annotate mercurial/helptext/internals/dirstate-v2.txt @ 48195:4d5a13253d34

dirstate-v2: Replace the 32-bit `mode` field with two bits

Previously we stored the entire value from `stat_result.st_mode`,
like dirstate-v1 does. However only the executable permission
and type of file (only symbolic links and normal files are supported)
are relevant to Mecurial.

So replace this field with two bits in the existing bitfield byte.
For now the unused space is left as padding, as it will be used
for something else soon.

Differential Revision: https://phab.mercurial-scm.org/D11635

The *dirstate* is what Mercurial uses internally to track
the state of files in the working directory,
such as set by commands like `hg add` and `hg rm`.
It also contains some cached data that help make `hg status` faster.
The name refers both to `.hg/dirstate` on the filesystem
and the corresponding data structure in memory while a Mercurial process
is running.

The original file format, retroactively dubbed `dirstate-v1`,
is described at https://www.mercurial-scm.org/wiki/DirState.
It is made of a flat sequence of unordered variable-size entries,
so accessing any information in it requires parsing all of it.
Similarly, saving changes requires rewriting the entire file.

The newer `dirsate-v2` file format is designed to fix these limitations
and make `hg status` faster.

User guide
==========

Compatibility
-------------

The file format is experimental and may still change.
Different versions of Mercurial may not be compatible with each other
when working on a local repository that uses this format.
When using an incompatible version with the experimental format,
anything can happen including data corruption.

Since the dirstate is entirely local and not relevant to the wire protocol,
`dirstate-v2` does not affect compatibility with remote Mercurial versions.

When `share-safe` is enabled, different repositories sharing the same store
can use different dirstate formats.

Enabling `dirsate-v2` for new local repositories
------------------------------------------------

When creating a new local repository such as with `hg init` or `hg clone`,
the `exp-dirstate-v2` boolean in the `format` configuration section
controls whether to use this file format.
This is disabled by default as of this writing.
To enable it for a single repository, run for example::

    $ hg init my-project --config format.exp-dirstate-v2=1

Checking the format of an existing local repsitory
--------------------------------------------------

The `debugformat` commands prints information about
which of multiple optional formats are used in the current repository,
including `dirstate-v2`::

    $ hg debugformat
    format-variant     repo
    fncache:            yes
    dirstate-v2:        yes
    […]

Upgrading or downgrading an existing local repository
-----------------------------------------------------

The `debugupgrade` command does various upgrades or downgrades
on a local repository
based on the current Mercurial version and on configuration.
The same `format.exp-dirstate-v2` configuration is used again.

Example to upgrade::

    $ hg debugupgrade --config format.exp-dirstate-v2=1

Example to downgrade to `dirstate-v1`::

    $ hg debugupgrade --config format.exp-dirstate-v2=0

Both of this commands do nothing but print a list of proposed changes,
which may include changes unrelated to the dirstate.
Those other changes are controlled by their own configuration keys.
Add `--run` to a command to actually apply the proposed changes.

Backups of `.hg/requires` and `.hg/dirstate` are created
in a `.hg/upgradebackup.*` directory.
If something goes wrong, restoring those files should undo the change.

Note that upgrading affects compatibility with older versions of Mercurial
as noted above.
This can be relevant when a repository’s files are on a USB drive
or some other removable media, or shared over the network, etc.

Internal filesystem representation
==================================

Requirements file
-----------------

The `.hg/requires` file indicates which of various optional file formats
are used by a given repository.
Mercurial aborts when seeing a requirement it does not know about,
which avoids older version accidentally messing up a respository
that uses a format that was introduced later.
For versions that do support a format, the presence or absence of
the corresponding requirement indicates whether to use that format.

When the file contains a `exp-dirstate-v2` line,
the `dirstate-v2` format is used.
With no such line `dirstate-v1` is used.

High level description
----------------------

Whereas `dirstate-v1` uses a single `.hg/disrtate` file,
in `dirstate-v2` that file is a "docket" file
that only contains some metadata
and points to separate data file named `.hg/dirstate.{ID}`,
where `{ID}` is a random identifier.

This separation allows making data files append-only
and therefore safer to memory-map.
Creating a new data file (occasionally to clean up unused data)
can be done with a different ID
without disrupting another Mercurial process
that could still be using the previous data file.

