2Data Integrity
   51. Introduction
   8Modern filesystems feature checksumming of data and metadata to
   9protect against data corruption.  However, the detection of the
  10corruption is done at read time which could potentially be months
  11after the data was written.  At that point the original data that the
  12application tried to write is most likely lost.
  14The solution is to ensure that the disk is actually storing what the
  15application meant it to.  Recent additions to both the SCSI family
  16protocols (SBC Data Integrity Field, SCC protection proposal) as well
  17as SATA/T13 (External Path Protection) try to remedy this by adding
  18support for appending integrity metadata to an I/O.  The integrity
  19metadata (or protection information in SCSI terminology) includes a
  20checksum for each sector as well as an incrementing counter that
  21ensures the individual sectors are written in the right order.  And
  22for some protection schemes also that the I/O is written to the right
  23place on disk.
  25Current storage controllers and devices implement various protective
  26measures, for instance checksumming and scrubbing.  But these
  27technologies are working in their own isolated domains or at best
  28between adjacent nodes in the I/O path.  The interesting thing about
  29DIF and the other integrity extensions is that the protection format
  30is well defined and every node in the I/O path can verify the
  31integrity of the I/O and reject it if corruption is detected.  This
  32allows not only corruption prevention but also isolation of the point
  33of failure.
  352. The Data Integrity Extensions
  38As written, the protocol extensions only protect the path between
  39controller and storage device.  However, many controllers actually
  40allow the operating system to interact with the integrity metadata
  41(IMD).  We have been working with several FC/SAS HBA vendors to enable
  42the protection information to be transferred to and from their
  45The SCSI Data Integrity Field works by appending 8 bytes of protection
  46information to each sector.  The data + integrity metadata is stored
  47in 520 byte sectors on disk.  Data + IMD are interleaved when
  48transferred between the controller and target.  The T13 proposal is
  51Because it is highly inconvenient for operating systems to deal with
  52520 (and 4104) byte sectors, we approached several HBA vendors and
  53encouraged them to allow separation of the data and integrity metadata
  54scatter-gather lists.
  56The controller will interleave the buffers on write and split them on
  57read.  This means that Linux can DMA the data buffers to and from
  58host memory without changes to the page cache.
  60Also, the 16-bit CRC checksum mandated by both the SCSI and SATA specs
  61is somewhat heavy to compute in software.  Benchmarks found that
  62calculating this checksum had a significant impact on system
  63performance for a number of workloads.  Some controllers allow a
  64lighter-weight checksum to be used when interfacing with the operating
  65system.  Emulex, for instance, supports the TCP/IP checksum instead.
  66The IP checksum received from the OS is converted to the 16-bit CRC
  67when writing and vice versa.  This allows the integrity metadata to be
  68generated by Linux or the application at very low cost (comparable to
  69software RAID5).
  71The IP checksum is weaker than the CRC in terms of detecting bit
  72errors.  However, the strength is really in the separation of the data
  73buffers and the integrity metadata.  These two distinct buffers must
  74match up for an I/O to complete.
  76The separation of the data and integrity metadata buffers as well as
  77the choice in checksums is referred to as the Data Integrity
  78Extensions.  As these extensions are outside the scope of the protocol
  79bodies (T10, T13), Oracle and its partners are trying to standardize
  80them within the Storage Networking Industry Association.
  823. Kernel Changes
  85The data integrity framework in Linux enables protection information
  86to be pinned to I/Os and sent to/received from controllers that
  87support it.
  89The advantage to the integrity extensions in SCSI and SATA is that
  90they enable us to protect the entire path from application to storage
  91device.  However, at the same time this is also the biggest
  92disadvantage. It means that the protection information must be in a
  93format that can be understood by the disk.
  95Generally Linux/POSIX applications are agnostic to the intricacies of
  96the storage devices they are accessing.  The virtual filesystem switch
  97and the block layer make things like hardware sector size and
  98transport protocols completely transparent to the application.
 100However, this level of detail is required when preparing the
 101protection information to send to a disk.  Consequently, the very
 102concept of an end-to-end protection scheme is a layering violation.
 103It is completely unreasonable for an application to be aware whether
 104it is accessing a SCSI or SATA disk.
 106The data integrity support implemented in Linux attempts to hide this
 107from the application.  As far as the application (and to some extent
 108the kernel) is concerned, the integrity metadata is opaque information
 109that's attached to the I/O.
 111The current implementation allows the block layer to automatically
 112generate the protection information for any I/O.  Eventually the
 113intent is to move the integrity metadata calculation to userspace for
 114user data.  Metadata and other I/O that originates within the kernel
 115will still use the automatic generation interface.
 117Some storage devices allow each hardware sector to be tagged with a
 11816-bit value.  The owner of this tag space is the owner of the block
 119device.  I.e. the filesystem in most cases.  The filesystem can use
 120this extra space to tag sectors as they see fit.  Because the tag
 121space is limited, the block interface allows tagging bigger chunks by
 122way of interleaving.  This way, 8*16 bits of information can be
 123attached to a typical 4KB filesystem block.
 125This also means that applications such as fsck and mkfs will need
 126access to manipulate the tags from user space.  A passthrough
 127interface for this is being worked on.
 1304. Block Layer Implementation Details
 1334.1 Bio
 136The data integrity patches add a new field to struct bio when
 137CONFIG_BLK_DEV_INTEGRITY is enabled.  bio_integrity(bio) returns a
 138pointer to a struct bip which contains the bio integrity payload.
