As mentioned in the thread starting at [1], trace2 counters should be
used to count events instead of ad-hoc static variables.
Convert the two fsync static variables to trace2 counters, reducing the
coupling between wrapper.c and the trace2 subsystem. Adjust t/t5351 to
match the trace2 counter output format.
The counters are not per-thread because the ones being replaced also
were not.
[1] https://lore.kernel.org/git/20230627195251.1973421-2-calvinwan@google.com/
Signed-off-by: Beat Bolli <dev+git@drbeat.li>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Extend SANITIZE=leak checking and declare more tests "currently leak-free".
* ab/leak-check:
CI: use "GIT_TEST_SANITIZE_LEAK_LOG=true" in linux-leaks
upload-pack: fix a memory leak in create_pack_file()
leak tests: mark passing SANITIZE=leak tests as leak-free
leak tests: don't skip some tests under SANITIZE=leak
test-lib: have the "check" mode for SANITIZE=leak consider leak logs
test-lib: add a GIT_TEST_PASSING_SANITIZE_LEAK=check mode
test-lib: simplify by removing test_external
tests: move copy/pasted PERL + Test::More checks to a lib-perl.sh
t/Makefile: don't remove test-results in "clean-except-prove-cache"
test-lib: add a SANITIZE=leak logging mode
t/README: reword the "GIT_TEST_PASSING_SANITIZE_LEAK" description
test-lib: add a --invert-exit-code switch
test-lib: fix GIT_EXIT_OK logic errors, use BAIL_OUT
test-lib: don't set GIT_EXIT_OK before calling test_atexit_handler
test-lib: use $1, not $@ in test_known_broken_{ok,failure}_
The `test_cmp` function is meant to provide nicer output than `cmp` when
expected and actual output of Git commands disagree. The implicit
assumption is that the output is line-based and human readable.
However, aaf81223f4 (unpack-objects: use stream_loose_object() to
unpack large objects, 2022-06-11) introduced a call that compares the
contents of pack files, which are distinctly not line-based nor human
readable.
This causes problems because on Windows, we hand off to the Bash
function `mingw_test_cmp` that compares the lines while ignoring line
ending differences. And this Bash function spends an insane amount of
cycles trying to read in that binary pack file, so that it is almost
indistinguishable from an infinite loop.
For example, t5351 took 1486 seconds in the CI run at
https://github.com/git/git/runs/7398490747?check_suite_focus=true#step:5:171,
to complete. And yes, that is almost half an hour.
Since Git's tests already use `cmp` consistently when comparing pack
files, let's change this instance to use `cmp` instead of `test_cmp`,
too, and fix that performance problem.
Now t5351 takes all of 22 seconds.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
On FreeBSD, this mode is not supported. But since 3a251bac0d (trace2:
only include "fsync" events if we git_fsync(), 2022-07-18) t5351 will
fail if this mode is unsupported.
Let's address this in the minimal fashion, by detecting that that mode
is unsupported and expecting a different count of hardware flushes in
that case.
This fixes the CI/PR builds on FreeBSD again.
Note: A better way would be to test only what is relevant in t5351.6
"unpack big object in stream (core.fsyncmethod=batch)" again instead of
blindly comparing the output against some exact text. But that would
pretty much revert the idea of above-mentioned commit, and that commit
was _just_ accepted into Git's main branch so one must assume that it
was intentional.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Mark those remaining tests that pass when run under SANITIZE=leak with
TEST_PASSES_SANITIZE_LEAK=true, these were either omitted in
f346fcb62a (Merge branch 'ab/mark-leak-free-tests-even-more',
2021-12-15) and 5a4f8381b6 (Merge branch 'ab/mark-leak-free-tests',
2021-10-25), or have had their memory leaks fixed since then.
With this change there's now a a one-to-one mapping between those
tests that we have opted-in via "TEST_PASSES_SANITIZE_LEAK=true", and
those that pass with the new "check" mode:
GIT_TEST_PASSING_SANITIZE_LEAK=check \
GIT_TEST_SANITIZE_LEAK_LOG=true \
make test SANITIZE=leak
Note that the "GIT_TEST_SANITIZE_LEAK_LOG=true" is needed due to the
edge cases noted in a preceding commit, i.e. in some cases we'd pass
the test itself, but still have outstanding leaks due to ignored exit
codes.
The "GIT_TEST_SANITIZE_LEAK_LOG=true" corrects for that, we're only
marking those tests as passing that really don't have any leaks,
whether that was reflected in their exit code or not.
Note that the change here to "t9100-git-svn-basic.sh" is marking that
test as passing under SANITIZE=leak, we're removing a
"TEST_FAILS_SANITIZE_LEAK=true" line, not
"TEST_PASSES_SANITIZE_LEAK=true". See 7a98d9ab00 (revisions API: have
release_revisions() release "cmdline", 2022-04-13) for the
introduction of that t/lib-git-svn.sh-specific variable.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Fix the overly verbose trace2 logging added in 9a4987677d (trace2:
add stats for fsync operations, 2022-03-30) (first released with
v2.36.0).
