| 1 | =======================
|
| 2 | Direct Access for files
|
| 3 | =======================
|
| 4 |
|
| 5 | Motivation
|
| 6 | ----------
|
| 7 |
|
| 8 | The page cache is usually used to buffer reads and writes to files.
|
| 9 | It is also used to provide the pages which are mapped into userspace
|
| 10 | by a call to mmap.
|
| 11 |
|
| 12 | For block devices that are memory-like, the page cache pages would be
|
| 13 | unnecessary copies of the original storage. The `DAX` code removes the
|
| 14 | extra copy by performing reads and writes directly to the storage device.
|
| 15 | For file mappings, the storage device is mapped directly into userspace.
|
| 16 |
|
| 17 |
|
| 18 | Usage
|
| 19 | -----
|
| 20 |
|
| 21 | If you have a block device which supports `DAX`, you can make a filesystem
|
| 22 | on it as usual. The `DAX` code currently only supports files with a block
|
| 23 | size equal to your kernel's `PAGE_SIZE`, so you may need to specify a block
|
| 24 | size when creating the filesystem.
|
| 25 |
|
| 26 | Currently 5 filesystems support `DAX`: ext2, ext4, xfs, virtiofs and erofs.
|
| 27 | Enabling `DAX` on them is different.
|
| 28 |
|
| 29 | Enabling DAX on ext2 and erofs
|
| 30 | ------------------------------
|
| 31 |
|
| 32 | When mounting the filesystem, use the ``-o dax`` option on the command line or
|
| 33 | add 'dax' to the options in ``/etc/fstab``. This works to enable `DAX` on all files
|
| 34 | within the filesystem. It is equivalent to the ``-o dax=always`` behavior below.
|
| 35 |
|
| 36 |
|
| 37 | Enabling DAX on xfs and ext4
|
| 38 | ----------------------------
|
| 39 |
|
| 40 | Summary
|
| 41 | -------
|
| 42 |
|
| 43 | 1. There exists an in-kernel file access mode flag `S_DAX` that corresponds to
|
| 44 | the statx flag `STATX_ATTR_DAX`. See the manpage for statx(2) for details
|
| 45 | about this access mode.
|
| 46 |
|
| 47 | 2. There exists a persistent flag `FS_XFLAG_DAX` that can be applied to regular
|
| 48 | files and directories. This advisory flag can be set or cleared at any
|
| 49 | time, but doing so does not immediately affect the `S_DAX` state.
|
| 50 |
|
| 51 | 3. If the persistent `FS_XFLAG_DAX` flag is set on a directory, this flag will
|
| 52 | be inherited by all regular files and subdirectories that are subsequently
|
| 53 | created in this directory. Files and subdirectories that exist at the time
|
| 54 | this flag is set or cleared on the parent directory are not modified by
|
| 55 | this modification of the parent directory.
|
| 56 |
|
| 57 | 4. There exist dax mount options which can override `FS_XFLAG_DAX` in the
|
| 58 | setting of the `S_DAX` flag. Given underlying storage which supports `DAX` the
|
| 59 | following hold:
|
| 60 |
|
| 61 | ``-o dax=inode`` means "follow `FS_XFLAG_DAX`" and is the default.
|
| 62 |
|
| 63 | ``-o dax=never`` means "never set `S_DAX`, ignore `FS_XFLAG_DAX`."
|
| 64 |
|
| 65 | ``-o dax=always`` means "always set `S_DAX` ignore `FS_XFLAG_DAX`."
|
| 66 |
|
| 67 | ``-o dax`` is a legacy option which is an alias for ``dax=always``.
|
| 68 |
|
| 69 | .. warning::
|
| 70 |
|
| 71 | The option ``-o dax`` may be removed in the future so ``-o dax=always`` is
|
| 72 | the preferred method for specifying this behavior.
|
| 73 |
|
| 74 | .. note::
|
| 75 |
|
| 76 | Modifications to and the inheritance behavior of `FS_XFLAG_DAX` remain
|
| 77 | the same even when the filesystem is mounted with a dax option. However,
|
| 78 | in-core inode state (`S_DAX`) will be overridden until the filesystem is
|
| 79 | remounted with dax=inode and the inode is evicted from kernel memory.
