- 31 Mar, 2020 1 commit
-
-
Darrick J. Wong authored
A customer reported rcu stalls and softlockup warnings on a computer with many CPU cores and many many more IO threads trying to write to a filesystem that is totally out of space. Subsequent analysis pointed to the many many IO threads calling xfs_flush_inodes -> sync_inodes_sb, which causes a lot of wb_writeback_work to be queued. The writeback worker spends so much time trying to wake the many many threads waiting for writeback completion that it trips the softlockup detector, and (in this case) the system automatically reboots. In addition, they complain that the lengthy xfs_flush_inodes scan traps all of those threads in uninterruptible sleep, which hampers their ability to kill the program or do anything else to escape the situation. If there's thousands of threads trying to write to files on a full filesystem, each of those threads will start separate copies of the inode flush scan. This is kind of pointless since we only need one scan, so rate limit the inode flush. Signed-off-by:
Darrick J. Wong <darrick.wong@oracle.com> Reviewed-by:
Dave Chinner <dchinner@redhat.com>
-
- 28 Mar, 2020 3 commits
-
-
Brian Foster authored
If the inode buffer backing a particular inode is locked, xfs_iflush() returns -EAGAIN and xfs_inode_item_push() skips the inode. It still returns success to xfsaild, however, which bypasses the xfsaild backoff heuristic. Update xfs_inode_item_push() to return locked status if the inode buffer couldn't be locked. Signed-off-by:
Brian Foster <bfoster@redhat.com> Reviewed-by:
Christoph Hellwig <hch@lst.de> Reviewed-by:
-
Brian Foster authored
A dquot flush currently blocks on the buffer lock for the underlying dquot buffer. In turn, this causes xfsaild to block rather than continue processing other items in the meantime. Update xfs_qm_dqflush() to trylock the buffer, similar to how inode buffers are handled, and return -EAGAIN if the lock fails. Fix up any callers that don't currently handle the error properly. Signed-off-by:
-
Kaixu Xia authored
Since the "no-allocation" reservations for file creations has been removed, the resblks value should be larger than zero, so remove unnecessary ternary conditional. Signed-off-by:
Kaixu Xia <kaixuxia@tencent.com> Reviewed-by:
-
- 27 Mar, 2020 18 commits
-
-
Darrick J. Wong authored
In commit f467cad9, I added the ability to force a recalculation of the filesystem summary counters if they seemed incorrect. This was done (not entirely correctly) by tweaking the log code to write an unmount record without the UMOUNT_TRANS flag set. At next mount, the log recovery code will fail to find the unmount record and go into recovery, which triggers the recalculation. What actually gets written to the log is what ought to be an unmount record, but without any flags set to indicate what kind of record it actually is. This worked to trigger the recalculation, but we shouldn't write bogus log records when we could simply write nothing. Fixes: f467cad9 ("xfs: force summary counter recalc at next mount") Signed-off-by:
-
Dave Chinner authored
There's lots of indent in this code which makes it a bit hard to follow. We are also going to completely rework the inode lookup code as part of the inode reclaim rework, so factor out the inode lookup code from the inode cluster freeing code. Based on prototype code from Christoph Hellwig. Signed-off-by:
-
Dave Chinner authored
We currently wake anything waiting on the log tail to move whenever the log item at the tail of the log is removed. Historically this was fine behaviour because there were very few items at any given LSN. But with delayed logging, there may be thousands of items at any given LSN, and we can't move the tail until they are all gone. Hence if we are removing them in near tail-first order, we might be waking up processes waiting on the tail LSN to change (e.g. log space waiters) repeatedly without them being able to make progress. This also occurs with the new sync push waiters, and can result in thousands of spurious wakeups every second when under heavy direct reclaim pressure. To fix this, check that the tail LSN has actually changed on the AIL before triggering wakeups. This will reduce the number of spurious wakeups when doing bulk AIL removal and make this code much more efficient. Signed-off-by:
