions were all3870modified to allow callers to specify owner numbers explicitly.3871 3872There is a downside to the reaping process -- if the system crashes during the3873reap phase and the fork extents are crosslinked, the iunlink processing will3874fail because freeing space will find the extra reverse mappings and abort.3875 3876Temporary files created for repair are similar to ``O_TMPFILE`` files created3877by userspace.3878They are not linked into a directory and the entire file will be reaped when3879the last reference to the file is lost.3880The key differences are that these files must have no access permission outside3881the kernel at all, they must be specially marked to prevent them from being3882opened by handle, and they must never be linked into the directory tree.3883 3884+--------------------------------------------------------------------------+3885| **Historical Sidebar**: |3886+--------------------------------------------------------------------------+3887| In the initial iteration of file metadata repair, the damaged metadata |3888| blocks would be scanned for salvageable data; the extents in the file |3889| fork would be reaped; and then a new structure would be built in its |3890| place. |3891| This strategy did not survive the introduction of the atomic repair |3892| requirement expressed earlier in this document. |3893| |3894| The second iteration explored building a second structure at a high |3895| offset in the fork from the salvage data, reaping the old extents, and |3896| using a ``COLLAPSE_RANGE`` operation to slide the new extents into |3897| place. |3898| |3899| This had many drawbacks: |3900| |3901| - Array structures are linearly addressed, and the regular filesystem |3902| codebase does not have the concept of a linear offset that could be |3903| applied to the record offset computation to build an alternate copy. |3904| |3905| - Extended attributes are allowed to use the entire attr fork offset |3906| address space. |3907| |3908| - Even if repair could build an alternate copy of a data structure in a |3909| different part of the fork address space, the atomic repair commit |3910| requirement means that online repair would have to be able to perform |3911| a log assisted ``COLLAPSE_RANGE`` operation to ensure that the old |3912| structure was completely replaced. |3913| |3914| - A crash after construction of the secondary tree but before the range |3915| collapse would leave unreachable blocks in the file fork. |3916| This would likely confuse things further. |3917| |3918| - Reaping blocks after a repair is not a simple operation, and |3919| initiating a reap operation from a restarted range collapse operation |3920| during log recovery is daunting. |3921| |3922| - Directory entry blocks and quota records record the file fork offset |3923| in the header area of each block. |3924| An atomic range collapse operation would have to rewrite this part of |3925| each block header. |3926| Rewriting a single field in block headers is not a huge problem, but |3927| it's something to be aware of. |3928| |3929| - Each block in a directory or extended attributes btree index contains |3930| sibling and child block pointers. |3931| Were the atomic commit to use a range collapse operation, each block |3932| would have to be rewritten very carefully to preserve the graph |3933| structure. |3934| Doing this as part of a range collapse means rewriting a large number |3935| of blocks repeatedly, which is not conducive to quick repairs. |3936| |3937| This lead to the introduction of temporary file staging. |3938+--------------------------------------------------------------------------+3939 3940Using a Temporary File3941``````````````````````3942 3943Online repair code should use the ``xrep_tempfile_create`` function to create a3944temporary file inside the filesystem.3945This allocates an inode, marks the in-core inode private, and attaches it to3946the scrub context.3947These files are hidden from userspace, may not be added to the directory tree,3948and must be kept private.3949 3950Temporary files only use two inode locks: the IOLOCK and the ILOCK.3951The MMAPLOCK is not needed here, because there must not be page faults from3952userspace for data fork blocks.3953The usage patterns of these two locks are the same as for any other XFS file --3954access to file data are controlled via the IOLOCK, and access to file metadata3955are controlled via the ILOCK.3956Locking helpers are provided so that the temporary file and its lock state can3957be cleaned up by the scrub context.3958To comply with the nested locking strategy laid out in the :ref:`inode3959locking<ilocking>` section, it is recommended that scrub functions use the3960xrep_tempfile_ilock*_nowait lock helpers.3961 3962Data can be written to a temporary file by two means:3963 39641. ``xrep_tempfile_copyin`` can be used to set the contents of a regular3965 temporary file from an xfile.3966 39672. The regular directory, symbolic link, and extended attribute functions can3968 be used to write to the temporary file.3969 3970Once a good copy of a data file has been constructed in a temporary file, it3971must be conveyed to the file being repaired, which is the topic of the next3972section.3973 3974The proposed patches are in the3975`repair temporary files3976<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-tempfiles>`_3977series.3978 3979Logged File Content Exchanges3980-----------------------------3981 3982Once repair builds a temporary file with a new data structure written into3983it, it must commit the new changes into the existing file.3984It is not possible to swap the inumbers of two files, so instead the new3985metadata must replace the old.3986This suggests the need for the ability to swap extents, but the existing extent3987swapping code used by the file defragmenting tool ``xfs_fsr`` is not sufficient3988for online repair because:3989 3990a. When the reverse-mapping btree is enabled, the swap code must keep the3991 reverse mapping information up to date with every exchange of mappings.3992 Therefore, it can only exchange one mapping per transaction, and each3993 transaction is independent.3994 3995b. Reverse-mapping is critical for the operation of online fsck, so the old3996 defragmentation code (which swapped entire extent forks in a single3997 operation) is not useful here.3998 3999c. Defragmentation is assumed to occur between two files with identical4000 contents.4001 For this use case, an incomplete exchange will not result in a user-visible4002 change in file contents, even if the operation is interrupted.4003 4004d. Online repair needs to swap the contents of two files that are by definition4005 *not* identical.4006 For directory and xattr repairs, the user-visible contents might be the4007 same, but the contents of individual blocks may be very different.4008 4009e. Old blocks in the file may be cross-linked with another structure and must4010 not reappear if the system goes down mid-repair.4011 4012These problems are overcome by creating a new deferred operation and a new type4013of log intent item to track the progress of an operation to exchange two file4014ranges.4015The new exchange operation type chains together the same transactions used by4016the reverse-mapping extent swap code, but records intermedia progress in the4017log so that operations can be restarted after a crash.4018This new functionality is called the file contents exchange (xfs_exchrange)4019code.4020The underlying implementation exchanges file fork mappings (xfs_exchmaps).4021The new log item records the progress of the exchange to ensure that once an4022exchange begins, it will always run to completion, even there are4023interruptions.4024The new ``XFS_SB_FEAT_INCOMPAT_EXCHRANGE`` incompatible feature flag4025in the superblock protects these new log item records from being replayed on4026old kernels.4027 4028The proposed patchset is the4029`file contents exchange4030<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=atomic-file-updates>`_4031series.4032 4033+--------------------------------------------------------------------------+4034| **Sidebar: Using Log-Incompatible Feature Flags** |4035+--------------------------------------------------------------------------+4036| Starting with XFS v5, the superblock contains a |4037| ``sb_features_log_incompat`` field to indicate that the log contains |4038| records that might not readable by all kernels that could mount this |4039| filesystem. |4040| In short, log incompat features protect the log contents against kernels |4041| that will not understand the contents. |4042| Unlike the other superblock feature bits, log incompat bits are |4043| ephemeral because an empty (clean) log does not need protection. |4044| The log cleans itself after its contents have been committed into the |4045| filesystem, either as part of an unmount or because the system is |4046| otherwise idle. |4047| Because upper level code can be working on a transaction at the same |4048| time that the log cleans itself, it is necessary for upper level code to |4049| communicate to the log when it is going to use a log incompatible |4050| feature. |4051| |4052| The log coordinates access to incompatible features through the use of |4053| one ``struct rw_semaphore`` for each feature. |4054| The log cleaning code tries to take this rwsem in exclusive mode to |4055| clear the bit; if the lock attempt fails, the feature bit remains set. |4056| The code supporting a log incompat feature should create wrapper |4057| functions to obtain the log feature and call |4058| ``xfs_add_incompat_log_feature`` to set the feature bits in the primary |4059| superblock. |4060| The