|
| 80 |
|
| 81 | 5. The `S_DAX` policy can be changed via:
|
| 82 |
|
| 83 | a) Setting the parent directory `FS_XFLAG_DAX` as needed before files are
|
| 84 | created
|
| 85 |
|
| 86 | b) Setting the appropriate dax="foo" mount option
|
| 87 |
|
| 88 | c) Changing the `FS_XFLAG_DAX` flag on existing regular files and
|
| 89 | directories. This has runtime constraints and limitations that are
|
| 90 | described in 6) below.
|
| 91 |
|
| 92 | 6. When changing the `S_DAX` policy via toggling the persistent `FS_XFLAG_DAX`
|
| 93 | flag, the change to existing regular files won't take effect until the
|
| 94 | files are closed by all processes.
|
| 95 |
|
| 96 |
|
| 97 | Details
|
| 98 | -------
|
| 99 |
|
| 100 | There are 2 per-file dax flags. One is a persistent inode setting (`FS_XFLAG_DAX`)
|
| 101 | and the other is a volatile flag indicating the active state of the feature
|
| 102 | (`S_DAX`).
|
| 103 |
|
| 104 | `FS_XFLAG_DAX` is preserved within the filesystem. This persistent config
|
| 105 | setting can be set, cleared and/or queried using the `FS_IOC_FS`[`GS`]`ETXATTR` ioctl
|
| 106 | (see ioctl_xfs_fsgetxattr(2)) or an utility such as 'xfs_io'.
|
| 107 |
|
| 108 | New files and directories automatically inherit `FS_XFLAG_DAX` from
|
| 109 | their parent directory **when created**. Therefore, setting `FS_XFLAG_DAX` at
|
| 110 | directory creation time can be used to set a default behavior for an entire
|
| 111 | sub-tree.
|
| 112 |
|
| 113 | To clarify inheritance, here are 3 examples:
|
| 114 |
|
| 115 | Example A:
|
| 116 |
|
| 117 | .. code-block:: shell
|
| 118 |
|
| 119 | mkdir -p a/b/c
|
| 120 | xfs_io -c 'chattr +x' a
|
| 121 | mkdir a/b/c/d
|
| 122 | mkdir a/e
|
| 123 |
|
| 124 | ------[outcome]------
|
| 125 |
|
| 126 | dax: a,e
|
| 127 | no dax: b,c,d
|
| 128 |
|
| 129 | Example B:
|
| 130 |
|
| 131 | .. code-block:: shell
|
| 132 |
|
| 133 | mkdir a
|
| 134 | xfs_io -c 'chattr +x' a
|
| 135 | mkdir -p a/b/c/d
|
| 136 |
|
| 137 | ------[outcome]------
|
| 138 |
|
| 139 | dax: a,b,c,d
|
| 140 | no dax:
|
| 141 |
|
| 142 | Example C:
|
| 143 |
|
| 144 | .. code-block:: shell
|
| 145 |
|
| 146 | mkdir -p a/b/c
|
| 147 | xfs_io -c 'chattr +x' c
|
| 148 | mkdir a/b/c/d
|
| 149 |
|
| 150 | ------[outcome]------
|
| 151 |
|
| 152 | dax: c,d
|
| 153 | no dax: a,b
|
| 154 |
|
| 155 | The current enabled state (`S_DAX`) is set when a file inode is instantiated in
|
| 156 | memory by the kernel. It is set based on the underlying media support, the
|
| 157 | value of `FS_XFLAG_DAX` and the filesystem's dax mount option.
|
| 158 |
|
| 159 | statx can be used to query `S_DAX`.
|
| 160 |
|
| 161 | .. note::
|
| 162 |
|
| 163 | That only regular files will ever have `S_DAX` set and therefore statx
|
| 164 | will never indicate that `S_DAX` is set on directories.
