- 20 Sep, 2017 40 commits
-
-
Brian Foster authored
commit 2dd3d709 upstream. The owner change bmbt scan that occurs during extent swap operations does not handle ordered buffer failures. Buffers that cannot be marked ordered must be physically logged so previously dirty ranges of the buffer can be relogged in the transaction. Since the bmbt scan may need to process and potentially log a large number of blocks, we can't expect to complete this operation in a single transaction. Update extent swap to use a permanent transaction with enough log reservation to physically log a buffer. Update the bmbt scan to physically log any buffers that cannot be ordered and to terminate the scan with -EAGAIN. On -EAGAIN, the caller rolls the transaction and restarts the scan. Finally, update the bmbt scan helper function to skip bmbt blocks that already match the expected owner so they are not reprocessed after scan restarts. Signed-off-by:
Brian Foster <bfoster@redhat.com> Reviewed-by:
Darrick J. Wong <darrick.wong@oracle.com> Reviewed-by:
Christoph Hellwig <hch@lst.de> [darrick: fix the xfs_trans_roll call] Signed-off-by:
Greg Kroah-Hartman <gregkh@linuxfoundation.org>
-
Brian Foster authored
commit a5814bce upstream. Ordered buffers are used in situations where the buffer is not physically logged but must pass through the transaction/logging pipeline for a particular transaction. As a result, ordered buffers are not unpinned and written back until the transaction commits to the log. Ordered buffers have a strict requirement that the target buffer must not be currently dirty and resident in the log pipeline at the time it is marked ordered. If a dirty+ordered buffer is committed, the buffer is reinserted to the AIL but not physically relogged at the LSN of the associated checkpoint. The buffer log item is assigned the LSN of the latest checkpoint and the AIL effectively releases the previously logged buffer content from the active log before the buffer has been written back. If the tail pushes forward and a filesystem crash occurs while in this state, an inconsistent filesystem could result. It is currently the caller responsibility to ensure an ordered buffer is not already dirty from a previous modification. This is unclear and error prone when not used in situations where it is guaranteed a buffer has not been previously modified (such as new metadata allocations). To facilitate general purpose use of ordered buffers, update xfs_trans_ordered_buf() to conditionally order the buffer based on state of the log item and return the status of the result. If the bli is dirty, do not order the buffer and return false. The caller must either physically log the buffer (having acquired the appropriate log reservation) or push it from the AIL to clean it before it can be marked ordered in the current transaction. Note that ordered buffers are currently only used in two situations: 1.) inode chunk allocation where previously logged buffers are not possible and 2.) extent swap which will be updated to handle ordered buffer failures in a separate patch. Signed-off-by:
-
Brian Foster authored
commit 6fb10d6d upstream. The extent swap operation currently resets bmbt block owners before the inode forks are swapped. The bmbt buffers are marked as ordered so they do not have to be physically logged in the transaction. This use of ordered buffers is not safe as bmbt buffers may have been previously physically logged. The bmbt owner change algorithm needs to be updated to physically log buffers that are already dirty when/if they are encountered. This means that an extent swap will eventually require multiple rolling transactions to handle large btrees. In addition, all inode related changes must be logged before the bmbt owner change scan begins and can roll the transaction for the first time to preserve fs consistency via log recovery. In preparation for such fixes to the bmbt owner change algorithm, refactor the bmbt scan out of the extent fork swap code to the last operation before the transaction is committed. Update xfs_swap_extent_forks() to only set the inode log flags when an owner change scan is necessary. Update xfs_swap_extents() to trigger the owner change based on the inode log flags. Note that since the owner change now occurs after the extent fork swap, the inode btrees must be fixed up with the inode number of the current inode (similar to log recovery). Signed-off-by:
-
Brian Foster authored
