Allocate msg only after fatal checks to avoid leaks

From: jinyaoguo <[email protected]>

In parse_reuse_arg, we previously called xmalloc and strbuf_init before resolving the ref and reading the object, leading to a leaked msg on die() paths. This change moves the allocation of struct note_msg until after repo_get_oid and repo_read_object_file succeed, ensuring no heap memory is held when a fatal error is triggered.

---
Allocate msg only after fatal checks to avoid leaks

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-1998%2Fmugitya03%2Fmlk-3-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-1998/mugitya03/mlk-3-v1
Pull-Request: https://github.com/git/git/pull/1998

 builtin/notes.c | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)

index a3f433ca4c0..6df8a7998fb 100644
@@ -308,7 +308,7 @@ static int parse_file_arg(const struct option *opt, const char *arg, int unset)
 static int parse_reuse_arg(const struct option *opt, const char *arg, int unset)
 {
 	struct note_data *d = opt->value;
-	struct note_msg *msg = xmalloc(sizeof(*msg));
+	struct note_msg *msg;
 	char *value;
 	struct object_id object;
 	enum object_type type;

@@ -316,17 +316,17 @@ static int parse_reuse_arg(const struct option *opt, const char *arg, int unset)
 
 	BUG_ON_OPT_NEG(unset);
 
-	strbuf_init(&msg->buf, 0);
 	if (repo_get_oid(the_repository, arg, &object))
 		die(_("failed to resolve '%s' as a valid ref."), arg);
 	if (!(value = repo_read_object_file(the_repository, &object, &type, &len)))
 		die(_("failed to read object '%s'."), arg);
-	if (type != OBJ_BLOB) {
-		strbuf_release(&msg->buf);
-		free(value);
-		free(msg);
-		die(_("cannot read note data from non-blob object '%s'."), arg);
-	}
+    if (type != OBJ_BLOB) {
+        free(value);
+        die(_("cannot read note data from non-blob object '%s'."), arg);
+    }
+
+    msg = xmalloc(sizeof(*msg));
+    strbuf_init(&msg->buf, 0);
 
 	strbuf_add(&msg->buf, value, len);
 	free(value);

base-commit: 9edff09aec9b5aaa3d5528129bb279a4d34cf5b3

gitgitgadget

Alex via GitGitGadget <[email protected]> writes：

A memory leak on the die() path shouldn't be considered a real leak, right? Since the OS will clean up all memory once the process terminates, explicitly freeing msg isn't necessary in this case.

Lidong

It may not matter in practice, but I think the leak checking machinery like sanitizers would still complain, so I view efforts on plugging such leaks in the error code paths more about decluttering the leak checker output to help us spot the real leaks.

Junio C Hamano <[email protected]> writes：

Makes sense. However, inserting a free-like statement in die() would be messier than using goto, since each die() has a unique message and we need to free stuff at each die().

I disagree here. These are not really leaks, and a leak-checker that complains about them is bad.

When we call die(), the pointer to the buffer is still on the stack, and thus the memory is still reachable and not leaked. Some tools like valgrind may still report these as "still reachable", but because they categorize them properly we can ignore them[1].

The one exception we've seen is that an optimizing compiler may reorder instructions to obliterate the stack (because it knows die() is marked with NORETURN), causing a false positive. We dealt with that via d3775de074 (Makefile: force -O0 when compiling with SANITIZE=leak, 2022-10-18). I don't think we've seen any recurrence since then.

And while it may be tempting to say "well, it does not hurt to free them on the die() path", in my opinion that way madness lies. You may have access to some local variables that can be freed, but there will be many other heap allocations that you don't even know about! Here's a toy example from a similar discussion a few years ago:

https://lore.kernel.org/git/[email protected]/

So I'd really prefer not to go down this route. And I think the existing code in this patch's pre-image that calls free() before die() only on one path should be simplified, so that all die() paths consistently do not worry about this.

I.e., this:

index cc1163242f..f3d5eda104 100644
@@ -321,12 +321,8 @@ static int parse_reuse_arg(const struct option *opt, const char *arg, int unset)
 		die(_("failed to resolve '%s' as a valid ref."), arg);
 	if (!(value = odb_read_object(the_repository->objects, &object, &type, &len)))
 		die(_("failed to read object '%s'."), arg);
-	if (type != OBJ_BLOB) {
-		strbuf_release(&msg->buf);
-		free(value);
-		free(msg);
+	if (type != OBJ_BLOB)
 		die(_("cannot read note data from non-blob object '%s'."), arg);
-	}
 
 	strbuf_add(&msg->buf, value, len);
 	free(value);

-Peff

[1] "still reachable" leaks _can_ be useful when returning from main, because they may show memory held in global structures that we might have been able to free at a more timely spot. But there are a lot of these in Git, most of which are not interesting (e.g., is freeing the_repository really worth caring about?), and I don't think there is a good way to tell the difference.

More pontificating from that earlier discussion:

https://lore.kernel.org/git/[email protected]/

Yeah, I recall that discussion and the example. Yes, we should not have to crawl up from a direct caller of die() and free everything these stack frames hold.