|
| 165 |
|
| 166 | Setting the `FS_XFLAG_DAX` flag (specifically or through inheritance) occurs even
|
| 167 | if the underlying media does not support dax and/or the filesystem is
|
| 168 | overridden with a mount option.
|
| 169 |
|
| 170 |
|
| 171 | Enabling DAX on virtiofs
|
| 172 | ----------------------------
|
| 173 | The semantic of DAX on virtiofs is basically equal to that on ext4 and xfs,
|
| 174 | except that when '-o dax=inode' is specified, virtiofs client derives the hint
|
| 175 | whether DAX shall be enabled or not from virtiofs server through FUSE protocol,
|
| 176 | rather than the persistent `FS_XFLAG_DAX` flag. That is, whether DAX shall be
|
| 177 | enabled or not is completely determined by virtiofs server, while virtiofs
|
| 178 | server itself may deploy various algorithm making this decision, e.g. depending
|
| 179 | on the persistent `FS_XFLAG_DAX` flag on the host.
|
| 180 |
|
| 181 | It is still supported to set or clear persistent `FS_XFLAG_DAX` flag inside
|
| 182 | guest, but it is not guaranteed that DAX will be enabled or disabled for
|
| 183 | corresponding file then. Users inside guest still need to call statx(2) and
|
| 184 | check the statx flag `STATX_ATTR_DAX` to see if DAX is enabled for this file.
|
| 185 |
|
| 186 |
|
| 187 | Implementation Tips for Block Driver Writers
|
| 188 | --------------------------------------------
|
| 189 |
|
| 190 | To support `DAX` in your block driver, implement the 'direct_access'
|
| 191 | block device operation. It is used to translate the sector number
|
| 192 | (expressed in units of 512-byte sectors) to a page frame number (pfn)
|
| 193 | that identifies the physical page for the memory. It also returns a
|
| 194 | kernel virtual address that can be used to access the memory.
|
| 195 |
|
| 196 | The direct_access method takes a 'size' parameter that indicates the
|
| 197 | number of bytes being requested. The function should return the number
|
| 198 | of bytes that can be contiguously accessed at that offset. It may also
|
| 199 | return a negative errno if an error occurs.
|
| 200 |
|
| 201 | In order to support this method, the storage must be byte-accessible by
|
| 202 | the CPU at all times. If your device uses paging techniques to expose
|
| 203 | a large amount of memory through a smaller window, then you cannot
|
| 204 | implement direct_access. Equally, if your device can occasionally
|
| 205 | stall the CPU for an extended period, you should also not attempt to
|
| 206 | implement direct_access.
|
| 207 |
|
| 208 | These block devices may be used for inspiration:
|
| 209 | - pmem: NVDIMM persistent memory driver
|
| 210 |
|
| 211 |
|
| 212 | Implementation Tips for Filesystem Writers
|
| 213 | ------------------------------------------
|
| 214 |
|
| 215 | Filesystem support consists of:
|
| 216 |
|
| 217 | * Adding support to mark inodes as being `DAX` by setting the `S_DAX` flag in
|
| 218 | i_flags
|
| 219 | * Implementing ->read_iter and ->write_iter operations which use
|
| 220 | :c:func:`dax_iomap_rw()` when inode has `S_DAX` flag set
|
| 221 | * Implementing an mmap file operation for `DAX` files which sets the
|
| 222 | `VM_MIXEDMAP` and `VM_HUGEPAGE` flags on the `VMA`, and setting the vm_ops to
|
| 223 | include handlers for fault, pmd_fault, page_mkwrite, pfn_mkwrite. These
|
| 224 | handlers should probably call :c:func:`dax_iomap_fault()` passing the
|
| 225 | appropriate fault size and iomap operations.
|
| 226 | * Calling :c:func:`iomap_zero_range()` passing appropriate iomap operations
|
| 227 | instead of :c:func:`block_truncate_page()` for `DAX` files
|
| 228 | * Ensuring that there is sufficient locking between reads, writes,
|
| 229 | truncates and page faults
|
| 230 |
|
| 231 | The iomap handlers for allocating blocks must make sure that allocated blocks
|
| 232 | are zeroed out and converted to written extents before being returned to avoid
|
| 233 | exposure of uninitialized data through mmap.
