From: Junio C Hamano Date: Thu, 21 Mar 2013 17:28:41 GMT Subject: Re: [PATCH v2 00/45] parse_pathspec and :(glob) magic Message-ID: <7v4ng4br1y.fsf@alter.siamese.dyndns.org> In-Reply-To: <20130321054349.GA18101@duynguyen-vnpc.dek-tpc.internal> Duy Nguyen writes: >> I still can't reproduce it. But I think I found a bug that >> miscalculates prefix length from absolute paths. Does this "fix" your >> test? >> ... > Nope, that one could cause more crashes. Try this > > -- 8< -- > diff --git a/setup.c b/setup.c > index 3584f22..3d8eb97 100644 > --- a/setup.c > +++ b/setup.c > @@ -14,6 +14,8 @@ char *prefix_path_gently(const char *prefix, int *p_len, const char *path) > const char *temp = real_path(path); > sanitized = xmalloc(len + strlen(temp) + 1); > strcpy(sanitized, temp); > + if (p_len) > + *p_len = 0; Yes, this one seems to. "$(pwd)/../src" was not handled correctly. The callchain to this locaiton would look like parse_pathspec() with prefix="docs/", prefixlen set to 5 -> prefix_pathspec(), &prefixlen passed down -> prefix_path_gently(), p_len points at the above prefixlen your "this should fix" patch sets *p_len to 0, original leaves *p_len as 5. -> normalize_path_copy_len() with p_len *p_len is used here. Why could the test pass for you without it? It doesn't look like a bug that depended on uninitialized memory or something from the above observation.