btrfs-kernel.md

Recommended Kernel Version for bees

First, a warning that is not specific to bees:

Kernel 5.1, 5.2, and 5.3 should not be used with btrfs due to a severe regression that can lead to fatal metadata corruption. This issue is fixed in kernel 5.4.14 and later.

Recommended kernel versions for bees are 4.19, 5.4, 5.10, 5.11, 5.15, 6.0, or 6.1, with recent LTS and -stable updates. The latest released kernel as of this writing is 6.4.1.

4.14, 4.9, and 4.4 LTS kernels with recent updates are OK with some issues. Older kernels will be slower (a little slower or a lot slower depending on which issues are triggered). Not all fixes are backported.

Obsolete non-LTS kernels have a variety of unfixed issues and should not be used with btrfs. For details see the table below.

bees requires btrfs kernel API version 4.2 or higher, and does not work at all on older kernels.

Some bees features rely on kernel 4.15 to work, and these features will not be available on older kernels. Currently, bees is still usable on older kernels with degraded performance or with options disabled, but support for older kernels may be removed.

Kernel Bug Tracking Table

These bugs are particularly popular among bees users, though not all are specifically relevant to bees:

First bad kernel	Last bad kernel	Issue Description	Fixed Kernel Versions	Fix Commit
-	4.10	garbage inserted in read data when reading compressed inline extent followed by a hole	3.18.89, 4.1.49, 4.4.107, 4.9.71, 4.11 and later	e1699d2d7bf6 btrfs: add missing memset while reading compressed inline extents
-	4.14	spurious warnings from fs/btrfs/backref.c in find_parent_nodes	3.16.57, 4.14.29, 4.15.12, 4.16 and later	c8195a7b1ad5 btrfs: remove spurious WARN_ON(ref->count < 0) in find_parent_nodes
4.15	4.18	compression ratio and performance regression on bees test corpus	improved in 4.19	4.14 performance not fully restored yet
-	5.0	silently corrupted data returned when reading compressed extents around a punched hole (bees dedupes all-zero data blocks with holes which can produce a similar effect to hole punching)	3.16.70, 3.18.137, 4.4.177, 4.9.165, 4.14.108, 4.19.31, 5.0.4, 5.1 and later	8e928218780e Btrfs: fix corruption reading shared and compressed extents after hole punching
-	5.0	deadlock when dedupe and rename are used simultaneously on the same files	5.0.4, 5.1 and later	4ea748e1d2c9 Btrfs: fix deadlock between clone/dedupe and rename
-	5.1	send failure or kernel crash while running send and dedupe on same snapshot at same time	5.0.18, 5.1.4, 5.2 and later	62d54f3a7fa2 Btrfs: fix race between send and deduplication that lead to failures and crashes
-	5.2	alternating send and dedupe results in incremental send failure	4.9.188, 4.14.137, 4.19.65, 5.2.7, 5.3 and later	b4f9a1a87a48 Btrfs: fix incremental send failure after deduplication
4.20	5.3	balance convert to single rejected with error on 32-bit CPUs	5.3.7, 5.4 and later	7a54789074a5 btrfs: fix balance convert to single on 32-bit host CPUs
-	5.3	kernel crash due to tree mod log issue #1 (often triggered by bees)	3.16.79, 4.4.195, 4.9.195, 4.14.147, 4.19.77, 5.2.19, 5.3.4, 5.4 and later	efad8a853ad2 Btrfs: fix use-after-free when using the tree modification log
-	5.4	kernel crash due to tree mod log issue #2 (often triggered by bees)	3.16.83, 4.4.208, 4.9.208, 4.14.161, 4.19.92, 5.4.7, 5.5 and later	6609fee8897a Btrfs: fix removal logic of the tree mod log that leads to use-after-free issues
5.1	5.4	metadata corruption resulting in loss of filesystem when a write operation occurs while balance starts a new block group. Do not use kernel 5.1 with btrfs. Kernel 5.2 and 5.3 have workarounds that may detect corruption in progress and abort before it becomes permanent, but do not prevent corruption from occurring. Also kernel crash due to tree mod log issue #4.	5.4.14, 5.5 and later	6282675e6708 btrfs: relocation: fix reloc_root lifespan and access
-	5.4	send performance failure when shared extents have too many references	4.9.207, 4.14.159, 4.19.90, 5.3.17, 5.4.4, 5.5 and later	fd0ddbe25095 Btrfs: send, skip backreference walking for extents with many references
5.0	5.5	dedupe fails to remove the last extent in a file if the file size is not a multiple of 4K	5.4.19, 5.5.3, 5.6 and later	831d2fa25ab8 Btrfs: make deduplication with range including the last block work
4.5, backported to 3.18.31, 4.1.22, 4.4.4	5.5	df incorrectly reports 0 free space while data space is available. Triggered by changes in metadata size, including those typical of large-scale dedupe. Occurs more often starting in 5.3 and especially 5.4	4.4.213, 4.9.213, 4.14.170, 4.19.102, 5.4.18, 5.5.2, 5.6 and later	d55966c4279b btrfs: do not zero f_bavail if we have available space
