dax.txt 10 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268
  1. Direct Access for files
  2. -----------------------
  3. Motivation
  4. ----------
  5. The page cache is usually used to buffer reads and writes to files.
  6. It is also used to provide the pages which are mapped into userspace
  7. by a call to mmap.
  8. For block devices that are memory-like, the page cache pages would be
  9. unnecessary copies of the original storage. The DAX code removes the
  10. extra copy by performing reads and writes directly to the storage device.
  11. For file mappings, the storage device is mapped directly into userspace.
  12. Usage
  13. -----
  14. If you have a block device which supports DAX, you can make a filesystem
  15. on it as usual. The DAX code currently only supports files with a block
  16. size equal to your kernel's PAGE_SIZE, so you may need to specify a block
  17. size when creating the filesystem.
  18. Currently 3 filesystems support DAX: ext2, ext4 and xfs. Enabling DAX on them
  19. is different.
  20. Enabling DAX on ext2
  21. -----------------------------
  22. When mounting the filesystem, use the "-o dax" option on the command line or
  23. add 'dax' to the options in /etc/fstab. This works to enable DAX on all files
  24. within the filesystem. It is equivalent to the '-o dax=always' behavior below.
  25. Enabling DAX on xfs and ext4
  26. ----------------------------
  27. Summary
  28. -------
  29. 1. There exists an in-kernel file access mode flag S_DAX that corresponds to
  30. the statx flag STATX_ATTR_DAX. See the manpage for statx(2) for details
  31. about this access mode.
  32. 2. There exists a persistent flag FS_XFLAG_DAX that can be applied to regular
  33. files and directories. This advisory flag can be set or cleared at any
  34. time, but doing so does not immediately affect the S_DAX state.
  35. 3. If the persistent FS_XFLAG_DAX flag is set on a directory, this flag will
  36. be inherited by all regular files and subdirectories that are subsequently
  37. created in this directory. Files and subdirectories that exist at the time
  38. this flag is set or cleared on the parent directory are not modified by
  39. this modification of the parent directory.
  40. 4. There exist dax mount options which can override FS_XFLAG_DAX in the
  41. setting of the S_DAX flag. Given underlying storage which supports DAX the
  42. following hold:
  43. "-o dax=inode" means "follow FS_XFLAG_DAX" and is the default.
  44. "-o dax=never" means "never set S_DAX, ignore FS_XFLAG_DAX."
  45. "-o dax=always" means "always set S_DAX ignore FS_XFLAG_DAX."
  46. "-o dax" is a legacy option which is an alias for "dax=always".
  47. This may be removed in the future so "-o dax=always" is
  48. the preferred method for specifying this behavior.
  49. NOTE: Modifications to and the inheritance behavior of FS_XFLAG_DAX remain
  50. the same even when the filesystem is mounted with a dax option. However,
  51. in-core inode state (S_DAX) will be overridden until the filesystem is
  52. remounted with dax=inode and the inode is evicted from kernel memory.
  53. 5. The S_DAX policy can be changed via:
  54. a) Setting the parent directory FS_XFLAG_DAX as needed before files are
  55. created
  56. b) Setting the appropriate dax="foo" mount option
  57. c) Changing the FS_XFLAG_DAX flag on existing regular files and
  58. directories. This has runtime constraints and limitations that are
  59. described in 6) below.
  60. 6. When changing the S_DAX policy via toggling the persistent FS_XFLAG_DAX flag,
  61. the change in behaviour for existing regular files may not occur
  62. immediately. If the change must take effect immediately, the administrator
  63. needs to:
  64. a) stop the application so there are no active references to the data set
  65. the policy change will affect
  66. b) evict the data set from kernel caches so it will be re-instantiated when
  67. the application is restarted. This can be achieved by:
  68. i. drop-caches
  69. ii. a filesystem unmount and mount cycle
  70. iii. a system reboot
  71. Details
  72. -------
  73. There are 2 per-file dax flags. One is a persistent inode setting (FS_XFLAG_DAX)
  74. and the other is a volatile flag indicating the active state of the feature
  75. (S_DAX).
  76. FS_XFLAG_DAX is preserved within the filesystem. This persistent config
  77. setting can be set, cleared and/or queried using the FS_IOC_FS[GS]ETXATTR ioctl
  78. (see ioctl_xfs_fsgetxattr(2)) or an utility such as 'xfs_io'.
  79. New files and directories automatically inherit FS_XFLAG_DAX from
  80. their parent directory _when_ _created_. Therefore, setting FS_XFLAG_DAX at
  81. directory creation time can be used to set a default behavior for an entire
  82. sub-tree.
  83. To clarify inheritance, here are 3 examples:
  84. Example A:
  85. mkdir -p a/b/c
  86. xfs_io -c 'chattr +x' a
  87. mkdir a/b/c/d
  88. mkdir a/e
  89. dax: a,e
  90. no dax: b,c,d
  91. Example B:
  92. mkdir a
  93. xfs_io -c 'chattr +x' a
  94. mkdir -p a/b/c/d
  95. dax: a,b,c,d
  96. no dax:
  97. Example C:
  98. mkdir -p a/b/c
  99. xfs_io -c 'chattr +x' c
  100. mkdir a/b/c/d
  101. dax: c,d
  102. no dax: a,b
  103. The current enabled state (S_DAX) is set when a file inode is instantiated in
  104. memory by the kernel. It is set based on the underlying media support, the
  105. value of FS_XFLAG_DAX and the filesystem's dax mount option.
