The Perl Toolchain Summit needs more sponsors. If your company depends on Perl, please support this very important event.
This file describes the design, layouts, and file formats of a
libsvn_fs_fs repository.

Design
------

In FSFS, each committed revision is represented as an immutable file
containing the new node-revisions, contents, and changed-path
information for the revision, plus a second, changeable file
containing the revision properties.

In contrast to the BDB back end, the contents of recent revision of
files are stored as deltas against earlier revisions, instead of the
other way around.  This is less efficient for common-case checkouts,
but brings greater simplicity and robustness, as well as the
flexibility to make commits work without write access to existing
revisions.  Skip-deltas and delta combination mitigate the checkout
cost.

In-progress transactions are represented with a prototype rev file
containing only the new text representations of files (appended to as
changed file contents come in), along with a separate file for each
node-revision, directory representation, or property representation
which has been changed or added in the transaction.  During the final
stage of the commit, these separate files are marshalled onto the end
of the prototype rev file to form the immutable revision file.

Layout of the FS directory
--------------------------

The layout of the FS directory (the "db" subdirectory of the
repository) is:

  revs/               Subdirectory containing revs
    <shard>/          Shard directory, if sharding is in use (see below)
      <revnum>        File containing rev <revnum>
    <shard>.pack      Pack directory, if the repo has been packed (see below)
      pack            Pack file, if the repository has been packed (see below)
      manifest        Pack manifest file, if a pack file exists (see below)
  revprops/           Subdirectory containing rev-props
    <shard>/          Shard directory, if sharding is in use (see below)
      <revnum>        File containing rev-props for <revnum>
  transactions/       Subdirectory containing transactions
    <txnid>.txn/      Directory containing transaction <txnid>
  txn-protorevs/      Subdirectory containing transaction proto-revision files
    <txnid>.rev       Proto-revision file for transaction <txnid>
    <txnid>.rev-lock  Write lock for proto-rev file
  txn-current         File containing the next transaction key
  locks/              Subdirectory containing locks
    <partial-digest>/ Subdirectory named for first 3 letters of an MD5 digest
      <digest>        File containing locks/children for path with <digest>
  node-origins/       Lazy cache of origin noderevs for nodes
    <partial-nodeid>  File containing noderev ID of origins of nodes
  current             File specifying current revision and next node/copy id
  fs-type             File identifying this filesystem as an FSFS filesystem
  write-lock          Empty file, locked to serialise writers
  txn-current-lock    Empty file, locked to serialise 'txn-current'
  uuid                File containing the UUID of the repository
  format              File containing the format number of this filesystem
  fsfs.conf           Configuration file
  min-unpacked-rev    File containing the oldest revision not in a pack file
  rep-cache.db        SQLite database mapping rep checksums to locations

Files in the revprops directory are in the hash dump format used by
svn_hash_write.

The format of the "current" file is:

 * Format 3 and above: a single line of the form
   "<youngest-revision>\n" giving the youngest revision for the
   repository.

 * Format 2 and below: a single line of the form "<youngest-revision>
   <next-node-id> <next-copy-id>\n" giving the youngest revision, the
   next unique node-ID, and the next unique copy-ID for the
   repository.

The "write-lock" file is an empty file which is locked before the
final stage of a commit and unlocked after the new "current" file has
been moved into place to indicate that a new revision is present.  It
is also locked during a revprop propchange while the revprop file is
read in, mutated, and written out again.  Note that readers are never
blocked by any operation - writers must ensure that the filesystem is
always in a consistent state.

The "txn-current" file is a file with a single line of text that
contains only a base-36 number.  The current value will be used in the
next transaction name, along with the revision number the transaction
is based on.  This sequence number ensures that transaction names are
not reused, even if the transaction is aborted and a new transaction
based on the same revision is begun.  The only operation that FSFS
performs on this file is "get and increment"; the "txn-current-lock"
file is locked during this operation.

"fsfs.conf" is a configuration file in the standard Subversion/Python
config format.  It is automatically generated when you create a new
repository; read the generated file for details on what it controls.

When representation sharing is enabled, the filesystem tracks
representation checksum and location mappings using a SQLite database in
"rep-cache.db".  The database has a single table, which stores the sha1
hash text as the primary key, mapped to the representation revision, offset,
size and expanded size.  A final field, reuse count, is currently used
for distinguishing representations which are shared.  This file is not
required, and may be removed at an abritrary time, with the subsequent
loss of rep-sharing capabilities.

Filesystem formats
------------------

The "format" file defines what features are permitted within the
filesystem, and indicates changes that are not backward-compatible.
It serves the same purpose as the repository file of the same name.

The filesystem format file was introduced in Subversion 1.2, and so
will not be present if the repository was created with an older
version of Subversion.  An absent format file should be interpreted as
indicating a format 1 filesystem.