superblock update is performed transactionally, so the wrapper to |4061| obtain log assistance must be called just prior to the creation of the |4062| transaction that uses the functionality. |4063| For a file operation, this step must happen after taking the IOLOCK |4064| and the MMAPLOCK, but before allocating the transaction. |4065| When the transaction is complete, the ``xlog_drop_incompat_feat`` |4066| function is called to release the feature. |4067| The feature bit will not be cleared from the superblock until the log |4068| becomes clean. |4069| |4070| Log-assisted extended attribute updates and file content exchanges bothe |4071| use log incompat features and provide convenience wrappers around the |4072| functionality. |4073+--------------------------------------------------------------------------+4074 4075Mechanics of a Logged File Content Exchange4076```````````````````````````````````````````4077 4078Exchanging contents between file forks is a complex task.4079The goal is to exchange all file fork mappings between two file fork offset4080ranges.4081There are likely to be many extent mappings in each fork, and the edges of4082the mappings aren't necessarily aligned.4083Furthermore, there may be other updates that need to happen after the exchange,4084such as exchanging file sizes, inode flags, or conversion of fork data to local4085format.4086This is roughly the format of the new deferred exchange-mapping work item:4087 4088.. code-block:: c4089 4090 struct xfs_exchmaps_intent {4091 /* Inodes participating in the operation. */4092 struct xfs_inode *xmi_ip1;4093 struct xfs_inode *xmi_ip2;4094 4095 /* File offset range information. */4096 xfs_fileoff_t xmi_startoff1;4097 xfs_fileoff_t xmi_startoff2;4098 xfs_filblks_t xmi_blockcount;4099 4100 /* Set these file sizes after the operation, unless negative. */4101 xfs_fsize_t xmi_isize1;4102 xfs_fsize_t xmi_isize2;4103 4104 /* XFS_EXCHMAPS_* log operation flags */4105 uint64_t xmi_flags;4106 };4107 4108The new log intent item contains enough information to track two logical fork4109offset ranges: ``(inode1, startoff1, blockcount)`` and ``(inode2, startoff2,4110blockcount)``.4111Each step of an exchange operation exchanges the largest file range mapping4112possible from one file to the other.4113After each step in the exchange operation, the two startoff fields are4114incremented and the blockcount field is decremented to reflect the progress4115made.4116The flags field captures behavioral parameters such as exchanging attr fork4117mappings instead of the data fork and other work to be done after the exchange.4118The two isize fields are used to exchange the file sizes at the end of the4119operation if the file data fork is the target of the operation.4120 4121When the exchange is initiated, the sequence of operations is as follows:4122 41231. Create a deferred work item for the file mapping exchange.4124 At the start, it should contain the entirety of the file block ranges to be4125 exchanged.4126 41272. Call ``xfs_defer_finish`` to process the exchange.4128 This is encapsulated in ``xrep_tempexch_contents`` for scrub operations.4129 This will log an extent swap intent item to the transaction for the deferred4130 mapping exchange work item.4131 41323. Until ``xmi_blockcount`` of the deferred mapping exchange work item is zero,4133 4134 a. Read the block maps of both file ranges starting at ``xmi_startoff1`` and4135 ``xmi_startoff2``, respectively, and compute the longest extent that can4136 be exchanged in a single step.4137 This is the minimum of the two ``br_blockcount`` s in the mappings.4138 Keep advancing through the file forks until at least one of the mappings4139 contains written blocks.4140 Mutual holes, unwritten extents, and extent mappings to the same physical4141 space are not exchanged.4142 4143 For the next few steps, this document will refer to the mapping that came4144 from file 1 as "map1", and the mapping that came from file 2 as "map2".4145 4146 b. Create a deferred block mapping update to unmap map1 from file 1.4147 4148 c. Create a deferred block mapping update to unmap map2 from file 2.4149 4150 d. Create a deferred block mapping update to map map1 into file 2.4151 4152 e. Create a deferred block mapping update to map map2 into file 1.4153 4154 f. Log the block, quota, and extent count updates for both files.4155 4156 g. Extend the ondisk size of either file if necessary.4157 4158 h. Log a mapping exchange done log item for th mapping exchange intent log4159 item that was read at the start of step 3.4160 4161 i. Compute the amount of file range that has just been covered.4162 This quantity is ``(map1.br_startoff + map1.br_blockcount -4163 xmi_startoff1)``, because step 3a could have skipped holes.4164 4165 j. Increase the starting offsets of ``xmi_startoff1`` and ``xmi_startoff2``4166 by the number of blocks computed in the previous step, and decrease4167 ``xmi_blockcount`` by the same quantity.4168 This advances the cursor.4169 4170 k. Log a new mapping exchange intent log item reflecting the advanced state4171 of the work item.4172 4173 l. Return the proper error code (EAGAIN) to the deferred operation manager4174 to inform it that there is more work to be done.4175 The operation manager completes the deferred work in steps 3b-3e before4176 moving back to the start of step 3.4177 41784. Perform any post-processing.4179 This will be discussed in more detail in subsequent sections.4180 4181If the filesystem goes down in the middle of an operation, log recovery will4182find the most recent unfinished maping exchange log intent item and restart4183from there.4184This is how atomic file mapping exchanges guarantees that an outside observer4185will either see the old broken structure or the new one, and never a mismash of4186both.4187 4188Preparation for File Content Exchanges4189``````````````````````````````````````4190 4191There are a few things that need to be taken care of before initiating an4192atomic file mapping exchange operation.4193First, regular files require the page cache to be flushed to disk before the4194operation begins, and directio writes to be quiesced.4195Like any filesystem operation, file mapping exchanges must determine the4196maximum amount of disk space and quota that can be consumed on behalf of both4197files in the operation, and reserve that quantity of resources to avoid an4198unrecoverable out of space failure once it starts dirtying metadata.4199The preparation step scans the ranges of both files to estimate:4200 4201- Data device blocks needed to handle the repeated updates to the fork4202 mappings.4203- Change in data and realtime block counts for both files.4204- Increase in quota usage for both files, if the two files do not share the4205 same set of quota ids.4206- The number of extent mappings that will be added to each file.4207- Whether or not there are partially written realtime extents.4208 User programs must never be able to access a realtime file extent that maps4209 to different extents on the realtime volume, which could happen if the4210 operation fails to run to completion.4211 4212The need for precise estimation increases the run time of the exchange4213operation, but it is very important to maintain correct accounting.4214The filesystem must not run completely out of free space, nor can the mapping4215exchange ever add more extent mappings to a fork than it can support.4216Regular users are required to abide the quota limits, though metadata repairs4217may exceed quota to resolve inconsistent metadata elsewhere.4218 4219Special Features for Exchanging Metadata File Contents4220``````````````````````````````````````````````````````4221 4222Extended attributes, symbolic links, and directories can set the fork format to4223"local" and treat the fork as a literal area for data storage.4224Metadata repairs must take extra steps to support these cases:4225 4226- If both forks are in local format and the fork areas are large enough, the4227 exchange is performed by copying the incore fork contents, logging both4228 forks, and committing.4229 The atomic file mapping exchange mechanism is not necessary, since this can4230 be done with a single transaction.4231 4232- If both forks map blocks, then the regular atomic file mapping exchange is4233 used.4234 4235- Otherwise, only one fork is in local format.4236 The contents of the local format fork are converted to a block to perform the4237 exchange.4238 The conversion to block format must be done in the same transaction that4239 logs the initial mapping exchange intent log item.4240 The regular atomic mapping exchange is used to exchange the metadata file4241 mappings.4242 Special flags are set on the exchange operation so that the transaction can4243 be rolled one more time to convert the second file's fork back to local4244 format so that the second file will be ready to go as soon as the ILOCK is4245 dropped.4246 4247Extended attributes and directories stamp the owning inode into every block,4248but the buffer verifiers do not actually check the inode number!4249Although there is no verification, it is still important to maintain4250referential integrity, so prior to performing the mapping exchange, online4251repair builds every block in the new data structure with the owner field of the4252file being repaired.4253 4254After a successful exchange operation, the repair operation must reap the old4255fork blocks by processing each fork mapping through the standard :ref:`file4256extent reaping <reaping>` mechanism that is done post-repair.4257If the filesystem should go down during the reap part of the repair, the4258iunlink processing at the end of recovery