|
| 234 |
|
| 235 | These filesystems may be used for inspiration:
|
| 236 |
|
| 237 | .. seealso::
|
| 238 |
|
| 239 | ext2: see Documentation/filesystems/ext2.rst
|
| 240 |
|
| 241 | .. seealso::
|
| 242 |
|
| 243 | xfs: see Documentation/admin-guide/xfs.rst
|
| 244 |
|
| 245 | .. seealso::
|
| 246 |
|
| 247 | ext4: see Documentation/filesystems/ext4/
|
| 248 |
|
| 249 |
|
| 250 | Handling Media Errors
|
| 251 | ---------------------
|
| 252 |
|
| 253 | The libnvdimm subsystem stores a record of known media error locations for
|
| 254 | each pmem block device (in gendisk->badblocks). If we fault at such location,
|
| 255 | or one with a latent error not yet discovered, the application can expect
|
| 256 | to receive a `SIGBUS`. Libnvdimm also allows clearing of these errors by simply
|
| 257 | writing the affected sectors (through the pmem driver, and if the underlying
|
| 258 | NVDIMM supports the clear_poison DSM defined by ACPI).
|
| 259 |
|
| 260 | Since `DAX` IO normally doesn't go through the ``driver/bio`` path, applications or
|
| 261 | sysadmins have an option to restore the lost data from a prior ``backup/inbuilt``
|
| 262 | redundancy in the following ways:
|
| 263 |
|
| 264 | 1. Delete the affected file, and restore from a backup (sysadmin route):
|
| 265 | This will free the filesystem blocks that were being used by the file,
|
| 266 | and the next time they're allocated, they will be zeroed first, which
|
| 267 | happens through the driver, and will clear bad sectors.
|
| 268 |
|
| 269 | 2. Truncate or hole-punch the part of the file that has a bad-block (at least
|
| 270 | an entire aligned sector has to be hole-punched, but not necessarily an
|
| 271 | entire filesystem block).
|
| 272 |
|
| 273 | These are the two basic paths that allow `DAX` filesystems to continue operating
|
| 274 | in the presence of media errors. More robust error recovery mechanisms can be
|
| 275 | built on top of this in the future, for example, involving redundancy/mirroring
|
| 276 | provided at the block layer through DM, or additionally, at the filesystem
|
| 277 | level. These would have to rely on the above two tenets, that error clearing
|
| 278 | can happen either by sending an IO through the driver, or zeroing (also through
|
| 279 | the driver).
|
| 280 |
|
| 281 |
|
| 282 | Shortcomings
|
| 283 | ------------
|
| 284 |
|
| 285 | Even if the kernel or its modules are stored on a filesystem that supports
|
| 286 | `DAX` on a block device that supports `DAX`, they will still be copied into RAM.
|
| 287 |
|
| 288 | The DAX code does not work correctly on architectures which have virtually
|
| 289 | mapped caches such as ARM, MIPS and SPARC.
|
| 290 |
|
| 291 | Calling :c:func:`get_user_pages()` on a range of user memory that has been
|
| 292 | mmapped from a `DAX` file will fail when there are no 'struct page' to describe
|
| 293 | those pages. This problem has been addressed in some device drivers
|
| 294 | by adding optional struct page support for pages under the control of
|
| 295 | the driver (see `CONFIG_NVDIMM_PFN` in ``drivers/nvdimm`` for an example of
|
| 296 | how to do this). In the non struct page cases `O_DIRECT` reads/writes to
|
| 297 | those memory ranges from a non-`DAX` file will fail
|
| 298 |
|
| 299 |
|
| 300 | .. note::
|
| 301 |
|
| 302 | `O_DIRECT` reads/writes _of a `DAX` file do work, it is the memory that
|
| 303 | is being accessed that is key here). Other things that will not work in
|
| 304 | the non struct page case include RDMA, :c:func:`sendfile()` and
|
| 305 | :c:func:`splice()`.
|