commit 99c794c6 upstream. Extent swap uses xfs_btree_visit_blocks() to fix up bmbt block owners on v5 (!rmapbt) filesystems. The bmbt scan uses xfs_btree_lookup_get_block() to read bmbt blocks which verifies the current owner of the block against the parent inode of the bmbt. This works during extent swap because the bmbt owners are updated to the opposite inode number before the inode extent forks are swapped. The modified bmbt blocks are marked as ordered buffers which allows everything to commit in a single transaction. If the transaction commits to the log and the system crashes such that recovery of the extent swap is required, log recovery restarts the bmbt scan to fix up any bmbt blocks that may have not been written back before the crash. The log recovery bmbt scan occurs after the inode forks have been swapped, however. This causes the bmbt block owner verification to fail, leads to log recovery failure and requires xfs_repair to zap the log to recover. Define a new invalid inode owner flag to inform the btree block lookup mechanism that the current inode may be invalid with respect to the current owner of the bmbt block. Set this flag on the cursor used for change owner scans to allow this operation to work at runtime and during log recovery. Signed-off-by:
-
Brian Foster authored
commit 8dc518df upstream. Ordered buffers are attached to transactions and pushed through the logging infrastructure just like normal buffers with the exception that they are not actually written to the log. Therefore, we don't need to log dirty ranges of ordered buffers. xfs_trans_log_buf() is called on ordered buffers to set up all of the dirty state on the transaction, buffer and log item and prepare the buffer for I/O. Now that xfs_trans_dirty_buf() is available, call it from xfs_trans_ordered_buf() so the latter is now mutually exclusive with xfs_trans_log_buf(). This reflects the implementation of ordered buffers and helps eliminate confusion over the need to log ranges of ordered buffers just to set up internal log state. Signed-off-by:
Allison Henderson <allison.henderson@oracle.com> Reviewed-by:
-
Brian Foster authored
commit 9684010d upstream. xfs_trans_log_buf() is responsible for logging the dirty segments of a buffer along with setting all of the necessary state on the transaction, buffer, bli, etc., to ensure that the associated items are marked as dirty and prepared for I/O. We have a couple use cases that need to to dirty a buffer in a transaction without actually logging dirty ranges of the buffer. One existing use case is ordered buffers, which are currently logged with arbitrary ranges to accomplish this even though the content of ordered buffers is never written to the log. Another pending use case is to relog an already dirty buffer across rolled transactions within the deferred operations infrastructure. This is required to prevent a held (XFS_BLI_HOLD) buffer from pinning the tail of the log. Refactor xfs_trans_log_buf() into a new function that contains all of the logic responsible to dirty the transaction, lidp, buffer and bli. This new function can be used in the future for the use cases outlined above. This patch does not introduce functional changes. Signed-off-by:
-
Brian Foster authored
commit e9385cc6 upstream. Ordered buffers pass through the logging infrastructure without ever being written to the log. The way this works is that the ordered buffer status is transferred to the log vector at commit time via the ->iop_size() callback. In xlog_cil_insert_format_items(), ordered log vectors bypass ->iop_format() processing altogether. Therefore it is unnecessary for xfs_buf_item_format() to handle ordered buffers. Remove the unnecessary logic and assert that an ordered buffer never reaches this point. Signed-off-by:
-
Brian Foster authored
commit 6453c65d upstream. xfs_buf_item_unlock() historically checked the dirty state of the buffer by manually checking the buffer log formats for dirty segments. The introduction of ordered buffers invalidated this check because ordered buffers have dirty bli's but no dirty (logged) segments. The check was updated to accommodate ordered buffers by looking at the bli state first and considering the blf only if the bli is clean. This logic is safe but unnecessary. There is no valid case where the bli is clean yet the blf has dirty segments. The bli is set dirty whenever the blf is logged (via xfs_trans_log_buf()) and the blf is cleared in the only place BLI_DIRTY is cleared (xfs_trans_binval()). Remove the conditional blf dirty checks and replace with an assert that should catch any discrepencies between bli and blf dirty states. Refactor the old blf dirty check into a helper function to be used by the assert. Signed-off-by:
-
Brian Foster authored
commit a4f6cf6b upstream. It checks a single flag and has one caller. It probably isn't worth its own function. Signed-off-by:
-
Omar Sandoval authored
commit f2e9ad21 upstream. After xfs_ifree_cluster() finds an inode in the radix tree and verifies that the inode number is what it expected, xfs_reclaim_inode() can swoop in and free it. xfs_ifree_cluster() will then happily continue working on the freed inode. Most importantly, it will mark the inode stale, which will probably be overwritten when the inode slab object is reallocated, but if it has already been reallocated then we can end up with an inode spuriously marked stale. In 8a17d7dd ("xfs: mark reclaimed inodes invalid earlier") we added a second check to xfs_iflush_cluster() to detect this race, but the similar RCU lookup in xfs_ifree_cluster() needs the same treatment. Signed-off-by:
Omar Sandoval <osandov@fb.com> Reviewed-by:
-
Darrick J. Wong authored