The format file is a single line of the form "<format number>\n",
followed by any number of lines specifying 'format options' -
additional information about the filesystem's format.  Each format
option line is of the form "<option>\n" or "<option> <parameters>\n".

Clients should raise an error if they encounter an option not
permitted by the format number in use.

The formats are:

  Format 1, understood by Subversion 1.1+
  Format 2, understood by Subversion 1.4+
  Format 3, understood by Subversion 1.5+
  Format 4, understood by Subversion 1.6+

The differences between the formats are:

Delta representation in revision files
  Format 1: svndiff0 only
  Formats 2-4: svndiff0 or svndiff1

Format options
  Formats 1-2: none permitted
  Format 3-4: "layout" option

Transaction name reuse
  Formats 1-2: transaction names may be reused
  Format 3-4: transaction names generated using txn-current file

Location of proto-rev file and its lock
  Formats 1-2: transactions/<txnid>/rev and
    transactions/<txnid>/rev-lock.
  Format 3-4: txn-protorevs/<txnid>.rev and
    txn-protorevs/<txnid>.rev-lock.

Node-ID and copy-ID generation
  Formats 1-2: Node-IDs and copy-IDs are guaranteed to form a
    monotonically increasing base36 sequence using the "current"
    file.
  Format 3-4: Node-IDs and copy-IDs use the new revision number to
    ensure uniqueness and the "current" file just contains the
    youngest revision.

Mergeinfo metadata:
  Format 1-2: minfo-here and minfo-count node-revision fields are not
    stored.  svn_fs_get_mergeinfo returns an error.
  Format 3-4: minfo-here and minfo-count node-revision fields are
    maintained.  svn_fs_get_mergeinfo works.

Revision changed paths list:
  Format 1-3: Does not contain the node's kind.
  Format 4: Contains the node's kind.


Filesystem format options
-------------------------

Currently, the only recognised format option is "layout", which
specifies the paths that will be used to store the revision files and
revision property files.

The "layout" option is followed by the name of the filesystem layout
and any required parameters.  The default layout, if no "layout"
keyword is specified, is the 'linear' layout.

The known layouts, and the parameters they require, are as follows:

"linear"
  Revision files and rev-prop files are named after the revision they
  represent, and are placed directly in the revs/ and revprops/
  directories.  r1234 will be represented by the revision file
  revs/1234 and the rev-prop file revprops/1234.

"sharded <max-files-per-directory>"
  Revision files and rev-prop files are named after the revision they
  represent, and are placed in a subdirectory of the revs/ and
  revprops/ directories named according to the 'shard' they belong to.

  Shards are numbered from zero and contain between one and the
  maximum number of files per directory specified in the layout's
  parameters.

  For the "sharded 1000" layout, r1234 will be represented by the
  revision file revs/1/1234 and rev-prop file revprops/1/1234.  The
  revs/0/ directory will contain revisions 0-999, revs/1/ will contain
  1000-1999, and so on.

Packing
-------

A repository can optionally be "packed" to conserve space on disk.  The
packing process concatenates all the revision files in each full shard to
create pack files.  A manifest file is also created for each shard which
records the indexes of the corresponding revision files in the pack file.
In addition, the original shard is removed, and reads are redirected to the
pack file.

The manifest file consists of a list of offsets, one for each revision in the
pack file.  The offsets are stored as ASCII decimal, and separated by a newline
character.

Node-revision IDs
-----------------

In order to support efficient lookup of node-revisions by their IDs
and to simplify the allocation of fresh node-IDs during a transaction,
we treat the fields of a node-ID in new and interesting ways.

Within a revision file, node-revs have a txn-id field of the form
"r<rev>/<offset>", to support easy lookup.  New node-revision IDs
assigned within a transaction have the txn-id field of "t<txnid>".

When a new node-id or copy-id is assigned in a transaction, the ID
used is a "_" followed by a base36 number unique to the transaction.
During the final phase of a commit, node-revision IDs are rewritten
to have repository-wide unique node-ID and copy-ID fields, and to have
"r<rev>/<offset>" txn-id fields.

In Format 3 and above, this uniqueness is done by changing a temporary
id of "_<base36>" to "<rev>-<base36>".  Note that this means that the
originating revision of a line of history or a copy can be determined
by looking at the node ID.

In Format 2 and below, the "current" file contains global base36
node-ID and copy-ID counters; during the commit, the counter value is
added to the transaction-specific base36 ID, and the value in
"current" is adjusted.

(It is legal for Format 3 repositories to contain Format 2-style IDs;
this just prevents I/O-less node-origin-rev lookup for those nodes.)