Both files have a format designed to reduce the need for parsing,
by using fixed-size binary components as much as possible.
For data that is not fixed-size,
references to other parts of a file can be made by storing "pseudo-pointers":
integers counted in bytes from the start of a file.
For read-only access no data structure is needed,
only a bytes buffer (possibly memory-mapped directly from the filesystem)
with specific parts read on demand.

The data file contains "nodes" organized in a tree.
Each node represents a file or directory inside the working directory
or its parent changeset.
This tree has the same structure as the filesystem,
so a node representing a directory has child nodes representing
the files and subdirectories contained directly in that directory.

The docket file format
----------------------

This is implemented in `rust/hg-core/src/dirstate_tree/on_disk.rs`
and `mercurial/dirstateutils/docket.py`.

Components of the docket file are found at fixed offsets,
counted in bytes from the start of the file:

* Offset 0:
  The 12-bytes marker string "dirstate-v2\n" ending with a newline character.
  This makes it easier to tell a dirstate-v2 file from a dirstate-v1 file,
  although it is not strictly necessary
  since `.hg/requires` determines which format to use.

* Offset 12:
  The changeset node ID on the first parent of the working directory,
  as up to 32 binary bytes.
  If a node ID is shorter (20 bytes for SHA-1),
  it is start-aligned and the rest of the bytes are set to zero.

* Offset 44:
  The changeset node ID on the second parent of the working directory,
  or all zeros if there isn’t one.
  Also 32 binary bytes.

* Offset 76:
  Tree metadata on 44 bytes, described below.
  Its separation in this documentation from the rest of the docket
  reflects a detail of the current implementation.
  Since tree metadata is also made of fields at fixed offsets, those could
  be inlined here by adding 76 bytes to each offset.

* Offset 120:
  The used size of the data file, as a 32-bit big-endian integer.
  The actual size of the data file may be larger
  (if another Mercurial processis in appending to it
  but has not updated the docket yet).
  That extra data must be ignored.

* Offset 124:
  The length of the data file identifier, as a 8-bit integer.

* Offset 125:
  The data file identifier.

* Any additional data is current ignored, and dropped when updating the file.

Tree metadata in the docket file
--------------------------------

Tree metadata is similarly made of components at fixed offsets.
These offsets are counted in bytes from the start of tree metadata,
which is 76 bytes after the start of the docket file.

This metadata can be thought of as the singular root of the tree
formed by nodes in the data file.

* Offset 0:
  Pseudo-pointer to the start of root nodes,
  counted in bytes from the start of the data file,
  as a 32-bit big-endian integer.
  These nodes describe files and directories found directly
  at the root of the working directory.

* Offset 4:
  Number of root nodes, as a 32-bit big-endian integer.

* Offset 8:
  Total number of nodes in the entire tree that "have a dirstate entry",
  as a 32-bit big-endian integer.
  Those nodes represent files that would be present at all in `dirstate-v1`.
  This is typically less than the total number of nodes.
  This counter is used to implement `len(dirstatemap)`.

* Offset 12:
  Number of nodes in the entire tree that have a copy source,
  as a 32-bit big-endian integer.
  At the next commit, these files are recorded
  as having been copied or moved/renamed from that source.
  (A move is recorded as a copy and separate removal of the source.)
  This counter is used to implement `len(dirstatemap.copymap)`.

* Offset 16:
  An estimation of how many bytes of the data file
  (within its used size) are unused, as a 32-bit big-endian integer.
  When appending to an existing data file,
  some existing nodes or paths can be unreachable from the new root
  but they still take up space.
  This counter is used to decide when to write a new data file from scratch
  instead of appending to an existing one,
  in order to get rid of that unreachable data
  and avoid unbounded file size growth.

* Offset 20:
  These four bytes are currently ignored
  and reset to zero when updating a docket file.
  This is an attempt at forward compatibility:
  future Mercurial versions could use this as a bit field
  to indicate that a dirstate has additional data or constraints.
  Finding a dirstate file with the relevant bit unset indicates that
  it was written by a then-older version
  which is not aware of that future change.

* Offset 24:
  Either 20 zero bytes, or a SHA-1 hash as 20 binary bytes.
  When present, the hash is of ignore patterns
  that were used for some previous run of the `status` algorithm.