will free both the temporary file and4259whatever blocks were not reaped.4260However, this iunlink processing omits the cross-link detection of online4261repair, and is not completely foolproof.4262 4263Exchanging Temporary File Contents4264``````````````````````````````````4265 4266To repair a metadata file, online repair proceeds as follows:4267 42681. Create a temporary repair file.4269 42702. Use the staging data to write out new contents into the temporary repair4271 file.4272 The same fork must be written to as is being repaired.4273 42743. Commit the scrub transaction, since the exchange resource estimation step4275 must be completed before transaction reservations are made.4276 42774. Call ``xrep_tempexch_trans_alloc`` to allocate a new scrub transaction with4278 the appropriate resource reservations, locks, and fill out a ``struct4279 xfs_exchmaps_req`` with the details of the exchange operation.4280 42815. Call ``xrep_tempexch_contents`` to exchange the contents.4282 42836. Commit the transaction to complete the repair.4284 4285.. _rtsummary:4286 4287Case Study: Repairing the Realtime Summary File4288^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^4289 4290In the "realtime" section of an XFS filesystem, free space is tracked via a4291bitmap, similar to Unix FFS.4292Each bit in the bitmap represents one realtime extent, which is a multiple of4293the filesystem block size between 4KiB and 1GiB in size.4294The realtime summary file indexes the number of free extents of a given size to4295the offset of the block within the realtime free space bitmap where those free4296extents begin.4297In other words, the summary file helps the allocator find free extents by4298length, similar to what the free space by count (cntbt) btree does for the data4299section.4300 4301The summary file itself is a flat file (with no block headers or checksums!)4302partitioned into ``log2(total rt extents)`` sections containing enough 32-bit4303counters to match the number of blocks in the rt bitmap.4304Each counter records the number of free extents that start in that bitmap block4305and can satisfy a power-of-two allocation request.4306 4307To check the summary file against the bitmap:4308 43091. Take the ILOCK of both the realtime bitmap and summary files.4310 43112. For each free space extent recorded in the bitmap:4312 4313 a. Compute the position in the summary file that contains a counter that4314 represents this free extent.4315 4316 b. Read the counter from the xfile.4317 4318 c. Increment it, and write it back to the xfile.4319 43203. Compare the contents of the xfile against the ondisk file.4321 4322To repair the summary file, write the xfile contents into the temporary file4323and use atomic mapping exchange to commit the new contents.4324The temporary file is then reaped.4325 4326The proposed patchset is the4327`realtime summary repair4328<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-rtsummary>`_4329series.4330 4331Case Study: Salvaging Extended Attributes4332^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^4333 4334In XFS, extended attributes are implemented as a namespaced name-value store.4335Values are limited in size to 64KiB, but there is no limit in the number of4336names.4337The attribute fork is unpartitioned, which means that the root of the attribute4338structure is always in logical block zero, but attribute leaf blocks, dabtree4339index blocks, and remote value blocks are intermixed.4340Attribute leaf blocks contain variable-sized records that associate4341user-provided names with the user-provided values.4342Values larger than a block are allocated separate extents and written there.4343If the leaf information expands beyond a single block, a directory/attribute4344btree (``dabtree``) is created to map hashes of attribute names to entries4345for fast lookup.4346 4347Salvaging extended attributes is done as follows:4348 43491. Walk the attr fork mappings of the file being repaired to find the attribute4350 leaf blocks.4351 When one is found,4352 4353 a. Walk the attr leaf block to find candidate keys.4354 When one is found,4355 4356 1. Check the name for problems, and ignore the name if there are.4357 4358 2. Retrieve the value.4359 If that succeeds, add the name and value to the staging xfarray and4360 xfblob.4361 43622. If the memory usage of the xfarray and xfblob exceed a certain amount of4363 memory or there are no more attr fork blocks to examine, unlock the file and4364 add the staged extended attributes to the temporary file.4365 43663. Use atomic file mapping exchange to exchange the new and old extended4367 attribute structures.4368 The old attribute blocks are now attached to the temporary file.4369 43704. Reap the temporary file.4371 4372The proposed patchset is the4373`extended attribute repair4374<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-xattrs>`_4375series.4376 4377Fixing Directories4378------------------4379 4380Fixing directories is difficult with currently available filesystem features,4381since directory entries are not redundant.4382The offline repair tool scans all inodes to find files with nonzero link count,4383and then it scans all directories to establish parentage of those linked files.4384Damaged files and directories are zapped, and files with no parent are4385moved to the ``/lost+found`` directory.4386It does not try to salvage anything.4387 4388The best that online repair can do at this time is to read directory data4389blocks and salvage any dirents that look plausible, correct link counts, and4390move orphans back into the directory tree.4391The salvage process is discussed in the case study at the end of this section.4392The :ref:`file link count fsck <nlinks>` code takes care of fixing link counts4393and moving orphans to the ``/lost+found`` directory.4394 4395Case Study: Salvaging Directories4396`````````````````````````````````4397 4398Unlike extended attributes, directory blocks are all the same size, so4399salvaging directories is straightforward:4400 44011. Find the parent of the directory.4402 If the dotdot entry is not unreadable, try to confirm that the alleged4403 parent has a child entry pointing back to the directory being repaired.4404 Otherwise, walk the filesystem to find it.4405 44062. Walk the first partition of data fork of the directory to find the directory4407 entry data blocks.4408 When one is found,4409 4410 a. Walk the directory data block to find candidate entries.4411 When an entry is found:4412 4413 i. Check the name for problems, and ignore the name if there are.4414 4415 ii. Retrieve the inumber and grab the inode.4416 If that succeeds, add the name, inode number, and file type to the4417 staging xfarray and xblob.4418 44193. If the memory usage of the xfarray and xfblob exceed a certain amount of4420 memory or there are no more directory data blocks to examine, unlock the4421 directory and add the staged dirents into the temporary directory.4422 Truncate the staging files.4423 44244. Use atomic file mapping exchange to exchange the new and old directory4425 structures.4426 The old directory blocks are now attached to the temporary file.4427 44285. Reap the temporary file.4429 4430**Future Work Question**: Should repair revalidate the dentry cache when4431rebuilding a directory?4432 4433*Answer*: Yes, it should.4434 4435In theory it is necessary to scan all dentry cache entries for a directory to4436ensure that one of the following apply:4437 44381. The cached dentry reflects an ondisk dirent in the new directory.4439 44402. The cached dentry no longer has a corresponding ondisk dirent in the new4441 directory and the dentry can be purged from the cache.4442 44433. The cached dentry no longer has an ondisk dirent but the dentry cannot be4444 purged.4445 This is the problem case.4446 4447Unfortunately, the current dentry cache design doesn't provide a means to walk4448every child dentry of a specific directory, which makes this a hard problem.4449There is no known solution.4450 4451The proposed patchset is the4452`directory repair4453<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-dirs>`_4454series.4455 4456Parent Pointers4457```````````````4458 4459A parent pointer is a piece of file metadata that enables a user to locate the4460file's parent directory without having to traverse the directory tree from the4461root.4462Without them, reconstruction of directory trees is hindered in much the same4463way that the historic lack of reverse space mapping information once hindered4464reconstruction of filesystem space metadata.4465The parent pointer feature, however, makes total directory reconstruction4466possible.4467 4468XFS parent pointers contain the information needed to identify the4469corresponding directory entry in the parent directory.4470In other words, child files use extended attributes to store pointers to4471parents in the form ``(dirent_name) → (parent_inum, parent_gen)``.4472The directory checking process can be strengthened to ensure that the target of4473each dirent also contains a parent pointer pointing back to the dirent.4474Likewise, each parent pointer can be checked by ensuring that the target of4475each parent pointer is a directory and that it contains a dirent matching4476the parent pointer.4477Both online and offline repair can use this strategy.4478 4479+--------------------------------------------------------------------------+4480| **Historical Sidebar**: |4481+--------------------------------------------------------------------------+4482| Directory parent pointers were first proposed as an XFS feature more |4483| than a decade ago by SGI. |4484| Each link from a parent directory to a child file is