To get the list of all promisor objects, we not only include all objects
in promisor packs, but also parse each of those objects to see which
objects they reference. After parsing a tree object, the tree->buffer
field will remain populated until we explicitly free it. So in a partial
clone of blob:none, for example, we are essentially reading every tree
in the repository (since they're all in the initial promisor pack), and
keeping all of their uncompressed contents in memory at once.
This patch frees the tree buffers after we've finished marking all of
their reachable objects. We shouldn't need to do this for any other
object type. While we are using some extra memory to store the structs,
no other object type stores the whole contents in its parsed form (we do
sometimes hold on to commit buffers, but less so these days due to
commit graphs, plus most commands which care about promisor objects turn
off the save_commit_buffer global).
Even for a moderate-sized repository like git.git, this patch drops the
peak heap (as measured by massif) for git-fsck from ~1.7GB to ~138MB.
Fsck is a good candidate for measuring here because it doesn't interact
with the promisor code except to call is_promisor_object(), so we can
isolate just this problem.
The added perf test shows only a tiny improvement on my machine for
git.git, since 1.7GB isn't enough to cause any real memory pressure:
Test HEAD^ HEAD
--------------------------------------------------------------------------------
5600.4: fsck 21.26(20.90+0.35) 20.84(20.79+0.04) -2.0%
With linux.git the absolute change is a bit bigger, though still a small
percentage:
Test HEAD^ HEAD
-----------------------------------------------------------------------------
5600.4: fsck 262.26(259.13+3.12) 254.92(254.62+0.29) -2.8%
I didn't have the patience to run it under massif with linux.git, but
it's probably on the order of about 14GB improvement, since that's the
sum of the sizes of all of the uncompressed trees (but still isn't
enough to create memory pressure on this particular machine, which has
64GB of RAM). Smaller machines would probably see a bigger effect on
runtime (and sadly our perf suite does not measure peak heap).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
An on-disk reverse-index to map the in-pack location of an object
back to its object name across multiple packfiles is introduced.
* tb/reverse-midx:
midx.c: improve cache locality in midx_pack_order_cmp()
pack-revindex: write multi-pack reverse indexes
pack-write.c: extract 'write_rev_file_order'
pack-revindex: read multi-pack reverse indexes
Documentation/technical: describe multi-pack reverse indexes
midx: make some functions non-static
midx: keep track of the checksum
midx: don't free midx_name early
midx: allow marking a pack as preferred
t/helper/test-read-midx.c: add '--show-objects'
builtin/multi-pack-index.c: display usage on unrecognized command
builtin/multi-pack-index.c: don't enter bogus cmd_mode
builtin/multi-pack-index.c: split sub-commands
builtin/multi-pack-index.c: define common usage with a macro
builtin/multi-pack-index.c: don't handle 'progress' separately
builtin/multi-pack-index.c: inline 'flags' with options
Implement reading for multi-pack reverse indexes, as described in the
previous patch.
Note that these functions don't yet have any callers, and won't until
multi-pack reachability bitmaps are introduced in a later patch series.
In the meantime, this patch implements some of the infrastructure
necessary to support multi-pack bitmaps.
There are three new functions exposed by the revindex API:
- load_midx_revindex(): loads the reverse index corresponding to the
given multi-pack index.
- midx_to_pack_pos() and pack_pos_to_midx(): these convert between the
multi-pack index and pseudo-pack order.
load_midx_revindex() and pack_pos_to_midx() are both relatively
straightforward.
load_midx_revindex() needs a few functions to be exposed from the midx
API. One to get the checksum of a midx, and another to get the .rev's
filename. Similar to recent changes in the packed_git struct, three new
fields are added to the multi_pack_index struct: one to keep track of
the size, one to keep track of the mmap'd pointer, and another to point
past the header and at the reverse index's data.
pack_pos_to_midx() simply reads the corresponding entry out of the
table.
midx_to_pack_pos() is the trickiest, since it needs to find an object's
position in the psuedo-pack order, but that order can only be recovered
in the .rev file itself. This mapping can be implemented with a binary
search, but note that the thing we're binary searching over isn't an
array of values, but rather a permuted order of those values.
So, when comparing two items, it's helpful to keep in mind the
difference. Instead of a traditional binary search, where you are
comparing two things directly, here we're comparing a (pack, offset)
tuple with an index into the multi-pack index. That index describes
another (pack, offset) tuple, and it is _those_ two tuples that are
compared.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git repack" so far has been only capable of repacking everything
under the sun into a single pack (or split by size). A cleverer
strategy to reduce the cost of repacking a repository has been
introduced.
