News and archive of git@vger.kernel.org, since April 2005
4 messages between Jul 19, 2026 and Jul 19, 2026, from brian m. carlson, Jeff King.
Plain Markdown or JSON for tools and agents. Diffs are folded; open one to read it.
Peff recently sent out a series to fix several memory leaks with our hashing code when not using the default block algorithm. This series follows up with a few fixes to our Rust hash code, which calls the C code, to fix various memory problems.
brian m. carlson (2): hash: initialize context before cloning rust: discard hash context when finished
src/hash.rs | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-)
[PATCH 2/2] rust: discard hash context when finished
When we allocate a context but then abandon it, we never discard it, which means that the underlying crypto library context may leak. This doesn't happen with our default block code, but it may with OpenSSL. Note that we do call git_hash_free, which frees the memory we called from git_hash_alloc, but doesn't discard the underlying context itself.
This can be seen with the following command when compiling with OpenSSL and running with nightly Rust:
RUSTFLAGS='-Z sanitizer=leak' cargo test
Discard the context in our context handler. Note that it is fine to do so even after finalizing the context, so our final functions which take self instead of &mut self will not mishandle memory.
Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net> --- src/hash.rs | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-)
diff --git a/src/hash.rs b/src/hash.rs index 4d14e4b4fa..e1f2d31fc3 100644 --- a/src/hash.rs +++ b/src/hash.rs @@ -194,7 +194,10 @@ impl Clone for CryptoHasher { impl Drop for CryptoHasher { fn drop(&mut self) { - unsafe { c::git_hash_free(self.ctx) }; + unsafe { + c::git_hash_discard(self.ctx); + c::git_hash_free(self.ctx); + }; } } @@ -356,6 +359,7 @@ pub mod c { pub fn git_hash_clone(dst: *mut c_void, src: *const c_void); pub fn git_hash_update(ctx: *mut c_void, inp: *const c_void, len: usize); pub fn git_hash_final(hash: *mut u8, ctx: *mut c_void); + pub fn git_hash_discard(ctx: *mut c_void); pub fn git_hash_final_oid(hash: *mut c_void, ctx: *mut c_void); } } @@ -450,6 +454,7 @@ mod tests { h.update(&data[2..]); let h2 = h.clone(); + let h3 = h2.clone(); let actual_oid = h.into_oid(); assert_eq!(**oid, actual_oid); @@ -463,6 +468,7 @@ mod tests { let actual_oid = h.into_oid(); assert_eq!(**oid, actual_oid); + std::mem::drop(h3); } } }
[PATCH 1/2] hash: initialize context before cloning
Our C-based clone helper requires that the context be initialized, but we neglect to do that in our Clone implementation for CryptoHasher. This does not matter when using our default block SHA-256 implementation, but it does cause a crash when using OpenSSL as the backend. Fix this by properly initializing the context before cloning into it.
Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net> --- src/hash.rs | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)
diff --git a/src/hash.rs b/src/hash.rs index dea2998de4..4d14e4b4fa 100644 --- a/src/hash.rs +++ b/src/hash.rs @@ -181,7 +181,10 @@ impl CryptoDigest for CryptoHasher { impl Clone for CryptoHasher { fn clone(&self) -> Self { let ctx = unsafe { c::git_hash_alloc() }; - unsafe { c::git_hash_clone(ctx, self.ctx) }; + unsafe { + c::git_hash_init(ctx, self.algo.hash_algo_ptr()); + c::git_hash_clone(ctx, self.ctx) + }; Self { algo: self.algo, ctx,
Re: [PATCH 0/2] Rust hash cleanups
On Sun, Jul 19, 2026 at 01:08:40AM +0000, brian m. carlson wrote:
> Peff recently sent out a series to fix several memory leaks with our > hashing code when not using the default block algorithm. This series > follows up with a few fixes to our Rust hash code, which calls the C > code, to fix various memory problems.
Both of these look good to me (modulo my almost-zero knowledge of the Rust bits).
I was worried at first that I had introduced new problems with my fixes, but I think these are both pre-existing issues (really just variants of the cleanups I did in the C code).
For patch 1, an alternative is to switch git_hash_clone() to _not_ require initialization. But it introduces the leak problem in the opposite direction. E.g., hashfile_truncate() wants to overwrite existing state, so it would now need to discard() before cloning. I doubt it's worth the effort or risk of regression to save the tiny bit of effort spent on a few init-then-overwrite cases.
So the approach taken here makes sense (and obviously this is just following the C code's lead anyway).
-Peff