Allison Collins <allison.henderson@oracle.com> Reviewed-by:
-
Dave Chinner authored
Factor the common AIL deletion code that does all the wakeups into a helper so we only have one copy of this somewhat tricky code to interface with all the wakeups necessary when the LSN of the log tail changes. Signed-off-by:
-
Dave Chinner authored
The XFS inode item slab actually reclaimed by inode shrinker callbacks from the memory reclaim subsystem. These should be marked as reclaimable so the mm subsystem has the full picture of how much memory it can actually reclaim from the XFS slab caches. Signed-off-by:
-
Dave Chinner authored
The buffer cache shrinker frees more than just the xfs_buf slab objects - it also frees the pages attached to the buffers. Make sure the memory reclaim code accounts for this memory being freed correctly, similar to how the inode shrinker accounts for pages freed from the page cache due to mapping invalidation. We also need to make sure that the mm subsystem knows these are reclaimable objects. We provide the memory reclaim subsystem with a a shrinker to reclaim xfs_bufs, so we should really mark the slab that way. We also have a lot of xfs_bufs in a busy system, spread them around like we do inodes. Signed-off-by:
-
Dave Chinner authored
Running metadata intensive workloads, I've been seeing the AIL pushing getting stuck on pinned buffers and triggering log forces. The log force is taking a long time to run because the log IO is getting throttled by wbt_wait() - the block layer writeback throttle. It's being throttled because there is a huge amount of metadata writeback going on which is filling the request queue. IOWs, we have a priority inversion problem here. Mark the log IO bios with REQ_IDLE so they don't get throttled by the block layer writeback throttle. When we are forcing the CIL, we are likely to need to to tens of log IOs, and they are issued as fast as they can be build and IO completed. Hence REQ_IDLE is appropriate - it's an indication that more IO will follow shortly. And because we also set REQ_SYNC, the writeback throttle will now treat log IO the same way it treats direct IO writes - it will not throttle them at all. Hence we solve the priority inversion problem caused by the writeback throttle being unable to distinguish between high priority log IO and background metadata writeback. Signed-off-by:
-
Dave Chinner authored
In certain situations the background CIL push can be indefinitely delayed. While we have workarounds from the obvious cases now, it doesn't solve the underlying issue. This issue is that there is no upper limit on the CIL where we will either force or wait for a background push to start, hence allowing the CIL to grow without bound until it consumes all log space. To fix this, add a new wait queue to the CIL which allows background pushes to wait for the CIL context to be switched out. This happens when the push starts, so it will allow us to block incoming transaction commit completion until the push has started. This will only affect processes that are running modifications, and only when the CIL threshold has been significantly overrun. This has no apparent impact on performance, and doesn't even trigger until over 45 million inodes had been created in a 16-way fsmark test on a 2GB log. That was limiting at 64MB of log space used, so the active CIL size is only about 3% of the total log in that case. The concurrent removal of those files did not trigger the background sleep at all. Signed-off-by:
-
Dave Chinner authored
The current CIL size aggregation limit is 1/8th the log size. This means for large logs we might be aggregating at least 250MB of dirty objects in memory before the CIL is flushed to the journal. With CIL shadow buffers sitting around, this means the CIL is often consuming >500MB of temporary memory that is all allocated under GFP_NOFS conditions. Flushing the CIL can take some time to do if there is other IO ongoing, and can introduce substantial log force latency by itself. It also pins the memory until the objects are in the AIL and can be written back and reclaimed by shrinkers. Hence this threshold also tends to determine the minimum amount of memory XFS can operate in under heavy