The temporary assignment of node-ID and copy-ID fields has
implications for svn_fs_compare_ids and svn_fs_check_related.  The IDs
_1.0.t1 is not related to the ID _1.0.t2 even though they have the
same node-ID, because temporary node-IDs are restricted in scope to
the transactions they belong to.

There is a lazily created cache mapping from node-IDs to the full
node-revision ID where they are created.  This is in the node-origins
directory; the file name is the node-ID without its last character (or
"0" for single-character node IDs) and the contents is a serialized
hash mapping from node-ID to node-revision ID.  This cache is only
used for node-IDs of the pre-Format 3 style.

Copy-IDs and copy roots
-----------------------

Copy-IDs are assigned in the same manner as they are in the BDB
implementation:

  * A node-rev resulting from a creation operation (with no copy
    history) receives the copy-ID of its parent directory.

  * A node-rev resulting from a copy operation receives a fresh
    copy-ID, as one would expect.

  * A node-rev resulting from a modification operation receives a
    copy-ID depending on whether its predecessor derives from a
    copy operation or whether it derives from a creation operation
    with no intervening copies:

      - If the predecessor does not derive from a copy, the new
        node-rev receives the copy-ID of its parent directory.  If the
        node-rev is being modified through its created-path, this will
        be the same copy-ID as the predecessor node-rev has; however,
        if the node-rev is being modified through a copied ancestor
        directory (i.e. we are performing a "lazy copy"), this will be
        a different copy-ID.

      - If the predecessor derives from a copy and the node-rev is
        being modified through its created-path, the new node-rev
        receives the copy-ID of the predecessor.

      - If the predecessor derives from a copy and the node-rev is not
        being modified through its created path, the new node-rev
        receives a fresh copy-ID.  This is called a "soft copy"
        operation, as distinct from a "true copy" operation which was
        actually requested through the svn_fs interface.  Soft copies
        exist to ensure that the same <node-ID,copy-ID> pair is not
        used twice within a transaction.

Unlike the BDB implementation, we do not have a "copies" table.
Instead, each node-revision record contains a "copyroot" field
identifying the node-rev resulting from the true copy operation most
proximal to the node-rev.  If the node-rev does not itself derive from
a copy operation, then the copyroot field identifies the copy of an
ancestor directory; if no ancestor directories derive from a copy
operation, then the copyroot field identifies the root directory of
rev 0.

Revision file format
--------------------

A revision file contains a concatenation of various kinds of data:

  * Text and property representations
  * Node-revisions
  * The changed-path data
  * Two offsets at the very end

A representation begins with a line containing either "PLAIN\n" or
"DELTA\n" or "DELTA <rev> <offset> <length>\n", where <rev>, <offset>,
and <length> give the location of the delta base of the representation
and the amount of data it contains (not counting the header or
trailer).  If no base location is given for a delta, the base is the
empty stream.  After the initial line comes raw svndiff data, followed
by a cosmetic trailer "ENDREP\n".

If the a representation is for the text contents of a directory node,
the expanded contents are in hash dump format mapping entry names to
"<type> <id>" pairs, where <type> is "file" or "dir" and <id> gives
the ID of the child node-rev.

If a representation is for a property list, the expanded contents are
in the form of a dumped hash map mapping property names to property
values.

The marshalling syntax for node-revs is a series of fields terminated
by a blank line.  Fields have the syntax "<name>: <value>\n", where
<name> is a symbolic field name (each symbolic name is used only once
in a given node-rev) and <value> is the value data.  Unrecognized
fields are ignored, for extensibility.  The following fields are
defined:

  id        The ID of the node-rev
  type      "file" or "dir"
  pred      The ID of the predecessor node-rev
  count     Count of node-revs since the base of the node
  text      "<rev> <offset> <length> <size> <digest>" for text rep
  props     "<rev> <offset> <length> <size> <digest>" for props rep
            <rev> and <offset> give location of rep
            <length> gives length of rep, sans header and trailer
            <size> gives size of expanded rep
            <digest> gives hex MD5 digest of expanded rep
  cpath     FS pathname node was created at
  copyfrom  "<rev> <path>" of copyfrom data
  copyroot  "<rev> <created-path>" of the root of this copy
  minfo-cnt The number of nodes under (and including) this node
             which have svn:mergeinfo.
  minfo-here Exists if if this node itself has svn:mergeinfo.

The predecessor of a node-rev crosses both soft and true copies;
together with the count field, it allows efficient determination of
the base for skip-deltas.  The first node-rev of a node contains no
"pred" field.  A node-revision with no properties may omit the "props"
field.  A node-revision with no contents (a zero-length file or an
empty directory) may omit the "text" field.  In a node-revision
resulting from a true copy operation, the "copyfrom" field gives the
copyfrom data.  The "copyroot" field identifies the root node-revision
of the copy; it may be omitted if the node-rev is its own copy root
(as is the case for node-revs with copy history, and for the root node
of revision 0).  Copy roots are identified by revision and
created-path, not by node-rev ID, because a copy root may be a
node-rev which exists later on within the same revision file, meaning
its offset is not yet known.

