When seeking a reftable record in a block we need to position the
iterator _before_ the sought-after record so that the next call to
`block_iter_next()` would yield that record. To achieve this, the loop
that performs the linear seeks to restore the previous position once it
has found the record.
This is done by advancing two `block_iter`s: one to check whether the
next record is our sought-after record, and one that we update after
every iteration. This of course involves quite a lot of copying and also
leads to needless memory allocations.
Refactor the code to get rid of the `next` iterator and the copying this
involves. Instead, we can restore the previous offset such that the call
to `next` will return the correct record.
Next to being simpler conceptually this also leads to a nice speedup.
The following benchmark parser 10k refs out of 100k existing refs via
`git-rev-list --no-walk`:
Benchmark 1: rev-list: print many refs (HEAD~)
Time (mean ± σ): 170.2 ms ± 1.7 ms [User: 86.1 ms, System: 83.6 ms]
Range (min … max): 166.4 ms … 180.3 ms 500 runs
Benchmark 2: rev-list: print many refs (HEAD~)
Time (mean ± σ): 161.6 ms ± 1.6 ms [User: 78.1 ms, System: 83.0 ms]
Range (min … max): 158.4 ms … 172.3 ms 500 runs
Summary
rev-list: print many refs (HEAD) ran
1.05 ± 0.01 times faster than rev-list: print many refs (HEAD~)
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When calling `inflateInit()` and `inflate()`, the zlib library will
allocate several data structures for the underlying `zstream` to keep
track of various information. Thus, when inflating repeatedly, it is
possible to optimize memory allocation patterns by reusing the `zstream`
and then calling `inflateReset()` on it to prepare it for the next chunk
of data to inflate.
This is exactly what the reftable code is doing: when iterating through
reflogs we need to potentially inflate many log blocks, but we discard
the `zstream` every single time. Instead, as we reuse the `block_reader`
for each of the blocks anyway, we can initialize the `zstream` once and
then reuse it for subsequent inflations.
Refactor the code to do so, which leads to a significant reduction in
the number of allocations. The following measurements were done when
iterating through 1 million reflog entries. Before:
HEAP SUMMARY:
in use at exit: 13,473 bytes in 122 blocks
total heap usage: 23,028 allocs, 22,906 frees, 162,813,552 bytes allocated
After:
HEAP SUMMARY:
in use at exit: 13,473 bytes in 122 blocks
total heap usage: 302 allocs, 180 frees, 88,352 bytes allocated
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The reftable format stores log blocks in a compressed format. Thus,
whenever we want to read such a block we first need to decompress it.
This is done by calling the convenience function `uncompress2()` of the
zlib library, which is a simple wrapper that manages the lifecycle of
the `zstream` structure for us.
While nice for one-off inflation of data, when iterating through reflogs
we will likely end up inflating many such log blocks. This requires us
to reallocate the state of the `zstream` every single time, which adds
up over time. It would thus be great to reuse the `zstream` instead of
discarding it after every inflation.
Open-code the call to `uncompress2()` such that we can start reusing the
`zstream` in the subsequent commit. Note that our open-coded variant is
different from `uncompress2()` in two ways:
- We do not loop around `inflate()` until we have processed all input.
As our input is limited by the maximum block size, which is 16MB, we
should not hit limits of `inflate()`.
- We use `Z_FINISH` instead of `Z_NO_FLUSH`. Quoting the `inflate()`
documentation: "inflate() should normally be called until it returns
Z_STREAM_END or an error. However if all decompression is to be
performed in a single step (a single call of inflate), the parameter
flush should be set to Z_FINISH."
Furthermore, "Z_FINISH also informs inflate to not maintain a
sliding window if the stream completes, which reduces inflate's
memory footprint."
Other than that this commit is expected to be functionally equivalent
and does not yet reuse the `zstream`.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The reftable backend stores reflog entries in a compressed format and
thus needs to uncompress blocks before one can read records from it.
For each reflog block we thus have to allocate an array that we can
decompress the block contents into. This block is being discarded
whenever the table iterator moves to the next block. Consequently, we
reallocate a new array on every block, which is quite wasteful.
