Am 28.11.2022 um 12:31 schrieb Ævar Arnfjörð Bjarmason: > > On Mon, Nov 28 2022, René Scharfe wrote: > >> Am 28.11.2022 um 11:03 schrieb Junio C Hamano: >>> René Scharfe <l.s.r@xxxxxx> writes: >>> >>>> This reverts commit 5cb28270a1ff94a0a23e67b479bbbec3bc993518. >>>> >>>> 5cb28270a1 (pack-objects: lazily set up "struct rev_info", don't leak, >>>> 2022-03-28) avoided leaking rev_info allocations in many cases by >>>> calling repo_init_revisions() only when the .filter member was actually >>>> needed, but then still leaking it. That was fixed later by 2108fe4a19 >>>> (revisions API users: add straightforward release_revisions(), >>>> 2022-04-13), making the reverted commit unnecessary. >>> >>> Hmph, with this merged, 'seen' breaks linux-leaks job in a strange >>> way. >>> >>> https://github.com/git/git/actions/runs/3563546608/jobs/5986458300#step:5:3917 >>> >>> Does anybody want to help looking into it? > > [I see we crossed E-Mails]: > https://lore.kernel.org/git/221128.868rjvmi3l.gmgdl@xxxxxxxxxxxxxxxxxxx/ > >> The patch exposes that release_revisions() leaks the diffopt allocations >> as we're yet to address the TODO added by 54c8a7c379 (revisions API: add >> a TODO for diff_free(&revs->diffopt), 2022-04-14). > > That's correct, and we have that leak in various places in our codebase, > but per the above side-thread I think this is primarily exposing that > we're setting up the "struct rev_info" with your change when we don't > need to. Why can't we just skip it? I have no idea how to stop get_object_list() from using struct rev_info. We could let it take a struct list_objects_filter_options pointer instead and have it build a struct rev_info internally, but that would just move the problem, not solve it. > Yeah, if we do set it up we'll run into an outstanding leak, and that > should also be fixed (I have some local patches...), but the other cases > I know of where we'll leak that data is where we're actually using the > "struct rev_info". > > I haven't tried tearing your change apart to poke at it myself, and > maybe there's some really good reason for why you can't separate getting > rid of the J.5.7 dependency and removing the lazy-init. > >> The patch below plugs it locally. >> >> --- >8 --- >> Subject: [PATCH 4/3] fixup! revision: free diffopt in release_revisions() >> >> Signed-off-by: René Scharfe <l.s.r@xxxxxx> >> --- >> builtin/pack-objects.c | 1 + >> 1 file changed, 1 insertion(+) >> >> diff --git a/builtin/pack-objects.c b/builtin/pack-objects.c >> index 3e74fbb0cd..a47a3f0fba 100644 >> --- a/builtin/pack-objects.c >> +++ b/builtin/pack-objects.c >> @@ -4462,6 +4462,7 @@ int cmd_pack_objects(int argc, const char **argv, const char *prefix) >> } else { >> get_object_list(&revs, rp.nr, rp.v); >> } >> + diff_free(&revs.diffopt); >> release_revisions(&revs); >> cleanup_preferred_base(); >> if (include_tag && nr_result) > > So, the main motivation for the change was paranoia that a compiler or > platform might show up without J.5.7 support and that would bite us, but > we're now adding a double-free-in-waiting? > > I think we're both a bit paranoid, but clearly have different > paranoia-priorities :) > > If we do end up with some hack like this instead of fixing the > underlying problem I'd much prefer that such a hack just be an UNLEAK() > here. > > I.e. we have a destructor for "revs.*" already, let's not bypass it and > start freeing things from under it, which will result in a double-free > if we forget this callsite once the TODO in 54c8a7c379 is addressed. Well, that TODO fix should remove this new diff_free() call, but I agree that this is fragile. Removing the "TEST_PASSES_SANITIZE_LEAK=true" line from affected tests is probably better. > As you'd see if you made release_revisions() simply call > diff_free(&revs.diffopt) doing so would reveal some really gnarly edge > cases. That was my first attempt; it breaks lots of tests due to double frees. > I haven't dug into this one, but offhand I'm not confident in saying > that this isn't exposing us to some aspect of that gnarlyness (maybe > not, it's been a while since I looked). I saw it as the way towards a release_revisions() that calls diff_free() itself: Add such calls to each of them, fix the "gnarlyness" individually, finally move them all into release_revisions(). The only problem is that there are 60+ callsites. > (IIRC some of the most gnarly edge cases will only show up as CI > failures on Windows, to do with the ordering of when we'll fclose() > files hanging off that "diffopt"). Fun. René