  106. statx can be used to query S_DAX. NOTE that only regular files will ever have
  107. S_DAX set and therefore statx will never indicate that S_DAX is set on
  108. directories.
  109. Setting the FS_XFLAG_DAX flag (specifically or through inheritance) occurs even
  110. if the underlying media does not support dax and/or the filesystem is
  111. overridden with a mount option.
  112. Implementation Tips for Block Driver Writers
  113. --------------------------------------------
  114. To support DAX in your block driver, implement the 'direct_access'
  115. block device operation. It is used to translate the sector number
  116. (expressed in units of 512-byte sectors) to a page frame number (pfn)
  117. that identifies the physical page for the memory. It also returns a
  118. kernel virtual address that can be used to access the memory.
  119. The direct_access method takes a 'size' parameter that indicates the
  120. number of bytes being requested. The function should return the number
  121. of bytes that can be contiguously accessed at that offset. It may also
  122. return a negative errno if an error occurs.
  123. In order to support this method, the storage must be byte-accessible by
  124. the CPU at all times. If your device uses paging techniques to expose
  125. a large amount of memory through a smaller window, then you cannot
  126. implement direct_access. Equally, if your device can occasionally
  127. stall the CPU for an extended period, you should also not attempt to
  128. implement direct_access.
  129. These block devices may be used for inspiration:
  130. - brd: RAM backed block device driver
  131. - dcssblk: s390 dcss block device driver
  132. - pmem: NVDIMM persistent memory driver
  133. Implementation Tips for Filesystem Writers
  134. ------------------------------------------
  135. Filesystem support consists of
  136. - adding support to mark inodes as being DAX by setting the S_DAX flag in
  137. i_flags
  138. - implementing ->read_iter and ->write_iter operations which use dax_iomap_rw()
  139. when inode has S_DAX flag set
  140. - implementing an mmap file operation for DAX files which sets the
  141. VM_MIXEDMAP and VM_HUGEPAGE flags on the VMA, and setting the vm_ops to
  142. include handlers for fault, pmd_fault, page_mkwrite, pfn_mkwrite. These
  143. handlers should probably call dax_iomap_fault() passing the appropriate
  144. fault size and iomap operations.
  145. - calling iomap_zero_range() passing appropriate iomap operations instead of
  146. block_truncate_page() for DAX files
  147. - ensuring that there is sufficient locking between reads, writes,
  148. truncates and page faults
  149. The iomap handlers for allocating blocks must make sure that allocated blocks
  150. are zeroed out and converted to written extents before being returned to avoid
  151. exposure of uninitialized data through mmap.
  152. These filesystems may be used for inspiration:
  153. - ext2: see Documentation/filesystems/ext2.rst
  154. - ext4: see Documentation/filesystems/ext4/
  155. - xfs: see Documentation/admin-guide/xfs.rst
  156. Handling Media Errors
  157. ---------------------
  158. The libnvdimm subsystem stores a record of known media error locations for
  159. each pmem block device (in gendisk->badblocks). If we fault at such location,
  160. or one with a latent error not yet discovered, the application can expect
  161. to receive a SIGBUS. Libnvdimm also allows clearing of these errors by simply
  162. writing the affected sectors (through the pmem driver, and if the underlying
  163. NVDIMM supports the clear_poison DSM defined by ACPI).
  164. Since DAX IO normally doesn't go through the driver/bio path, applications or
  165. sysadmins have an option to restore the lost data from a prior backup/inbuilt
  166. redundancy in the following ways:
  167. 1. Delete the affected file, and restore from a backup (sysadmin route):
  168. This will free the filesystem blocks that were being used by the file,
  169. and the next time they're allocated, they will be zeroed first, which
  170. happens through the driver, and will clear bad sectors.
  171. 2. Truncate or hole-punch the part of the file that has a bad-block (at least
  172. an entire aligned sector has to be hole-punched, but not necessarily an
  173. entire filesystem block).
  174. These are the two basic paths that allow DAX filesystems to continue operating
  175. in the presence of media errors. More robust error recovery mechanisms can be
  176. built on top of this in the future, for example, involving redundancy/mirroring
  177. provided at the block layer through DM, or additionally, at the filesystem
  178. level. These would have to rely on the above two tenets, that error clearing
  179. can happen either by sending an IO through the driver, or zeroing (also through
  180. the driver).
  181. Shortcomings
  182. ------------
  183. Even if the kernel or its modules are stored on a filesystem that supports
  184. DAX on a block device that supports DAX, they will still be copied into RAM.
  185. The DAX code does not work correctly on architectures which have virtually
  186. mapped caches such as ARM, MIPS and SPARC.
  187. Calling get_user_pages() on a range of user memory that has been mmaped
  188. from a DAX file will fail when there are no 'struct page' to describe
  189. those pages. This problem has been addressed in some device drivers
  190. by adding optional struct page support for pages under the control of
  191. the driver (see CONFIG_NVDIMM_PFN in drivers/nvdimm for an example of
  192. how to do this). In the non struct page cases O_DIRECT reads/writes to
  193. those memory ranges from a non-DAX file will fail (note that O_DIRECT
  194. reads/writes _of a DAX file_ do work, it is the memory that is being
  195. accessed that is key here). Other things that will not work in the
  196. non struct page case include RDMA, sendfile() and splice().