-	5.5	kernel crash due to tree mod log issue #3 (often triggered by bees)	3.16.84, 4.4.214, 4.9.214, 4.14.171, 4.19.103, 5.4.19, 5.5.3, 5.6 and later	7227ff4de55d Btrfs: fix race between adding and putting tree mod seq elements and nodes
-	5.6	deadlock when enumerating file references to physical extent addresses while some references still exist in deleted subvols	5.7 and later	39dba8739c4e btrfs: do not resolve backrefs for roots that are being deleted
-	5.6	deadlock when many extent reference updates are pending and available memory is low	4.14.177, 4.19.116, 5.4.33, 5.5.18, 5.6.5, 5.7 and later	351cbf6e4410 btrfs: use nofs allocations for running delayed items
-	5.6	excessive CPU usage in LOGICAL_INO and FIEMAP ioctl and increased btrfs write latency in other processes when bees translates from extent physical address to list of referencing files and offsets. Also affects other tools like duperemove and btrfs send	5.4.96, 5.7 and later	b25b0b871f20 btrfs: backref, use correct count to resolve normal data refs, plus 3 parent commits. Some improvements also in earlier kernels.
-	5.7	filesystem becomes read-only if out of space while deleting snapshot	4.9.238, 4.14.200, 4.19.149, 5.4.69, 5.8 and later	7c09c03091ac btrfs: don't force read-only after error in drop snapshot
5.1	5.7	balance, device delete, or filesystem shrink operations loop endlessly on a single block group without decreasing extent count	5.4.54, 5.7.11, 5.8 and later	1dae7e0e58b4 btrfs: reloc: clear DEAD_RELOC_TREE bit for orphan roots to prevent runaway balance
-	5.8	deadlock in TREE_SEARCH ioctl (core component of bees filesystem scanner), followed by regression in deadlock fix	4.4.237, 4.9.237, 4.14.199, 4.19.146, 5.4.66, 5.8.10 and later	a48b73eca4ce btrfs: fix potential deadlock in the search ioctl, 1c78544eaa46 btrfs: fix wrong address when faulting in pages in the search ioctl
5.7	5.10	kernel crash if balance receives fatal signal e.g. Ctrl-C	5.4.93, 5.10.11, 5.11 and later	18d3bff411c8 btrfs: don't get an EINTR during drop_snapshot for reloc
5.10	5.10	20x write performance regression	5.10.8, 5.11 and later	e076ab2a2ca7 btrfs: shrink delalloc pages instead of full inodes
5.4	5.11	spurious tree checker failures on extent ref hash	5.4.125, 5.10.43, 5.11.5, 5.12 and later	1119a72e223f btrfs: tree-checker: do not error out if extent ref hash doesn't match
-	5.11	tree mod log issue #5	4.4.263, 4.9.263, 4.14.227, 4.19.183, 5.4.108, 5.10.26, 5.11.9, 5.12 and later	dbcc7d57bffc btrfs: fix race when cloning extent buffer during rewind of an old root
-	5.12	tree mod log issue #6	4.14.233, 4.19.191, 5.4.118, 5.10.36, 5.11.20, 5.12.3, 5.13 and later	f9690f426b21 btrfs: fix race when picking most recent mod log operation for an old root
4.15	5.16	spurious warnings from fs/fs-writeback.c when flushoncommit is enabled	5.15.27, 5.16.13, 5.17 and later	a0f0cf8341e3 btrfs: get rid of warning on transaction commit when using flushoncommit
-	5.17	crash during device removal can make filesystem unmountable	5.15.54, 5.16.20, 5.17.3, 5.18 and later	bbac58698a55 btrfs: remove device item and update super block in the same transaction
-	5.18	wrong superblock num_devices makes filesystem unmountable	4.14.283, 4.19.247, 5.4.198, 5.10.121, 5.15.46, 5.17.14, 5.18.3, 5.19 and later	d201238ccd2f btrfs: repair super block num_devices automatically
5.18	5.19	parent transid verify failed during log tree replay after a crash during a rename operation	5.18.18, 5.19.2, 6.0 and later	723df2bcc9e1 btrfs: join running log transaction when logging new name
5.12	6.0	space cache corruption and potential double allocations	5.15.65, 5.19.6, 6.0 and later	ced8ecf026fd btrfs: fix space cache corruption and potential double allocations
6.0	6.5	suboptimal allocation in multi-device filesystems due to chunk allocator regression	6.1.60, 6.5.9, 6.6 and later	8a540e990d7d btrfs: fix stripe length calculation for non-zoned data chunk allocation
6.3, backported to 5.15.107, 6.1.24, 6.2.11	6.3	vmalloc error, failed to allocate pages	6.3.10, 6.4 and later. Bug (f349b15e183d "mm: vmalloc: avoid warn_alloc noise caused by fatal signal" in v6.3-rc6) backported to 6.1.24, 6.2.11, and 5.15.107.	95a301eefa82 mm/vmalloc: do not output a spurious warning when huge vmalloc() fails
6.2	6.3	IGNORE_OFFSET flag ignored in LOGICAL_INO ioctl	6.2.16, 6.3.3, 6.4 and later	0cad8f14d70c btrfs: fix backref walking not returning all inode refs
6.10	6.11	adding refs to an existing tree ref, failed to run delayed ref, then read-only	6.11.10, 6.12 and later	7d493a5ecc26 btrfs: fix incorrect comparison for delayed refs
5.4	-	kernel hang when multiple threads are running LOGICAL_INO and dedupe ioctl on the same extent	-	workaround: avoid doing that

