When an on-disk reverse index exists, there is no need to generate one
in memory. In fact, doing so can be slow, and require large amounts of
the heap.
Let's make sure that we treat the on-disk reverse index with precedence
(i.e., that when it exists, we don't bother trying to generate an
equivalent one in memory) by teaching Git how to conditionally die()
when generating a reverse index in memory.
Then, add a test to ensure that when (a) an on-disk reverse index
exists, and (b) when setting GIT_TEST_REV_INDEX_DIE_IN_MEMORY, that we
do not die, implying that we read from the on-disk one.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add a new option that unconditionally enables the pack.writeReverseIndex
setting in order to run the whole test suite in a mode that generates
on-disk reverse indexes. Additionally, enable this mode in the second
run of tests under linux-gcc in 'ci/run-build-and-tests.sh'.
Once on-disk reverse indexes are proven out over several releases, we
can change the default value of that configuration to 'true', and drop
this patch.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
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>
Now that all spots outside of pack-revindex.c that reference 'struct
revindex_entry' directly have been removed, it is safe to hide the
implementation by moving it from pack-revindex.h to pack-revindex.c.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Now that all 'find_revindex_position()' callers have been removed (and
converted to the more descriptive 'offset_to_pack_pos()'), it is almost
safe to get rid of 'find_revindex_position()' entirely. Almost, except
for the fact that 'offset_to_pack_pos()' calls
'find_revindex_position()'.
Inline 'find_revindex_position()' into 'offset_to_pack_pos()', and
then remove 'find_revindex_position()' entirely.
This is a straightforward refactoring with one minor snag.
'offset_to_pack_pos()' used to load the index before calling
'find_revindex_position()'. That means that by the time
'find_revindex_position()' starts executing, 'p->num_objects' can be
safely read. After inlining, be careful to not read 'p->num_objects'
until _after_ 'load_pack_revindex()' (which loads the index as a
side-effect) has been called.
Another small fix that is included is converting the upper- and
lower-bounds to be unsigned's instead of ints. This dates back to
92e5c77c37 (revindex: export new APIs, 2013-10-24)--ironically, the last
time we introduced new APIs here--but this unifies the types.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Now that no callers of 'find_pack_revindex()' remain, remove the
function's declaration and implementation entirely.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In the next several patches, we will prepare for loading a reverse index
either in memory (mapping the inverse of the .idx's contents in-core),
or directly from a yet-to-be-introduced on-disk format. To prepare for
that, we'll introduce an API that avoids the caller explicitly indexing
the revindex pointer in the packed_git structure.
There are four ways to interact with the reverse index. Accordingly,
four functions will be exported from 'pack-revindex.h' by the time that
the existing API is removed. A caller may:
1. Load the pack's reverse index. This involves opening up the index,
generating an array, and then sorting it. Since opening the index
can fail, this function ('load_pack_revindex()') returns an int.
Accordingly, it takes only a single argument: the 'struct
packed_git' the caller wants to build a reverse index for.
This function is well-suited for both the current and new API.
Callers will have to continue to open the reverse index explicitly,
but this function will eventually learn how to detect and load a
reverse index from the on-disk format, if one exists. Otherwise, it
will fallback to generating one in memory from scratch.
2. Convert a pack position into an offset. This operation is now
called `pack_pos_to_offset()`. It takes a pack and a position, and
returns the corresponding off_t.
Any error simply calls BUG(), since the callers are not well-suited
to handle a failure and keep going.
3. Convert a pack position into an index position. Same as above; this
takes a pack and a position, and returns a uint32_t. This operation
is known as `pack_pos_to_index()`. The same thinking about error
conditions applies here as well.
4. Find the pack position for a given offset. This operation is now
known as `offset_to_pack_pos()`. It takes a pack, an offset, and a
pointer to a uint32_t where the position is written, if an object
exists at that offset. Otherwise, -1 is returned to indicate
failure.
Unlike some of the callers that used to access '->offset' and '->nr'
directly, the error checking around this call is somewhat more
robust. This is important since callers should always pass an offset
which points at the boundary of two objects. The API, unlike direct
access, enforces that that is the case.