mirrored with an |4485| extended attribute in the child that could be used to identify the |4486| parent directory. |4487| Unfortunately, this early implementation had major shortcomings and was |4488| never merged into Linux XFS: |4489| |4490| 1. The XFS codebase of the late 2000s did not have the infrastructure to |4491| enforce strong referential integrity in the directory tree. |4492| It did not guarantee that a change in a forward link would always be |4493| followed up with the corresponding change to the reverse links. |4494| |4495| 2. Referential integrity was not integrated into offline repair. |4496| Checking and repairs were performed on mounted filesystems without |4497| taking any kernel or inode locks to coordinate access. |4498| It is not clear how this actually worked properly. |4499| |4500| 3. The extended attribute did not record the name of the directory entry |4501| in the parent, so the SGI parent pointer implementation cannot be |4502| used to reconnect the directory tree. |4503| |4504| 4. Extended attribute forks only support 65,536 extents, which means |4505| that parent pointer attribute creation is likely to fail at some |4506| point before the maximum file link count is achieved. |4507| |4508| The original parent pointer design was too unstable for something like |4509| a file system repair to depend on. |4510| Allison Henderson, Chandan Babu, and Catherine Hoang are working on a |4511| second implementation that solves all shortcomings of the first. |4512| During 2022, Allison introduced log intent items to track physical |4513| manipulations of the extended attribute structures. |4514| This solves the referential integrity problem by making it possible to |4515| commit a dirent update and a parent pointer update in the same |4516| transaction. |4517| Chandan increased the maximum extent counts of both data and attribute |4518| forks, thereby ensuring that the extended attribute structure can grow |4519| to handle the maximum hardlink count of any file. |4520| |4521| For this second effort, the ondisk parent pointer format as originally |4522| proposed was ``(parent_inum, parent_gen, dirent_pos) → (dirent_name)``. |4523| The format was changed during development to eliminate the requirement |4524| of repair tools needing to to ensure that the ``dirent_pos`` field |4525| always matched when reconstructing a directory. |4526| |4527| There were a few other ways to have solved that problem: |4528| |4529| 1. The field could be designated advisory, since the other three values |4530| are sufficient to find the entry in the parent. |4531| However, this makes indexed key lookup impossible while repairs are |4532| ongoing. |4533| |4534| 2. We could allow creating directory entries at specified offsets, which |4535| solves the referential integrity problem but runs the risk that |4536| dirent creation will fail due to conflicts with the free space in the |4537| directory. |4538| |4539| These conflicts could be resolved by appending the directory entry |4540| and amending the xattr code to support updating an xattr key and |4541| reindexing the dabtree, though this would have to be performed with |4542| the parent directory still locked. |4543| |4544| 3. Same as above, but remove the old parent pointer entry and add a new |4545| one atomically. |4546| |4547| 4. Change the ondisk xattr format to |4548| ``(parent_inum, name) → (parent_gen)``, which would provide the attr |4549| name uniqueness that we require, without forcing repair code to |4550| update the dirent position. |4551| Unfortunately, this requires changes to the xattr code to support |4552| attr names as long as 263 bytes. |4553| |4554| 5. Change the ondisk xattr format to ``(parent_inum, hash(name)) → |4555| (name, parent_gen)``. |4556| If the hash is sufficiently resistant to collisions (e.g. sha256) |4557| then this should provide the attr name uniqueness that we require. |4558| Names shorter than 247 bytes could be stored directly. |4559| |4560| 6. Change the ondisk xattr format to ``(dirent_name) → (parent_ino, |4561| parent_gen)``. This format doesn't require any of the complicated |4562| nested name hashing of the previous suggestions. However, it was |4563| discovered that multiple hardlinks to the same inode with the same |4564| filename caused performance problems with hashed xattr lookups, so |4565| the parent inumber is now xor'd into the hash index. |4566| |4567| In the end, it was decided that solution #6 was the most compact and the |4568| most performant. A new hash function was designed for parent pointers. |4569+--------------------------------------------------------------------------+4570 4571 4572Case Study: Repairing Directories with Parent Pointers4573^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^4574 4575Directory rebuilding uses a :ref:`coordinated inode scan <iscan>` and4576a :ref:`directory entry live update hook <liveupdate>` as follows:4577 45781. Set up a temporary directory for generating the new directory structure,4579 an xfblob for storing entry names, and an xfarray for stashing the fixed4580 size fields involved in a directory update: ``(child inumber, add vs.4581 remove, name cookie, ftype)``.4582 45832. Set up an inode scanner and hook into the directory entry code to receive4584 updates on directory operations.4585 45863. For each parent pointer found in each file scanned, decide if the parent4587 pointer references the directory of interest.4588 If so:4589 4590 a. Stash the parent pointer name and an addname entry for this dirent in the4591 xfblob and xfarray, respectively.4592 4593 b. When finished scanning that file or the kernel memory consumption exceeds4594 a threshold, flush the stashed updates to the temporary directory.4595 45964. For each live directory update received via the hook, decide if the child4597 has already been scanned.4598 If so:4599 4600 a. Stash the parent pointer name an addname or removename entry for this4601 dirent update in the xfblob and xfarray for later.4602 We cannot write directly to the temporary directory because hook4603 functions are not allowed to modify filesystem metadata.4604 Instead, we stash updates in the xfarray and rely on the scanner thread4605 to apply the stashed updates to the temporary directory.4606 46075. When the scan is complete, replay any stashed entries in the xfarray.4608 46096. When the scan is complete, atomically exchange the contents of the temporary4610 directory and the directory being repaired.4611 The temporary directory now contains the damaged directory structure.4612 46137. Reap the temporary directory.4614 4615The proposed patchset is the4616`parent pointers directory repair4617<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=pptrs-fsck>`_4618series.4619 4620Case Study: Repairing Parent Pointers4621^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^4622 4623Online reconstruction of a file's parent pointer information works similarly to4624directory reconstruction:4625 46261. Set up a temporary file for generating a new extended attribute structure,4627 an xfblob for storing parent pointer names, and an xfarray for stashing the4628 fixed size fields involved in a parent pointer update: ``(parent inumber,4629 parent generation, add vs. remove, name cookie)``.4630 46312. Set up an inode scanner and hook into the directory entry code to receive4632 updates on directory operations.4633 46343. For each directory entry found in each directory scanned, decide if the4635 dirent references the file of interest.4636 If so:4637 4638 a. Stash the dirent name and an addpptr entry for this parent pointer in the4639 xfblob and xfarray, respectively.4640 4641 b. When finished scanning the directory or the kernel memory consumption4642 exceeds a threshold, flush the stashed updates to the temporary file.4643 46444. For each live directory update received via the hook, decide if the parent4645 has already been scanned.4646 If so:4647 4648 a. Stash the dirent name and an addpptr or removepptr entry for this dirent4649 update in the xfblob and xfarray for later.4650 We cannot write parent pointers directly to the temporary file because4651 hook functions are not allowed to modify filesystem metadata.4652 Instead, we stash updates in the xfarray and rely on the scanner thread4653 to apply the stashed parent pointer updates to the temporary file.4654 46555. When the scan is complete, replay any stashed entries in the xfarray.4656 46576. Copy all non-parent pointer extended attributes to the temporary file.4658 46597. When the scan is complete, atomically exchange the mappings of the attribute4660 forks of the temporary file and the file being repaired.4661 The temporary file now contains the damaged extended attribute structure.4662 46638. Reap the temporary file.4664 4665The proposed patchset is the4666`parent pointers repair4667<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=pptrs-fsck>`_4668series.4669 4670Digression: Offline Checking of Parent Pointers4671^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^4672 4673Examining parent pointers in offline repair works differently because corrupt4674files are erased long before directory tree connectivity checks are performed.4675Parent pointer checks are therefore a second pass to be added to the existing4676connectivity checks:4677 46781. After the set of surviving files has been established (phase 6),4679 walk the surviving directories of each AG in the filesystem.4680 This is already performed as part of the connectivity checks.4681 46822. For each directory entry found,4683 4684 a. If the name has already been stored in the xfblob, then use that cookie4685 and skip the next step.4686 4687 