commit 799ea9e9 upstream. When we introduced the bmap redo log items, we set MS_ACTIVE on the mountpoint and XFS_IRECOVERY on the inode to prevent unlinked inodes from being truncated prematurely during log recovery. This also had the effect of putting linked inodes on the lru instead of evicting them. Unfortunately, we neglected to find all those unreferenced lru inodes and evict them after finishing log recovery, which means that we leak them if anything goes wrong in the rest of xfs_mountfs, because the lru is only cleaned out on unmount. Therefore, evict unreferenced inodes in the lru list immediately after clearing MS_ACTIVE. Fixes: 17c12bcd ("xfs: when replaying bmap operations, don't let unlinked inodes get reaped") Signed-off-by:
-
Carlos Maiolino authored
commit 2d32311c upstream. In a filesystem without finobt, the Space manager selects an AG to alloc a new inode, where xfs_dialloc_ag_inobt() will search the AG for the free slot chunk. When the new inode is in the same AG as its parent, the btree will be searched starting on the parent's record, and then retried from the top if no slot is available beyond the parent's record. To exit this loop though, xfs_dialloc_ag_inobt() relies on the fact that the btree must have a free slot available, once its callers relied on the agi->freecount when deciding how/where to allocate this new inode. In the case when the agi->freecount is corrupted, showing available inodes in an AG, when in fact there is none, this becomes an infinite loop. Add a way to stop the loop when a free slot is not found in the btree, making the function to fall into the whole AG scan which will then, be able to detect the corruption and shut the filesystem down. As pointed by Brian, this might impact performance, giving the fact we don't reset the search distance anymore when we reach the end of the tree, giving it fewer tries before falling back to the whole AG search, but it will only affect searches that start within 10 records to the end of the tree. Signed-off-by:
Carlos Maiolino <cmaiolino@redhat.com> Reviewed-by:
-
Brian Foster authored
commit e67d3d42 upstream. Torn write detection and tail overwrite detection can shift the log head and tail respectively in the event of CRC mismatch or corruption errors. Add a high-level log recovery tracepoint to dump the final log head/tail and make those values easily attainable in debug/diagnostic situations. Signed-off-by:
-
Brian Foster authored
commit a4c9b34d upstream. Torn write and tail overwrite detection both trigger only on -EFSBADCRC errors. While this is the most likely failure scenario for each condition, -EFSCORRUPTED is still possible in certain cases depending on what ends up on disk when a torn write or partial tail overwrite occurs. For example, an invalid log record h_len can lead to an -EFSCORRUPTED error when running the log recovery CRC pass. Therefore, update log head and tail verification to trigger the associated head/tail fixups in the event of -EFSCORRUPTED errors along with -EFSBADCRC. Also, -EFSCORRUPTED can currently be returned from xlog_do_recovery_pass() before rhead_blk is initialized if the first record encountered happens to be corrupted. This leads to an incorrect 'first_bad' return value. Initialize rhead_blk earlier in the function to address that problem as well. Signed-off-by:
-
Brian Foster authored
commit 4a4f66ea upstream. If we consider the case where the tail (T) of the log is pinned long enough for the head (H) to push and block behind the tail, we can end up blocked in the following state without enough free space (f) in the log to satisfy a transaction reservation: 0 phys. log N [-------HffT---H'--T'---] The last good record in the log (before H) refers to T. The tail eventually pushes forward (T') leaving more free space in the log for writes to H. At this point, suppose space frees up in the log for the maximum of 8 in-core log buffers to start flushing out to the log. If this pushes the head from H to H', these next writes overwrite the previous tail T. This is safe because the items logged from T to T' have been written back and removed from the AIL. If the next log writes (H -> H') happen to fail and result in partial records in the log, the filesystem shuts down having overwritten T with invalid data. Log recovery correctly locates H on the subsequent mount, but H still refers to the now corrupted tail T. This results in log corruption errors and recovery failure. Since the tail overwrite results from otherwise correct runtime behavior, it is up to log recovery to try and deal with this situation. Update log recovery tail verification to run a CRC pass from the first record past the tail to the head. This facilitates error detection at T and moves the recovery tail to the first good record past H' (similar to truncating the head on torn write detection). If corruption is detected beyond the range possibly affected by the max number of iclogs, the log is legitimately corrupted and log recovery failure is expected. Signed-off-by:
-
Brian Foster authored
