mbox series

[00/12] Fix all leaks in tests t0002-t0099: Part 1

Message ID pull.929.git.1617994052.gitgitgadget@gmail.com (mailing list archive)
Headers show
Series Fix all leaks in tests t0002-t0099: Part 1 | expand

Message

Derrick Stolee via GitGitGadget April 9, 2021, 6:47 p.m. UTC
This series fixes approximately half of the real leaks I've found while
running t0002 - t0099 under LSAN.

2 more series will likely be needed to allow t0000-t0099 to pass with LSAN
enabled. (I have all the necessary fixes ready and tested on my machine,
although I want to revisit some of my changes before I'm happy enough to
send them out for review - either way I figure it's easiest to deal with one
batch at a time, so I'll hold off on sending those out for now. One series
is going to consist almost entirely of UNLEAK annotations that are boring
and not worth merging until the real leaks are fixed.)

The exciting news is that once we succeed in getting t000-t0099 to run leak
free, we'll be a significant step closer to being able to run the entire
test-suite leak-free:

 * before the merging of ah/plugleaks (fixing leaks in t0001): 53% of test
   cases fail when LSAN is enabled (12386/23330).
 * with ah/plugleaks + this series + my 2 currently unpublished series: 34%
   of test cases fail when LSAN is enabled (7829/23342).

(I haven't bothered to test most of the intermediate stages, but ISTR that
ah/plugleaks which only had a marginal effect - somewhere on the order of
51-52% test cases were failing after that work merged.)

On the topic of avoiding regressions: I've started running a subset of the
test-suite with LSAN enabled (in addition to the full test-suite with ASAN
and UBSAN) on my Github fork of git, automatically on a daily basis. This
should hopefully help catch any new leaks that appear (and also new
ASAN/UBSAN issues). [The entire test-suite takes around 35 minutes with ASAN
or UBSAN enabled, which isn't too bad compared to the default
linux-gcc/linux-clang jobs which take a similar amount of time - although
they run the test-suite twice with 2 configurations.]

Andrzej Hunt (12):
  revision: free remainder of old commit list in limit_list
  wt-status: fix multiple small leaks
  ls-files: free max_prefix when done
  bloom: clear each bloom_key after use
  branch: FREE_AND_NULL instead of NULL'ing real_ref
  builtin/bugreport: don't leak prefixed filename
  builtin/check-ignore: clear_pathspec before returning
  builtin/checkout: clear pending objects after diffing
  mailinfo: also free strbuf lists when clearing mailinfo
  builtin/for-each-ref: free filter and UNLEAK sorting.
  builtin/rebase: release git_format_patch_opt too
  builtin/rm: avoid leaking pathspec and seen

 bloom.c                |  1 +
 branch.c               |  2 +-
 builtin/bugreport.c    |  8 +++++---
 builtin/check-ignore.c |  1 +
 builtin/checkout.c     |  1 +
 builtin/for-each-ref.c |  3 +++
 builtin/ls-files.c     |  1 +
 builtin/rebase.c       |  1 +
 builtin/rm.c           |  2 ++
 mailinfo.c             | 14 +++-----------
 revision.c             |  1 +
 wt-status.c            |  4 ++++
 12 files changed, 24 insertions(+), 15 deletions(-)


base-commit: 89b43f80a514aee58b662ad606e6352e03eaeee4
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-929%2Fahunt%2Fleaksan-100-part1-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-929/ahunt/leaksan-100-part1-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/929