* (Offset 44: end of tree metadata)

Optional hash of ignore patterns
--------------------------------

The implementation of `status` at `rust/hg-core/src/dirstate_tree/status.rs`
has been optimized such that its run time is dominated by calls
to `stat` for reading the filesystem metadata of a file or directory,
and to `readdir` for listing the contents of a directory.
In some cases the algorithm can skip calls to `readdir`
(saving significant time)
because the dirstate already contains enough of the relevant information
to build the correct `status` results.

The default configuration of `hg status` is to list unknown files
but not ignored files.
In this case, it matters for the `readdir`-skipping optimization
if a given file used to be ignored but became unknown
because `.hgignore` changed.
To detect the possibility of such a change,
the tree metadata contains an optional hash of all ignore patterns.

We define:

* "Root" ignore files as:

  - `.hgignore` at the root of the repository if it exists
  - And all files from `ui.ignore.*` config.

  This set of files is sorted by the string representation of their path.

* The "expanded contents" of an ignore files is the byte string made
  by the concatenation of its contents followed by the "expanded contents"
  of other files included with `include:` or `subinclude:` directives,
  in inclusion order. This definition is recursive, as included files can
  themselves include more files.

This hash is defined as the SHA-1 of the concatenation (in sorted
order) of the "expanded contents" of each "root" ignore file.
(Note that computing this does not require actually concatenating
into a single contiguous byte sequence.
Instead a SHA-1 hasher object can be created
and fed separate chunks one by one.)

The data file format
--------------------

This is implemented in `rust/hg-core/src/dirstate_tree/on_disk.rs`
and `mercurial/dirstateutils/v2.py`.

The data file contains two types of data: paths and nodes.

Paths and nodes can be organized in any order in the file, except that sibling
nodes must be next to each other and sorted by their path.
Contiguity lets the parent refer to them all
by their count and a single pseudo-pointer,
instead of storing one pseudo-pointer per child node.
Sorting allows using binary seach to find a child node with a given name
in `O(log(n))` byte sequence comparisons.

The current implemention writes paths and child node before a given node
for ease of figuring out the value of pseudo-pointers by the time the are to be
written, but this is not an obligation and readers must not rely on it.

A path is stored as a byte string anywhere in the file, without delimiter.
It is refered to by one or more node by a pseudo-pointer to its start, and its
length in bytes. Since there is no delimiter,
when a path is a substring of another the same bytes could be reused,
although the implementation does not exploit this as of this writing.

A node is stored on 43 bytes with components at fixed offsets. Paths and
child nodes relevant to a node are stored externally and referenced though
pseudo-pointers.

All integers are stored in big-endian. All pseudo-pointers are 32-bit integers
counting bytes from the start of the data file. Path lengths and positions
are 16-bit integers, also counted in bytes.

Node components are:

* Offset 0:
  Pseudo-pointer to the full path of this node,
  from the working directory root.

* Offset 4:
  Length of the full path.

* Offset 6:
  Position of the last `/` path separator within the full path,
  in bytes from the start of the full path,
  or zero if there isn’t one.
  The part of the full path after this position is the "base name".
  Since sibling nodes have the same parent, only their base name vary
  and needs to be considered when doing binary search to find a given path.

* Offset 8:
  Pseudo-pointer to the "copy source" path for this node,
  or zero if there is no copy source.

* Offset 12:
  Length of the copy source path, or zero if there isn’t one.

* Offset 14:
  Pseudo-pointer to the start of child nodes.

* Offset 18:
  Number of child nodes, as a 32-bit integer.
  They occupy 43 times this number of bytes
  (not counting space for paths, and further descendants).

* Offset 22:
  Number as a 32-bit integer of descendant nodes in this subtree,
  not including this node itself,
  that "have a dirstate entry".
  Those nodes represent files that would be present at all in `dirstate-v1`.
  This is typically less than the total number of descendants.
  This counter is used to implement `has_dir`.

* Offset 26:
  Number as a 32-bit integer of descendant nodes in this subtree,
  not including this node itself,
  that represent files tracked in the working directory.
  (For example, `hg rm` makes a file untracked.)
  This counter is used to implement `has_tracked_dir`.

