mirror/ git
1
0
Fork 0
mirror of https://github.com/git/git.git synced 2024-05-31 02:36:12 +02:00
Code
git/Documentation/technical/pack-format.txt
Taylor Blau 2f4ba2a867 packfile: prepare for the existence of '*.rev' files 
Specify the format of the on-disk reverse index 'pack-*.rev' file, as
well as prepare the code for the existence of such files.

The reverse index maps from pack relative positions (i.e., an index into
the array of object which is sorted by their offsets within the
packfile) to their position within the 'pack-*.idx' file. Today, this is
done by building up a list of (off_t, uint32_t) tuples for each object
(the off_t corresponding to that object's offset, and the uint32_t
corresponding to its position in the index). To convert between pack and
index position quickly, this array of tuples is radix sorted based on
its offset.

This has two major drawbacks:

First, the in-memory cost scales linearly with the number of objects in
a pack.  Each 'struct revindex_entry' is sizeof(off_t) +
sizeof(uint32_t) + padding bytes for a total of 16.

To observe this, force Git to load the reverse index by, for e.g.,
running 'git cat-file --batch-check="%(objectsize:disk)"'. When asking
for a single object in a fresh clone of the kernel, Git needs to
allocate 120+ MB of memory in order to hold the reverse index in memory.

Second, the cost to sort also scales with the size of the pack.
Luckily, this is a linear function since 'load_pack_revindex()' uses a
radix sort, but this cost still must be paid once per pack per process.

As an example, it takes ~60x longer to print the _size_ of an object as
it does to print that entire object's _contents_:

  Benchmark #1: git.compile cat-file --batch <obj
    Time (mean ± σ):       3.4 ms ±   0.1 ms    [User: 3.3 ms, System: 2.1 ms]
    Range (min … max):     3.2 ms …   3.7 ms    726 runs

  Benchmark #2: git.compile cat-file --batch-check="%(objectsize:disk)" <obj
    Time (mean ± σ):     210.3 ms ±   8.9 ms    [User: 188.2 ms, System: 23.2 ms]
    Range (min … max):   193.7 ms … 224.4 ms    13 runs

Instead, avoid computing and sorting the revindex once per process by
writing it to a file when the pack itself is generated.

The format is relatively straightforward. It contains an array of
uint32_t's, the length of which is equal to the number of objects in the
pack.  The ith entry in this table contains the index position of the
ith object in the pack, where "ith object in the pack" is determined by
pack offset.

One thing that the on-disk format does _not_ contain is the full (up to)
eight-byte offset corresponding to each object. This is something that
the in-memory revindex contains (it stores an off_t in 'struct
revindex_entry' along with the same uint32_t that the on-disk format
has). Omit it in the on-disk format, since knowing the index position
for some object is sufficient to get a constant-time lookup in the
pack-*.idx file to ask for an object's offset within the pack.

This trades off between the on-disk size of the 'pack-*.rev' file for
runtime to chase down the offset for some object. Even though the lookup
is constant time, the constant is heavier, since it can potentially
involve two pointer walks in v2 indexes (one to access the 4-byte offset
table, and potentially a second to access the double wide offset table).

Consider trying to map an object's pack offset to a relative position
within that pack. In a cold-cache scenario, more page faults occur while
switching between binary searching through the reverse index and
searching through the *.idx file for an object's offset. Sure enough,
with a cold cache (writing '3' into '/proc/sys/vm/drop_caches' after
'sync'ing), printing out the entire object's contents is still
marginally faster than printing its size:

  Benchmark #1: git.compile cat-file --batch-check="%(objectsize:disk)" <obj >/dev/null
    Time (mean ± σ):      22.6 ms ±   0.5 ms    [User: 2.4 ms, System: 7.9 ms]
    Range (min … max):    21.4 ms …  23.5 ms    41 runs

  Benchmark #2: git.compile cat-file --batch <obj >/dev/null
    Time (mean ± σ):      17.2 ms ±   0.7 ms    [User: 2.8 ms, System: 5.5 ms]
    Range (min … max):    15.6 ms …  18.2 ms    45 runs

(Numbers taken in the kernel after cheating and using the next patch to
generate a reverse index). There are a couple of approaches to improve
cold cache performance not pursued here:

  - We could include the object offsets in the reverse index format.
    Predictably, this does result in fewer page faults, but it triples
    the size of the file, while simultaneously duplicating a ton of data
    already available in the .idx file. (This was the original way I
    implemented the format, and it did show
    `--batch-check='%(objectsize:disk)'` winning out against `--batch`.)

