1Direct Access for files 2----------------------- 3 4Motivation 5---------- 6 7The page cache is usually used to buffer reads and writes to files. 8It is also used to provide the pages which are mapped into userspace 9by a call to mmap. 10 11For block devices that are memory-like, the page cache pages would be 12unnecessary copies of the original storage. The DAX code removes the 13extra copy by performing reads and writes directly to the storage device. 14For file mappings, the storage device is mapped directly into userspace. 15 16 17Usage 18----- 19 20If you have a block device which supports DAX, you can make a filesystem 21on it as usual. The DAX code currently only supports files with a block 22size equal to your kernel's PAGE_SIZE, so you may need to specify a block 23size when creating the filesystem. When mounting it, use the "-o dax" 24option on the command line or add 'dax' to the options in /etc/fstab. 25 26 27Implementation Tips for Block Driver Writers 28-------------------------------------------- 29 30To support DAX in your block driver, implement the 'direct_access' 31block device operation. It is used to translate the sector number 32(expressed in units of 512-byte sectors) to a page frame number (pfn) 33that identifies the physical page for the memory. It also returns a 34kernel virtual address that can be used to access the memory. 35 36The direct_access method takes a 'size' parameter that indicates the 37number of bytes being requested. The function should return the number 38of bytes that can be contiguously accessed at that offset. It may also 39return a negative errno if an error occurs. 40 41In order to support this method, the storage must be byte-accessible by 42the CPU at all times. If your device uses paging techniques to expose 43a large amount of memory through a smaller window, then you cannot 44implement direct_access. Equally, if your device can occasionally 45stall the CPU for an extended period, you should also not attempt to 46implement direct_access. 47 48These block devices may be used for inspiration: 49- brd: RAM backed block device driver 50- dcssblk: s390 dcss block device driver 51- pmem: NVDIMM persistent memory driver 52 53 54Implementation Tips for Filesystem Writers 55------------------------------------------ 56 57Filesystem support consists of 58- adding support to mark inodes as being DAX by setting the S_DAX flag in 59 i_flags 60- implementing ->read_iter and ->write_iter operations which use dax_iomap_rw() 61 when inode has S_DAX flag set 62- implementing an mmap file operation for DAX files which sets the 63 VM_MIXEDMAP and VM_HUGEPAGE flags on the VMA, and setting the vm_ops to 64 include handlers for fault, pmd_fault, page_mkwrite, pfn_mkwrite. These 65 handlers should probably call dax_iomap_fault() passing the appropriate 66 fault size and iomap operations. 67- calling iomap_zero_range() passing appropriate iomap operations instead of 68 block_truncate_page() for DAX files 69- ensuring that there is sufficient locking between reads, writes, 70 truncates and page faults 71 72The iomap handlers for allocating blocks must make sure that allocated blocks 73are zeroed out and converted to written extents before being returned to avoid 74exposure of uninitialized data through mmap. 75 76These filesystems may be used for inspiration: 77- ext2: see Documentation/filesystems/ext2.txt 78- ext4: see Documentation/filesystems/ext4/ 79- xfs: see Documentation/admin-guide/xfs.rst 80 81 82Handling Media Errors 83--------------------- 84 85The libnvdimm subsystem stores a record of known media error locations for 86each pmem block device (in gendisk->badblocks). If we fault at such location, 87or one with a latent error not yet discovered, the application can expect 88to receive a SIGBUS. Libnvdimm also allows clearing of these errors by simply 89writing the affected sectors (through the pmem driver, and if the underlying 90NVDIMM supports the clear_poison DSM defined by ACPI). 91 92Since DAX IO normally doesn't go through the driver/bio path, applications or 93sysadmins have an option to restore the lost data from a prior backup/inbuilt 94redundancy in the following ways: 95 961. Delete the affected file, and restore from a backup (sysadmin route): 97 This will free the file system blocks that were being used by the file, 98 and the next time they're allocated, they will be zeroed first, which 99 happens through the driver, and will clear bad sectors. 100 1012. Truncate or hole-punch the part of the file that has a bad-block (at least 102 an entire aligned sector has to be hole-punched, but not necessarily an 103 entire filesystem block). 104 105These are the two basic paths that allow DAX filesystems to continue operating 106in the presence of media errors. More robust error recovery mechanisms can be 107built on top of this in the future, for example, involving redundancy/mirroring 108provided at the block layer through DM, or additionally, at the filesystem 109level. These would have to rely on the above two tenets, that error clearing 110can happen either by sending an IO through the driver, or zeroing (also through 111the driver). 112 113 114Shortcomings 115------------ 116 117Even if the kernel or its modules are stored on a filesystem that supports 118DAX on a block device that supports DAX, they will still be copied into RAM. 119 120The DAX code does not work correctly on architectures which have virtually 121mapped caches such as ARM, MIPS and SPARC. 122 123Calling get_user_pages() on a range of user memory that has been mmaped 124from a DAX file will fail when there are no 'struct page' to describe 125those pages. This problem has been addressed in some device drivers 126by adding optional struct page support for pages under the control of 127the driver (see CONFIG_NVDIMM_PFN in drivers/nvdimm for an example of 128how to do this). In the non struct page cases O_DIRECT reads/writes to 129those memory ranges from a non-DAX file will fail (note that O_DIRECT 130reads/writes _of a DAX file_ do work, it is the memory that is being 131accessed that is key here). Other things that will not work in the 132non struct page case include RDMA, sendfile() and splice(). 133