* tb/geometric-repack:
builtin/pack-objects.c: ignore missing links with --stdin-packs
builtin/repack.c: reword comment around pack-objects flags
builtin/repack.c: be more conservative with unsigned overflows
builtin/repack.c: assign pack split later
t7703: test --geometric repack with loose objects
builtin/repack.c: do not repack single packs with --geometric
builtin/repack.c: add '--geometric' option
packfile: add kept-pack cache for find_kept_pack_entry()
builtin/pack-objects.c: rewrite honor-pack-keep logic
p5303: measure time to repack with keep
p5303: add missing &&-chains
builtin/pack-objects.c: add '--stdin-packs' option
revision: learn '--no-kept-objects'
packfile: introduce 'find_kept_pack_entry()'
Add and apply a semantic patch for converting code that open-codes
CALLOC_ARRAY to use it instead. It shortens the code and infers the
element size automatically.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In a recent patch we added a function 'find_kept_pack_entry()' to look
for an object only among kept packs.
While this function avoids doing any lookup work in non-kept packs, it
is still linear in the number of packs, since we have to traverse the
linked list of packs once per object. Let's cache a reduced version of
that list to save us time.
Note that this cache will last the lifetime of the program. We could
invalidate it on reprepare_packed_git(), but there's not much point in
being rigorous here:
- we might already fail to notice new .keep packs showing up after the
program starts. We only reprepare_packed_git() when we fail to find
an object. But adding a new pack won't cause that to happen.
Somebody repacking could add a new pack and delete an old one, but
most of the time we'd have a descriptor or mmap open to the old
pack anyway, so we might not even notice.
- in pack-objects we already cache the .keep state at startup, since
56dfeb6263 (pack-objects: compute local/ignore_pack_keep early,
2016-07-29). So this is just extending that concept further.
- we don't have to worry about any packed_git being removed; we always
keep the old structs around, even after reprepare_packed_git()
We do defensively invalidate the cache in case the set of kept packs
being asked for changes (e.g., only in-core kept packs were cached, but
suddenly the caller also wants on-disk kept packs, too). In theory we
could build all three caches and switch between them, but it's not
necessary, since this patch (and series) never changes the set of kept
packs that it wants to inspect from the cache.
So that "optimization" is more about being defensive in the face of
future changes than it is about asking for multiple kinds of kept packs
in this patch.
Here are p5303 results (as always, measured against the kernel):
Test HEAD^ HEAD
-----------------------------------------------------------------------------------------------
5303.5: repack (1) 57.34(54.66+10.88) 56.98(54.36+10.98) -0.6%
5303.6: repack with kept (1) 57.38(54.83+10.49) 57.17(54.97+10.26) -0.4%
5303.11: repack (50) 71.70(88.99+4.74) 71.62(88.48+5.08) -0.1%
5303.12: repack with kept (50) 72.58(89.61+4.78) 71.56(88.80+4.59) -1.4%
5303.17: repack (1000) 217.19(491.72+14.25) 217.31(490.82+14.53) +0.1%
5303.18: repack with kept (1000) 246.12(520.07+14.93) 217.08(490.37+15.10) -11.8%
and the --stdin-packs case, which scales a little bit better (although
not by that much even at 1,000 packs):
5303.7: repack with --stdin-packs (1) 0.00(0.00+0.00) 0.00(0.00+0.00) =
5303.13: repack with --stdin-packs (50) 3.43(11.75+0.24) 3.43(11.69+0.30) +0.0%
5303.19: repack with --stdin-packs (1000) 130.50(307.15+7.66) 125.13(301.36+8.04) -4.1%
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Future callers will want a function to fill a 'struct pack_entry' for a
given object id but _only_ from its position in any kept pack(s).
In particular, an new 'git repack' mode which ensures the resulting
packs form a geometric progress by object count will mark packs that it
does not want to repack as "kept in-core", and it will want to halt a
reachability traversal as soon as it visits an object in any of the kept
packs. But, it does not want to halt the traversal at non-kept, or
.keep packs.
The obvious alternative is 'find_pack_entry()', but this doesn't quite
suffice since it only returns the first pack it finds, which may or may
not be kept (and the mru cache makes it unpredictable which one you'll
get if there are options).
Short of that, you could walk over all packs looking for the object in
each one, but it scales with the number of packs, which may be
prohibitive.
Introduce 'find_kept_pack_entry()', a function which is like
'find_pack_entry()', but only fills in objects in the kept packs.