b. Otherwise, record the name in an xfblob, and remember the xfblob cookie.4688 Unique mappings are critical for4689 4690 1. Deduplicating names to reduce memory usage, and4691 4692 2. Creating a stable sort key for the parent pointer indexes so that the4693 parent pointer validation described below will work.4694 4695 c. Store ``(child_ag_inum, parent_inum, parent_gen, name_hash, name_len,4696 name_cookie)`` tuples in a per-AG in-memory slab. The ``name_hash``4697 referenced in this section is the regular directory entry name hash, not4698 the specialized one used for parent pointer xattrs.4699 47003. For each AG in the filesystem,4701 4702 a. Sort the per-AG tuple set in order of ``child_ag_inum``, ``parent_inum``,4703 ``name_hash``, and ``name_cookie``.4704 Having a single ``name_cookie`` for each ``name`` is critical for4705 handling the uncommon case of a directory containing multiple hardlinks4706 to the same file where all the names hash to the same value.4707 4708 b. For each inode in the AG,4709 4710 1. Scan the inode for parent pointers.4711 For each parent pointer found,4712 4713 a. Validate the ondisk parent pointer.4714 If validation fails, move on to the next parent pointer in the4715 file.4716 4717 b. If the name has already been stored in the xfblob, then use that4718 cookie and skip the next step.4719 4720 c. Record the name in a per-file xfblob, and remember the xfblob4721 cookie.4722 4723 d. Store ``(parent_inum, parent_gen, name_hash, name_len,4724 name_cookie)`` tuples in a per-file slab.4725 4726 2. Sort the per-file tuples in order of ``parent_inum``, ``name_hash``,4727 and ``name_cookie``.4728 4729 3. Position one slab cursor at the start of the inode's records in the4730 per-AG tuple slab.4731 This should be trivial since the per-AG tuples are in child inumber4732 order.4733 4734 4. Position a second slab cursor at the start of the per-file tuple slab.4735 4736 5. Iterate the two cursors in lockstep, comparing the ``parent_ino``,4737 ``name_hash``, and ``name_cookie`` fields of the records under each4738 cursor:4739 4740 a. If the per-AG cursor is at a lower point in the keyspace than the4741 per-file cursor, then the per-AG cursor points to a missing parent4742 pointer.4743 Add the parent pointer to the inode and advance the per-AG4744 cursor.4745 4746 b. If the per-file cursor is at a lower point in the keyspace than4747 the per-AG cursor, then the per-file cursor points to a dangling4748 parent pointer.4749 Remove the parent pointer from the inode and advance the per-file4750 cursor.4751 4752 c. Otherwise, both cursors point at the same parent pointer.4753 Update the parent_gen component if necessary.4754 Advance both cursors.4755 47564. Move on to examining link counts, as we do today.4757 4758The proposed patchset is the4759`offline parent pointers repair4760<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfsprogs-dev.git/log/?h=pptrs-fsck>`_4761series.4762 4763Rebuilding directories from parent pointers in offline repair would be very4764challenging because xfs_repair currently uses two single-pass scans of the4765filesystem during phases 3 and 4 to decide which files are corrupt enough to be4766zapped.4767This scan would have to be converted into a multi-pass scan:4768 47691. The first pass of the scan zaps corrupt inodes, forks, and attributes4770 much as it does now.4771 Corrupt directories are noted but not zapped.4772 47732. The next pass records parent pointers pointing to the directories noted4774 as being corrupt in the first pass.4775 This second pass may have to happen after the phase 4 scan for duplicate4776 blocks, if phase 4 is also capable of zapping directories.4777 47783. The third pass resets corrupt directories to an empty shortform directory.4779 Free space metadata has not been ensured yet, so repair cannot yet use the4780 directory building code in libxfs.4781 47824. At the start of phase 6, space metadata have been rebuilt.4783 Use the parent pointer information recorded during step 2 to reconstruct4784 the dirents and add them to the now-empty directories.4785 4786This code has not yet been constructed.4787 4788.. _dirtree:4789 4790Case Study: Directory Tree Structure4791^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^4792 4793As mentioned earlier, the filesystem directory tree is supposed to be a4794directed acylic graph structure.4795However, each node in this graph is a separate ``xfs_inode`` object with its4796own locks, which makes validating the tree qualities difficult.4797Fortunately, non-directories are allowed to have multiple parents and cannot4798have children, so only directories need to be scanned.4799Directories typically constitute 5-10% of the files in a filesystem, which4800reduces the amount of work dramatically.4801 4802If the directory tree could be frozen, it would be easy to discover cycles and4803disconnected regions by running a depth (or breadth) first search downwards4804from the root directory and marking a bitmap for each directory found.4805At any point in the walk, trying to set an already set bit means there is a4806cycle.4807After the scan completes, XORing the marked inode bitmap with the inode4808allocation bitmap reveals disconnected inodes.4809However, one of online repair's design goals is to avoid locking the entire4810filesystem unless it's absolutely necessary.4811Directory tree updates can move subtrees across the scanner wavefront on a live4812filesystem, so the bitmap algorithm cannot be applied.4813 4814Directory parent pointers enable an incremental approach to validation of the4815tree structure.4816Instead of using one thread to scan the entire filesystem, multiple threads can4817walk from individual subdirectories upwards towards the root.4818For this to work, all directory entries and parent pointers must be internally4819consistent, each directory entry must have a parent pointer, and the link4820counts of all directories must be correct.4821Each scanner thread must be able to take the IOLOCK of an alleged parent4822directory while holding the IOLOCK of the child directory to prevent either4823directory from being moved within the tree.4824This is not possible since the VFS does not take the IOLOCK of a child4825subdirectory when moving that subdirectory, so instead the scanner stabilizes4826the parent -> child relationship by taking the ILOCKs and installing a dirent4827update hook to detect changes.4828 4829The scanning process uses a dirent hook to detect changes to the directories4830mentioned in the scan data.4831The scan works as follows:4832 48331. For each subdirectory in the filesystem,4834 4835 a. For each parent pointer of that subdirectory,4836 4837 1. Create a path object for that parent pointer, and mark the4838 subdirectory inode number in the path object's bitmap.4839 4840 2. Record the parent pointer name and inode number in a path structure.4841 4842 3. If the alleged parent is the subdirectory being scrubbed, the path is4843 a cycle.4844 Mark the path for deletion and repeat step 1a with the next4845 subdirectory parent pointer.4846 4847 4. Try to mark the alleged parent inode number in a bitmap in the path4848 object.4849 If the bit is already set, then there is a cycle in the directory4850 tree.4851 Mark the path as a cycle and repeat step 1a with the next subdirectory4852 parent pointer.4853 4854 5. Load the alleged parent.4855 If the alleged parent is not a linked directory, abort the scan4856 because the parent pointer information is inconsistent.4857 4858 6. For each parent pointer of this alleged ancestor directory,4859 4860 a. Record the parent pointer name and inode number in the path object4861 if no parent has been set for that level.4862 4863 b. If an ancestor has more than one parent, mark the path as corrupt.4864 Repeat step 1a with the next subdirectory parent pointer.4865 4866 c. Repeat steps 1a3-1a6 for the ancestor identified in step 1a6a.4867 This repeats until the directory tree root is reached or no parents4868 are found.4869 4870 7. If the walk terminates at the root directory, mark the path as ok.4871 4872 8. If the walk terminates without reaching the root, mark the path as4873 disconnected.4874 48752. If the directory entry update hook triggers, check all paths already found4876 by the scan.4877 If the entry matches part of a path, mark that path and the scan stale.4878 When the scanner thread sees that the scan has been marked stale, it deletes4879 all scan data and starts over.4880 4881Repairing the directory tree works as follows:4882 48831. Walk each path of the target subdirectory.4884 4885 a. Corrupt paths and cycle paths are counted as suspect.4886 4887 b. Paths already marked for deletion are counted as bad.4888 4889 c. Paths that reached the root are counted as good.4890 48912. If the subdirectory is either the root directory or has zero link count,4892 delete all incoming directory entries in the immediate parents.4893 Repairs are complete.4894 48953. If the subdirectory has exactly one path, set the dotdot entry to the4896 parent and exit.4897 48984. If the subdirectory has at least one good path, delete all the other4899 incoming directory entries in the immediate parents.4900 49015. If the subdirectory has no good paths and more than one suspect path, delete4902 all the other incoming directory entries in the immediate parents.4903 49046. If the subdirectory has zero paths, attach it to the lost and found.4905 4906The proposed patches are in the4907`directory tree repair4908<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=scrub-directory-tree>`_4909series.4910 4911 4912.. _orphanage:4913 4914The Orphanage4915-------------4916 4917Filesystems present files as