xref: /linux/Documentation/admin-guide/xfs.rst (revision e445fba2d76369d72b497ecadf6b9787930693d9)
1.. SPDX-License-Identifier: GPL-2.0
2
3======================
4The SGI XFS Filesystem
5======================
6
7XFS is a high performance journaling filesystem which originated
8on the SGI IRIX platform.  It is completely multi-threaded, can
9support large files and large filesystems, extended attributes,
10variable block sizes, is extent based, and makes extensive use of
11Btrees (directories, extents, free space) to aid both performance
12and scalability.
13
14Refer to the documentation at https://xfs.wiki.kernel.org/
15for further details.  This implementation is on-disk compatible
16with the IRIX version of XFS.
17
18
19Mount Options
20=============
21
22When mounting an XFS filesystem, the following options are accepted.
23
24  allocsize=size
25	Sets the buffered I/O end-of-file preallocation size when
26	doing delayed allocation writeout (default size is 64KiB).
27	Valid values for this option are page size (typically 4KiB)
28	through to 1GiB, inclusive, in power-of-2 increments.
29
30	The default behaviour is for dynamic end-of-file
31	preallocation size, which uses a set of heuristics to
32	optimise the preallocation size based on the current
33	allocation patterns within the file and the access patterns
34	to the file. Specifying a fixed ``allocsize`` value turns off
35	the dynamic behaviour.
36
37  discard or nodiscard (default)
38	Enable/disable the issuing of commands to let the block
39	device reclaim space freed by the filesystem.  This is
40	useful for SSD devices, thinly provisioned LUNs and virtual
41	machine images, but may have a performance impact.
42
43	Note: It is currently recommended that you use the ``fstrim``
44	application to ``discard`` unused blocks rather than the ``discard``
45	mount option because the performance impact of this option
46	is quite severe.
47
48  grpid/bsdgroups or nogrpid/sysvgroups (default)
49	These options define what group ID a newly created file
50	gets.  When ``grpid`` is set, it takes the group ID of the
51	directory in which it is created; otherwise it takes the
52	``fsgid`` of the current process, unless the directory has the
53	``setgid`` bit set, in which case it takes the ``gid`` from the
54	parent directory, and also gets the ``setgid`` bit set if it is
55	a directory itself.
56
57  filestreams
58	Make the data allocator use the filestreams allocation mode
59	across the entire filesystem rather than just on directories
60	configured to use it.
61
62  inode32 or inode64 (default)
63	When ``inode32`` is specified, it indicates that XFS limits
64	inode creation to locations which will not result in inode
65	numbers with more than 32 bits of significance.
66
67	When ``inode64`` is specified, it indicates that XFS is allowed
68	to create inodes at any location in the filesystem,
69	including those which will result in inode numbers occupying
70	more than 32 bits of significance.
71
72	``inode32`` is provided for backwards compatibility with older
73	systems and applications, since 64 bits inode numbers might
74	cause problems for some applications that cannot handle
75	large inode numbers.  If applications are in use which do
76	not handle inode numbers bigger than 32 bits, the ``inode32``
77	option should be specified.
78
79  largeio or nolargeio (default)
80	If ``nolargeio`` is specified, the optimal I/O reported in
81	``st_blksize`` by **stat(2)** will be as small as possible to allow
82	user applications to avoid inefficient read/modify/write
83	I/O.  This is typically the page size of the machine, as
84	this is the granularity of the page cache.
85
86	If ``largeio`` is specified, a filesystem that was created with a
87	``swidth`` specified will return the ``swidth`` value (in bytes)
88	in ``st_blksize``. If the filesystem does not have a ``swidth``
89	specified but does specify an ``allocsize`` then ``allocsize``
90	(in bytes) will be returned instead. Otherwise the behaviour
91	is the same as if ``nolargeio`` was specified.
92
93  logbufs=value
94	Set the number of in-memory log buffers.  Valid numbers
95	range from 2-8 inclusive.
96
97	The default value is 8 buffers.
98
99	If the memory cost of 8 log buffers is too high on small
100	systems, then it may be reduced at some cost to performance
101	on metadata intensive workloads. The ``logbsize`` option below
102	controls the size of each buffer and so is also relevant to
103	this case.
104
105  lifetime (default) or nolifetime
106	Enable data placement based on write life time hints provided
107	by the user. This turns on co-allocation of data of similar
108	life times when statistically favorable to reduce garbage
109	collection cost.