This will become important in a subsequent patch where a caller
which does not but could check the return value treats the signed
`-1` from `find_revindex_position()` as an index into the 'revindex'
array.
Two design warts are carried over into the new API:
- Asking for the index position of an out-of-bounds object will result
in a BUG() (since no such object exists), but asking for the offset
of the non-existent object at the end of the pack returns the total
size of the pack.
This makes it convenient for callers who always want to take the
difference of two adjacent object's offsets (to compute the on-disk
size) but don't want to worry about boundaries at the end of the
pack.
- offset_to_pack_pos() lazily loads the reverse index, but
pack_pos_to_index() doesn't (callers of the former are well-suited
to handle errors, but callers of the latter are not).
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We can't create a pack revindex if we haven't actually looked at the
index. Normally we would never get as far as creating a revindex without
having already been looking in the pack, so this code never bothered to
double-check that pack->index_data had been loaded.
But with the new multi-pack-index feature, many code paths might not
load the individual pack .idx at all (they'd find objects via the midx
and then open the .pack, but not its index).
This can't yet be triggered in practice, because a bug in the midx code
means we accidentally open up the individual .idx files anyway. But in
preparation for fixing that, let's have the revindex code check that
everything it needs has been loaded.
In most cases this will just be a quick noop. But note that this does
introduce a possibility of error (if we have to open the index and it's
corrupt), so load_pack_revindex() now returns a result code, and callers
need to handle the error.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
A pack_revindex struct has two elements: the revindex
entries themselves, and a pointer to the packed_git. We need
both to do lookups, because only the latter knows things
like the number of objects in the pack.
Now that packed_git contains the pack_revindex struct it's
just as easy to pass around the packed_git itself, and we do
not need the extra back-pointer.
We can instead just store the entries directly in the pack.
All functions which took a pack_revindex now just take a
packed_git. We still lazy-load in find_pack_revindex, so
most callers are unaffected.
The exception is the bitmap code, which computes the
revindex and caches the pointer when we load the bitmaps. We
can continue to load, drop the extra cache pointer, and just
access bitmap_git.pack.revindex directly.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When an object lookup fails, we re-read the objects/pack
directory to pick up any new packfiles that may have been
created since our last read. We also discard any pack
revindex structs we've allocated.
The discarding is a problem for the pack-bitmap code, which keeps
a pointer to the revindex for the bitmapped pack. After the
discard, the pointer is invalid, and we may read free()d
memory.
Other revindex users do not keep a bare pointer to the
revindex; instead, they always access it through
revindex_for_pack(), which lazily builds the revindex. So
one solution is to teach the pack-bitmap code a similar
trick. It would be slightly less efficient, but probably not
all that noticeable.
However, it turns out this discarding is not actually
necessary. When we call reprepare_packed_git, we do not
throw away our old pack list. We keep the existing entries,
and only add in new ones. So there is no safety problem; we
will still have the pack struct that matches each revindex.
The packfile itself may go away, of course, but we are
already prepared to handle that, and it may happen outside
of reprepare_packed_git anyway.
Throwing away the revindex may save some RAM if the pack
never gets reused (about 12 bytes per object). But it also
wastes some CPU time (to regenerate the index) if the pack
does get reused. It's hard to say which is more valuable,
but in either case, it happens very rarely (only when we
race with a simultaneous repack). Just leaving the revindex
in place is simple and safe both for current and future
code.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Allow users to efficiently lookup consecutive entries that are expected
to be found on the same revindex by exporting `find_revindex_position`:
this function takes a pointer to revindex itself, instead of looking up
the proper revindex for a given packfile on each call.
Signed-off-by: Vicent Marti <tanoku@gmail.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This is needed to fix verify-pack -v with multiple pack arguments.
Also, in theory, revindex data (if any) must be discarded whenever
reprepare_packed_git() is called. In practice this is hard to trigger
though.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This makes life much easier for next patch, as well as being more efficient
when the revindex is actually not used.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
No functional change. This is needed to fix verify-pack in a later patch.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