Handle packs which have .keep files, as well as in-core kept packs
separately, since certain callers will want to distinguish one from the
other. (Though on-disk and in-core kept packs share the adjective
"kept", it is best to think of the two sets as independent.)
There is a gotcha when looking up objects that are duplicated in kept
and non-kept packs, particularly when the MIDX stores the non-kept
version and the caller asked for kept objects only. This could be
resolved by teaching the MIDX to resolve duplicates by always favoring
the kept pack (if one exists), but this breaks an assumption in existing
MIDXs, and so it would require a format change.
The benefit to changing the MIDX in this way is marginal, so we instead
have a more thorough check here which is explained with a comment.
Callers will be added in subsequent patches.
Co-authored-by: Jeff King <peff@peff.net>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Reviewed-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
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>
Abstract accesses to in-core revindex that allows enumerating
objects stored in a packfile in the order they appear in the pack,
in preparation for introducing an on-disk precomputed revindex.
* tb/pack-revindex-api: (21 commits)
for_each_object_in_pack(): clarify pack vs index ordering
pack-revindex.c: avoid direct revindex access in 'offset_to_pack_pos()'
pack-revindex: hide the definition of 'revindex_entry'
pack-revindex: remove unused 'find_revindex_position()'
pack-revindex: remove unused 'find_pack_revindex()'
builtin/gc.c: guess the size of the revindex
for_each_object_in_pack(): convert to new revindex API
unpack_entry(): convert to new revindex API
packed_object_info(): convert to new revindex API
retry_bad_packed_offset(): convert to new revindex API
get_delta_base_oid(): convert to new revindex API
rebuild_existing_bitmaps(): convert to new revindex API
try_partial_reuse(): convert to new revindex API
get_size_by_pos(): convert to new revindex API
show_objects_for_type(): convert to new revindex API
bitmap_position_packfile(): convert to new revindex API
check_object(): convert to new revindex API
write_reused_pack_verbatim(): convert to new revindex API
write_reused_pack_one(): convert to new revindex API
write_reuse_object(): convert to new revindex API
...
We may return objects in one of two orders: how they appear in the .idx
(sorted by object id) or how they appear in the packfile itself. To
further complicate matters, we have two ordering variables, "i" and
"pos", and it is not clear to which order they apply.
Let's clarify this by using an unambiguous name where possible, and
leaving a comment for the variable that does double-duty.
Signed-off-by: Jeff King <peff@peff.net>
Acked-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Avoid looking at the 'revindex' pointer directly and instead call
'pack_pos_to_index()'.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Remove direct manipulation of the 'struct revindex_entry' type as well
as calls to the deprecated API in 'packfile.c:unpack_entry()'. Usual
clean-up is performed (replacing '->nr' with calls to
'pack_pos_to_index()' and so on).
Add an additional check to make sure that 'obj_offset()' points at a
valid object. In the case this check is violated, we cannot call
'mark_bad_packed_object()' because we don't know the OID. At the top of
the call stack is do_oid_object_info_extended() (via
packed_object_info()), which does mark the object.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Convert another call of 'find_pack_revindex()' to its replacement
'pack_pos_to_offset()'. Likewise:
- Avoid manipulating `struct packed_git`'s `revindex` pointer directly
by removing the pointer-as-array indexing.
- Add an additional guard to check that the offset 'obj_offset()'
points to a real object. This should be the case with well-behaved
callers to 'packed_object_info()', but isn't guarenteed.
Other blocks that fill in various other values from the 'struct
object_info' request handle bad inputs by setting the type to
'OBJ_BAD' and jumping to 'out'. Do the same when given a bad offset
here.
The previous code would have segfaulted when given a bad
'obj_offset' value, since 'find_pack_revindex()' would return
'NULL', and then the line that fills 'oi->disk_sizep' would try to
access 'NULL[1]' with a stride of 16 bytes (the width of 'struct
revindex_entry)'.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Perform exactly the same conversion as in the previous commit to another
caller within 'packfile.c'.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Replace direct accesses to the 'struct revindex' type with a call to
'pack_pos_to_index()'.
Likewise drop the old-style 'find_pack_revindex()' with its replacement
'offset_to_pack_pos()' (while continuing to perform the same error
checking).
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Change all remnants of "sha1" in hash-lookup.c and .h and rename them to
reflect that we're not just able to handle SHA-1 these days.
Signed-off-by: Martin Ågren <martin.agren@gmail.com>
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Processes that access packdata while the .idx file gets removed
(e.g. while repacking) did not fail or fall back gracefully as they
could.