110
111	These options are only available for zoned rt file systems.
112
113  logbsize=value
114	Set the size of each in-memory log buffer.  The size may be
115	specified in bytes, or in kilobytes with a "k" suffix.
116	Valid sizes for version 1 and version 2 logs are 16384 (16k)
117	and 32768 (32k).  Valid sizes for version 2 logs also
118	include 65536 (64k), 131072 (128k) and 262144 (256k). The
119	logbsize must be an integer multiple of the log
120	stripe unit configured at **mkfs(8)** time.
121
122	The default value for version 1 logs is 32768, while the
123	default value for version 2 logs is MAX(32768, log_sunit).
124
125  logdev=device and rtdev=device
126	Use an external log (metadata journal) and/or real-time device.
127	An XFS filesystem has up to three parts: a data section, a log
128	section, and a real-time section.  The real-time section is
129	optional, and the log section can be separate from the data
130	section or contained within it.
131
132  max_atomic_write=value
133	Set the maximum size of an atomic write.  The size may be
134	specified in bytes, in kilobytes with a "k" suffix, in megabytes
135	with a "m" suffix, or in gigabytes with a "g" suffix.  The size
136	cannot be larger than the maximum write size, larger than the
137	size of any allocation group, or larger than the size of a
138	remapping operation that the log can complete atomically.
139
140	The default value is to set the maximum I/O completion size
141	to allow each CPU to handle one at a time.
142
143  max_open_zones=value
144	Specify the max number of zones to keep open for writing on a
145	zoned rt device. Many open zones aids file data separation
146	but may impact performance on HDDs.
147
148	If ``max_open_zones`` is not specified, the value is determined
149	by the capabilities and the size of the zoned rt device.
150
151  noalign
152	Data allocations will not be aligned at stripe unit
153	boundaries. This is only relevant to filesystems created
154	with non-zero data alignment parameters (``sunit``, ``swidth``) by
155	**mkfs(8)**.
156
157  norecovery
158	The filesystem will be mounted without running log recovery.
159	If the filesystem was not cleanly unmounted, it is likely to
160	be inconsistent when mounted in ``norecovery`` mode.
161	Some files or directories may not be accessible because of this.
162	Filesystems mounted ``norecovery`` must be mounted read-only or
163	the mount will fail.
164
165  nouuid
166	Don't check for double mounted file systems using the file
167	system ``uuid``.  This is useful to mount LVM snapshot volumes,
168	and often used in combination with ``norecovery`` for mounting
169	read-only snapshots.
170
171  noquota
172	Forcibly turns off all quota accounting and enforcement
173	within the filesystem.
174
175  uquota/usrquota/uqnoenforce/quota
176	User disk quota accounting enabled, and limits (optionally)
177	enforced.  Refer to **xfs_quota(8)** for further details.
178
179  gquota/grpquota/gqnoenforce
180	Group disk quota accounting enabled and limits (optionally)
181	enforced.  Refer to **xfs_quota(8)** for further details.
182
183  pquota/prjquota/pqnoenforce
184	Project disk quota accounting enabled and limits (optionally)
185	enforced.  Refer to **xfs_quota(8)** for further details.
186
187  sunit=value and swidth=value
188	Used to specify the stripe unit and width for a RAID device
189	or a stripe volume.  "value" must be specified in 512-byte
190	block units. These options are only relevant to filesystems
191	that were created with non-zero data alignment parameters.
192
193	The ``sunit`` and ``swidth`` parameters specified must be compatible
194	with the existing filesystem alignment characteristics.  In
195	general, that means the only valid changes to ``sunit`` are
196	increasing it by a power-of-2 multiple. Valid ``swidth`` values
197	are any integer multiple of a valid ``sunit`` value.
198
199	Typically the only time these mount options are necessary if
200	after an underlying RAID device has had its geometry
201	modified, such as adding a new disk to a RAID5 lun and
202	reshaping it.
203
204  swalloc
205	Data allocations will be rounded up to stripe width boundaries
206	when the current end of file is being extended and the file
207	size is larger than the stripe width size.
208
209  wsync
210	When specified, all filesystem namespace operations are
211	executed synchronously. This ensures that when the namespace
212	operation (create, unlink, etc) completes, the change to the
213	namespace is on stable storage. This is useful in HA setups
214	where failover must not result in clients seeing
215	inconsistent namespace presentation during or after a
216	failover event.
