1.. SPDX-License-Identifier: GPL-2.0
   4XFS Self Describing Metadata
  10The largest scalability problem facing XFS is not one of algorithmic
  11scalability, but of verification of the filesystem structure. Scalabilty of the
  12structures and indexes on disk and the algorithms for iterating them are
  13adequate for supporting PB scale filesystems with billions of inodes, however it
  14is this very scalability that causes the verification problem.
  16Almost all metadata on XFS is dynamically allocated. The only fixed location
  17metadata is the allocation group headers (SB, AGF, AGFL and AGI), while all
  18other metadata structures need to be discovered by walking the filesystem
  19structure in different ways. While this is already done by userspace tools for
  20validating and repairing the structure, there are limits to what they can
  21verify, and this in turn limits the supportable size of an XFS filesystem.
  23For example, it is entirely possible to manually use xfs_db and a bit of
  24scripting to analyse the structure of a 100TB filesystem when trying to
  25determine the root cause of a corruption problem, but it is still mainly a
  26manual task of verifying that things like single bit errors or misplaced writes
  27weren't the ultimate cause of a corruption event. It may take a few hours to a
  28few days to perform such forensic analysis, so for at this scale root cause
  29analysis is entirely possible.
  31However, if we scale the filesystem up to 1PB, we now have 10x as much metadata
  32to analyse and so that analysis blows out towards weeks/months of forensic work.
  33Most of the analysis work is slow and tedious, so as the amount of analysis goes
  34up, the more likely that the cause will be lost in the noise.  Hence the primary
  35concern for supporting PB scale filesystems is minimising the time and effort
  36required for basic forensic analysis of the filesystem structure.
  39Self Describing Metadata
  42One of the problems with the current metadata format is that apart from the
  43magic number in the metadata block, we have no other way of identifying what it
  44is supposed to be. We can't even identify if it is the right place. Put simply,
  45you can't look at a single metadata block in isolation and say "yes, it is
  46supposed to be there and the contents are valid".
  48Hence most of the time spent on forensic analysis is spent doing basic
  49verification of metadata values, looking for values that are in range (and hence
  50not detected by automated verification checks) but are not correct. Finding and
  51understanding how things like cross linked block lists (e.g. sibling
  52pointers in a btree end up with loops in them) are the key to understanding what
  53went wrong, but it is impossible to tell what order the blocks were linked into
  54each other or written to disk after the fact.
  56Hence we need to record more information into the metadata to allow us to
  57quickly determine if the metadata is intact and can be ignored for the purpose
  58of analysis. We can't protect against every possible type of error, but we can
  59ensure that common types of errors are easily detectable.  Hence the concept of
  60self describing metadata.
  62The first, fundamental requirement of self describing metadata is that the
  63metadata object contains some form of unique identifier in a well known
  64location. This allows us to identify the expected contents of the block and
  65hence parse and verify the metadata object. IF we can't independently identify
  66the type of metadata in the object, then the metadata doesn't describe itself
  67very well at all!
  69Luckily, almost all XFS metadata has magic numbers embedded already - only the
  70AGFL, remote symlinks and remote attribute blocks do not contain identifying
  71magic numbers. Hence we can change the on-disk format of all these objects to
  72add more identifying information and detect this simply by changing the magic
  73numbers in the metadata objects. That is, if it has the current magic number,
  74the metadata isn't self identifying. If it contains a new magic number, it is
  75self identifying and we can do much more expansive automated verification of the
  76metadata object at runtime, during forensic analysis or repair.
  78As a primary concern, self describing metadata needs some form of overall
  79integrity checking. We cannot trust the metadata if we cannot verify that it has
  80not been changed as a result of external influences. Hence we need some form of
  81integrity check, and this is done by adding CRC32c validation to the metadata
  82block. If we can verify the block contains the metadata it was intended to
  83contain, a large amount of the manual verification work can be skipped.