modification without triggering the OOM killer. Modify the CIL space limit to prevent such huge amounts of pinned metadata from aggregating. We can have 2MB of log IO in flight at once, so limit aggregation to 16x this size. This threshold was chosen as it little impact on performance (on 16-way fsmark) or log traffic but pins a lot less memory on large logs especially under heavy memory pressure. An aggregation limit of 8x had 5-10% performance degradation and a 50% increase in log throughput for the same workload, so clearly that was too small for highly concurrent workloads on large logs. This was found via trace analysis of AIL behaviour. e.g. insertion from a single CIL flush: xfs_ail_insert: old lsn 0/0 new lsn 1/3033090 type XFS_LI_INODE flags IN_AIL $ grep xfs_ail_insert /mnt/scratch/s.t |grep "new lsn 1/3033090" |wc -l 1721823 $ So there were 1.7 million objects inserted into the AIL from this CIL checkpoint, the first at 2323.392108, the last at 2325.667566 which was the end of the trace (i.e. it hadn't finished). Clearly a major problem. Signed-off-by:
-
Dave Chinner authored
Signed-off-by:
-
Dave Chinner authored
Separate out the unmount record writing from the rest of the ticket and log state futzing necessary to make it work. This is a no-op, just makes the code cleaner and places the unmount record formatting and writing alongside the commit record formatting and writing code. We can also get rid of the ticket flag clearing before the xlog_write() call because it no longer cares about the state of XLOG_TIC_INITED. Signed-off-by:
-
Dave Chinner authored
xlog_write_done() is just a thin wrapper around xlog_commit_record(), so they can be merged together easily. Signed-off-by:
-
Christoph Hellwig authored
Remove xlog_ticket_done and just call the renamed low-level helpers for ungranting or regranting log space directly. To make that a little the reference put on the ticket and all tracing is moved into the actual helpers. Signed-off-by:
-
Dave Chinner authored
It is not longer used or checked by anything, so remove the last traces from the log ticket code. Signed-off-by:
-
Dave Chinner authored
xfs_log_done() does two separate things. Firstly, it triggers commit records to be written for permanent transactions, and secondly it releases or regrants transaction reservation space. Since delayed logging was introduced, transactions no longer write directly to the log, hence they never have the XLOG_TIC_INITED flag cleared on them. Hence transactions never write commit records to the log and only need to modify reservation space. Split up xfs_log_done into two parts, and only call the parts of the operation needed for the context xfs_log_done() is currently being called from. Signed-off-by:
-
Dave Chinner authored
Commit and unmount records records do not need start records to be written, so rearrange the logic in xlog_write() to remove the need to check for XLOG_TIC_INITED to determine if we should account for the space used by a start record. Signed-off-by:
-
Dave Chinner authored
The xlog_write() function iterates over iclogs until it completes writing all the log vectors passed in. The ticket tracks whether a start record has been written or not, so only the first iclog gets a start record. We only ever pass single use tickets to xlog_write() so we only ever need to write a start record once per xlog_write() call. Hence we don't need to store whether we should write a start record in the ticket as the callers provide all the information we need to determine if a start record should be written. For the moment, we have to ensure that we clear the XLOG_TIC_INITED appropriately so the code in xfs_log_done() still works correctly for committing transactions. (darrick: Note the slight behavior change that we always deduct the size of the op header from the ticket, even for unmount records) Signed-off-by:
-
Darrick J. Wong authored
Validate the geometry of the realtime geometry when we mount the filesystem, so that we don't abruptly shut down the filesystem later on. Signed-off-by:
-
- 26 Mar, 2020 5 commits
-
-
Darrick J. Wong authored