commit 5297ac1f upstream. Log tail verification currently only occurs when torn writes are detected at the head of the log. This was introduced because a change in the head block due to torn writes can lead to a change in the tail block (each log record header references the current tail) and the tail block should be verified before log recovery proceeds. Tail corruption is possible outside of torn write scenarios, however. For example, partial log writes can be detected and cleared during the initial head/tail block discovery process. If the partial write coincides with a tail overwrite, the log tail is corrupted and recovery fails. To facilitate correct handling of log tail overwites, update log recovery to always perform tail verification. This is necessary to detect potential tail overwrite conditions when torn writes may not have occurred. This changes normal (i.e., no torn writes) recovery behavior slightly to detect and return CRC related errors near the tail before actual recovery starts. Signed-off-by:
-
Brian Foster authored
commit 284f1c2c upstream. The high-level log recovery algorithm consists of two loops that walk the physical log and process log records from the tail to the head. The first loop handles the case where the tail is beyond the head and processes records up to the end of the physical log. The subsequent loop processes records from the beginning of the physical log to the head. Because log records can wrap around the end of the physical log, the first loop mentioned above must handle this case appropriately. Records are processed from in-core buffers, which means that this algorithm must split the reads of such records into two partial I/Os: 1.) from the beginning of the record to the end of the log and 2.) from the beginning of the log to the end of the record. This is further complicated by the fact that the log record header and log record data are read into independent buffers. The current handling of each buffer correctly splits the reads when either the header or data starts before the end of the log and wraps around the end. The data read does not correctly handle the case where the prior header read wrapped or ends on the physical log end boundary. blk_no is incremented to or beyond the log end after the header read to point to the record data, but the split data read logic triggers, attempts to read from an invalid log block and ultimately causes log recovery to fail. This can be reproduced fairly reliably via xfstests tests generic/047 and generic/388 with large iclog sizes (256k) and small (10M) logs. If the record header read has pushed beyond the end of the physical log, the subsequent data read is actually contiguous. Update the data read logic to detect the case where blk_no has wrapped, mod it against the log size to read from the correct address and issue one contiguous read for the log data buffer. The log record is processed as normal from the buffer(s), the loop exits after the current iteration and the subsequent loop picks up with the first new record after the start of the log. Signed-off-by:
-
Carlos Maiolino authored
commit d3a304b6 upstream. When a buffer has been failed during writeback, the inode items into it are kept flush locked, and are never resubmitted due the flush lock, so, if any buffer fails to be written, the items in AIL are never written to disk and never unlocked. This causes unmount operation to hang due these items flush locked in AIL, but this also causes the items in AIL to never be written back, even when the IO device comes back to normal. I've been testing this patch with a DM-thin device, creating a filesystem larger than the real device. When writing enough data to fill the DM-thin device, XFS receives ENOSPC errors from the device, and keep spinning on xfsaild (when 'retry forever' configuration is set). At this point, the filesystem can not be unmounted because of the flush locked items in AIL, but worse, the items in AIL are never retried at all (once xfs_inode_item_push() will skip the items that are flush locked), even if the underlying DM-thin device is expanded to the proper size. This patch fixes both cases, retrying any item that has been failed previously, using the infra-structure provided by the previous patch. Reviewed-by:
-
Carlos Maiolino authored
commit 0b80ae6e upstream. With the current code, XFS never re-submit a failed buffer for IO, because the failed item in the buffer is kept in the flush locked state forever. To be able to resubmit an log item for IO, we need a way to mark an item as failed, if, for any reason the buffer which the item belonged to failed during writeback. Add a new log item callback to be used after an IO completion failure and make the needed clean ups. Reviewed-by:
-
Christoph Hellwig authored
commit 27af1bbf upstream. xfs_iflush_done uses an on-stack variable length array to pass the log items to be deleted to xfs_trans_ail_delete_bulk. On-stack VLAs are a nasty gcc extension that can lead to unbounded stack allocations, but fortunately we can easily avoid them by simply open coding xfs_trans_ail_delete_bulk in xfs_iflush_done, which is the only caller of it except for the single-item xfs_trans_ail_delete. Signed-off-by:
-
Eric Sandeen authored
commit 6f4a1eef upstream. When we do log recovery on a readonly mount, unlinked inode processing does not happen due to the readonly checks in xfs_inactive(), which are trying to prevent any I/O on a readonly mount. This is misguided - we do I/O on readonly mounts all the time, for consistency; for example, log recovery. So do the same RDONLY flag twiddling around xfs_log_mount_finish() as we do around xfs_log_mount(), for the same reason. This all cries out for a big rework but for now this is a simple fix to an obvious problem. Signed-off-by:
Eric Sandeen <sandeen@redhat.com> Reviewed-by:
-
Eric Sandeen authored
commit 757a69ef upstream. There are dueling comments in the xfs code about intent for log writes when unmounting a readonly filesystem. In xfs_mountfs, we see the intent: /* * Now the log is fully replayed, we can transition to full read-only * mode for read-only mounts. This will sync all the metadata and clean * the log so that the recovery we just performed does not have to be * replayed again on the next mount. */ and it calls xfs_quiesce_attr(), but by the time we get to xfs_log_unmount_write(), it returns early for a RDONLY mount: * Don't write out unmount record on read-only mounts. Because of this, sequential ro mounts of a filesystem with a dirty log will replay the log each time, which seems odd. Fix this by writing an unmount record even for RO mounts, as long as norecovery wasn't specified (don't write a clean log record if a dirty log may still be there!) and the log device is writable. Signed-off-by:
-
Christoph Hellwig authored
commit e28ae8e4 upstream. Fix the min_t calls in the zeroing and dirtying helpers to perform the comparisms on 64-bit types, which prevents them from incorrectly being truncated, and larger zeroing operations being stuck in a never ending loop. Special thanks to Markus Stockhausen for spotting the bug. Reported-by:
Paul Menzel <pmenzel@molgen.mpg.de> Tested-by:
-
Darrick J. Wong authored
commit 77aff8c7 upstream. If we fail a mount on account of cow recovery errors, it's possible that a previous quotacheck left some dquots in memory. The bailout clause of xfs_mountfs forgets to purge these, and so we leak them. Fix that. Signed-off-by:
-
Darrick J. Wong authored
commit 8204f8dd upstream. Way back when we established inode block-map redo log items, it was discovered that we needed to prevent the VFS from evicting inodes during log recovery because any given inode might be have bmap redo items to replay even if the inode has no link count and is ultimately deleted, and any eviction of an unlinked inode causes the inode to be truncated and freed too early. To make this possible, we set MS_ACTIVE so that inodes would not be torn down immediately upon release. Unfortunately, this also results in the quota inodes not being released at all if a later part of the mount process should fail, because we never reclaim the inodes. So, set MS_ACTIVE right before we do the last part of log recovery and clear it immediately after we finish the log recovery so that everything will be torn down properly if we abort the mount. Fixes: 17c12bcd ("xfs: when replaying bmap operations, don't let unlinked inodes get reaped") Signed-off-by:
-
Omar Sandoval authored
commit c44245b3 upstream. When we try to allocate a free inode by searching the inobt, we try to find the inode nearest the parent inode by searching chunks both left and right of the chunk containing the parent. As an optimization, we cache the leftmost and rightmost records that we previously searched; if we do another allocation with the same parent inode, we'll pick up the search where it last left off. There's a bug in the case where we found a free inode to the left of the parent's chunk: we need to update the cached left and right records, but because we already reassigned the right record to point to the left, we end up assigning the left record to both the cached left and right records. This isn't a correctness problem strictly, but it can result in the next allocation rechecking chunks unnecessarily or allocating inodes further away from the parent than it needs to. Fix it by swapping the record pointer after we update the cached left and right records. Fixes: bd169565 ("xfs: speed up free inode search") Signed-off-by:
-
Lukas Czerner authored
commit 56bdf855 upstream. According to the commit that implemented per-inode DAX flag: commit 58f88ca2 ("xfs: introduce per-inode DAX enablement") the flag is supposed to act as "inherit flag". Currently this only works in the situations where parent directory already has a flag in di_flags set, otherwise inheritance does not work. This is because setting the XFS_DIFLAG2_DAX flag is done in a wrong branch designated for di_flags, not di_flags2. Fix this by moving the code to branch designated for setting di_flags2, which does test for flags in di_flags2. Fixes: 58f88ca2 ("xfs: introduce per-inode DAX enablement") Signed-off-by:
Lukas Czerner <lczerner@redhat.com> Reviewed-by:
-