The changed-path data is represented as a series of changed-path
items, each consisting of two lines.  The first line has the format
"<id> <action> <text-mod> <prop-mod> <path>\n", where <id> is the
node-rev ID of the new node-rev, <action> is "add", "delete",
"replace", or "modify", <text-mod> and <prop-mod> are "true" or
"false" indicating whether the text and/or properties changed, and
<path> is the changed pathname.  For deletes, <id> is the node-rev ID
of the deleted node-rev, and <text-mod> and <prop-mod> are always
"false".  The second line has the format "<rev> <path>\n" containing
the node-rev's copyfrom information if it has any; if it does not, the
second line is blank.

Starting with FS format 4, <action> may contain the kind ("file" or
"dir") of the node, after a hyphen; for example, an added directory
may be represented as "add-dir".

At the very end of a rev file is a pair of lines containing
"\n<root-offset> <cp-offset>\n", where <root-offset> is the offset of
the root directory node revision and <cp-offset> is the offset of the
changed-path data.

All numbers in the rev file format are unsigned and are represented as
ASCII decimal.

Transaction layout
------------------

A transaction directory has the following layout:

  props                      Transaction props
  next-ids                   Next temporary node-ID and copy-ID
  changes                    Changed-path information so far
  node.<nid>.<cid>           New node-rev data for node
  node.<nid>.<cid>.props     Props for new node-rev, if changed
  node.<nid>.<cid>.children  Directory contents for node-rev

In FS formats 1 and 2, it also contains:

  rev                        Prototype rev file with new text reps
  rev-lock                   Lockfile for writing to the above

In newer formats, these files are in the txn-protorevs/ directory.

The prototype rev file is used to store the text representations as
they are received from the client.  To ensure that only one client is
writing to the file at a given time, the "rev-lock" file is locked for
the duration of each write.

The two kinds of props files are all in hash dump format.  The "props"
file will always be present.  The "node.<nid>.<cid>.props" file will
only be present if the node-rev properties have been changed.

The "next-ids" file contains a single line "<next-temp-node-id>
<next-temp-copy-id>\n" giving the next temporary node-ID and copy-ID
assignments (without the leading underscores).  The next node-ID is
also used as a uniquifier for representations which may share the same
underlying rep.

The "children" file for a node-revision begins with a copy of the hash
dump representation of the directory entries from the old node-rev (or
a dump of the empty hash for new directories), and then an incremental
hash dump entry for each change made to the directory.

The "changes" file contains changed-path entries in the same form as
the changed-path entries in a rev file, except that <id> and <action>
may both be "reset" (in which case <text-mod> and <prop-mod> are both
always "false") to indicate that all changes to a path should be
considered undone.  Reset entries are only used during the final merge
phase of a transaction.  Actions in the "changes" file always contain
a node kind, even if the FS format is older than format 4.

The node-rev files have the same format as node-revs in a revision
file, except that the "text" and "props" fields are augmented as
follows:

  * The "props" field may have the value "-1" if properties have
    been changed and are contained in a "props" file within the
    node-rev subdirectory.

  * For directory node-revs, the "text" field may have the value
    "-1" if entries have been changed and are contained in a
    "contents" file in the node-rev subdirectory.

  * For the directory node-rev representing the root of the
    transaction, the "is-fresh-txn-root" field indicates that it has
    not been made mutable yet (see Issue #2608).

  * For file node-revs, the "text" field may have the value "-1
    <offset> <length> <size> <digest>" if the text representation is
    within the prototype rev file.

  * The "copyroot" field may have the value "-1 <created-path>" if the
    copy root of the node-rev is part of the transaction in process.

Locks layout
------------

Locks in FSFS are stored in serialized hash format in files whose
names are MD5 digests of the FS path which the lock is associated
with.  For the purposes of keeping directory inode usage down, these
digest files live in subdirectories of the main lock directory whose
names are the first 3 characters of the digest filename.

Also stored in the digest file for a given FS path are pointers to
other digest files which contain information associated with other FS
paths that are our path's immediate children.

To answer the question, "Does path FOO have a lock associated with
it?", one need only generate the MD5 digest of FOO's
absolute-in-the-FS path (say, 3b1b011fed614a263986b5c4869604e8), look
for a file located like so:

   /path/to/repos/locks/3b1/3b1b011fed614a263986b5c4869604e8

And then see if that file contains lock information.

To inquire about locks on children of the path FOO, you would
reference the same path as above, but look for a list of children in
that file (instead of lock information).  Children are listed as MD5
digests, too, so you would simply iterate over those digests and
consult the files they reference, and so on, recursively.