* tb/idx-midx-race-fix:
midx.c: protect against disappearing packs
packfile.c: protect against disappearing indexes
In 17c35c8969 (packfile: skip loading index if in multi-pack-index,
2018-07-12) we stopped loading the .idx file for packs that are
contained within a multi-pack index.
This saves us the effort of loading an .idx and doing some lightweight
validity checks by way of 'packfile.c:load_idx()', but introduces a race
between processes that need to load the index (e.g., to generate a
reverse index) and processes that can delete the index.
For example, running the following in your shell:
$ git init repo && cd repo
$ git commit --allow-empty -m 'base'
$ git repack -ad && git multi-pack-index write
followed by:
$ rm -f .git/objects/pack/pack-*.idx
$ git rev-parse HEAD | git cat-file --batch-check='%(objectsize:disk)'
will result in a segfault prior to this patch. What's happening here is
that we notice that the pack is in the multi-pack index, and so don't
check that it still has a .idx. When we then try and load that index to
generate a reverse index, we don't have it, so the call to
'find_pack_revindex()' in 'packfile.c:packed_object_info()' returns
NULL, and then dereferencing it causes a segfault.
Of course, we don't ever expect someone to remove the index file by
hand, or to be in a state where we never wrote it to begin with (yet
find that pack in the multi-pack-index). But, this can happen in a
timing race with 'git repack -ad', which removes all existing packs
after writing a new pack containing all of their objects.
Avoid this by reverting the hunk of 17c35c8969 which stops loading the
index when the pack is contained in a MIDX. This makes the latter half
of 17c35c8969 useless, since we'll always have a non-NULL
'p->index_data', in which case that if statement isn't guarding
anything.
These two together effectively revert 17c35c8969, and avoid the race
explained above.
Co-authored-by: Jeff King <peff@peff.net>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In load_idx(), we check that the .idx file is sized appropriately for
the number of objects it claims to have. We recently fixed the case
where the number of objects caused our expected size to overflow a
32-bit unsigned int, and we switched to size_t.
On a 64-bit system, this is fine; our size_t covers any expected size.
On a 32-bit system, though, it won't. The file may claim to have 2^31
objects, which will overflow even a size_t.
This doesn't hurt us at all for a well-formed idx file. A 32-bit system
would already have failed to mmap such a file, since it would be too
big. But an .idx file which _claims_ to have 2^31 objects but is
actually much smaller would fool our check.
This is a broken file, and for the most part we don't care that much
what happens. But:
- it's a little friendlier to notice up front "woah, this file is
broken" than it is to get nonsense results
- later access of the data assumes that the loading function
sanity-checked that we have at least enough bytes for the regular
object-id table. A malformed .idx file could lead to an
out-of-bounds read.
So let's use our overflow-checking functions to make sure that we're not
fooled by a malformed file.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We sometimes store the offset into a pack .idx file as an "unsigned
long", but the mmap'd size of a pack .idx file can exceed 4GB. This is
sufficient on LP64 systems like Linux, but will be too small on LLP64
systems like Windows, where "unsigned long" is still only 32 bits. Let's
use size_t, which is a better type for an offset into a memory buffer.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
A pack and its matching .idx file are limited to 2^32 objects, because
the pack format contains a 32-bit field to store the number of objects.
Hence we use uint32_t in the code.
But the byte count of even a .idx file can be much larger than that,
because it stores at least a hash and an offset for each object. So
using SHA-1, a v2 .idx file will cross the 4GB boundary at 153,391,650
objects. This confuses load_idx(), which computes the minimum size like
this:
unsigned long min_size = 8 + 4*256 + nr*(hashsz + 4 + 4) + hashsz + hashsz;
Even though min_size will be big enough on most 64-bit platforms, the
actual arithmetic is done as a uint32_t, resulting in a truncation. We
actually exceed that min_size, but then we do:
unsigned long max_size = min_size;
if (nr)
max_size += (nr - 1)*8;
to account for the variable-sized table. That computation doesn't
overflow quite so low, but with the truncation for min_size, we end up
with a max_size that is much smaller than our actual size. So we
complain that the idx is invalid, and can't find any of its objects.
We can fix this case by casting "nr" to a size_t, which will do the
multiplication in 64-bits (assuming you're on a 64-bit platform; this
will never work on a 32-bit system since we couldn't map the whole .idx
anyway). Likewise, we don't have to worry about further additions,
because adding a smaller number to a size_t will convert the other side
to a size_t.
A few notes:
- obviously we could just declare "nr" as a size_t in the first place
(and likewise, packed_git.num_objects). But it's conceptually a
uint32_t because of the on-disk format, and we correctly treat it
that way in other contexts that don't need to compute byte offsets
(e.g., iterating over the set of objects should and generally does
use a uint32_t). Switching to size_t would make all of those other
cases look wrong.
