From: Junio C Hamano <gitster@pobox.com>
Date: Wed, 16 Mar 2011 20:18:04 GMT
Subject: Re: [PATCH] valgrind: ignore SSE-based strlen invalid reads
Message-ID: <7vr5a67rb7.fsf@alter.siamese.dyndns.org>
In-Reply-To: <1300275961-5798-1-git-send-email-cmn@elego.de>

Carlos Martín Nieto <cmn@elego.de> writes:

> Some versions of strlen use SSE to speed up the calculation and load 4
> bytes at a time, even if it means reading past the end of the
> allocated memory. This read is safe and when the strlen function is
> inlined, it is not replaced by valgrind, which reports a
> false-possitive.
>
> Tell valgrind to ignore this particular error, as the read is, in
> fact, safe. Current upstream-released version 2.6.1 is affected. Some
> distributions have this fixed in their latest versions.
>
> Signed-off-by: Carlos Martín Nieto <cmn@elego.de>
> ---
>
>>>  I think 3.6.1 doesn't need it, as Debian's 1:3.5.0+3.6.0svn20100609-1
>>> version is reportedly fixed.
>>
>>Ah, nice.  A phrase like "some versions of valgrind before 3.6.1"
>>would be fine with me fwiw. :)
>
> I just downloaded and compiled the upstream release 2.6.1 and it's
> still affected (it does fix some other related
> false-positives). Fedorea rawhide has the fix in, according to their
> bug tracker. I haven't tested the reportedly-fixed version in Debian
> yet.

I take it that you meant 3.6.1 in both places above?

This somehow reminds me of my past life where I saw a buggy implementation
of strlen() in the C library loaded one word too many from memory, and
segfaulted even when the string ended before the end of a mapped page when
the next page was unmapped.

Anyway, nice digging.