    On the other hand, this increase in size also results in a large
    block-cache footprint, which could potentially hurt other workloads.

  - We could store the mapping from pack to index position in more
    cache-friendly way, like constructing a binary search tree from the
    table and writing the values in breadth-first order. This would
    result in much better locality, but the price you pay is trading
    O(1) lookup in 'pack_pos_to_index()' for an O(log n) one (since you
    can no longer directly index the table).

So, neither of these approaches are taken here. (Thankfully, the format
is versioned, so we are free to pursue these in the future.) But, cold
cache performance likely isn't interesting outside of one-off cases like
asking for the size of an object directly. In real-world usage, Git is
often performing many operations in the revindex (i.e., asking about
many objects rather than a single one).

The trade-off is worth it, since we will avoid the vast majority of the
cost of generating the revindex that the extra pointer chase will look
like noise in the following patch's benchmarks.

This patch describes the format and prepares callers (like in
pack-revindex.c) to be able to read *.rev files once they exist. An
implementation of the writer will appear in the next patch, and callers
will gradually begin to start using the writer in the patches that
follow after that.

Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-01-25 18:32:43 -08:00

379 lines
13 KiB
Plaintext
Raw Blame History

Git pack format
===============
== Checksums and object IDs
In a repository using the traditional SHA-1, pack checksums, index checksums,
and object IDs (object names) mentioned below are all computed using SHA-1.
Similarly, in SHA-256 repositories, these values are computed using SHA-256.
== pack-*.pack files have the following format:
- A header appears at the beginning and consists of the following:
4-byte signature:
The signature is: {'P', 'A', 'C', 'K'}
4-byte version number (network byte order):
Git currently accepts version number 2 or 3 but
generates version 2 only.
4-byte number of objects contained in the pack (network byte order)
Observation: we cannot have more than 4G versions ;-) and
more than 4G objects in a pack.
- The header is followed by number of object entries, each of
which looks like this:
(undeltified representation)
n-byte type and length (3-bit type, (n-1)*7+4-bit length)
compressed data
(deltified representation)
n-byte type and length (3-bit type, (n-1)*7+4-bit length)
base object name if OBJ_REF_DELTA or a negative relative
offset from the delta object's position in the pack if this
is an OBJ_OFS_DELTA object
compressed delta data
Observation: length of each object is encoded in a variable
length format and is not constrained to 32-bit or anything.
- The trailer records a pack checksum of all of the above.
=== Object types
Valid object types are:
- OBJ_COMMIT (1)
- OBJ_TREE (2)
- OBJ_BLOB (3)
- OBJ_TAG (4)
- OBJ_OFS_DELTA (6)
- OBJ_REF_DELTA (7)
Type 5 is reserved for future expansion. Type 0 is invalid.
=== Size encoding
This document uses the following "size encoding" of non-negative
integers: From each byte, the seven least significant bits are
used to form the resulting integer. As long as the most significant
bit is 1, this process continues; the byte with MSB 0 provides the
last seven bits. The seven-bit chunks are concatenated. Later
values are more significant.
This size encoding should not be confused with the "offset encoding",
which is also used in this document.
=== Deltified representation
Conceptually there are only four object types: commit, tree, tag and
blob. However to save space, an object could be stored as a "delta" of
another "base" object. These representations are assigned new types
ofs-delta and ref-delta, which is only valid in a pack file.
Both ofs-delta and ref-delta store the "delta" to be applied to
another object (called 'base object') to reconstruct the object. The
difference between them is, ref-delta directly encodes base object
name. If the base object is in the same pack, ofs-delta encodes
the offset of the base object in the pack instead.
The base object could also be deltified if it's in the same pack.
Ref-delta can also refer to an object outside the pack (i.e. the
so-called "thin pack"). When stored on disk however, the pack should
be self contained to avoid cyclic dependency.
The delta data starts with the size of the base object and the
size of the object to be reconstructed. These sizes are
encoded using the size encoding from above. The remainder of
the delta data is a sequence of instructions to reconstruct the object
from the base object. If the base object is deltified, it must be