- it could be argued that the proper type is off_t to represent the
file offset. But in practice the .idx file must fit within memory,
because we mmap the whole thing. And the rest of the code (including
the idx_size variable we're comparing against) uses size_t.
- we'll add the same cast to the max_size arithmetic line. Even though
we're adding to a larger type, which will convert our result, the
multiplication is still done as a 32-bit value and can itself
overflow. I didn't check this with my test case, since it would need
an even larger pack (~530M objects), but looking at compiler output
shows that it works this way. The standard should agree, but I
couldn't find anything explicit in 6.3.1.8 ("usual arithmetic
conversions").
The case in load_idx() was the most immediate one that I was able to
trigger. After fixing it, looking up actual objects (including the very
last one in sha1 order) works in a test repo with 153,725,110 objects.
That's because bsearch_hash() works with uint32_t entry indices, and the
actual byte access:
int cmp = hashcmp(table + mi * stride, sha1);
is done with "stride" as a size_t, causing the uint32_t "mi" to be
promoted to a size_t. This is the way most code will access the index
data.
However, I audited all of the other byte-wise accesses of
packed_git.index_data, and many of the others are suspect (they are
similar to the max_size one, where we are adding to a properly sized
offset or directly to a pointer, but the multiplication in the
sub-expression can overflow). I didn't trigger any of these in practice,
but I believe they're potential problems, and certainly adding in the
cast is not going to hurt anything here.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
A race that leads to an access to a free'd data was corrected in
the codepath that reads pack files.
* mt/delta-base-cache-races:
packfile: fix memory leak in add_delta_base_cache()
packfile: fix race condition on unpack_entry()
When add_delta_base_cache() is called with a base that is already in the
cache, no operation is performed. But the check is done after allocating
space for a new entry, so we end up leaking memory on the early return.
In addition, the caller never free()'s the base as it expects the
function to take ownership of it. But the base is not released when we
skip insertion, so it also gets leaked. To fix these problems, move the
allocation of a new entry further down in add_delta_base_cache(), and
free() the base on early return.
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The third phase of unpack_entry() performs the following sequence in a
loop, until all the deltas enumerated in phase one are applied and the
entry is fully reconstructed:
1. Add the current base entry to the delta base cache
2. Unpack the next delta
3. Patch the unpacked delta on top of the base
When the optional object reading lock is enabled, the above steps will
be performed while holding the lock. However, step 2. momentarily
releases it so that inflation can be performed in parallel for increased
performance. Because the `base` buffer inserted in the cache at 1. is
not duplicated, another thread can potentially free() it while the lock
is released at 2. (e.g. when there is no space left in the cache to
insert another entry). In this case, the later attempt to dereference
`base` at 3. will cause a segmentation fault. This problem was observed
during a multithreaded git-grep execution on a repository with large
objects.
To fix the race condition (and later segmentation fault), let's reorder
the aforementioned steps so that `base` is only added to the cache at
the end. This will prevent the buffer from being released by another
thread while it is still in use. An alternative solution which would not
require the reordering would be to duplicate `base` before inserting it
in the cache. However, as Phil Hord mentioned, memcpy()'ing large bases
can negatively affect performance: in his experiments, this alternative
approach slowed git-grep down by 10% to 20%.
Reported-by: Phil Hord <phil.hord@gmail.com>
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There is a logic to estimate how many objects are in the
repository, which is mean to run once per process invocation, but
it ran every time the estimated value was requested.
* jk/dont-count-existing-objects-twice:
packfile: actually set approximate_object_count_valid
The approximate_object_count() function tries to compute the count only
once per process. But ever since it was introduced in 8e3f52d778
(find_unique_abbrev: move logic out of get_short_sha1(), 2016-10-03), we
failed to actually set the "valid" flag, meaning we'd compute it fresh
on every call.
This turns out not to be _too_ bad, because we're only iterating through
the packed_git list, and not making any system calls. But since it may
get called for every abbreviated hash we output, even this can add up if
you have many packs.