  85CRC32c was selected as metadata cannot be more than 64k in length in XFS and
  86hence a 32 bit CRC is more than sufficient to detect multi-bit errors in
  87metadata blocks. CRC32c is also now hardware accelerated on common CPUs so it is
  88fast. So while CRC32c is not the strongest of possible integrity checks that
  89could be used, it is more than sufficient for our needs and has relatively
  90little overhead. Adding support for larger integrity fields and/or algorithms
  91does really provide any extra value over CRC32c, but it does add a lot of
  92complexity and so there is no provision for changing the integrity checking
  95Self describing metadata needs to contain enough information so that the
  96metadata block can be verified as being in the correct place without needing to
  97look at any other metadata. This means it needs to contain location information.
  98Just adding a block number to the metadata is not sufficient to protect against
  99mis-directed writes - a write might be misdirected to the wrong LUN and so be
 100written to the "correct block" of the wrong filesystem. Hence location
 101information must contain a filesystem identifier as well as a block number.
 103Another key information point in forensic analysis is knowing who the metadata
 104block belongs to. We already know the type, the location, that it is valid
 105and/or corrupted, and how long ago that it was last modified. Knowing the owner
 106of the block is important as it allows us to find other related metadata to
 107determine the scope of the corruption. For example, if we have a extent btree
 108object, we don't know what inode it belongs to and hence have to walk the entire
 109filesystem to find the owner of the block. Worse, the corruption could mean that
 110no owner can be found (i.e. it's an orphan block), and so without an owner field
 111in the metadata we have no idea of the scope of the corruption. If we have an
 112owner field in the metadata object, we can immediately do top down validation to
 113determine the scope of the problem.
 115Different types of metadata have different owner identifiers. For example,
 116directory, attribute and extent tree blocks are all owned by an inode, while
 117freespace btree blocks are owned by an allocation group. Hence the size and
 118contents of the owner field are determined by the type of metadata object we are
 119looking at.  The owner information can also identify misplaced writes (e.g.
 120freespace btree block written to the wrong AG).
 122Self describing metadata also needs to contain some indication of when it was
 123written to the filesystem. One of the key information points when doing forensic
 124analysis is how recently the block was modified. Correlation of set of corrupted
 125metadata blocks based on modification times is important as it can indicate
 126whether the corruptions are related, whether there's been multiple corruption
 127events that lead to the eventual failure, and even whether there are corruptions
 128present that the run-time verification is not detecting.
 130For example, we can determine whether a metadata object is supposed to be free
 131space or still allocated if it is still referenced by its owner by looking at
 132when the free space btree block that contains the block was last written
 133compared to when the metadata object itself was last written.  If the free space
 134block is more recent than the object and the object's owner, then there is a
 135very good chance that the block should have been removed from the owner.
 137To provide this "written timestamp", each metadata block gets the Log Sequence
 138Number (LSN) of the most recent transaction it was modified on written into it.
 139This number will always increase over the life of the filesystem, and the only
 140thing that resets it is running xfs_repair on the filesystem. Further, by use of
 141the LSN we can tell if the corrupted metadata all belonged to the same log
 142checkpoint and hence have some idea of how much modification occurred between
 143the first and last instance of corrupt metadata on disk and, further, how much
 144modification occurred between the corruption being written and when it was
 147Runtime Validation
 150Validation of self-describing metadata takes place at runtime in two places:
 152        - immediately after a successful read from disk
 153        - immediately prior to write IO submission
 155The verification is completely stateless - it is done independently of the
 156modification process, and seeks only to check that the metadata is what it says
 157it is and that the metadata fields are within bounds and internally consistent.