* Offset 30:
  Some boolean values packed as bits of a single byte.
  Starting from least-significant, bit masks are::

    WDIR_TRACKED = 1 << 0
    P1_TRACKED = 1 << 1
    P2_INFO = 1 << 2
    HAS_MODE_AND_SIZE = 1 << 3
    HAS_MTIME = 1 << 4
    MODE_EXEC_PERM = 1 << 5
    MODE_IS_SYMLINK = 1 << 7


  Other bits are unset. The meaning of these bits are:

  `WDIR_TRACKED`
      Set if the working directory contains a tracked file at this node’s path.
      This is typically set and unset by `hg add` and `hg rm`.

  `P1_TRACKED`
      set if the working directory’s first parent changeset
      (whose node identifier is found in tree metadata)
      contains a tracked file at this node’s path.
      This is a cache to reduce manifest lookups.

  `P2_INFO`
      Set if the file has been involved in some merge operation.
      Either because it was actually merged,
      or because the version in the second parent p2 version was ahead,
      or because some rename moved it there.
      In either case `hg status` will want it displayed as modified.

  Files that would be mentioned at all in the `dirstate-v1` file format
  have a node with at least one of the above three bits set in `dirstate-v2`.
  Let’s call these files "tracked anywhere",
  and "untracked" the nodes with all three of these bits unset.
  Untracked nodes are typically for directories:
  they hold child nodes and form the tree structure.
  Additional untracked nodes may also exist.
  Although implementations should strive to clean up nodes
  that are entirely unused, other untracked nodes may also exist.
  For example, a future version of Mercurial might in some cases
  add nodes for untracked files or/and ignored files in the working directory
  in order to optimize `hg status`
  by enabling it to skip `readdir` in more cases.

  When a node is for a file tracked anywhere:
  - If `HAS_MODE_AND_SIZE` is set, the file is expected
    to be a symbolic link or a normal file based on `MODE_IS_SYMLINK`.
  - If `HAS_MODE_AND_SIZE` is set, the file’s owner is expected
    to have execute permission or not based on `MODE_EXEC_PERM`.
  - If `HAS_MODE_AND_SIZE` is unset,
    the expected type of file and permission are unknown.
  The rest of the node data is three fields:

  * Offset 31:
    4 unused bytes, set to zero

  * Offset 35:
    If `HAS_MODE_AND_SIZE` is unset, four zero bytes.
    Otherwise, a 32-bit integer for expected size of the file
    truncated to its 31 least-significant bits.
    Unlike in dirstate-v1, negative values are not used.

  * Offset 39:
    If `HAS_MTIME` is unset, four zero bytes.
    Otherwise, a 32-bit integer for expected modified time of the file
    (as in `stat_result.st_mtime`),
    truncated to its 31 least-significant bits.
    Unlike in dirstate-v1, negative values are not used.

  If an untracked node `HAS_MTIME` *unset*, this space is unused:

  * Offset 31:
    12 unused bytes, set to zero

  If an untracked node `HAS_MTIME` *set*,
  what follows is the modification time of a directory
  represented similarly to the C `timespec` struct:

  * Offset 31:
    4 unused bytes, set to zero

  * Offset 35:
    The number of seconds elapsed since the Unix epoch,
    truncated to its lower 31 bits,
    as a 32-bit integer.

  * Offset 39:
    The sub-second number of nanoseconds elapsed since the Unix epoch,
    as 32-bit integer.
    Always greater than or equal to zero, and strictly less than a billion.

  The presence of a directory modification time means that at some point,
  this path in the working directory was observed:

  - To be a directory
  - With the given modification time
  - That time was already strictly in the past when observed,
    meaning that later changes cannot happen in the same clock tick
    and must cause a different modification time
    (unless the system clock jumps back and we get unlucky,
    which is not impossible but deemed unlikely enough).
  - All direct children of this directory
    (as returned by `std::fs::read_dir`)
    either have a corresponding dirstate node,
    or are ignored by ignore patterns whose hash is in tree metadata.

  This means that if `std::fs::symlink_metadata` later reports
  the same modification time
  and ignored patterns haven’t changed,
  a run of status that is not listing ignored files
  can skip calling `std::fs::read_dir` again for this directory,
  and iterate child dirstate nodes instead.


* (Offset 43: end of this node)