a directed, and hopefully acyclic, graph.4918In other words, a tree.4919The root of the filesystem is a directory, and each entry in a directory points4920downwards either to more subdirectories or to non-directory files.4921Unfortunately, a disruption in the directory graph pointers result in a4922disconnected graph, which makes files impossible to access via regular path4923resolution.4924 4925Without parent pointers, the directory parent pointer online scrub code can4926detect a dotdot entry pointing to a parent directory that doesn't have a link4927back to the child directory and the file link count checker can detect a file4928that isn't pointed to by any directory in the filesystem.4929If such a file has a positive link count, the file is an orphan.4930 4931With parent pointers, directories can be rebuilt by scanning parent pointers4932and parent pointers can be rebuilt by scanning directories.4933This should reduce the incidence of files ending up in ``/lost+found``.4934 4935When orphans are found, they should be reconnected to the directory tree.4936Offline fsck solves the problem by creating a directory ``/lost+found`` to4937serve as an orphanage, and linking orphan files into the orphanage by using the4938inumber as the name.4939Reparenting a file to the orphanage does not reset any of its permissions or4940ACLs.4941 4942This process is more involved in the kernel than it is in userspace.4943The directory and file link count repair setup functions must use the regular4944VFS mechanisms to create the orphanage directory with all the necessary4945security attributes and dentry cache entries, just like a regular directory4946tree modification.4947 4948Orphaned files are adopted by the orphanage as follows:4949 49501. Call ``xrep_orphanage_try_create`` at the start of the scrub setup function4951 to try to ensure that the lost and found directory actually exists.4952 This also attaches the orphanage directory to the scrub context.4953 49542. If the decision is made to reconnect a file, take the IOLOCK of both the4955 orphanage and the file being reattached.4956 The ``xrep_orphanage_iolock_two`` function follows the inode locking4957 strategy discussed earlier.4958 49593. Use ``xrep_adoption_trans_alloc`` to reserve resources to the repair4960 transaction.4961 49624. Call ``xrep_orphanage_compute_name`` to compute the new name in the4963 orphanage.4964 49655. If the adoption is going to happen, call ``xrep_adoption_reparent`` to4966 reparent the orphaned file into the lost and found and invalidate the dentry4967 cache.4968 49696. Call ``xrep_adoption_finish`` to commit any filesystem updates, release the4970 orphanage ILOCK, and clean the scrub transaction. Call4971 ``xrep_adoption_commit`` to commit the updates and the scrub transaction.4972 49737. If a runtime error happens, call ``xrep_adoption_cancel`` to release all4974 resources.4975 4976The proposed patches are in the4977`orphanage adoption4978<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=repair-orphanage>`_4979series.4980 49816. Userspace Algorithms and Data Structures4982===========================================4983 4984This section discusses the key algorithms and data structures of the userspace4985program, ``xfs_scrub``, that provide the ability to drive metadata checks and4986repairs in the kernel, verify file data, and look for other potential problems.4987 4988.. _scrubcheck:4989 4990Checking Metadata4991-----------------4992 4993Recall the :ref:`phases of fsck work<scrubphases>` outlined earlier.4994That structure follows naturally from the data dependencies designed into the4995filesystem from its beginnings in 1993.4996In XFS, there are several groups of metadata dependencies:4997 4998a. Filesystem summary counts depend on consistency within the inode indices,4999 the allocation group space btrees, and the realtime volume space5000 information.5001 5002b. Quota resource counts depend on consistency within the quota file data5003 forks, inode indices, inode records, and the forks of every file on the5004 system.5005 5006c. The naming hierarchy depends on consistency within the directory and5007 extended attribute structures.5008 This includes file link counts.5009 5010d. Directories, extended attributes, and file data depend on consistency within5011 the file forks that map directory and extended attribute data to physical5012 storage media.5013 5014e. The file forks depends on consistency within inode records and the space5015 metadata indices of the allocation groups and the realtime volume.5016 This includes quota and realtime metadata files.5017 5018f. Inode records depends on consistency within the inode metadata indices.5019 5020g. Realtime space metadata depend on the inode records and data forks of the5021 realtime metadata inodes.5022 5023h. The allocation group metadata indices (free space, inodes, reference count,5024 and reverse mapping btrees) depend on consistency within the AG headers and5025 between all the AG metadata btrees.5026 5027i. ``xfs_scrub`` depends on the filesystem being mounted and kernel support5028 for online fsck functionality.5029 5030Therefore, a metadata dependency graph is a convenient way to schedule checking5031operations in the ``xfs_scrub`` program:5032 5033- Phase 1 checks that the provided path maps to an XFS filesystem and detect5034 the kernel's scrubbing abilities, which validates group (i).5035 5036- Phase 2 scrubs groups (g) and (h) in parallel using a threaded workqueue.5037 5038- Phase 3 scans inodes in parallel.5039 For each inode, groups (f), (e), and (d) are checked, in that order.5040 5041- Phase 4 repairs everything in groups (i) through (d) so that phases 5 and 65042 may run reliably.5043 5044- Phase 5 starts by checking groups (b) and (c) in parallel before moving on5045 to checking names.5046 5047- Phase 6 depends on groups (i) through (b) to find file data blocks to verify,5048 to read them, and to report which blocks of which files are affected.5049 5050- Phase 7 checks group (a), having validated everything else.5051 5052Notice that the data dependencies between groups are enforced by the structure5053of the program flow.5054 5055Parallel Inode Scans5056--------------------5057 5058An XFS filesystem can easily contain hundreds of millions of inodes.5059Given that XFS targets installations with large high-performance storage,5060it is desirable to scrub inodes in parallel to minimize runtime, particularly5061if the program has been invoked manually from a command line.5062This requires careful scheduling to keep the threads as evenly loaded as5063possible.5064 5065Early iterations of the ``xfs_scrub`` inode scanner naïvely created a single5066workqueue and scheduled a single workqueue item per AG.5067Each workqueue item walked the inode btree (with ``XFS_IOC_INUMBERS``) to find5068inode chunks and then called bulkstat (``XFS_IOC_BULKSTAT``) to gather enough5069information to construct file handles.5070The file handle was then passed to a function to generate scrub items for each5071metadata object of each inode.5072This simple algorithm leads to thread balancing problems in phase 3 if the5073filesystem contains one AG with a few large sparse files and the rest of the5074AGs contain many smaller files.5075The inode scan dispatch function was not sufficiently granular; it should have5076been dispatching at the level of individual inodes, or, to constrain memory5077consumption, inode btree records.5078 5079Thanks to Dave Chinner, bounded workqueues in userspace enable ``xfs_scrub`` to5080avoid this problem with ease by adding a second workqueue.5081Just like before, the first workqueue is seeded with one workqueue item per AG,5082and it uses INUMBERS to find inode btree chunks.5083The second workqueue, however, is configured with an upper bound on the number5084of items that can be waiting to be run.5085Each inode btree chunk found by the first workqueue's workers are queued to the5086second workqueue, and it is this second workqueue that queries BULKSTAT,5087creates a file handle, and passes it to a function to generate scrub items for5088each metadata object of each inode.5089If the second workqueue is too full, the workqueue add function blocks the5090first workqueue's workers until the backlog eases.5091This doesn't completely solve the balancing problem, but reduces it enough to5092move on to more pressing issues.5093 5094The proposed patchsets are the scrub5095`performance tweaks5096<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfsprogs-dev.git/log/?h=scrub-performance-tweaks>`_5097and the5098`inode scan rebalance5099<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfsprogs-dev.git/log/?h=scrub-iscan-rebalance>`_5100series.5101 5102.. _scrubrepair:5103 5104Scheduling Repairs5105------------------5106 5107During phase 2, corruptions and inconsistencies reported in any AGI header or5108inode btree are repaired immediately, because phase 3 relies on proper5109functioning of the inode indices to find inodes to scan.5110Failed repairs are rescheduled to phase 4.5111Problems reported in any other space metadata are deferred to phase 4.5112Optimization opportunities are always deferred to phase 4, no matter their5113origin.5114 5115During phase 3, corruptions and inconsistencies reported in any part of a5116file's metadata are repaired immediately if all space metadata were validated5117during phase 2.5118Repairs that fail or cannot be repaired immediately are scheduled for phase 4.5119 5120In the original design of ``xfs_scrub``, it was thought that repairs would be5121so infrequent that the ``struct xfs_scrub_metadata`` objects used to5122communicate with the kernel could also be used