do not write outside heap buffer

* parsing.c (substr): Handle tail < head.

This started when I noticed some cgit segfaults on savannah.gnu.org.
Finding the offending URL/commit and then constructing a stand-alone
reproducer were far more time-consuming than writing the actual patch.

The problem arises with a commit like this, in which the user name
part of the "Author" field is empty:

    $ git log -1
    commit 6f3f41d73393278f3ede68a2cb1e7a2a23fa3421
    Author: <T at h.or>
    Date:   Mon Apr 23 22:29:16 2012 +0200

Here's what happens:

(this is due to buf=malloc(0); strncpy (buf, head, -1);
 where "head" may point to plenty of attacker-specified non-NUL bytes,
 so we can overwrite a zero-length heap buffer with arbitrary data)

 Invalid write of size 1
    at 0x4A09361: strncpy (mc_replace_strmem.c:463)
    by 0x408977: substr (parsing.c:61)
    by 0x4089EF: parse_user (parsing.c:73)
    by 0x408D10: cgit_parse_commit (parsing.c:153)
    by 0x40A540: cgit_mk_refinfo (shared.c:171)
    by 0x40A581: cgit_refs_cb (shared.c:181)
    by 0x43DEB3: do_for_each_ref (refs.c:690)
    by 0x41075E: cgit_print_branches (ui-refs.c:191)
    by 0x416EF2: cgit_print_summary (ui-summary.c:56)
    by 0x40780A: summary_fn (cmd.c:120)
    by 0x40667A: process_request (cgit.c:544)
    by 0x404078: cache_process (cache.c:322)
  Address 0x4c718d0 is 0 bytes after a block of size 0 alloc'd
    at 0x4A0884D: malloc (vg_replace_malloc.c:263)
    by 0x455C85: xmalloc (wrapper.c:35)
    by 0x40894C: substr (parsing.c:60)
    by 0x4089EF: parse_user (parsing.c:73)
    by 0x408D10: cgit_parse_commit (parsing.c:153)
    by 0x40A540: cgit_mk_refinfo (shared.c:171)
    by 0x40A581: cgit_refs_cb (shared.c:181)
    by 0x43DEB3: do_for_each_ref (refs.c:690)
    by 0x41075E: cgit_print_branches (ui-refs.c:191)
    by 0x416EF2: cgit_print_summary (ui-summary.c:56)
    by 0x40780A: summary_fn (cmd.c:120)
    by 0x40667A: process_request (cgit.c:544)

 Invalid write of size 1
    at 0x4A09400: strncpy (mc_replace_strmem.c:463)
    by 0x408977: substr (parsing.c:61)
    by 0x4089EF: parse_user (parsing.c:73)
    by 0x408D10: cgit_parse_commit (parsing.c:153)
    by 0x40A540: cgit_mk_refinfo (shared.c:171)
    by 0x40A581: cgit_refs_cb (shared.c:181)
    by 0x43DEB3: do_for_each_ref (refs.c:690)
    by 0x41075E: cgit_print_branches (ui-refs.c:191)
    by 0x416EF2: cgit_print_summary (ui-summary.c:56)
    by 0x40780A: summary_fn (cmd.c:120)
    by 0x40667A: process_request (cgit.c:544)
    by 0x404078: cache_process (cache.c:322)
  Address 0x4c7192b is not stack'd, malloc'd or (recently) free'd

 Invalid write of size 1
    at 0x4A0940E: strncpy (mc_replace_strmem.c:463)
    by 0x408977: substr (parsing.c:61)
    by 0x4089EF: parse_user (parsing.c:73)
    by 0x408D10: cgit_parse_commit (parsing.c:153)
    by 0x40A540: cgit_mk_refinfo (shared.c:171)
    by 0x40A581: cgit_refs_cb (shared.c:181)
    by 0x43DEB3: do_for_each_ref (refs.c:690)
    by 0x41075E: cgit_print_branches (ui-refs.c:191)
    by 0x416EF2: cgit_print_summary (ui-summary.c:56)
    by 0x40780A: summary_fn (cmd.c:120)
    by 0x40667A: process_request (cgit.c:544)
    by 0x404078: cache_process (cache.c:322)
  Address 0x4c7192d is not stack'd, malloc'd or (recently) free'd

 Process terminating with default action of signal 11 (SIGSEGV)
  Access not within mapped region at address 0x502F000
    at 0x4A09400: strncpy (mc_replace_strmem.c:463)
    by 0x408977: substr (parsing.c:61)
    by 0x4089EF: parse_user (parsing.c:73)
    by 0x408D10: cgit_parse_commit (parsing.c:153)
    by 0x40A540: cgit_mk_refinfo (shared.c:171)
    by 0x40A581: cgit_refs_cb (shared.c:181)
    by 0x43DEB3: do_for_each_ref (refs.c:690)
    by 0x41075E: cgit_print_branches (ui-refs.c:191)
    by 0x416EF2: cgit_print_summary (ui-summary.c:56)
    by 0x40780A: summary_fn (cmd.c:120)
    by 0x40667A: process_request (cgit.c:544)
    by 0x404078: cache_process (cache.c:322)

This happens when tail - head == -1 here:
(parsing.c)

  char *substr(const char *head, const char *tail)
  {
          char *buf;

          buf = xmalloc(tail - head + 1);
          strncpy(buf, head, tail - head);
          buf[tail - head] = '\0';
          return buf;
  }

  char *parse_user(char *t, char **name, char **email, unsigned long *date)
  {
          char *p = t;
          int mode = 1;

          while (p && *p) {
                  if (mode == 1 && *p == '<') {
                          *name = substr(t, p - 1);
                          t = p;
                          mode++;
                  } else if (mode == 1 && *p == '\n') {

The fix is to handle the case of (tail < head) before calling xmalloc,
thus avoiding passing an invalid value to xmalloc.

And here's the reproducer:
It was tricky to reproduce, because git prohibits use of an empty "name"
in a commit ID.  To construct the offending commit, I had to resort to
using "git hash-object".

git init -q foo &&
( cd foo &&
  echo a > j && git add . && git ci -q --author='au <T at h.or>' -m. . &&
  h=$(git cat-file commit HEAD|sed 's/au //' \
    |git hash-object -t commit -w --stdin) &&
  git co -q -b test $h &&
  git br -q -D master &&
  git br -q -m test master)
git clone -q --bare foo foo.git

cat <<EOF > in
repo.url=foo.git
repo.path=foo.git
EOF
CGIT_CONFIG=in QUERY_STRING=url=foo.git valgrind ./cgit

The valgrind output is what you see above.

AFAICS, this is not exploitable thanks (ironically) to the use of strncpy.
Since that -1 translates to SIZE_MAX and this is strncpy, not only does it
copy whatever is in "head" (up to first NUL), but it also writes
SIZE_MAX - strlen(head) NUL bytes into the destination buffer, and that
latter is guaranteed to evoke a segfault.  Since cgit is single-threaded,
AFAICS, there is no way that the buffer clobbering can be turned into
an exploit.

parsing.c

@@ -56,6 +56,8 @@ char *substr(const char *head, const char *tail)
 {
 	char *buf;
 
+	if (tail < head)
+		return xstrdup("");
 	buf = xmalloc(tail - head + 1);
 	strncpy(buf, head, tail - head);
 	buf[tail - head] = '\0';