I noticed that fsfreeze can take a very long time to freeze an XFS if there happens to be a GETFSMAP caller running in the background. I also happened to notice the following in dmesg: ------------[ cut here ]------------ WARNING: CPU: 2 PID: 43492 at fs/xfs/xfs_super.c:853 xfs_quiesce_attr+0x83/0x90 [xfs] Modules linked in: xfs libcrc32c ip6t_REJECT nf_reject_ipv6 ipt_REJECT nf_reject_ipv4 ip_set_hash_ip ip_set_hash_net xt_tcpudp xt_set ip_set_hash_mac ip_set nfnetlink ip6table_filter ip6_tables bfq iptable_filter sch_fq_codel ip_tables x_tables nfsv4 af_packet [last unloaded: xfs] CPU: 2 PID: 43492 Comm: xfs_io Not tainted 5.6.0-rc4-djw #rc4 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.10.2-1ubuntu1 04/01/2014 RIP: 0010:xfs_quiesce_attr+0x83/0x90 [xfs] Code: 7c 07 00 00 85 c0 75 22 48 89 df 5b e9 96 c1 00 00 48 c7 c6 b0 2d 38 a0 48 89 df e8 57 64 ff ff 8b 83 7c 07 00 00 85 c0 74 de <0f> 0b 48 89 df 5b e9 72 c1 00 00 66 90 0f 1f 44 00 00 41 55 41 54 RSP: 0018:ffffc900030f3e28 EFLAGS: 00010202 RAX: 0000000000000001 RBX: ffff88802ac54000 RCX: 0000000000000000 RDX: 0000000000000000 RSI: ffffffff81e4a6f0 RDI: 00000000ffffffff RBP: ffff88807859f070 R08: 0000000000000001 R09: 0000000000000000 R10: 0000000000000000 R11: 0000000000000010 R12: 0000000000000000 R13: ffff88807859f388 R14: ffff88807859f4b8 R15: ffff88807859f5e8 FS: 00007fad1c6c0fc0(0000) GS:ffff88807e000000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007f0c7d237000 CR3: 0000000077f01003 CR4: 00000000001606a0 Call Trace: xfs_fs_freeze+0x25/0x40 [xfs] freeze_super+0xc8/0x180 do_vfs_ioctl+0x70b/0x750 ? __fget_files+0x135/0x210 ksys_ioctl+0x3a/0xb0 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x50/0x1a0 entry_SYSCALL_64_after_hwframe+0x49/0xbe These two things appear to be related. The assertion trips when another thread initiates a fsmap request (which uses an empty transaction) after the freezer waited for m_active_trans to hit zero but before the the freezer executes the WARN_ON just prior to calling xfs_log_quiesce. The lengthy delays in freezing happen because the freezer calls xfs_wait_buftarg to clean out the buffer lru list. Meanwhile, the GETFSMAP caller is continuing to grab and release buffers, which means that it can take a very long time for the buffer lru list to empty out. We fix both of these races by calling sb_start_write to obtain freeze protection while using empty transactions for GETFSMAP and for metadata scrubbing. The other two users occur during mount, during which time we cannot fs freeze. Signed-off-by:
-
Brian Foster authored
If the bio_add_page() call fails, we proceed to write out a partially constructed log buffer. This corrupts the physical log such that log recovery is not possible. Worse, persistent occurrences of this error eventually lead to a BUG_ON() failure in bio_split() as iclogs wrap the end of the physical log, which triggers log recovery on subsequent mount. Rather than warn about writing out a corrupted log buffer, shutdown the fs as is done for any log I/O related error. This preserves the consistency of the physical log such that log recovery succeeds on a subsequent mount. Note that this was observed on a 64k page debug kernel without upstream commit 59bb4798 ("mm, sl[aou]b: guarantee natural alignment for kmalloc(power-of-two)"), which demonstrated frequent iclog bio overflows due to unaligned (slab allocated) iclog data buffers. Signed-off-by:
-
Darrick J. Wong authored
When we're checking bestfree information in directory blocks, always drop the block buffer at the end of the function. We should always release resources when we're done using them. Signed-off-by:
-
Darrick J. Wong authored
The dirattr btree checking code uses the altpath substructure of the dirattr state structure to check the sibling pointers of dir/attr tree blocks. At the end of sibling checks, xfs_da3_path_shift could have changed multiple levels of buffer pointers in the altpath structure. Although we release the leaf level buffer, this isn't enough -- we also need to release the node buffers that are unique to the altpath. Not releasing all of the altpath buffers leaves them locked to the transaction. This is suboptimal because we should release resources when we don't need them anymore. Fix the function to loop all levels of the altpath, and fix the return logic so that we always run the loop. Signed-off-by:
-
Darrick J. Wong authored