as the primary object to5123schedule repairs.5124With recent increases in the number of optimizations possible for a given5125filesystem object, it became much more memory-efficient to track all eligible5126repairs for a given filesystem object with a single repair item.5127Each repair item represents a single lockable object -- AGs, metadata files,5128individual inodes, or a class of summary information.5129 5130Phase 4 is responsible for scheduling a lot of repair work in as quick a5131manner as is practical.5132The :ref:`data dependencies <scrubcheck>` outlined earlier still apply, which5133means that ``xfs_scrub`` must try to complete the repair work scheduled by5134phase 2 before trying repair work scheduled by phase 3.5135The repair process is as follows:5136 51371. Start a round of repair with a workqueue and enough workers to keep the CPUs5138 as busy as the user desires.5139 5140 a. For each repair item queued by phase 2,5141 5142 i. Ask the kernel to repair everything listed in the repair item for a5143 given filesystem object.5144 5145 ii. Make a note if the kernel made any progress in reducing the number5146 of repairs needed for this object.5147 5148 iii. If the object no longer requires repairs, revalidate all metadata5149 associated with this object.5150 If the revalidation succeeds, drop the repair item.5151 If not, requeue the item for more repairs.5152 5153 b. If any repairs were made, jump back to 1a to retry all the phase 2 items.5154 5155 c. For each repair item queued by phase 3,5156 5157 i. Ask the kernel to repair everything listed in the repair item for a5158 given filesystem object.5159 5160 ii. Make a note if the kernel made any progress in reducing the number5161 of repairs needed for this object.5162 5163 iii. If the object no longer requires repairs, revalidate all metadata5164 associated with this object.5165 If the revalidation succeeds, drop the repair item.5166 If not, requeue the item for more repairs.5167 5168 d. If any repairs were made, jump back to 1c to retry all the phase 3 items.5169 51702. If step 1 made any repair progress of any kind, jump back to step 1 to start5171 another round of repair.5172 51733. If there are items left to repair, run them all serially one more time.5174 Complain if the repairs were not successful, since this is the last chance5175 to repair anything.5176 5177Corruptions and inconsistencies encountered during phases 5 and 7 are repaired5178immediately.5179Corrupt file data blocks reported by phase 6 cannot be recovered by the5180filesystem.5181 5182The proposed patchsets are the5183`repair warning improvements5184<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfsprogs-dev.git/log/?h=scrub-better-repair-warnings>`_,5185refactoring of the5186`repair data dependency5187<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfsprogs-dev.git/log/?h=scrub-repair-data-deps>`_5188and5189`object tracking5190<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfsprogs-dev.git/log/?h=scrub-object-tracking>`_,5191and the5192`repair scheduling5193<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfsprogs-dev.git/log/?h=scrub-repair-scheduling>`_5194improvement series.5195 5196Checking Names for Confusable Unicode Sequences5197-----------------------------------------------5198 5199If ``xfs_scrub`` succeeds in validating the filesystem metadata by the end of5200phase 4, it moves on to phase 5, which checks for suspicious looking names in5201the filesystem.5202These names consist of the filesystem label, names in directory entries, and5203the names of extended attributes.5204Like most Unix filesystems, XFS imposes the sparest of constraints on the5205contents of a name:5206 5207- Slashes and null bytes are not allowed in directory entries.5208 5209- Null bytes are not allowed in userspace-visible extended attributes.5210 5211- Null bytes are not allowed in the filesystem label.5212 5213Directory entries and attribute keys store the length of the name explicitly5214ondisk, which means that nulls are not name terminators.5215For this section, the term "naming domain" refers to any place where names are5216presented together -- all the names in a directory, or all the attributes of a5217file.5218 5219Although the Unix naming constraints are very permissive, the reality of most5220modern-day Linux systems is that programs work with Unicode character code5221points to support international languages.5222These programs typically encode those code points in UTF-8 when interfacing5223with the C library because the kernel expects null-terminated names.5224In the common case, therefore, names found in an XFS filesystem are actually5225UTF-8 encoded Unicode data.5226 5227To maximize its expressiveness, the Unicode standard defines separate control5228points for various characters that render similarly or identically in writing5229systems around the world.5230For example, the character "Cyrillic Small Letter A" U+0430 "а" often renders5231identically to "Latin Small Letter A" U+0061 "a".5232 5233The standard also permits characters to be constructed in multiple ways --5234either by using a defined code point, or by combining one code point with5235various combining marks.5236For example, the character "Angstrom Sign U+212B "Å" can also be expressed5237as "Latin Capital Letter A" U+0041 "A" followed by "Combining Ring Above"5238U+030A "◌̊".5239Both sequences render identically.5240 5241Like the standards that preceded it, Unicode also defines various control5242characters to alter the presentation of text.5243For example, the character "Right-to-Left Override" U+202E can trick some5244programs into rendering "moo\\xe2\\x80\\xaegnp.txt" as "mootxt.png".5245A second category of rendering problems involves whitespace characters.5246If the character "Zero Width Space" U+200B is encountered in a file name, the5247name will render identically to a name that does not have the zero width5248space.5249 5250If two names within a naming domain have different byte sequences but render5251identically, a user may be confused by it.5252The kernel, in its indifference to upper level encoding schemes, permits this.5253Most filesystem drivers persist the byte sequence names that are given to them5254by the VFS.5255 5256Techniques for detecting confusable names are explained in great detail in5257sections 4 and 5 of the5258`Unicode Security Mechanisms <https://unicode.org/reports/tr39/>`_5259document.5260When ``xfs_scrub`` detects UTF-8 encoding in use on a system, it uses the5261Unicode normalization form NFD in conjunction with the confusable name5262detection component of5263`libicu <https://github.com/unicode-org/icu>`_5264to identify names with a directory or within a file's extended attributes that5265could be confused for each other.5266Names are also checked for control characters, non-rendering characters, and5267mixing of bidirectional characters.5268All of these potential issues are reported to the system administrator during5269phase 5.5270 5271Media Verification of File Data Extents5272---------------------------------------5273 5274The system administrator can elect to initiate a media scan of all file data5275blocks.5276This scan after validation of all filesystem metadata (except for the summary5277counters) as phase 6.5278The scan starts by calling ``FS_IOC_GETFSMAP`` to scan the filesystem space map5279to find areas that are allocated to file data fork extents.5280Gaps between data fork extents that are smaller than 64k are treated as if5281they were data fork extents to reduce the command setup overhead.5282When the space map scan accumulates a region larger than 32MB, a media5283verification request is sent to the disk as a directio read of the raw block5284device.5285 5286If the verification read fails, ``xfs_scrub`` retries with single-block reads5287to narrow down the failure to the specific region of the media and recorded.5288When it has finished issuing verification requests, it again uses the space5289mapping ioctl to map the recorded media errors back to metadata structures5290and report what has been lost.5291For media errors in blocks owned by files, parent pointers can be used to5292construct file paths from inode numbers for user-friendly reporting.5293 52947. Conclusion and Future Work5295=============================5296 5297It is hoped that the reader of this document has followed the designs laid out5298in this document and now has some familiarity with how XFS performs online5299rebuilding of its metadata indices, and how filesystem users can interact with5300that functionality.5301Although the scope of this work is daunting, it is hoped that this guide will5302make it easier for code readers to understand what has been built, for whom it5303has been built, and why.5304Please feel free to contact the XFS mailing list with questions.5305 5306XFS_IOC_EXCHANGE_RANGE5307----------------------5308 5309As discussed earlier, a second frontend to the atomic file mapping exchange5310mechanism is a new ioctl call that userspace programs can use to commit updates5311to files atomically.5312This frontend has been out for review for several years now, though the5313necessary refinements to online repair and lack of customer demand mean that5314the proposal has not been pushed very hard.5315 5316File Content Exchanges with Regular User Files5317``````````````````````````````````````````````5318 5319As mentioned earlier, XFS has long had the ability to swap extents between5320files, which is used almost exclusively by ``xfs_fsr`` to defragment files.5321The earliest form of this was the fork swap mechanism, where the entire5322contents of data forks could be exchanged between two