Refactor the code to reuse the uncompressed block data when moving the
block reader to a new block. This significantly reduces the number of
allocations when iterating through many compressed blocks. The following
measurements are done with `git reflog list` when listing 100k reflogs.
Before:
HEAP SUMMARY:
in use at exit: 13,473 bytes in 122 blocks
total heap usage: 45,755 allocs, 45,633 frees, 254,779,456 bytes allocated
After:
HEAP SUMMARY:
in use at exit: 13,473 bytes in 122 blocks
total heap usage: 23,028 allocs, 22,906 frees, 162,813,547 bytes allocated
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The table iterator has to iterate towards the next block once it has
yielded all records of the current block. This is done by creating a new
table iterator, initializing it to the next block, releasing the old
iterator and then copying over the data.
Refactor the code to instead advance the table iterator in place. This
is simpler and unlocks some optimizations in subsequent patches. Also,
it allows us to avoid some allocations.
The following measurements show a single matching ref out of 1 million
refs. Before this change:
HEAP SUMMARY:
in use at exit: 13,603 bytes in 125 blocks
total heap usage: 7,235 allocs, 7,110 frees, 301,481 bytes allocated
After:
HEAP SUMMARY:
in use at exit: 13,603 bytes in 125 blocks
total heap usage: 315 allocs, 190 frees, 107,027 bytes allocated
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The table iterator allows the caller to iterate through all records in a
reftable table. To do so it iterates through all blocks of the desired
type one by one, where for each block it creates a new block iterator
and yields all its entries.
One of the things that is somewhat confusing in this context is who owns
the block reader that is being used to read the blocks and pass them to
the block iterator. Intuitively, as the table iterator is responsible
for iterating through the blocks, one would assume that this iterator is
also responsible for managing the lifecycle of the reader. And while it
somewhat is, the block reader is ultimately stored inside of the block
iterator.
Refactor the code such that the block reader is instead fully managed by
the table iterator. Instead of passing the reader to the block iterator,
we now only end up passing the block data to it. Despite clearing up the
lifecycle of the reader, it will also allow for better reuse of the
reader in subsequent patches.
The following benchmark prints a single matching ref out of 1 million
refs. Before:
HEAP SUMMARY:
in use at exit: 13,603 bytes in 125 blocks
total heap usage: 6,607 allocs, 6,482 frees, 509,635 bytes allocated
After:
HEAP SUMMARY:
in use at exit: 13,603 bytes in 125 blocks
total heap usage: 7,235 allocs, 7,110 frees, 301,481 bytes allocated
Note that while there are more allocation and free calls now, the
overall number of bytes allocated is significantly lower. The number of
allocations will be reduced significantly by the next patch though.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Introduce a new function `block_reader_release()` that releases
resources acquired by the block reader. This function will be extended
in a subsequent commit.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Function definitions and declaration of `struct block_reader` and
`struct block_iter` are somewhat mixed up, making it hard to see which
functions belong together. Rearrange them.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The function `block_iter_seek()` is merely a simple wrapper around
`block_reader_seek()`. Merge those two functions into a new function
`block_iter_seek_key()` that more clearly says what it is actually
doing.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The function `block_reader_start()` does not really apply to the block
reader, but to the block iterator. It's name is thus somewhat confusing.
Rename it to `block_iter_seek_start()` to clarify.
We will rename `block_reader_seek()` in similar spirit in the next
commit.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When searching over restart points in a block we decode the key of each
of the records, which results in a memory allocation. This is quite
pointless though given that records it restart points will never use
prefix compression and thus store their keys verbatim in the block.
Refactor the code so that we can avoid decoding the keys, which saves us
some allocations.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When doing the binary search over restart points in a block we need to
decode the record keys. This decoding step can result in an error when
the block is corrupted, which we indicate to the caller of the binary
search by setting `args.error = 1`. But the only caller that exists
mishandles this because it in fact performs the error check before
calling `binsearch()`.
Fix this bug by checking for errors at the right point in time.
Furthermore, refactor `binsearch()` so that it aborts the search in case
the callback function returns a negative value so that we don't
needlessly continue to search the block.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When seeking a record in our block reader we perform a binary search
over the block's restart points so that we don't have to do a linear
scan over the whole block. The logic to do so is quite intricate though,
which makes it hard to understand.
