Hi,

While analyzing push performance on gitlab.com, I've been at times
wondering what git-receive-pack(1) is doing for so long. For some repos
which have loads of references (~880k), even tiny pushes of less than 10
objects took dozens of seconds to get accepted.

One of the issues I've found is the object connectivity check, which may
run for a significant amount of time. The root cause here is that we're
computing connectivity via `git rev-list --not --all`: if we've got many
refs in the repository, computing `--not --all` is hugely expensive.

This commit series thus implements an alternative way of computing
reachability, which reuses information from the object quarantine
environment. Instead of doing a refwalk, we just look iterate over all
packed and loose quarantined objects any for each of them, we determine
whether their immediate references are all satisfied.

This reimplementation is paying out quite well for repos which have many
refs. The following benchmarks for git-receive-pack(1) (added in patch
2/8) have been performed in linux-stable.git:

Test                                     v2.32.0-rc0             HEAD
--------------------------------------------------------------------------------------------
5400.3: receive-pack clone create        1.27(1.11+0.16)         0.02(0.01+0.01) -98.4%
5400.5: receive-pack clone update        1.27(1.13+0.13)         0.02(0.02+0.00) -98.4%
5400.7: receive-pack clone reset         0.13(0.11+0.02)         0.02(0.01+0.01) -84.6%
5400.9: receive-pack clone delete        0.02(0.01+0.01)         0.03(0.02+0.01) +50.0%
5400.11: receive-pack extrarefs create   33.01(18.80+14.43)      9.00(4.30+4.65) -72.7%
5400.13: receive-pack extrarefs update   33.13(18.85+14.50)      9.01(4.28+4.67) -72.8%
5400.15: receive-pack extrarefs reset    32.90(18.82+14.32)      9.04(4.26+4.77) -72.5%
5400.17: receive-pack extrarefs delete   9.13(4.35+4.77)         8.94(4.29+4.64) -2.1%
5400.19: receive-pack empty create       223.35(640.63+127.74)   227.55(651.75+130.94) +1.9%

These rather clearly show that the previous rev-walk has been a major
bottleneck in the implementation.

Patrick

Patrick Steinhardt (8):
  perf: fix when running with TEST_OUTPUT_DIRECTORY
  p5400: add perf tests for git-receive-pack(1)
  tmp-objdir: expose function to retrieve path
  packfile: have `for_each_file_in_pack_dir()` return error codes
  object-file: allow reading loose objects without reading their
    contents
  connected: implement connectivity check via temporary object dirs
  receive-pack: skip connectivity checks on delete-only commands
  receive-pack: check connectivity via quarantined objects

 builtin/receive-pack.c       |  57 +++++++----
 connected.c                  | 192 +++++++++++++++++++++++++++++++++++
 connected.h                  |  19 ++++
 midx.c                       |  22 ++--
 object-file.c                |   9 +-
 packfile.c                   |  26 +++--
 packfile.h                   |  10 +-
 t/perf/aggregate.perl        |   8 +-
 t/perf/p5400-receive-pack.sh |  74 ++++++++++++++
 t/perf/perf-lib.sh           |   4 +-
 t/perf/run                   |  25 +++--
 tmp-objdir.c                 |   7 ++
 tmp-objdir.h                 |   5 +
 13 files changed, 401 insertions(+), 57 deletions(-)
 create mode 100755 t/perf/p5400-receive-pack.sh

-- 
2.31.1