When quotacheck runs, it zeroes all the timer fields in every dquot. Unfortunately, it also does this to the root dquot, which erases any preconfigured grace intervals and warning limits that the administrator may have set. Worse yet, the incore copies of those variables remain set. This cache coherence problem manifests itself as the grace interval mysteriously being reset back to the defaults at the /next/ mount. Fix it by not resetting the root disk dquot's timer and warning fields. Signed-off-by:
-
- 23 Mar, 2020 8 commits
-
-
Christoph Hellwig authored
Open code the xlog_state_want_sync logic in its two callers given that this function is a trivial wrapper around xlog_state_switch_iclogs. Move the lockdep assert into xlog_state_switch_iclogs to not lose this debugging aid, and improve the comment that documents xlog_state_switch_iclogs as well. Signed-off-by:
-
Christoph Hellwig authored
Use the shutdown flag in the log to bypass xlog_state_clean_iclog entirely in case of a shut down log. Signed-off-by:
-
Christoph Hellwig authored
Factor out a few self-contained helpers from xlog_state_clean_iclog, and update the documentation so it primarily documents why things happens instead of how. Signed-off-by:
-
Christoph Hellwig authored
We can just check for a shut down log all the way down in xlog_cil_committed instead of passing the parameter. This means a slight behavior change in that we now also abort log items if the shutdown came in halfway into the I/O completion processing, which actually is the right thing to do. Signed-off-by:
-
Christoph Hellwig authored
There is no need to check for the ioerror state before the lock, as the shutdown case is not a fast path. Also remove the call to force shutdown the file system, as it must have been shut down already for an iclog to be in the ioerror state. Also clean up the flow of the function a bit. Signed-off-by:
-
Christoph Hellwig authored
The only caller of xfs_log_release_iclog doesn't care about the return value, so remove it. Also don't bother passing the mount pointer, given that we can trivially derive it from the iclog. Signed-off-by:
-
Christoph Hellwig authored
Factor out the shared code to wait for a log force into a new helper. This helper uses the XLOG_FORCED_SHUTDOWN check previous only used by the unmount code over the equivalent iclog ioerror state used by the other two functions. There is a slight behavior change in that the force of the unmount record is now accounted in the log force statistics. Signed-off-by:
-
Christoph Hellwig authored
xlog_cil_push is only called by xlog_cil_push_work, so merge the two functions. Signed-off-by:
-
- 19 Mar, 2020 5 commits
-
-
Christoph Hellwig authored
We know the version is 3 if on a v5 file system. For earlier file systems formats we always upgrade the remaining v1 inodes to v2 and thus only use v2 inodes. Use the xfs_sb_version_has_large_dinode helper to check if we deal with small or large dinodes, and thus remove the need for the di_version field in struct icdinode. Signed-off-by:
Chandan Rajendra <chandanrlinux@gmail.com> Reviewed-by:
-
Christoph Hellwig authored
Only v5 file systems can have the reflink feature, and those will always use the large dinode format. Remove the extra check for the inode version. Signed-off-by:
-
Christoph Hellwig authored
di_flags2 is initialized to zero for v4 and earlier file systems. This means di_flags2 can only be non-zero for a v5 file systems, in which case both the parent and child inodes can store the field. Remove the extra di_version check, and also remove the rather pointless local di_flags2 variable while at it. Signed-off-by:
-
Christoph Hellwig authored
The size of the dinode structure is only dependent on the file system version, so instead of checking the individual inode version just use the newly added xfs_sb_version_has_large_dinode helper, and simplify various calling conventions. Signed-off-by:
-
Christoph Hellwig authored
Add a new wrapper to check if a file system supports the v3 inode format with a larger dinode core. Previously we used xfs_sb_version_hascrc for that, which is technically correct but a little confusing to read. Also move xfs_dinode_good_version next to xfs_sb_version_has_v3inode so that we have one place that documents the superblock version to inode version relationship. Signed-off-by:
-