files by exchanging the5323raw bytes in each inode fork's immediate area.5324When XFS v5 came along with self-describing metadata, this old mechanism grew5325some log support to continue rewriting the owner fields of BMBT blocks during5326log recovery.5327When the reverse mapping btree was later added to XFS, the only way to maintain5328the consistency of the fork mappings with the reverse mapping index was to5329develop an iterative mechanism that used deferred bmap and rmap operations to5330swap mappings one at a time.5331This mechanism is identical to steps 2-3 from the procedure above except for5332the new tracking items, because the atomic file mapping exchange mechanism is5333an iteration of an existing mechanism and not something totally novel.5334For the narrow case of file defragmentation, the file contents must be5335identical, so the recovery guarantees are not much of a gain.5336 5337Atomic file content exchanges are much more flexible than the existing swapext5338implementations because it can guarantee that the caller never sees a mix of5339old and new contents even after a crash, and it can operate on two arbitrary5340file fork ranges.5341The extra flexibility enables several new use cases:5342 5343- **Atomic commit of file writes**: A userspace process opens a file that it5344 wants to update.5345 Next, it opens a temporary file and calls the file clone operation to reflink5346 the first file's contents into the temporary file.5347 Writes to the original file should instead be written to the temporary file.5348 Finally, the process calls the atomic file mapping exchange system call5349 (``XFS_IOC_EXCHANGE_RANGE``) to exchange the file contents, thereby5350 committing all of the updates to the original file, or none of them.5351 5352.. _exchrange_if_unchanged:5353 5354- **Transactional file updates**: The same mechanism as above, but the caller5355 only wants the commit to occur if the original file's contents have not5356 changed.5357 To make this happen, the calling process snapshots the file modification and5358 change timestamps of the original file before reflinking its data to the5359 temporary file.5360 When the program is ready to commit the changes, it passes the timestamps5361 into the kernel as arguments to the atomic file mapping exchange system call.5362 The kernel only commits the changes if the provided timestamps match the5363 original file.5364 A new ioctl (``XFS_IOC_COMMIT_RANGE``) is provided to perform this.5365 5366- **Emulation of atomic block device writes**: Export a block device with a5367 logical sector size matching the filesystem block size to force all writes5368 to be aligned to the filesystem block size.5369 Stage all writes to a temporary file, and when that is complete, call the5370 atomic file mapping exchange system call with a flag to indicate that holes5371 in the temporary file should be ignored.5372 This emulates an atomic device write in software, and can support arbitrary5373 scattered writes.5374 5375Vectorized Scrub5376----------------5377 5378As it turns out, the :ref:`refactoring <scrubrepair>` of repair items mentioned5379earlier was a catalyst for enabling a vectorized scrub system call.5380Since 2018, the cost of making a kernel call has increased considerably on some5381systems to mitigate the effects of speculative execution attacks.5382This incentivizes program authors to make as few system calls as possible to5383reduce the number of times an execution path crosses a security boundary.5384 5385With vectorized scrub, userspace pushes to the kernel the identity of a5386filesystem object, a list of scrub types to run against that object, and a5387simple representation of the data dependencies between the selected scrub5388types.5389The kernel executes as much of the caller's plan as it can until it hits a5390dependency that cannot be satisfied due to a corruption, and tells userspace5391how much was accomplished.5392It is hoped that ``io_uring`` will pick up enough of this functionality that5393online fsck can use that instead of adding a separate vectored scrub system5394call to XFS.5395 5396The relevant patchsets are the5397`kernel vectorized scrub5398<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=vectorized-scrub>`_5399and5400`userspace vectorized scrub5401<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfsprogs-dev.git/log/?h=vectorized-scrub>`_5402series.5403 5404Quality of Service Targets for Scrub5405------------------------------------5406 5407One serious shortcoming of the online fsck code is that the amount of time that5408it can spend in the kernel holding resource locks is basically unbounded.5409Userspace is allowed to send a fatal signal to the process which will cause5410``xfs_scrub`` to exit when it reaches a good stopping point, but there's no way5411for userspace to provide a time budget to the kernel.5412Given that the scrub codebase has helpers to detect fatal signals, it shouldn't5413be too much work to allow userspace to specify a timeout for a scrub/repair5414operation and abort the operation if it exceeds budget.5415However, most repair functions have the property that once they begin to touch5416ondisk metadata, the operation cannot be cancelled cleanly, after which a QoS5417timeout is no longer useful.5418 5419Defragmenting Free Space5420------------------------5421 5422Over the years, many XFS users have requested the creation of a program to5423clear a portion of the physical storage underlying a filesystem so that it5424becomes a contiguous chunk of free space.5425Call this free space defragmenter ``clearspace`` for short.5426 5427The first piece the ``clearspace`` program needs is the ability to read the5428reverse mapping index from userspace.5429This already exists in the form of the ``FS_IOC_GETFSMAP`` ioctl.5430The second piece it needs is a new fallocate mode5431(``FALLOC_FL_MAP_FREE_SPACE``) that allocates the free space in a region and5432maps it to a file.5433Call this file the "space collector" file.5434The third piece is the ability to force an online repair.5435 5436To clear all the metadata out of a portion of physical storage, clearspace5437uses the new fallocate map-freespace call to map any free space in that region5438to the space collector file.5439Next, clearspace finds all metadata blocks in that region by way of5440``GETFSMAP`` and issues forced repair requests on the data structure.5441This often results in the metadata being rebuilt somewhere that is not being5442cleared.5443After each relocation, clearspace calls the "map free space" function again to5444collect any newly freed space in the region being cleared.5445 5446To clear all the file data out of a portion of the physical storage, clearspace5447uses the FSMAP information to find relevant file data blocks.5448Having identified a good target, it uses the ``FICLONERANGE`` call on that part5449of the file to try to share the physical space with a dummy file.5450Cloning the extent means that the original owners cannot overwrite the5451contents; any changes will be written somewhere else via copy-on-write.5452Clearspace makes its own copy of the frozen extent in an area that is not being5453cleared, and uses ``FIEDEUPRANGE`` (or the :ref:`atomic file content exchanges5454<exchrange_if_unchanged>` feature) to change the target file's data extent5455mapping away from the area being cleared.5456When all other mappings have been moved, clearspace reflinks the space into the5457space collector file so that it becomes unavailable.5458 5459There are further optimizations that could apply to the above algorithm.5460To clear a piece of physical storage that has a high sharing factor, it is5461strongly desirable to retain this sharing factor.5462In fact, these extents should be moved first to maximize sharing factor after5463the operation completes.5464To make this work smoothly, clearspace needs a new ioctl5465(``FS_IOC_GETREFCOUNTS``) to report reference count information to userspace.5466With the refcount information exposed, clearspace can quickly find the longest,5467most shared data extents in the filesystem, and target them first.5468 5469**Future Work Question**: How might the filesystem move inode chunks?5470 5471*Answer*: To move inode chunks, Dave Chinner constructed a prototype program5472that creates a new file with the old contents and then locklessly runs around5473the filesystem updating directory entries.5474The operation cannot complete if the filesystem goes down.5475That problem isn't totally insurmountable: create an inode remapping table5476hidden behind a jump label, and a log item that tracks the kernel walking the5477filesystem to update directory entries.5478The trouble is, the kernel can't do anything about open files, since it cannot5479revoke them.5480 5481**Future Work Question**: Can static keys be used to minimize the cost of5482supporting ``revoke()`` on XFS files?5483 5484*Answer*: Yes.5485Until the first revocation, the bailout code need not be in the call path at5486all.5487 5488The relevant patchsets are the5489`kernel freespace defrag5490<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/log/?h=defrag-freespace>`_5491and5492`userspace freespace defrag5493<https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfsprogs-dev.git/log/?h=defrag-freespace>`_5494series.5495 5496Shrinking Filesystems5497---------------------5498 5499Removing the end of the filesystem ought to be a simple matter of evacuating5500the data and metadata at the end of the filesystem, and handing the freed space5501to the shrink code.5502That requires an evacuation of the space at end of the filesystem, which is a5503use of free space defragmentation!5504