"Last bad kernel" refers to that version's last stable update from kernel.org. Distro kernels may backport additional fixes. Consult your distro's kernel support for details.

When the same version appears in both "last bad kernel" and "fixed kernel version" columns, it means the bug appears in the .0 release and is fixed in the stated .y release. e.g. a "last bad kernel" of 5.4 and a "fixed kernel version" of 5.4.14 has the bug in kernel versions 5.4.0 through 5.4.13 inclusive.

A "-" for "first bad kernel" indicates the bug has been present since the relevant feature first appeared in btrfs.

A "-" for "last bad kernel" indicates the bug has not yet been fixed in current kernels (see top of this page for which kernel version that is).

In cases where issues are fixed by commits spread out over multiple kernel versions, "fixed kernel version" refers to the version that contains the last committed component of the fix.

Workarounds for known kernel bugs

• Hangs with concurrent LOGICAL_INO and dedupe: on all kernel versions so far, multiple threads running LOGICAL_INO and dedupe ioctls at the same time on the same inodes or extents can lead to a kernel hang. The kernel enters an infinite loop in add_all_parents, where count is 0, ref->count is 1, and btrfs_next_item or btrfs_next_old_item never find a matching ref).

  bees has two workarounds for this bug: 1. schedule work so that multiple threads do not simultaneously access the same inode or the same extent, and 2. use a brute-force global lock within bees that prevents any thread from running LOGICAL_INO while any other thread is running dedupe.

  Workaround #1 isn't really a workaround, since we want to do the same thing for unrelated performance reasons. If multiple threads try to perform dedupe operations on the same extent or inode, btrfs will make all the threads wait for the same locks anyway, so it's better to have bees find some other inode or extent to work on while waiting for btrfs to finish.

  Workaround #2 doesn't seem to be needed after implementing workaround #1, but it's better to be slightly slower than to hang one CPU core and the filesystem until the kernel is rebooted.

  It is still theoretically possible to trigger the kernel bug when running bees at the same time as other dedupers, or other programs that use LOGICAL_INO like btdu; however, it's extremely difficult to reproduce the bug without closely cooperating threads.

• Slow backrefs (aka toxic extents): Under certain conditions, if the number of references to a single shared extent grows too high, the kernel consumes more and more CPU while also holding locks that delay write access to the filesystem. bees avoids this bug by measuring the time the kernel spends performing LOGICAL_INO operations and permanently blacklisting any extent or hash involved where the kernel starts to get slow. In the bees log, such blocks are labelled as 'toxic' hash/block addresses. Toxic extents are rare (about 1 in 100,000 extents become toxic), but toxic extents can become 8 orders of magnitude more expensive to process than the fastest non-toxic extents. This seems to affect all dedupe agents on btrfs; at this time of writing only bees has a workaround for this bug.

  This workaround is less necessary for kernels 5.4.96, 5.7 and later, though the bees workaround can still be triggered on newer kernels by changes in btrfs since kernel version 5.1.

• dedupe breaks btrfs send in old kernels. The bees option --workaround-btrfs-send prevents any modification of read-only subvols in order to avoid breaking btrfs send.

  This workaround is no longer necessary to avoid kernel crashes and send performance failure on kernel 4.9.207, 4.14.159, 4.19.90, 5.3.17, 5.4.4, 5.5 and later; however, some conflict between send and dedupe still remains, so the workaround is still useful.

  btrfs receive is not and has never been affected by this issue.

Unfixed kernel bugs

• The kernel does not permit btrfs send and dedupe to run at the same time. Recent kernels no longer crash, but now refuse one operation with an error if the other operation was already running.

  bees has not been updated to handle the new dedupe behavior optimally. Optimal behavior is to defer dedupe operations when send is detected, and resume after the send is finished. Current bees behavior is to complain loudly about each individual dedupe failure in log messages, and abandon duplicate data references in the snapshot that send is processing. A future bees version shall have better handling for this situation.

  Workaround: send SIGSTOP to bees, or terminate the bees process, before running btrfs send.

  This workaround is not strictly required if snapshot is deleted after sending. In that case, any duplicate data blocks that were not removed by dedupe will be removed by snapshot delete instead. The workaround still saves some IO.

  btrfs receive is not affected by this issue.