Here are before-and-after timings for a new perf test which just asks
rev-list to abbreviate each commit hash (the test repo is linux.git,
with commit-graphs):
Test origin HEAD
----------------------------------------------------------------------------
5303.3: rev-list (1) 28.91(28.46+0.44) 29.03(28.65+0.38) +0.4%
5303.4: abbrev-commit (1) 1.18(1.06+0.11) 1.17(1.02+0.14) -0.8%
5303.7: rev-list (50) 28.95(28.56+0.38) 29.50(29.17+0.32) +1.9%
5303.8: abbrev-commit (50) 3.67(3.56+0.10) 3.57(3.42+0.15) -2.7%
5303.11: rev-list (1000) 30.34(29.89+0.43) 30.82(30.35+0.46) +1.6%
5303.12: abbrev-commit (1000) 86.82(86.52+0.29) 77.82(77.59+0.22) -10.4%
5303.15: load 10,000 packs 0.08(0.02+0.05) 0.08(0.02+0.06) +0.0%
It doesn't help at all when we have 1 pack (5303.4), but we get a 10%
speedup when there are 1000 packs (5303.12). That's a modest speedup for
a case that's already slow and we'd hope to avoid in general (note how
slow it is even after, because we have to look in each of those packs
for abbreviations). But it's a one-line change that clearly matches the
original intent, so it seems worth doing.
The included perf test may also be useful for keeping an eye on any
regressions in the overall abbreviation code.
Reported-by: Rasmus Villemoes <rv@rasmusvillemoes.dk>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When a repository has an alternate object directory configured, callers
can traverse through each alternate's MIDX by walking the '->next'
pointer.
But, when 'prepare_multi_pack_index_one()' loads multiple MIDXs, it
places the new ones at the front of this pointer chain, not at the end.
This can be confusing for callers such as 'git repack -ad', causing test
failures like in t7700.6 with 'GIT_TEST_MULTI_PACK_INDEX=1'.
The occurs when dropping a pack known to the local MIDX with alternates
configured that have their own MIDX. Since the alternate's MIDX is
returned via 'get_multi_pack_index()', 'midx_contains_pack()' returns
true (which is correct, since it traverses through the '->next' pointer
to find the MIDX in the chain that does contain the requested object).
But, we call 'clear_midx_file()' on 'the_repository', which drops the
MIDX at the path of the first MIDX in the chain, which (in the case of
t7700.6 is the one in the alternate).
This patch addresses that by:
- placing the local MIDX first in the chain when calling
'prepare_multi_pack_index_one()', and
- introducing a new 'get_local_multi_pack_index()', which explicitly
returns the repository-local MIDX, if any.
Don't impose an additional order on the MIDX's '->next' pointer beyond
that the first item in the chain must be local if one exists so that we
avoid a quadratic insertion.
Likewise, use 'get_local_multi_pack_index()' in
'remove_redundant_pack()' to fix the formerly broken t7700.6 when run
with 'GIT_TEST_MULTI_PACK_INDEX=1'.
Finally, note that the MIDX ordering invariant is only preserved by the
insertion order in 'prepare_packed_git()', which traverses through the
ODB's '->next' pointer, meaning we visit the local object store first.
This fragility makes this an undesirable long-term solution if more
callers are added, but it is acceptable for now since this is the only
caller.
Helped-by: Jeff King <peff@peff.net>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Both v2 pack index files and the v3 format specified as part of the
NewHash work have similar data starting at the CRC table. Much of the
existing code wants to read either this table or the offset entries
following it, and in doing so computes the offset each time.
In order to share as much code between v2 and v3, compute the offset of
the CRC table and store it when the pack is opened. Use this value to
compute offsets to not only the CRC table, but to the offset entries
beyond it.
Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Once upon a time, nth_packed_object_sha1() was the primary way to get
the oid of a packfile's index position. But these days we have the more
type-safe nth_packed_object_id() wrapper, and all callers have been
converted.
Let's drop the "sha1" version (turning the safer wrapper into a single
function) so that nobody is tempted to introduce new callers.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The previous commit changed the public interface of packed_object_info()
to return a struct object_id rather than a bare hash. That enables us to
convert our internal helper, as well. We can use nth_packed_object_id()
directly for OFS_DELTA, but we'll still have to use oidread() to pull
the hash for a REF_DELTA out of the packfile.
There should be no additional cost, since we're copying directly into
the object_id the caller provided us (just as we did before; it's just
happening now via nth_packed_object_id()).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If a caller sets the object_info.delta_base_sha1 to a non-NULL pointer,
we'll write the oid of the object's delta base to it. But we can
increase our type safety by switching this to a real object_id struct.
All of our callers are just pointing into the hash member of an
object_id anyway, so there's no inconvenience.
Note that we do still keep it as a pointer-to-struct, because the NULL
sentinel value tells us whether the caller is even interested in the
information.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Our nth_packed_object_sha1() function returns NULL for error. So when we
wrapped it with nth_packed_object_oid(), we kept the same semantics. But
it's a bit funny, because the caller actually passes in an out
parameter, and the pointer we return is just that same struct they
passed to us (or NULL).
