On Sat, Nov 14, 2020 at 07:37:03PM +0100, René Scharfe wrote:
> do_diff_cache() builds a struct rev_info to hand to diff_cache() from
> scratch by initializing it using repo_init_revisions() and then
> replacing its diffopt and prune_data members.
>
> The diffopt member is initialized to a heap-allocated list of options,
> though. Release it using diff_setup_done() before overwriting it.
Makes sense. This whole "rewrite the options as a heap-allocated list"
thing is pretty gross, but is probably the least-bad solution to the
problem. I wondered if there might be other unpaired diff_setup() /
diff_setup_done() calls. Curiously, there are more of the latter:
$ git grep 'diff_setup(' | wc -l
22
$ git grep 'diff_setup_done(' | wc -l
35
I think because repo_init_revisions() makes an implicit call to
diff_setup().
> The initial value of the prune_data member doesn't need to be released,
> but the copy created using copy_pathspec() does. Clear it after use.
I suspect there are more elements of rev_info that could be allocated
(e.g., in a traversal without "--objects", I think trees and blobs are
left sitting in the pending array). It's a prime candidate for UNLEAK()
in most cases where we do a single traversal and then exit the program.
But for sub-functions like this, we perhaps should bite the bullet and
just make a rev_info_clear() function that can be used everywhere.
(I'm not opposed to your patch here in the meantime, though).
-Peff