converted to canonical form first. Each instruction appends more and
more data to the target object until it's complete. There are two
supported instructions so far: one for copy a byte range from the
source object and one for inserting new data embedded in the
instruction itself.
Each instruction has variable length. Instruction type is determined
by the seventh bit of the first octet. The following diagrams follow
the convention in RFC 1951 (Deflate compressed data format).
==== Instruction to copy from base object
+----------+---------+---------+---------+---------+-------+-------+-------+
| 1xxxxxxx | offset1 | offset2 | offset3 | offset4 | size1 | size2 | size3 |
+----------+---------+---------+---------+---------+-------+-------+-------+
This is the instruction format to copy a byte range from the source
object. It encodes the offset to copy from and the number of bytes to
copy. Offset and size are in little-endian order.
All offset and size bytes are optional. This is to reduce the
instruction size when encoding small offsets or sizes. The first seven
bits in the first octet determines which of the next seven octets is
present. If bit zero is set, offset1 is present. If bit one is set
offset2 is present and so on.
Note that a more compact instruction does not change offset and size
encoding. For example, if only offset2 is omitted like below, offset3
still contains bits 16-23. It does not become offset2 and contains
bits 8-15 even if it's right next to offset1.
+----------+---------+---------+
| 10000101 | offset1 | offset3 |
+----------+---------+---------+
In its most compact form, this instruction only takes up one byte
(0x80) with both offset and size omitted, which will have default
values zero. There is another exception: size zero is automatically
converted to 0x10000.
==== Instruction to add new data
+----------+============+
| 0xxxxxxx | data |
+----------+============+
This is the instruction to construct target object without the base
object. The following data is appended to the target object. The first
seven bits of the first octet determines the size of data in
bytes. The size must be non-zero.
==== Reserved instruction
+----------+============
| 00000000 |
+----------+============
This is the instruction reserved for future expansion.
== Original (version 1) pack-*.idx files have the following format:
- The header consists of 256 4-byte network byte order
integers. N-th entry of this table records the number of
objects in the corresponding pack, the first byte of whose
object name is less than or equal to N. This is called the
'first-level fan-out' table.
- The header is followed by sorted 24-byte entries, one entry
per object in the pack. Each entry is:
4-byte network byte order integer, recording where the
object is stored in the packfile as the offset from the
beginning.
one object name of the appropriate size.
- The file is concluded with a trailer:
A copy of the pack checksum at the end of the corresponding
packfile.
Index checksum of all of the above.
Pack Idx file:
-- +--------------------------------+
fanout | fanout[0] = 2 (for example) |-.
table +--------------------------------+ |
| fanout[1] | |
+--------------------------------+ |
| fanout[2] | |
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
| fanout[255] = total objects |---.
-- +--------------------------------+ | |
main | offset | | |
index | object name 00XXXXXXXXXXXXXXXX | | |
table +--------------------------------+ | |
| offset | | |
| object name 00XXXXXXXXXXXXXXXX | | |
+--------------------------------+<+ |
.-| offset | |
| | object name 01XXXXXXXXXXXXXXXX | |
| +--------------------------------+ |
| | offset | |
| | object name 01XXXXXXXXXXXXXXXX | |
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
| | offset | |
| | object name FFXXXXXXXXXXXXXXXX | |
--| +--------------------------------+<--+
trailer | | packfile checksum |
| +--------------------------------+
| | idxfile checksum |
| +--------------------------------+
.-------.
|
Pack file entry: <+
packed object header:
1-byte size extension bit (MSB)
type (next 3 bit)
size0 (lower 4-bit)
n-byte sizeN (as long as MSB is set, each 7-bit)
size0..sizeN form 4+7+7+..+7 bit integer, size0
is the least significant part, and sizeN is the
most significant part.
packed object data:
If it is not DELTA, then deflated bytes (the size above
is the size before compression).
If it is REF_DELTA, then
base object name (the size above is the
size of the delta data that follows).
delta data, deflated.
If it is OFS_DELTA, then
n-byte offset (see below) interpreted as a negative
offset from the type-byte of the header of the
ofs-delta entry (the size above is the size of
the delta data that follows).