Christoph Hellwig authored
commit 5b094d6d upstream. Just like in the allocator we must avoid touching multiple AGs out of order when freeing blocks, as freeing still locks the AGF and can cause the same AB-BA deadlocks as in the allocation path. Signed-off-by:
Nikolay Borisov <n.borisov.lkml@gmail.com> Reviewed-by:
-
Brian Foster authored
commit cfaf2d03 upstream. If a dquot has an id of U32_MAX, the next lookup index increment overflows the uint32_t back to 0. This starts the lookup sequence over from the beginning, repeats indefinitely and results in a livelock. Update xfs_qm_dquot_walk() to explicitly check for the lookup overflow and exit the loop. Signed-off-by:
-
Darrick J. Wong authored
commit 10479e2d upstream. In some circumstances, _alloc_read_agf can return an error code of zero but also a null AGF buffer pointer. Check for this and jump out. Fixes-coverity-id: 1415250 Fixes-coverity-id: 1415320 Signed-off-by:
-
Darrick J. Wong authored
commit 4c1a67bd upstream. We must initialize the firstfsb parameter to _bmapi_write so that it doesn't incorrectly treat stack garbage as a restriction on which AGs it can search for free space. Fixes-coverity-id: 1402025 Fixes-coverity-id: 1415167 Signed-off-by:
-
Darrick J. Wong authored
commit 1e86eabe upstream. Check the _btree_check_block return value for the firstrec and lastrec functions, since we have the ability to signal that the repositioning did not succeed. Fixes-coverity-id: 114067 Fixes-coverity-id: 114068 Signed-off-by:
-
Darrick J. Wong authored
commit cd87d867 upstream. In quite a few places we call xfs_da_read_buf with a mappedbno that we don't control, then assume that the function passes back either an error code or a buffer pointer. Unfortunately, if mappedbno == -2 and bno maps to a hole, we get a return code of zero and a NULL buffer, which means that we crash if we actually try to use that buffer pointer. This happens immediately when we set the buffer type for transaction context. Therefore, check that we have no error code and a non-NULL bp before trying to use bp. This patch is a follow-up to an incomplete fix in 96a3aefb ("xfs: don't crash if reading a directory results in an unexpected hole"). Signed-off-by:
-
Brian Foster authored
commit cf2cb784 upstream. XFS runs an eofblocks reclaim scan before returning an ENOSPC error to userspace for buffered writes. This facilitates aggressive speculative preallocation without causing user visible side effects such as premature ENOSPC. Run a cowblocks scan in the same situation to reclaim lingering COW fork preallocation throughout the filesystem. Signed-off-by:
-
Brian Foster authored
commit 39775431 upstream. Log recovery allocates in-core transaction and member item data structures on-demand as it processes the on-disk log. Transactions are allocated on first encounter on-disk and stored in a hash table structure where they are easily accessible for subsequent lookups. Transaction items are also allocated on demand and are attached to the associated transactions. When a commit record is encountered in the log, the transaction is committed to the fs and the in-core structures are freed. If a filesystem crashes or shuts down before all in-core log buffers are flushed to the log, however, not all transactions may have commit records in the log. As expected, the modifications in such an incomplete transaction are not replayed to the fs. The in-core data structures for the partial transaction are never freed, however, resulting in a memory leak. Update xlog_do_recovery_pass() to first correctly initialize the hash table array so empty lists can be distinguished from populated lists on function exit. Update xlog_recover_free_trans() to always remove the transaction from the list prior to freeing the associated memory. Finally, walk the hash table of transaction lists as the last step before it goes out of scope and free any transactions that may remain on the lists. This prevents a memory leak of partial transactions in the log. Signed-off-by:
-
Darrick J. Wong authored
commit 61d819e7 upstream. bmap returns a dumb LBA address but not the block device that goes with that LBA. Swapfiles don't care about this and will blindly assume that the data volume is the correct blockdev, which is totally bogus for files on the rt subvolume. This results in the swap code doing IOs to arbitrary locations on the data device(!) if the passed in mapping is a realtime file, so just turn off bmap for rt files. Signed-off-by:
-
Brian Foster authored