It's not too terrible, but it does make the interface a little
non-idiomatic. Let's switch to our usual "0 for success, negative for
error" return value. Most callers either don't check it, or are
trivially converted. The one that requires the biggest change is
actually improved, as we can ditch an extra aliased pointer variable.
Since we are changing the interface in a subtle way that the compiler
wouldn't catch, let's also change the name to catch any topics in
flight. We can drop the 'o' and make it nth_packed_object_id(). That's
slightly shorter, but also less redundant since the 'o' stands for
"object" already.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Traditionally, we avoided threaded grep while searching in objects
(as opposed to files in the working tree) as accesses to the object
layer is not thread-safe. This limitation is getting lifted.
* mt/threaded-grep-in-object-store:
grep: use no. of cores as the default no. of threads
grep: move driver pre-load out of critical section
grep: re-enable threads in non-worktree case
grep: protect packed_git [re-]initialization
grep: allow submodule functions to run in parallel
submodule-config: add skip_if_read option to repo_read_gitmodules()
grep: replace grep_read_mutex by internal obj read lock
object-store: allow threaded access to object reading
replace-object: make replace operations thread-safe
grep: fix racy calls in grep_objects()
grep: fix race conditions at grep_submodule()
grep: fix race conditions on userdiff calls
Some fields in struct raw_object_store are lazy initialized by the
thread-unsafe packfile.c:prepare_packed_git(). Although this function is
present in the call stack of git-grep threads, all paths to it are
currently protected by obj_read_lock() (and the main thread usually
indirectly calls it before firing the worker threads, anyway). However,
it's possible that future modifications add new unprotected paths to it,
introducing a race condition. Because errors derived from it wouldn't
happen often, it could be hard to detect. So to prevent future
headaches, let's force eager initialization of packed_git when setting
git-grep up. There'll be a small overhead in the cases where we didn't
really need to prepare packed_git during execution but this shouldn't be
very noticeable.
Also, packed_git may be re-initialized by
packfile.c:reprepare_packed_git(). Again, all paths to it in git-grep
are already protected by obj_read_lock() but it may suffer from the same
problem in the future. So let's also internally protect it with
obj_read_lock() (which is a recursive mutex).
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Allow object reading to be performed by multiple threads protecting it
with an internal lock, the obj_read_mutex. The lock usage can be toggled
with enable_obj_read_lock() and disable_obj_read_lock(). Currently, the
functions which can be safely called in parallel are:
read_object_file_extended(), repo_read_object_file(),
read_object_file(), read_object_with_reference(), read_object(),
oid_object_info() and oid_object_info_extended(). It's also possible
to use obj_read_lock() and obj_read_unlock() to protect other sections
that cannot execute in parallel with object reading.
Probably there are many spots in the functions listed above that could
be executed unlocked (and thus, in parallel). But, for now, we are most
interested in allowing parallel access to zlib inflation. This is one of
the sections where object reading spends most of the time in (e.g. up to
one-third of git-grep's execution time in the chromium repo corresponds
to inflation) and it's already thread-safe. So, to take advantage of
that, the obj_read_mutex is released when calling git_inflate() and
re-acquired right after, for every calling spot in
oid_object_info_extended()'s call chain. We may refine this lock to also
exploit other possible parallel spots in the future, but for now,
threaded zlib inflation should already give great speedups for threaded
object reading callers.
Note that add_delta_base_cache() was also modified to skip adding
already present entries to the cache. This wasn't possible before, but
it would be now, with the parallel inflation. Take for example the
following situation, where two threads - A and B - are executing the
code at unpack_entry():
1. Thread A is performing the decompression of a base O (which is not
yet in the cache) at PHASE II. Thread B is simultaneously trying to
unpack O, but just starting at PHASE I.
2. Since O is not yet in the cache, B will go to PHASE II to also
perform the decompression.
3. When they finish decompressing, one of them will get the object
reading mutex and go to PHASE III while the other waits for the
mutex. Let’s say A got the mutex first.
4. Thread A will add O to the cache, go throughout the rest of PHASE III
and return.
5. Thread B gets the mutex, also add O to the cache (if the check wasn't
there) and returns.
Finally, it is also important to highlight that the object reading lock
can only ensure thread-safety in the mentioned functions thanks to two
complementary mechanisms: the use of 'struct raw_object_store's
replace_mutex, which guards sections in the object reading machinery
that would otherwise be thread-unsafe; and the 'struct pack_window's
inuse_cnt, which protects window reading operations (such as the one
performed during the inflation of a packed object), allowing them to
execute without the acquisition of the obj_read_mutex.