 158As such, we cannot catch all types of corruption that can occur within a block
 159as there may be certain limitations that operational state enforces of the
 160metadata, or there may be corruption of interblock relationships (e.g. corrupted
 161sibling pointer lists). Hence we still need stateful checking in the main code
 162body, but in general most of the per-field validation is handled by the
 165For read verification, the caller needs to specify the expected type of metadata
 166that it should see, and the IO completion process verifies that the metadata
 167object matches what was expected. If the verification process fails, then it
 168marks the object being read as EFSCORRUPTED. The caller needs to catch this
 169error (same as for IO errors), and if it needs to take special action due to a
 170verification error it can do so by catching the EFSCORRUPTED error value. If we
 171need more discrimination of error type at higher levels, we can define new
 172error numbers for different errors as necessary.
 174The first step in read verification is checking the magic number and determining
 175whether CRC validating is necessary. If it is, the CRC32c is calculated and
 176compared against the value stored in the object itself. Once this is validated,
 177further checks are made against the location information, followed by extensive
 178object specific metadata validation. If any of these checks fail, then the
 179buffer is considered corrupt and the EFSCORRUPTED error is set appropriately.
 181Write verification is the opposite of the read verification - first the object
 182is extensively verified and if it is OK we then update the LSN from the last
 183modification made to the object, After this, we calculate the CRC and insert it
 184into the object. Once this is done the write IO is allowed to continue. If any
 185error occurs during this process, the buffer is again marked with a EFSCORRUPTED
 186error for the higher layers to catch.
 191A typical on-disk structure needs to contain the following information::
 193    struct xfs_ondisk_hdr {
 194            __be32  magic;              /* magic number */
 195            __be32  crc;                /* CRC, not logged */
 196            uuid_t  uuid;               /* filesystem identifier */
 197            __be64  owner;              /* parent object */
 198            __be64  blkno;              /* location on disk */
 199            __be64  lsn;                /* last modification in log, not logged */
 200    };
 202Depending on the metadata, this information may be part of a header structure
 203separate to the metadata contents, or may be distributed through an existing
 204structure. The latter occurs with metadata that already contains some of this
 205information, such as the superblock and AG headers.
 207Other metadata may have different formats for the information, but the same
 208level of information is generally provided. For example:
 210        - short btree blocks have a 32 bit owner (ag number) and a 32 bit block
 211          number for location. The two of these combined provide the same
 212          information as @owner and @blkno in eh above structure, but using 8
 213          bytes less space on disk.
 215        - directory/attribute node blocks have a 16 bit magic number, and the
 216          header that contains the magic number has other information in it as
 217          well. hence the additional metadata headers change the overall format
 218          of the metadata.
 220A typical buffer read verifier is structured as follows::
 222    #define XFS_FOO_CRC_OFF             offsetof(struct xfs_ondisk_hdr, crc)
 224    static void
 225    xfs_foo_read_verify(
 226            struct xfs_buf      *bp)
 227    {
 228        struct xfs_mount *mp = bp->b_mount;
 230            if ((xfs_sb_version_hascrc(&mp->m_sb) &&
 231                !xfs_verify_cksum(bp->b_addr, BBTOB(bp->b_length),
 232                                            XFS_FOO_CRC_OFF)) ||
 233                !xfs_foo_verify(bp)) {
 234                    XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp, bp->b_addr);
 235                    xfs_buf_ioerror(bp, EFSCORRUPTED);
 236            }
 237    }
 239The code ensures that the CRC is only checked if the filesystem has CRCs enabled
 240by checking the superblock of the feature bit, and then if the CRC verifies OK
 241(or is not needed) it verifies the actual contents of the block.