217
218Deprecation of V4 Format
219========================
220
221The V4 filesystem format lacks certain features that are supported by
222the V5 format, such as metadata checksumming, strengthened metadata
223verification, and the ability to store timestamps past the year 2038.
224Because of this, the V4 format is deprecated.  All users should upgrade
225by backing up their files, reformatting, and restoring from the backup.
226
227Administrators and users can detect a V4 filesystem by running xfs_info
228against a filesystem mountpoint and checking for a string containing
229"crc=".  If no such string is found, please upgrade xfsprogs to the
230latest version and try again.
231
232The deprecation will take place in two parts.  Support for mounting V4
233filesystems can now be disabled at kernel build time via Kconfig option.
234These options were changed to default to no in September 2025.  In
235September 2030, support will be removed from the codebase entirely.
236
237Note: Distributors may choose to withdraw V4 format support earlier than
238the dates listed above.
239
240Deprecated Mount Options
241========================
242
243============================    ================
244  Name				Removal Schedule
245============================    ================
246Mounting with V4 filesystem     September 2030
247Mounting ascii-ci filesystem    September 2030
248============================    ================
249
250
251Removed Mount Options
252=====================
253
254===========================     =======
255  Name				Removed
256===========================	=======
257  delaylog/nodelaylog		v4.0
258  ihashsize			v4.0
259  irixsgid			v4.0
260  osyncisdsync/osyncisosync	v4.0
261  barrier			v4.19
262  nobarrier			v4.19
263  ikeep/noikeep			v6.18
264  attr2/noattr2			v6.18
265===========================     =======
266
267sysctls
268=======
269
270The following sysctls are available for the XFS filesystem:
271
272  fs.xfs.stats_clear		(Min: 0  Default: 0  Max: 1)
273	Setting this to "1" clears accumulated XFS statistics
274	in /proc/fs/xfs/stat.  It then immediately resets to "0".
275
276  fs.xfs.xfssyncd_centisecs	(Min: 100  Default: 3000  Max: 720000)
277	The interval at which the filesystem flushes metadata
278	out to disk and runs internal cache cleanup routines.
279
280  fs.xfs.filestream_centisecs	(Min: 1  Default: 3000  Max: 360000)
281	The interval at which the filesystem ages filestreams cache
282	references and returns timed-out AGs back to the free stream
283	pool.
284
285  fs.xfs.speculative_prealloc_lifetime
286	(Units: seconds   Min: 1  Default: 300  Max: 86400)
287	The interval at which the background scanning for inodes
288	with unused speculative preallocation runs. The scan
289	removes unused preallocation from clean inodes and releases
290	the unused space back to the free pool.
291
292  fs.xfs.error_level		(Min: 0  Default: 3  Max: 11)
293	A volume knob for error reporting when internal errors occur.
294	This will generate detailed messages & backtraces for filesystem
295	shutdowns, for example.  Current threshold values are:
296
297		XFS_ERRLEVEL_OFF:       0
298		XFS_ERRLEVEL_LOW:       1
299		XFS_ERRLEVEL_HIGH:      5
300
301  fs.xfs.panic_mask		(Min: 0  Default: 0  Max: 511)
302	Causes certain error conditions to call BUG(). Value is a bitmask;
303	OR together the tags which represent errors which should cause panics:
304
305		XFS_NO_PTAG                     0
306		XFS_PTAG_IFLUSH                 0x00000001
307		XFS_PTAG_LOGRES                 0x00000002
308		XFS_PTAG_AILDELETE              0x00000004
309		XFS_PTAG_ERROR_REPORT           0x00000008
310		XFS_PTAG_SHUTDOWN_CORRUPT       0x00000010
311		XFS_PTAG_SHUTDOWN_IOERROR       0x00000020
312		XFS_PTAG_SHUTDOWN_LOGERROR      0x00000040
313		XFS_PTAG_FSBLOCK_ZERO           0x00000080
314		XFS_PTAG_VERIFIER_ERROR         0x00000100
315
316	This option is intended for debugging only.
317
318  fs.xfs.inherit_sync		(Min: 0  Default: 1  Max: 1)
319	Setting this to "1" will cause the "sync" flag set
320	by the **xfs_io(8)** chattr command on a directory to be
321	inherited by files in that directory.
322
323  fs.xfs.inherit_nodump		(Min: 0  Default: 1  Max: 1)
324	Setting this to "1" will cause the "nodump" flag set
325	by the **xfs_io(8)** chattr command on a directory to be
326	inherited by files in that directory.