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We already have pread emulation for portability, so there's
there's no reason to make two syscalls where one suffices.
Furthermore, readers of the packfile will be using mmap
(or pread to emulate mmap), anyways, so the file description
offset does not matter in this case.
Signed-off-by: Eric Wong <e@80x24.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
git_open sets close-on-exec since cd66ada065
("sha1_file: open window into packfiles with O_CLOEXEC").
There's no reason to keep using fcntl to set the close-on-exec
flag, anymore.
Signed-off-by: Eric Wong <e@80x24.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When loading packfiles on start-up, we traverse the internal packfile
list once per file to avoid reloading packfiles that have already
been loaded. This check runs in quadratic time, so for poorly
maintained repos with a large number of packfiles, it can be pretty
slow.
Add a hashmap containing the packfile names as we load them so that
the average runtime cost of checking for already-loaded packs becomes
constant.
Add a perf test to p5303 to show speed-up.
The existing p5303 test runtimes are dominated by other factors and do
not show an appreciable speed-up. The new test in p5303 clearly exposes
a speed-up in bad cases. In this test we create 10,000 packfiles and
measure the start-up time of git rev-parse, which does little else
besides load in the packs.
Here are the numbers for the new p5303 test:
Test HEAD^ HEAD
---------------------------------------------------------------------
5303.12: load 10,000 packs 1.03(0.92+0.10) 0.12(0.02+0.09) -88.3%
Signed-off-by: Colin Stolley <cstolley@runbox.com>
Helped-by: Jeff King <peff@peff.net>
[jc: squashed the change to call hashmap in install_packed_git() by peff]
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Preparation for SHA-256 upgrade continues.
* bc/object-id-part17: (26 commits)
midx: switch to using the_hash_algo
builtin/show-index: replace sha1_to_hex
rerere: replace sha1_to_hex
builtin/receive-pack: replace sha1_to_hex
builtin/index-pack: replace sha1_to_hex
packfile: replace sha1_to_hex
wt-status: convert struct wt_status to object_id
cache: remove null_sha1
builtin/worktree: switch null_sha1 to null_oid
builtin/repack: write object IDs of the proper length
pack-write: use hash_to_hex when writing checksums
sequencer: convert to use the_hash_algo
bisect: switch to using the_hash_algo
sha1-lookup: switch hard-coded constants to the_hash_algo
config: use the_hash_algo in abbrev comparison
combine-diff: replace GIT_SHA1_HEXSZ with the_hash_algo
bundle: switch to use the_hash_algo
connected: switch GIT_SHA1_HEXSZ to the_hash_algo
show-index: switch hard-coded constants to the_hash_algo
blame: remove needless comparison with GIT_SHA1_HEXSZ
...
Another step in eliminating the requirement of hashmap_entry
being the first member of a struct.
Signed-off-by: Eric Wong <e@80x24.org>
Reviewed-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Update callers to use hashmap_get_entry, hashmap_get_entry_from_hash
or container_of as appropriate.
This is another step towards eliminating the requirement of
hashmap_entry being the first field in a struct.
Signed-off-by: Eric Wong <e@80x24.org>
Reviewed-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This is less error-prone than "const void *" as the compiler
now detects invalid types being passed.
Signed-off-by: Eric Wong <e@80x24.org>
Reviewed-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This is less error-prone than "void *" as the compiler now
detects invalid types being passed.
Signed-off-by: Eric Wong <e@80x24.org>
Reviewed-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
C compilers do type checking to make life easier for us. So
rely on that and update all hashmap_entry_init callers to take
"struct hashmap_entry *" to avoid future bugs while improving
safety and readability.
Signed-off-by: Eric Wong <e@80x24.org>
Reviewed-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This hashmap_entry_init function is intended to take a
hashmap_entry struct pointer, not a hashmap struct pointer.
This was not noticed because hashmap_entry_init takes a "void *"
arg instead of "struct hashmap_entry *", and the hashmap struct
is larger and can be cast into a hashmap_entry struct without
data corruption.
This has the beneficial side effect of reducing the size of
a delta_base_cache_entry from 104 bytes to 72 bytes on 64-bit
systems.
Signed-off-by: Eric Wong <e@80x24.org>
Reviewed-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Found with "git grep '^#include ' '*.c' | sort | uniq -d".
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Jeff King
Junio C Hamano
Taylor Blau
René Scharfe
Martin Ågren
Matheus Tavares
brian m. carlson
Eric Wong
Colin Stolley