 243The verifier function will take a couple of different forms, depending on
 244whether the magic number can be used to determine the format of the block. In
 245the case it can't, the code is structured as follows::
 247    static bool
 248    xfs_foo_verify(
 249            struct xfs_buf              *bp)
 250    {
 251            struct xfs_mount    *mp = bp->b_mount;
 252            struct xfs_ondisk_hdr       *hdr = bp->b_addr;
 254            if (hdr->magic != cpu_to_be32(XFS_FOO_MAGIC))
 255                    return false;
 257            if (!xfs_sb_version_hascrc(&mp->m_sb)) {
 258                    if (!uuid_equal(&hdr->uuid, &mp->m_sb.sb_uuid))
 259                            return false;
 260                    if (bp->b_bn != be64_to_cpu(hdr->blkno))
 261                            return false;
 262                    if (hdr->owner == 0)
 263                            return false;
 264            }
 266            /* object specific verification checks here */
 268            return true;
 269    }
 271If there are different magic numbers for the different formats, the verifier
 272will look like::
 274    static bool
 275    xfs_foo_verify(
 276            struct xfs_buf              *bp)
 277    {
 278            struct xfs_mount    *mp = bp->b_mount;
 279            struct xfs_ondisk_hdr       *hdr = bp->b_addr;
 281            if (hdr->magic == cpu_to_be32(XFS_FOO_CRC_MAGIC)) {
 282                    if (!uuid_equal(&hdr->uuid, &mp->m_sb.sb_uuid))
 283                            return false;
 284                    if (bp->b_bn != be64_to_cpu(hdr->blkno))
 285                            return false;
 286                    if (hdr->owner == 0)
 287                            return false;
 288            } else if (hdr->magic != cpu_to_be32(XFS_FOO_MAGIC))
 289                    return false;
 291            /* object specific verification checks here */
 293            return true;
 294    }
 296Write verifiers are very similar to the read verifiers, they just do things in
 297the opposite order to the read verifiers. A typical write verifier::
 299    static void
 300    xfs_foo_write_verify(
 301            struct xfs_buf      *bp)
 302    {
 303            struct xfs_mount    *mp = bp->b_mount;
 304            struct xfs_buf_log_item     *bip = bp->b_fspriv;
 306            if (!xfs_foo_verify(bp)) {
 307                    XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp, bp->b_addr);
 308                    xfs_buf_ioerror(bp, EFSCORRUPTED);
 309                    return;
 310            }
 312            if (!xfs_sb_version_hascrc(&mp->m_sb))
 313                    return;
 316            if (bip) {
 317                    struct xfs_ondisk_hdr       *hdr = bp->b_addr;
 318                    hdr->lsn = cpu_to_be64(bip->bli_item.li_lsn);
 319            }
 320            xfs_update_cksum(bp->b_addr, BBTOB(bp->b_length), XFS_FOO_CRC_OFF);
 321    }
 323This will verify the internal structure of the metadata before we go any
 324further, detecting corruptions that have occurred as the metadata has been
 325modified in memory. If the metadata verifies OK, and CRCs are enabled, we then
 326update the LSN field (when it was last modified) and calculate the CRC on the
 327metadata. Once this is done, we can issue the IO.
 329Inodes and Dquots
 332Inodes and dquots are special snowflakes. They have per-object CRC and
 333self-identifiers, but they are packed so that there are multiple objects per
 334buffer. Hence we do not use per-buffer verifiers to do the work of per-object
 335verification and CRC calculations. The per-buffer verifiers simply perform basic
 336identification of the buffer - that they contain inodes or dquots, and that
 337there are magic numbers in all the expected spots. All further CRC and
 338verification checks are done when each inode is read from or written back to the
 341The structure of the verifiers and the identifiers checks is very similar to the
 342buffer code described above. The only difference is where they are called. For
 343example, inode read verification is done in xfs_inode_from_disk() when the inode
 344is first read out of the buffer and the struct xfs_inode is instantiated. The
 345inode is already extensively verified during writeback in xfs_iflush_int, so the
 346only addition here is to add the LSN and CRC to the inode as it is copied back
 347into the buffer.
 349XXX: inode unlinked list modification doesn't recalculate the inode CRC! None of
 350the unlinked list modifications check or update CRCs, neither during unlink nor
 351log recovery. So, it's gone unnoticed until now. This won't matter immediately -
 352repair will probably complain about it - but it needs to be fixed.