327
328  fs.xfs.inherit_noatime	(Min: 0  Default: 1  Max: 1)
329	Setting this to "1" will cause the "noatime" flag set
330	by the **xfs_io(8)** chattr command on a directory to be
331	inherited by files in that directory.
332
333  fs.xfs.inherit_nosymlinks	(Min: 0  Default: 1  Max: 1)
334	Setting this to "1" will cause the "nosymlinks" flag set
335	by the **xfs_io(8)** chattr command on a directory to be
336	inherited by files in that directory.
337
338  fs.xfs.inherit_nodefrag	(Min: 0  Default: 1  Max: 1)
339	Setting this to "1" will cause the "nodefrag" flag set
340	by the **xfs_io(8)** chattr command on a directory to be
341	inherited by files in that directory.
342
343  fs.xfs.rotorstep		(Min: 1  Default: 1  Max: 256)
344	In "inode32" allocation mode, this option determines how many
345	files the allocator attempts to allocate in the same allocation
346	group before moving to the next allocation group.  The intent
347	is to control the rate at which the allocator moves between
348	allocation groups when allocating extents for new files.
349
350Deprecated Sysctls
351==================
352
353None currently.
354
355Removed Sysctls
356===============
357
358==========================================   =======
359  Name                                       Removed
360==========================================   =======
361  fs.xfs.xfsbufd_centisec                    v4.0
362  fs.xfs.age_buffer_centisecs                v4.0
363  fs.xfs.irix_symlink_mode                   v6.18
364  fs.xfs.irix_sgid_inherit                   v6.18
365  fs.xfs.speculative_cow_prealloc_lifetime   v6.18
366==========================================   =======
367
368Error handling
369==============
370
371XFS can act differently according to the type of error found during its
372operation. The implementation introduces the following concepts to the error
373handler:
374
375 -failure speed:
376	Defines how fast XFS should propagate an error upwards when a specific
377	error is found during the filesystem operation. It can propagate
378	immediately, after a defined number of retries, after a set time period,
379	or simply retry forever.
380
381 -error classes:
382	Specifies the subsystem the error configuration will apply to, such as
383	metadata IO or memory allocation. Different subsystems will have
384	different error handlers for which behaviour can be configured.
385
386 -error handlers:
387	Defines the behavior for a specific error.
388
389The filesystem behavior during an error can be set via ``sysfs`` files. Each
390error handler works independently - the first condition met by an error handler
391for a specific class will cause the error to be propagated rather than reset and
392retried.
393
394The action taken by the filesystem when the error is propagated is context
395dependent - it may cause a shut down in the case of an unrecoverable error,
396it may be reported back to userspace, or it may even be ignored because
397there's nothing useful we can with the error or anyone we can report it to (e.g.
398during unmount).
399
400The configuration files are organized into the following hierarchy for each
401mounted filesystem:
402
403  /sys/fs/xfs/<dev>/error/<class>/<error>/
404
405Where:
406  <dev>
407	The short device name of the mounted filesystem. This is the same device
408	name that shows up in XFS kernel error messages as "XFS(<dev>): ..."
409
410  <class>
411	The subsystem the error configuration belongs to. As of 4.9, the defined
412	classes are:
413
414		- "metadata": applies metadata buffer write IO
415
416  <error>
417	The individual error handler configurations.
418
419
420Each filesystem has "global" error configuration options defined in their top
421level directory:
422
423  /sys/fs/xfs/<dev>/error/
424
425  fail_at_unmount		(Min:  0  Default:  1  Max: 1)
426	Defines the filesystem error behavior at unmount time.
427
428	If set to a value of 1, XFS will override all other error configurations
429	during unmount and replace them with "immediate fail" characteristics.
430	i.e. no retries, no retry timeout. This will always allow unmount to
431	succeed when there are persistent errors present.
432
433	If set to 0, the configured retry behaviour will continue until all
434	retries and/or timeouts have been exhausted. This will delay unmount
435	completion when there are persistent errors, and it may prevent the
436	filesystem from ever unmounting fully in the case of "retry forever"
437	handler configurations.
438
439	Note: there is no guarantee that fail_at_unmount can be set while an
440	unmount is in progress. It is possible that the ``sysfs`` entries are
441	removed by the unmounting filesystem before a "retry forever" error
442	handler configuration causes unmount to hang, and hence the filesystem
443	must be configured appropriately before unmount begins to prevent
444	unmount hangs.