Improve documentation and rename some of the functions and variables so
that the code becomes easier to understand overall. This refactoring
should not result in any change in behaviour.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The `binsearch()` function can be used to find the first element for
which a callback functions returns a truish value. But while the array
size is of type `size_t`, the function in fact returns an `int` that is
supposed to index into that array.
Fix the function signature to return a `size_t`. This conversion does
not change any semantics given that the function would only ever return
a value in the range `[0, sz]` anyway.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The code to iterate over reflogs in the reftable has been optimized
to reduce memory allocation and deallocation.
Reviewed-by: Josh Steadmon <steadmon@google.com>
cf. <Ze9eX-aaWoVaqsPP@google.com>
* ps/reftable-reflog-iteration-perf:
refs/reftable: track last log record name via strbuf
reftable/record: use scratch buffer when decoding records
reftable/record: reuse message when decoding log records
reftable/record: reuse refnames when decoding log records
reftable/record: avoid copying author info
reftable/record: convert old and new object IDs to arrays
refs/reftable: reload correct stack when creating reflog iter
The reftable code has its own custom binary search function whose
comparison callback has an unusual interface, which caused the
binary search to degenerate into a linear search, which has been
corrected.
* ps/reftable-block-search-fix:
reftable/block: fix binary search over restart counter
reftable/record: fix memory leak when decoding object records
Records store their keys prefix-compressed. As many records will share a
common prefix (e.g. "refs/heads/"), this can end up saving quite a bit
of disk space. The downside of this is that it is not possible to just
seek into the middle of a block and consume the corresponding record
because it may depend on prefixes read from preceding records.
To help with this usecase, the reftable format writes every n'th record
without using prefix compression, which is called a "restart". The list
of restarts is stored at the end of each block so that a reader can
figure out entry points at which to read a full record without having to
read all preceding records.
This allows us to do a binary search over the records in a block when
searching for a particular key by iterating through the restarts until
we have found the section in which our record must be located. From
thereon we perform a linear search to locate the desired record.
This mechanism is broken though. In `block_reader_seek()` we call
`binsearch()` over the count of restarts in the current block. The
function we pass to compare records with each other computes the key at
the current index and then compares it to our search key by calling
`strbuf_cmp()`, returning its result directly. But `binsearch()` expects
us to return a truish value that indicates whether the current index is
smaller than the searched-for key. And unless our key exactly matches
the value at the restart counter we always end up returning a truish
value.
The consequence is that `binsearch()` essentially always returns 0,
indicacting to us that we must start searching right at the beginning of
the block. This works by chance because we now always do a linear scan
from the start of the block, and thus we would still end up finding the
desired record. But needless to say, this makes the optimization quite
useless.
Fix this bug by returning whether the current key is smaller than the
searched key. As the current behaviour was correct it is not possible to
write a test. Furthermore it is also not really possible to demonstrate
in a benchmark that this fix speeds up seeking records.
This may cause the reader to question whether this binary search makes
sense in the first place if it doesn't even help with performance. But
it would end up helping if we were to read a reftable with a much larger
block size. Blocks can be up to 16MB in size, in which case it will
become much more important to avoid the linear scan. We are not yet
ready to read or write such larger blocks though, so we have to live
without a benchmark demonstrating this.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When decoding log records we need a temporary buffer to decode the
reflog entry's name, mail address and message. As this buffer is local
to the function we thus have to reallocate it for every single log
record which we're about to decode, which is inefficient.
Refactor the code such that callers need to pass in a scratch buffer,
which allows us to reuse it for multiple decodes. This reduces the
number of allocations when iterating through reflogs. Before:
HEAP SUMMARY:
in use at exit: 13,473 bytes in 122 blocks
total heap usage: 2,068,487 allocs, 2,068,365 frees, 305,122,946 bytes allocated
After:
HEAP SUMMARY:
in use at exit: 13,473 bytes in 122 blocks
total heap usage: 1,068,485 allocs, 1,068,363 frees, 281,122,886 bytes allocated
Note that this commit also drop some redundant calls to `strbuf_reset()`
right before calling `decode_string()`. The latter already knows to
reset the buffer, so there is no need for these.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When reading a record from a block, we need to decode the record's key.
