git.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
* [PATCH 0/17] removing questionable uses of git_path
@ 2015-08-10  9:27 Jeff King
  2015-08-10  9:32 ` [PATCH 01/17] cache.h: clarify documentation for git_path, et al Jeff King
                   ` (19 more replies)
  0 siblings, 20 replies; 34+ messages in thread
From: Jeff King @ 2015-08-10  9:27 UTC (permalink / raw)
  To: git; +Cc: Michael Haggerty

Recently Michael and I were working on a patch series (not yet
published), which did something like:


  const char *path = git_path("foo");

  ... do stuff with path ...

  for_each_ref(some_callback, NULL);

  ... do some other stuff ...

  unlink(path);

Clever readers may have spotted the bug immediately, but we did not,
until we found that random loose refs were being deleted from the
repository.

The problem is that git_path uses a static buffer that gets overwritten
by subsequent calls. The ref code uses it to iterate over all of the
loose refs in a directory, so our original path is trashed before
for_each_ref returns. Except to make it even more exciting, git_path
actually has a ring of _four_ buffers, so any trivial test you write
will probably work just fine; it's only when you use a real repository
that it causes problems (and then, only if the code path is such that
the loose refs were not previously accessed and cached!).

Michael likened git_path to "a hand-grenade with the pin pulled out",
and I tend to agree. On the other hand, it's pretty darn useful to be
able to get a quick path without having to deal with memory allocation
and ownership.  This patch series tries to document the danger, and
remove some of the more questionable uses. I don't know whether this is
fixing any actual latent bugs; I traced a number of the code paths
manually, but never found a bug. There were some near misses, though,
which make me believe that seemingly-unrelated refactoring could
introduce a bug.

I stopped short of trying to eradicate git_path entirely, and settled
for:

  git grep -E '[^_](git_|mk)path\('

producing a fairly tame-looking set of function calls. It's OK to pass
the result of git_path() to a system call, or something that is a thin
wrapper around one (e.g., strbuf_read_file).

I think this takes us most of the way there. I left out a few cases
where introducing allocations would have been awkward, and I verified
that there were no bugs (e.g., rerere_path). And I left out a few spots
that conflict with topics in "next" (and luckily, in all cases what is
in next makes the problem go away, so we do not have to follow-up for
those sites).

Along the way, there are a few cleanups (e.g., I polished off the recent
hold_lock_file_for_append topic which was on the list, as it had some
problematic calls).

  [01/17]: cache.h: clarify documentation for git_path, et al
  [02/17]: cache.h: complete set of git_path_submodule helpers
  [03/17]: t5700: modernize style
  [04/17]: add_to_alternates_file: don't add duplicate entries
  [05/17]: remove hold_lock_file_for_append
  [06/17]: prefer git_pathdup to git_path in some possibly-dangerous cases
  [07/17]: prefer mkpathdup to mkpath in assignments
  [08/17]: remote.c: drop extraneous local variable from migrate_file
  [09/17]: refs.c: remove extra git_path calls from read_loose refs
  [10/17]: path.c: drop git_path_submodule
  [11/17]: refs.c: simplify strbufs in reflog setup and writing
  [12/17]: refs.c: avoid repeated git_path calls in rename_tmp_log
  [13/17]: refs.c: avoid git_path assignment in lock_ref_sha1_basic
  [14/17]: refs.c: remove_empty_directories can take a strbuf
  [15/17]: find_hook: keep our own static buffer
  [16/17]: get_repo_path: refactor path-allocation
  [17/17]: memoize common git-path "constant" files

-Peff

^ permalink raw reply	[flat|nested] 34+ messages in thread

end of thread, other threads:[~2015-08-15  9:06 UTC | newest]

Thread overview: 34+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2015-08-10  9:27 [PATCH 0/17] removing questionable uses of git_path Jeff King
2015-08-10  9:32 ` [PATCH 01/17] cache.h: clarify documentation for git_path, et al Jeff King
2015-08-10  9:32 ` [PATCH 02/17] cache.h: complete set of git_path_submodule helpers Jeff King
2015-08-10  9:32 ` [PATCH 03/17] t5700: modernize style Jeff King
2015-08-10  9:34 ` [PATCH 04/17] add_to_alternates_file: don't add duplicate entries Jeff King
2015-08-11  4:00   ` Michael Haggerty
2015-08-11  9:54     ` Jeff King
2015-08-10  9:35 ` [PATCH 05/17] remove hold_lock_file_for_append Jeff King
2015-08-10 22:36   ` Junio C Hamano
2015-08-11  9:38     ` Jeff King
2015-08-10  9:35 ` [PATCH 06/17] prefer git_pathdup to git_path in some possibly-dangerous cases Jeff King
2015-08-10  9:35 ` [PATCH 07/17] prefer mkpathdup to mkpath in assignments Jeff King
2015-08-10  9:35 ` [PATCH 08/17] remote.c: drop extraneous local variable from migrate_file Jeff King
2015-08-10  9:36 ` [PATCH 09/17] refs.c: remove extra git_path calls from read_loose_refs Jeff King
2015-08-10  9:36 ` [PATCH 10/17] path.c: drop git_path_submodule Jeff King
2015-08-10 22:50   ` Junio C Hamano
2015-08-10 22:57     ` Junio C Hamano
2015-08-10 23:52       ` Junio C Hamano
2015-08-11  9:53       ` Jeff King
2015-08-10  9:36 ` [PATCH 11/17] refs.c: simplify strbufs in reflog setup and writing Jeff King
2015-08-10 10:34   ` Michael Haggerty
2015-08-10 12:26     ` Jeff King
2015-08-10  9:36 ` [PATCH 12/17] refs.c: avoid repeated git_path calls in rename_tmp_log Jeff King
2015-08-10  9:37 ` [PATCH 13/17] refs.c: avoid git_path assignment in lock_ref_sha1_basic Jeff King
2015-08-10  9:37 ` [PATCH 14/17] refs.c: remove_empty_directories can take a strbuf Jeff King
2015-08-10  9:37 ` [PATCH 15/17] find_hook: keep our own static buffer Jeff King
2015-08-10  9:37 ` [PATCH 16/17] get_repo_path: refactor path-allocation Jeff King
2015-08-10  9:38 ` [PATCH 17/17] memoize common git-path "constant" files Jeff King
2015-08-10 12:05   ` Michael Haggerty
2015-08-10 12:30     ` Jeff King
2015-08-10 12:06 ` [PATCH 0/17] removing questionable uses of git_path Michael Haggerty
2015-08-10 17:31 ` Junio C Hamano
2015-08-10 17:47   ` Jeff King
2015-08-15  9:05 ` Duy Nguyen

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).