delta data, deflated.
offset encoding:
n bytes with MSB set in all but the last one.
The offset is then the number constructed by
concatenating the lower 7 bit of each byte, and
for n >= 2 adding 2^7 + 2^14 + ... + 2^(7*(n-1))
to the result.
== Version 2 pack-*.idx files support packs larger than 4 GiB, and
have some other reorganizations. They have the format:
- A 4-byte magic number '\377tOc' which is an unreasonable
fanout[0] value.
- A 4-byte version number (= 2)
- A 256-entry fan-out table just like v1.
- A table of sorted object names. These are packed together
without offset values to reduce the cache footprint of the
binary search for a specific object name.
- A table of 4-byte CRC32 values of the packed object data.
This is new in v2 so compressed data can be copied directly
from pack to pack during repacking without undetected
data corruption.
- A table of 4-byte offset values (in network byte order).
These are usually 31-bit pack file offsets, but large
offsets are encoded as an index into the next table with
the msbit set.
- A table of 8-byte offset entries (empty for pack files less
than 2 GiB). Pack files are organized with heavily used
objects toward the front, so most object references should
not need to refer to this table.
- The same trailer as a v1 pack file:
A copy of the pack checksum at the end of
corresponding packfile.
Index checksum of all of the above.
== pack-*.rev files have the format:
- A 4-byte magic number '0x52494458' ('RIDX').
- A 4-byte version identifier (= 1).
- A 4-byte hash function identifier (= 1 for SHA-1, 2 for SHA-256).
- A table of index positions (one per packed object, num_objects in
total, each a 4-byte unsigned integer in network order), sorted by
their corresponding offsets in the packfile.
- A trailer, containing a:
checksum of the corresponding packfile, and
a checksum of all of the above.
All 4-byte numbers are in network order.
== multi-pack-index (MIDX) files have the following format:
The multi-pack-index files refer to multiple pack-files and loose objects.
In order to allow extensions that add extra data to the MIDX, we organize
the body into "chunks" and provide a lookup table at the beginning of the
body. The header includes certain length values, such as the number of packs,
the number of base MIDX files, hash lengths and types.
All 4-byte numbers are in network order.
HEADER:
4-byte signature:
The signature is: {'M', 'I', 'D', 'X'}
1-byte version number:
Git only writes or recognizes version 1.
1-byte Object Id Version
We infer the length of object IDs (OIDs) from this value:
1 => SHA-1
2 => SHA-256
If the hash type does not match the repository's hash algorithm,
the multi-pack-index file should be ignored with a warning
presented to the user.
1-byte number of "chunks"
1-byte number of base multi-pack-index files:
This value is currently always zero.
4-byte number of pack files
CHUNK LOOKUP:
(C + 1) * 12 bytes providing the chunk offsets:
First 4 bytes describe chunk id. Value 0 is a terminating label.
Other 8 bytes provide offset in current file for chunk to start.
(Chunks are provided in file-order, so you can infer the length
using the next chunk position if necessary.)
The remaining data in the body is described one chunk at a time, and
these chunks may be given in any order. Chunks are required unless
otherwise specified.
CHUNK DATA:
Packfile Names (ID: {'P', 'N', 'A', 'M'})
Stores the packfile names as concatenated, null-terminated strings.
Packfiles must be listed in lexicographic order for fast lookups by
name. This is the only chunk not guaranteed to be a multiple of four
bytes in length, so should be the last chunk for alignment reasons.
OID Fanout (ID: {'O', 'I', 'D', 'F'})
The ith entry, F[i], stores the number of OIDs with first
byte at most i. Thus F[255] stores the total
number of objects.
OID Lookup (ID: {'O', 'I', 'D', 'L'})
The OIDs for all objects in the MIDX are stored in lexicographic
order in this chunk.
Object Offsets (ID: {'O', 'O', 'F', 'F'})
Stores two 4-byte values for every object.
1: The pack-int-id for the pack storing this object.
2: The offset within the pack.
If all offsets are less than 2^32, then the large offset chunk
will not exist and offsets are stored as in IDX v1.
If there is at least one offset value larger than 2^32-1, then
the large offset chunk must exist, and offsets larger than
2^31-1 must be stored in it instead. If the large offset chunk
exists and the 31st bit is on, then removing that bit reveals
the row in the large offsets containing the 8-byte offset of
this object.
[Optional] Object Large Offsets (ID: {'L', 'O', 'F', 'F'})
8-byte offsets into large packfiles.
TRAILER:
Index checksum of the above contents.
View Git Blame Copy Permalink