As reftable keys are prefix-compressed, meaning they reuse a prefix from
the preceding record's key, this is a bit more involved than just having
to copy the relevant bytes: we need to figure out the prefix and suffix
lengths, copy the prefix from the preceding record and finally copy the
suffix from the current record.
This is done by passing three buffers to `reftable_decode_key()`: one
buffer that holds the result, one buffer that holds the last key, and
one buffer that points to the current record. The final key is then
assembled by calling `strbuf_add()` twice to copy over the prefix and
suffix.
Performing two memory copies is inefficient though. And we can indeed do
better by decoding keys in place. Instead of providing two buffers, the
caller may only call a single buffer that is already pre-populated with
the last key. Like this, we only have to call `strbuf_setlen()` to trim
the record to its prefix and then `strbuf_add()` to add the suffix.
This refactoring leads to a noticeable performance bump when iterating
over 1 million refs:
Benchmark 1: show-ref: single matching ref (revision = HEAD~)
Time (mean ± σ): 112.2 ms ± 3.9 ms [User: 109.3 ms, System: 2.8 ms]
Range (min … max): 109.2 ms … 149.6 ms 1000 runs
Benchmark 2: show-ref: single matching ref (revision = HEAD)
Time (mean ± σ): 106.0 ms ± 3.5 ms [User: 103.2 ms, System: 2.7 ms]
Range (min … max): 103.2 ms … 133.7 ms 1000 runs
Summary
show-ref: single matching ref (revision = HEAD) ran
1.06 ± 0.05 times faster than show-ref: single matching ref (revision = HEAD~)
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The code to iterate over refs with the reftable backend has seen
some optimization.
* ps/reftable-iteration-perf:
reftable/reader: add comments to `table_iter_next()`
reftable/record: don't try to reallocate ref record name
reftable/block: swap buffers instead of copying
reftable/pq: allocation-less comparison of entry keys
reftable/merged: skip comparison for records of the same subiter
reftable/merged: allocation-less dropping of shadowed records
reftable/record: introduce function to compare records by key
When iterating towards the next record in a reftable block we need to
keep track of the key that the last record had. This is required because
reftable records use prefix compression, where subsequent records may
reuse parts of their preceding record's key.
This key is stored in the `block_iter::last_key`, which we update after
every call to `block_iter_next()`: we simply reset the buffer and then
add the current key to it.
This is a bit inefficient though because it requires us to copy over the
key on every iteration, which adds up when iterating over many records.
Instead, we can make use of the fact that the `block_iter::key` buffer
is basically only a scratch buffer. So instead of copying over contents,
we can just swap both buffers.
The following benchmark prints a single ref matching a specific pattern
out of 1 million refs via git-show-ref(1):
Benchmark 1: show-ref: single matching ref (revision = HEAD~)
Time (mean ± σ): 155.7 ms ± 5.0 ms [User: 152.1 ms, System: 3.4 ms]
Range (min … max): 150.8 ms … 185.7 ms 1000 runs
Benchmark 2: show-ref: single matching ref (revision = HEAD)
Time (mean ± σ): 150.8 ms ± 4.2 ms [User: 147.1 ms, System: 3.5 ms]
Range (min … max): 145.1 ms … 180.7 ms 1000 runs
Summary
show-ref: single matching ref (revision = HEAD) ran
1.03 ± 0.04 times faster than show-ref: single matching ref (revision = HEAD~)
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
According to our usual coding style, the `reftable_new_record()`
function would indicate that it is allocating a new record. This is not
the case though as the function merely initializes records without
allocating any memory.
Replace `reftable_new_record()` with a new `reftable_record_init()`
function that takes a record pointer as input and initializes it
accordingly.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Similar to the preceding commit, let's carry over macros to allocate
arrays with `REFTABLE_ALLOC_ARRAY()` and `REFTABLE_CALLOC_ARRAY()`. This
requires us to change the signature of `reftable_calloc()`, which only
takes a single argument right now and thus puts the burden on the caller
to calculate the final array's size. This is a net improvement though as
it means that we can now provide proper overflow checks when multiplying
the array size with the member size.