445
446Each filesystem has specific error class handlers that define the error
447propagation behaviour for specific errors. There is also a "default" error
448handler defined, which defines the behaviour for all errors that don't have
449specific handlers defined. Where multiple retry constraints are configured for
450a single error, the first retry configuration that expires will cause the error
451to be propagated. The handler configurations are found in the directory:
452
453  /sys/fs/xfs/<dev>/error/<class>/<error>/
454
455  max_retries			(Min: -1  Default: Varies  Max: INTMAX)
456	Defines the allowed number of retries of a specific error before
457	the filesystem will propagate the error. The retry count for a given
458	error context (e.g. a specific metadata buffer) is reset every time
459	there is a successful completion of the operation.
460
461	Setting the value to "-1" will cause XFS to retry forever for this
462	specific error.
463
464	Setting the value to "0" will cause XFS to fail immediately when the
465	specific error is reported.
466
467	Setting the value to "N" (where 0 < N < Max) will make XFS retry the
468	operation "N" times before propagating the error.
469
470  retry_timeout_seconds		(Min:  -1  Default:  Varies  Max: 1 day)
471	Define the amount of time (in seconds) that the filesystem is
472	allowed to retry its operations when the specific error is
473	found.
474
475	Setting the value to "-1" will allow XFS to retry forever for this
476	specific error.
477
478	Setting the value to "0" will cause XFS to fail immediately when the
479	specific error is reported.
480
481	Setting the value to "N" (where 0 < N < Max) will allow XFS to retry the
482	operation for up to "N" seconds before propagating the error.
483
484**Note:** The default behaviour for a specific error handler is dependent on both
485the class and error context. For example, the default values for
486"metadata/ENODEV" are "0" rather than "-1" so that this error handler defaults
487to "fail immediately" behaviour. This is done because ENODEV is a fatal,
488unrecoverable error no matter how many times the metadata IO is retried.
489
490Workqueue Concurrency
491=====================
492
493XFS uses kernel workqueues to parallelize metadata update processes.  This
494enables it to take advantage of storage hardware that can service many IO
495operations simultaneously.  This interface exposes internal implementation
496details of XFS, and as such is explicitly not part of any userspace API/ABI
497guarantee the kernel may give userspace.  These are undocumented features of
498the generic workqueue implementation XFS uses for concurrency, and they are
499provided here purely for diagnostic and tuning purposes and may change at any
500time in the future.
501
502The control knobs for a filesystem's workqueues are organized by task at hand
503and the short name of the data device.  They all can be found in:
504
505  /sys/bus/workqueue/devices/${task}!${device}
506
507================  ===========
508  Task            Description
509================  ===========
510  xfs_iwalk-$pid  Inode scans of the entire filesystem. Currently limited to
511                  mount time quotacheck.
512  xfs-gc          Background garbage collection of disk space that have been
513                  speculatively allocated beyond EOF or for staging copy on
514                  write operations.
515================  ===========
516
517For example, the knobs for the quotacheck workqueue for /dev/nvme0n1 would be
518found in /sys/bus/workqueue/devices/xfs_iwalk-1111!nvme0n1/.
519
520The interesting knobs for XFS workqueues are as follows:
521
522============     ===========
523  Knob           Description
524============     ===========
525  max_active     Maximum number of background threads that can be started to
526                 run the work.
527  cpumask        CPUs upon which the threads are allowed to run.
528  nice           Relative priority of scheduling the threads.  These are the
529                 same nice levels that can be applied to userspace processes.
530============     ===========
531
532Zoned Filesystems
533=================
534
535For zoned file systems, the following attributes are exposed in:
536
537  /sys/fs/xfs/<dev>/zoned/
538
539  max_open_zones		(Min:  1  Default:  Varies  Max:  UINTMAX)
540	This read-only attribute exposes the maximum number of open zones
541	available for data placement. The value is determined at mount time and
542	is limited by the capabilities of the backing zoned device, file system
543	size and the max_open_zones mount option.
544
545  zonegc_low_space		(Min:  0  Default:  0  Max:  100)
546	Define a percentage for how much of the unused space that GC should keep
547	available for writing. A high value will reclaim more of the space
548	occupied by unused blocks, creating a larger buffer against write
549	bursts at the cost of increased write amplification.  Regardless
550	of this value, garbage collection will always aim to free a minimum
551	amount of blocks to keep max_open_zones open for data placement purposes.
552