Re: [PATCH 0/4] plugging some mmap() leaks

From: Ramsay Jones <ramsay@ramsayjones.plus.com>
Date: Mar 6, 2026, 23:25 UTC
Message-ID: <c3ae9ff6-8577-48de-8473-9ec8d22ebc71@ramsayjones.plus.com>
In-Reply-To: <xmqqjyvoy5p2.fsf@gitster.g>

On 06/03/2026 10:05 pm, Junio C Hamano wrote:

Show 38 quoted lines

> Ramsay Jones <ramsay@ramsayjones.plus.com> writes:
> 
>> On 06/03/2026 6:37 pm, Junio C Hamano wrote:
>>> Ramsay Jones <ramsay@ramsayjones.plus.com> writes:
>>>
>>>> When compiling with the NO_MMAP build variable set, the built-in
>>>> 'git_mmap()' and 'git_munmap()' compatability routines use simple
>>>> memory allocation and file I/O to emulate the required behaviour.
>>>> The current implementation is vunerable to the "double-delete" bug
>>>> (where the pointer returned by malloc() is passed to free() two or
>>>> more times), should the mapped memory block address be passed to
>>>> munmap() multiple times.
>>>
>>> Sorry if I am missing something glaringly obvious, but quite
>>> honestly I am confused.  Wouldn't it be a bug to call munmap() again
>>> on the same region of memory obtained from mmap() and then already
>>> unmapped by calling munmap()?
>>
>> Yes. The (second) call to munmap() with the (already unmapped) memory
>> region would return -1 with errno set to EINVAL.
>>
>> The emulation layer does not detect this situation and simply calls
>> free() on the given pointer. Hence the 'double-delete' bug.
>>
>>> Or can the emulation layer cause such a second free() even if the
>>> munmap() is done once and only once per memory region obtained from
>>> a single mmap()?
>>
>> No. If you only git_munmap() once for a given memory region, everything
>> is fine.
> 
> Hmph.  You make it sound as if we have some code that calls munmap()
> on something that we are not sure if we have unmapped just in case,
> trusting that it won't crash us and instead give us EINVAL, and that
> is very much deliberate?  Unless we have such a code, bending over
> backwards to track what has already been unmapped and return -1 with
> EINVAL from munmap() for a second call is of dubious value, no?  I
> still must be missing something...

Sorry to confuse you, I'm not trying to, honest! :)

First, forget that patch. It was not meant to be applied by anyone to anything. It was just to (hopefully) support the explanation (to Jeff specifically) of a bug which happened to me long ago.

In particular, valgrind demonstrated a bug, which has probably been fixed in the intervening years, which directly implied that some code called munmap() on a memory region which had already been unmapped.

Given that this was on cygwin and, at that time, was built with NO_MMAP, this meant that free() was called twice on the same malloc()ed memory. (Indeed this is what valgrind reported).

If you look at the (currently 31) calls to munmap(), only one seems to look at the return (in refs/packed-backend.c:183). So, calling munmap() twice on the same memory region will probably go unnoticed when NO_MMAP is not set. I have no idea why munmap() was called twice on the same memory region, since I didn't track down the code responsible.

It was just an FYI about a _potential_ lurking bug when using the mmap compat routines.

Have I cleared that up, or confused you more. :)

ATB, Ramsay Jones