Fixes a bug likely introduced in
d396819101 (in 1.8.1) whereby
‘proxy_settings.on_status’ would be left uninitialized when using the
‘http-parser’ backend, eventually leading to a segfault in
‘http_parser_execute’. Valgrind would report use of the uninitialized
value like so:
Conditional jump or move depends on uninitialised value(s)
at 0x50CD533: http_parser_execute (http_parser.c:910)
by 0x4928504: git_http_parser_execute (httpparser.c:82)
by 0x4925C42: client_read_and_parse (httpclient.c:1178)
by 0x4926F27: git_http_client_read_response (httpclient.c:1458)
by 0x49255FE: http_stream_read (http.c:427)
by 0x4929B90: git_smart__recv (smart.c:29)
by 0x492C147: git_smart__store_refs (smart_protocol.c:58)
by 0x4929F6C: git_smart__connect (smart.c:171)
by 0x4904DCE: git_remote_connect_ext (remote.c:963)
by 0x48A15D2: clone_into (clone.c:449)
by 0x48A15D2: git__clone (clone.c:546)
by 0x4010E9: main (libgit2-proxy.c:20)
Description: In an older version of libgit2 in git_object_lookup_prefix was a check that repo is valid, but now there is repo->oid_type in the git_object_lookup and must be checked as well
realpath(3) _may_ allocate strings (if the second param is NULL) using
the system allocator. However, callers need an assurance that they can
free memory using git__free. If we made realpath do an allocation, then
make sure that we strdup it into our allocator's memory.
More importantly, avoid this behavior by always providing a buffer to
p_realpath invocations.
Instead of tweaking the `stdalloc` allocator when
`GIT_DEBUG_STRICT_ALLOC` is defined, actually create a debugging
allocator. This allows us to ensure that we are strict about things like
not expecting `malloc(0)` to do something useful, but we can also
introduce an excessively pedantic `realloc` implementation that _always_
creates a new buffer, throws away its original `ptr`, and overwrites the
data that's there with garbage. This may be helpful to identify places
that make assumptions about realloc.
Avoid sloppy aliasing in our (re-)allocation, which is undefined
behavior. This has been problematic before and was helped by `volatile`
(see b62a6a13b2) but that is not
technically correct, and some compilers / architectures do not
understand that `ptr` is changing due to its aliasing.
Just make `git_array_alloc` behave like `realloc`, taking a `void *` and
returning a `void *`.
In attempting to make a clever change that added fetch options as flags,
while keeping ABI and API compatibility, we screwed up. Bitfields in
structs are implementation-specific and are not necessarily ABI
compatible across compilers.
Make `update_fetchhead` a flags value which is weirdly named, but
that's a future problem. This removes the `report_unchanged` option from
API.
Since 3618a2a, `git_error_last` does not return NULL when there was no
error. Adapt to this when testing for a user cancelling a callback but
not setting an error message, testing for `klass` of `GIT_ERROR_NONE`.