commit 3d4b4a3e upstream. When a buffer is modified, logged and committed, it ultimately ends up sitting on the AIL with a dirty bli waiting for metadata writeback. If another transaction locks and invalidates the buffer (freeing an inode chunk, for example) in the meantime, the bli is flagged as stale, the dirty state is cleared and the bli remains in the AIL. If a shutdown occurs before the transaction that has invalidated the buffer is committed, the transaction is ultimately aborted. The log items are flagged as such and ->iop_unlock() handles the aborted items. Because the bli is clean (due to the invalidation), ->iop_unlock() unconditionally releases it. The log item may still reside in the AIL, however, which means the I/O completion handler may still run and attempt to access it. This results in assert failure due to the release of the bli while still present in the AIL and a subsequent NULL dereference and panic in the buffer I/O completion handling. This can be reproduced by running generic/388 in repetition. To avoid this problem, update xfs_buf_item_unlock() to first check whether the bli is aborted and if so, remove it from the AIL before it is released. This ensures that the bli is no longer accessed during the shutdown sequence after it has been freed. Signed-off-by:
-
Brian Foster authored
commit 79e641ce upstream. If a filesystem shutdown occurs with a buffer log item in the CIL and a log force occurs, the ->iop_unpin() handler is generally expected to tear down the bli properly. This entails freeing the bli memory and releasing the associated hold on the buffer so it can be released and the filesystem unmounted. If this sequence occurs while ->bli_refcount is elevated (i.e., another transaction is open and attempting to modify the buffer), however, ->iop_unpin() may not be responsible for releasing the bli. Instead, the transaction may release the final ->bli_refcount reference and thus xfs_trans_brelse() is responsible for tearing down the bli. While xfs_trans_brelse() does drop the reference count, it only attempts to release the bli if it is clean (i.e., not in the CIL/AIL). If the filesystem is shutdown and the bli is sitting dirty in the CIL as noted above, this ends up skipping the last opportunity to release the bli. In turn, this leaves the hold on the buffer and causes an unmount hang. This can be reproduced by running generic/388 in repetition. Update xfs_trans_brelse() to handle this shutdown corner case correctly. If the final bli reference is dropped and the filesystem is shutdown, remove the bli from the AIL (if necessary) and release the bli to drop the buffer hold and ensure an unmount does not hang. Signed-off-by:
-
Darrick J. Wong authored
commit e1a4e37c upstream. In a pathological scenario where we are trying to bunmapi a single extent in which every other block is shared, it's possible that trying to unmap the entire large extent in a single transaction can generate so many EFIs that we overflow the transaction reservation. Therefore, use a heuristic to guess at the number of blocks we can safely unmap from a reflink file's data fork in an single transaction. This should prevent problems such as the log head slamming into the tail and ASSERTs that trigger because we've exceeded the transaction reservation. Note that since bunmapi can fail to unmap the entire range, we must also teach the deferred unmap code to roll into a new transaction whenever we get low on reservation. Signed-off-by:
-
Brian Foster authored
commit 7912e7fe upstream. Reclaim during quotacheck can lead to deadlocks on the dquot flush lock: - Quotacheck populates a local delwri queue with the physical dquot buffers. - Quotacheck performs the xfs_qm_dqusage_adjust() bulkstat and dirties all of the dquots. - Reclaim kicks in and attempts to flush a dquot whose buffer is already queud on the quotacheck queue. The flush succeeds but queueing to the reclaim delwri queue fails as the backing buffer is already queued. The flush unlock is now deferred to I/O completion of the buffer from the quotacheck queue. - The dqadjust bulkstat continues and dirties the recently flushed dquot once again. - Quotacheck proceeds to the xfs_qm_flush_one() walk which requires the flush lock to update the backing buffers with the in-core recalculated values. It deadlocks on the redirtied dquot as the flush lock was already acquired by reclaim, but the buffer resides on the local delwri queue which isn't submitted until the end of quotacheck. This is reproduced by running quotacheck on a filesystem with a couple million inodes in low memory (512MB-1GB) situations. This is a regression as of commit 43ff2122 ("xfs: on-stack delayed write buffer lists"), which removed a trylock and buffer I/O submission from the quotacheck dquot flush sequence. Quotacheck first resets and collects the physical dquot buffers in a delwri queue. Then, it traverses the filesystem inodes via bulkstat, updates the in-core dquots, flushes the corrected dquots to the backing buffers and finally submits the delwri queue for I/O. Since the backing buffers are queued across the entire quotacheck operation, dquot reclaim cannot possibly complete a dquot flush before quotacheck completes. Therefore, quotacheck must submit the buffer for I/O in order to cycle the flush lock and flush the dirty in-core dquot to the buffer. Add a delwri queue buffer push mechanism to submit an individual buffer for I/O without losing the delwri queue status and use it from quotacheck to avoid the deadlock. This restores quotacheck behavior to as before the regression was introduced. Reported-by:
Martin Svec <martin.svec@zoner.cz> Signed-off-by:
-