Since that change every single "git" command invocation has included
these "data" events, even though we'll only make use of these with
core.fsyncMethod=batch, and even then only have non-zero values if
we're writing object data to disk. See c0f4752ed2 (core.fsyncmethod:
batched disk flushes for loose-objects, 2022-04-04) for that feature.
As we're needing to indent the trace2_data_intmax() lines let's
introduce helper variables to ensure that our resulting lines (which
were already too) don't exceed the recommendations of the
CodingGuidelines. Doing that requires either wrapping them twice, or
introducing short throwaway variable names, let's do the latter.
The result was that e.g. "git version" would previously emit a total
of 6 trace2 events with the GIT_TRACE2_EVENT target (version, start,
cmd_ancestry, cmd_name, exit, atexit), but afterwards would emit
8. We'd emit 2 "data" events before the "exit" event.
The reason we didn't catch this was that the trace2 unit tests added
in a15860dca3 (trace2: t/helper/test-trace2, t0210.sh, t0211.sh,
t0212.sh, 2019-02-22) would omit any "data" events that weren't the
ones it cared about. Before this change to the C code 6/7 of our
"t/t0212-trace2-event.sh" tests would fail if this change was applied
to "t/t0212/parse_events.perl".
Let's make the trace2 testing more strict, and further append any new
events types we don't know about in "t/t0212/parse_events.perl". Since
we only invoke the "test-tool trace2" there's no guarantee that we'll
catch other overly verbose events in the future, but we'll at least
notice if we start emitting new events that are issues every time we
log anything with trace2's JSON target.
We exclude the "data_json" event type, we'd otherwise would fail on
both "win test" and "win+VS test" CI due to the logging added in
353d3d77f4 (trace2: collect Windows-specific process information,
2019-02-22). It looks like that logging should really be using
trace2_cmd_ancestry() instead, which was introduced later in
2f732bf15e (tr2: log parent process name, 2021-07-21), but let's
leave it for now.
The fix-up to aaf81223f4 (unpack-objects: use stream_loose_object()
to unpack large objects, 2022-06-11) is needed because we're changing
the behavior of these events as discussed above. Since we'd always
emit a "hardware-flush" event the test added in aaf81223f4 wasn't
testing anything except that this trace2 data was unconditionally
logged. Even if "core.fsyncMethod" wasn't set to "batch" we'd pass the
test.
Now we'll check the expected number of "writeout" v.s. "flush" calls
under "core.fsyncMethod=batch", but note that this doesn't actually
test if we carried out the sync using that method, on a platform where
we'd have to fall back to fsync() each of those "writeout" would
really be a "flush" (i.e. a full fsync()).
But in this case what we're testing is that the logic in
"unpack-objects" behaves as expected, not the OS-specific question of
whether we actually were able to use the "bulk" method.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Make use of the stream_loose_object() function introduced in the
preceding commit to unpack large objects. Before this we'd need to
malloc() the size of the blob before unpacking it, which could cause
OOM with very large blobs.
We could use the new streaming interface to unpack all blobs, but
doing so would be much slower, as demonstrated e.g. with this
benchmark using git-hyperfine[0]:
rm -rf /tmp/scalar.git &&
git clone --bare https://github.com/Microsoft/scalar.git /tmp/scalar.git &&
mv /tmp/scalar.git/objects/pack/*.pack /tmp/scalar.git/my.pack &&
git hyperfine \
-r 2 --warmup 1 \
-L rev origin/master,HEAD -L v "10,512,1k,1m" \
-s 'make' \
-p 'git init --bare dest.git' \
-c 'rm -rf dest.git' \
'./git -C dest.git -c core.bigFileThreshold={v} unpack-objects </tmp/scalar.git/my.pack'
Here we'll perform worse with lower core.bigFileThreshold settings
with this change in terms of speed, but we're getting lower memory use
in return:
Summary
'./git -C dest.git -c core.bigFileThreshold=10 unpack-objects </tmp/scalar.git/my.pack' in 'origin/master' ran
1.01 ± 0.01 times faster than './git -C dest.git -c core.bigFileThreshold=1k unpack-objects </tmp/scalar.git/my.pack' in 'origin/master'
1.01 ± 0.01 times faster than './git -C dest.git -c core.bigFileThreshold=1m unpack-objects </tmp/scalar.git/my.pack' in 'origin/master'
1.01 ± 0.02 times faster than './git -C dest.git -c core.bigFileThreshold=1m unpack-objects </tmp/scalar.git/my.pack' in 'HEAD'
1.02 ± 0.00 times faster than './git -C dest.git -c core.bigFileThreshold=512 unpack-objects </tmp/scalar.git/my.pack' in 'origin/master'
1.09 ± 0.01 times faster than './git -C dest.git -c core.bigFileThreshold=1k unpack-objects </tmp/scalar.git/my.pack' in 'HEAD'
1.10 ± 0.00 times faster than './git -C dest.git -c core.bigFileThreshold=512 unpack-objects </tmp/scalar.git/my.pack' in 'HEAD'
1.11 ± 0.00 times faster than './git -C dest.git -c core.bigFileThreshold=10 unpack-objects </tmp/scalar.git/my.pack' in 'HEAD'
A better benchmark to demonstrate the benefits of that this one, which
creates an artificial repo with a 1, 25, 50, 75 and 100MB blob:
rm -rf /tmp/repo &&
git init /tmp/repo &&
(
cd /tmp/repo &&
for i in 1 25 50 75 100
do
dd if=/dev/urandom of=blob.$i count=$(($i*1024)) bs=1024
done &&
git add blob.* &&
git commit -mblobs &&
git gc &&
PACK=$(echo .git/objects/pack/pack-*.pack) &&
cp "$PACK" my.pack
) &&
git hyperfine \
--show-output \
-L rev origin/master,HEAD -L v "512,50m,100m" \
-s 'make' \
-p 'git init --bare dest.git' \
-c 'rm -rf dest.git' \
'/usr/bin/time -v ./git -C dest.git -c core.bigFileThreshold={v} unpack-objects </tmp/repo/my.pack 2>&1 | grep Maximum'
Using this test we'll always use >100MB of memory on
origin/master (around ~105MB), but max out at e.g. ~55MB if we set
core.bigFileThreshold=50m.