 139Essentially a bip is a trimmed down struct bio which holds a bio_vec
 140containing the integrity metadata and the required housekeeping
 141information (bvec pool, vector count, etc.)
 143A kernel subsystem can enable data integrity protection on a bio by
 144calling bio_integrity_alloc(bio).  This will allocate and attach the
 145bip to the bio.
 147Individual pages containing integrity metadata can subsequently be
 148attached using bio_integrity_add_page().
 150bio_free() will automatically free the bip.
 1534.2 Block Device
 156Because the format of the protection data is tied to the physical
 157disk, each block device has been extended with a block integrity
 158profile (struct blk_integrity).  This optional profile is registered
 159with the block layer using blk_integrity_register().
 161The profile contains callback functions for generating and verifying
 162the protection data, as well as getting and setting application tags.
 163The profile also contains a few constants to aid in completing,
 164merging and splitting the integrity metadata.
 166Layered block devices will need to pick a profile that's appropriate
 167for all subdevices.  blk_integrity_compare() can help with that.  DM
 168and MD linear, RAID0 and RAID1 are currently supported.  RAID4/5/6
 169will require extra work due to the application tag.
 1725.0 Block Layer Integrity API
 1755.1 Normal Filesystem
 178    The normal filesystem is unaware that the underlying block device
 179    is capable of sending/receiving integrity metadata.  The IMD will
 180    be automatically generated by the block layer at submit_bio() time
 181    in case of a WRITE.  A READ request will cause the I/O integrity
 182    to be verified upon completion.
 184    IMD generation and verification can be toggled using the::
 186      /sys/block/<bdev>/integrity/write_generate
 188    and::
 190      /sys/block/<bdev>/integrity/read_verify
 192    flags.
 1955.2 Integrity-Aware Filesystem
 198    A filesystem that is integrity-aware can prepare I/Os with IMD
 199    attached.  It can also use the application tag space if this is
 200    supported by the block device.
 203    `bool bio_integrity_prep(bio);`
 205      To generate IMD for WRITE and to set up buffers for READ, the
 206      filesystem must call bio_integrity_prep(bio).
 208      Prior to calling this function, the bio data direction and start
 209      sector must be set, and the bio should have all data pages
 210      added.  It is up to the caller to ensure that the bio does not
 211      change while I/O is in progress.
 212      Complete bio with error if prepare failed for some reson.
 2155.3 Passing Existing Integrity Metadata
 218    Filesystems that either generate their own integrity metadata or
 219    are capable of transferring IMD from user space can use the
 220    following calls:
 223    `struct bip * bio_integrity_alloc(bio, gfp_mask, nr_pages);`
 225      Allocates the bio integrity payload and hangs it off of the bio.
 226      nr_pages indicate how many pages of protection data need to be
 227      stored in the integrity bio_vec list (similar to bio_alloc()).
 229      The integrity payload will be freed at bio_free() time.
 232    `int bio_integrity_add_page(bio, page, len, offset);`
 234      Attaches a page containing integrity metadata to an existing
 235      bio.  The bio must have an existing bip,
 236      i.e. bio_integrity_alloc() must have been called.  For a WRITE,
 237      the integrity metadata in the pages must be in a format
 238      understood by the target device with the notable exception that
 239      the sector numbers will be remapped as the request traverses the
 240      I/O stack.  This implies that the pages added using this call
 241      will be modified during I/O!  The first reference tag in the
 242      integrity metadata must have a value of bip->bip_sector.
 244      Pages can be added using bio_integrity_add_page() as long as
 245      there is room in the bip bio_vec array (nr_pages).
 247      Upon completion of a READ operation, the attached pages will
 248      contain the integrity metadata received from the storage device.
 249      It is up to the receiver to process them and verify data
 250      integrity upon completion.
 2535.4 Registering A Block Device As Capable Of Exchanging Integrity Metadata
 256    To enable integrity exchange on a block device the gendisk must be
 257    registered as capable:
 259    `int blk_integrity_register(gendisk, blk_integrity);`
 261      The blk_integrity struct is a template and should contain the
 262      following::
 264        static struct blk_integrity my_profile = {
 265            .name                   = "STANDARDSBODY-TYPE-VARIANT-CSUM",
 266            .generate_fn            = my_generate_fn,
 267            .verify_fn              = my_verify_fn,
 268            .tuple_size             = sizeof(struct my_tuple_size),
 269            .tag_size               = <tag bytes per hw sector>,
 270        };
 272      'name' is a text string which will be visible in sysfs.  This is
 273      part of the userland API so chose it carefully and never change
 274      it.  The format is standards body-type-variant.
 275      E.g. T10-DIF-TYPE1-IP or T13-EPP-0-CRC.
 277      'generate_fn' generates appropriate integrity metadata (for WRITE).
 279      'verify_fn' verifies that the data buffer matches the integrity
 280      metadata.
 282      'tuple_size' must be set to match the size of the integrity
 283      metadata per sector.  I.e. 8 for DIF and EPP.
 285      'tag_size' must be set to identify how many bytes of tag space
 286      are available per hardware sector.  For DIF this is either 2 or
 287      0 depending on the value of the Control Mode Page ATO bit.
 2912007-12-24 Martin K. Petersen <>