Convert callsites of `reftable_calloc()` to the new signature and start
using the new macros where possible.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Throughout the reftable library we have many cases where we need to grow
arrays. In order to avoid too many reallocations, we roughly double the
capacity of the array on each iteration. The resulting code pattern is
duplicated across many sites.
We have similar patterns in our main codebase, which is why we have
eventually introduced an `ALLOC_GROW()` macro to abstract it away and
avoid some code duplication. We cannot easily reuse this macro here
though because `ALLOC_GROW()` uses `REALLOC_ARRAY()`, which in turn will
call realloc(3P) to grow the array. The reftable code is structured as a
library though (even if the boundaries are fuzzy), and one property this
brings with it is that it is possible to plug in your own allocators. So
instead of using realloc(3P), we need to use `reftable_realloc()` that
knows to use the user-provided implementation.
So let's introduce two new macros `REFTABLE_REALLOC_ARRAY()` and
`REFTABLE_ALLOC_GROW()` that mirror what we do in our main codebase,
with two modifications:
- They use `reftable_realloc()`, as explained above.
- They use a different growth factor of `2 * cap + 1` instead of `(cap
+ 16) * 3 / 2`.
The second change is because we know a bit more about the allocation
patterns in the reftable library. In most cases, we end up only having a
handful of items in the array and don't end up growing them. The initial
capacity that our normal growth factor uses (which is 24) would thus end
up over-allocating in a lot of code paths. This effect is measurable:
- Before change:
HEAP SUMMARY:
in use at exit: 671,983 bytes in 152 blocks
total heap usage: 3,843,446 allocs, 3,843,294 frees, 223,761,402 bytes allocated
- After change with a growth factor of `(2 * alloc + 1)`:
HEAP SUMMARY:
in use at exit: 671,983 bytes in 152 blocks
total heap usage: 3,843,446 allocs, 3,843,294 frees, 223,761,410 bytes allocated
- After change with a growth factor of `(alloc + 16)* 2 / 3`:
HEAP SUMMARY:
in use at exit: 671,983 bytes in 152 blocks
total heap usage: 3,833,673 allocs, 3,833,521 frees, 4,728,251,742 bytes allocated
While the total heap usage is roughly the same, we do end up allocating
significantly more bytes with our usual growth factor (in fact, roughly
21 times as many).
Convert the reftable library to use these new macros.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When iterating over entries in the block iterator we compute the key of
each of the entries and write it into a buffer. We do not reuse the
buffer though and thus re-allocate it on every iteration, which is
wasteful.
Refactor the code to reuse the buffer.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There are a bunch of locations where we initialize members of `struct
block_iter`, which makes it harder than necessary to expand this struct
to have additional members. Unify the logic via a new `BLOCK_ITER_INIT`
macro that initializes all members.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The public interface (reftable_writer) already ensures that keys are
written in strictly increasing order, and an empty key by definition
fails this check.
However, by also enforcing this at the block layer, it is easier to
verify that records (which are written into blocks) never have to
consider the possibility of empty keys.
Signed-off-by: Han-Wen Nienhuys <hanwen@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This reduces the amount of glue code, because we don't need a void
pointer or vtable within the structure.
The only snag is that reftable_index_record contain a strbuf, so it
cannot be zero-initialized. To address this, use reftable_new_record()
to return fresh instance, given a record type. Since
reftable_new_record() doesn't cause heap allocation anymore, it should
be balanced with reftable_record_release() rather than
reftable_record_destroy().
Thanks to Peff for the suggestion.
Helped-by: Jeff King <peff@peff.net>
Signed-off-by: Han-Wen Nienhuys <hanwen@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add test coverage for corrupt zlib data. Fix memory leaks demonstrated by
unittest.
This problem was discovered by a Coverity scan.
Signed-off-by: Han-Wen Nienhuys <hanwen@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The reftable format is structured as a sequence of block. Within a block,
records are prefix compressed, with an index of offsets for fully expand keys to
enable binary search within blocks.
This commit provides the logic to read and write these blocks.
Helped-by: Carlo Marcelo Arenas Belón <carenas@gmail.com>
Signed-off-by: Han-Wen Nienhuys <hanwen@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