The relevant "Maximum resident set size" lines were manually added
below the relevant benchmark:
'/usr/bin/time -v ./git -C dest.git -c core.bigFileThreshold=50m unpack-objects </tmp/repo/my.pack 2>&1 | grep Maximum' in 'origin/master' ran
Maximum resident set size (kbytes): 107080
1.02 ± 0.78 times faster than '/usr/bin/time -v ./git -C dest.git -c core.bigFileThreshold=512 unpack-objects </tmp/repo/my.pack 2>&1 | grep Maximum' in 'origin/master'
Maximum resident set size (kbytes): 106968
1.09 ± 0.79 times faster than '/usr/bin/time -v ./git -C dest.git -c core.bigFileThreshold=100m unpack-objects </tmp/repo/my.pack 2>&1 | grep Maximum' in 'origin/master'
Maximum resident set size (kbytes): 107032
1.42 ± 1.07 times faster than '/usr/bin/time -v ./git -C dest.git -c core.bigFileThreshold=100m unpack-objects </tmp/repo/my.pack 2>&1 | grep Maximum' in 'HEAD'
Maximum resident set size (kbytes): 107072
1.83 ± 1.02 times faster than '/usr/bin/time -v ./git -C dest.git -c core.bigFileThreshold=50m unpack-objects </tmp/repo/my.pack 2>&1 | grep Maximum' in 'HEAD'
Maximum resident set size (kbytes): 55704
2.16 ± 1.19 times faster than '/usr/bin/time -v ./git -C dest.git -c core.bigFileThreshold=512 unpack-objects </tmp/repo/my.pack 2>&1 | grep Maximum' in 'HEAD'
Maximum resident set size (kbytes): 4564
This shows that if you have enough memory this new streaming method is
slower the lower you set the streaming threshold, but the benefit is
more bounded memory use.
An earlier version of this patch introduced a new
"core.bigFileStreamingThreshold" instead of re-using the existing
"core.bigFileThreshold" variable[1]. As noted in a detailed overview
of its users in [2] using it has several different meanings.
Still, we consider it good enough to simply re-use it. While it's
possible that someone might want to e.g. consider objects "small" for
the purposes of diffing but "big" for the purposes of writing them
such use-cases are probably too obscure to worry about. We can always
split up "core.bigFileThreshold" in the future if there's a need for
that.
0. https://github.com/avar/git-hyperfine/
1. https://lore.kernel.org/git/20211210103435.83656-1-chiyutianyi@gmail.com/
2. https://lore.kernel.org/git/20220120112114.47618-5-chiyutianyi@gmail.com/
Helped-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Helped-by: Derrick Stolee <stolee@gmail.com>
Helped-by: Jiang Xin <zhiyou.jx@alibaba-inc.com>
Signed-off-by: Han Xin <chiyutianyi@gmail.com>
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
As the name implies, "get_data(size)" will allocate and return a given
amount of memory. Allocating memory for a large blob object may cause the
system to run out of memory. Before preparing to replace calling of
"get_data()" to unpack large blob objects in latter commits, refactor
"get_data()" to reduce memory footprint for dry_run mode.
Because in dry_run mode, "get_data()" is only used to check the
integrity of data, and the returned buffer is not used at all, we can
allocate a smaller buffer and use it as zstream output. Make the function
return NULL in the dry-run mode, as no callers use the returned buffer.
The "find [...]objects/?? -type f | wc -l" test idiom being used here
is adapted from the same "find" use added to another test in
d9545c7f46 (fast-import: implement unpack limit, 2016-04-25).
Suggested-by: Jiang Xin <zhiyou.jx@alibaba-inc.com>
Signed-off-by: Han Xin <chiyutianyi@gmail.com>
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
