[PATCH 1/7] config: avoid "write_in_full(fd, buf, len) < len" pattern

The return type of write_in_full() is a signed ssize_t, because we may return "-1" on failure (even if we succeeded in writing some bytes). But "len" itself may be an unsigned type (the function takes a size_t, but of course we may have something else in the calling function). So while it seems like:

  if (write_in_full(fd, buf, len) < len)
	die_errno("write error");

would notice an error, it won't if "len" is unsigned. The compiler sees a signed/unsigned comparison and promotes the signed value, resulting in (size_t)-1, the highest possible size_t (or again, whatever type the caller has). This cannot possibly be smaller than "len", and so the conditional can never trigger.

I scoured the code base for cases of this, but it turns out that these two in git_config_set_multivar_in_file_gently() are the only ones. This case is actually quite interesting: we don't have a size_t, but rather use the subtraction of two pointers. Which you might think would be a signed ptrdiff_t, but clearly both gcc and clang treat it as unsigned (possibly because the conditional just above guarantees that the result is greater than zero).

We can avoid the whole question by just checking for a negative return value directly, as write_in_full() will never return any value except -1 or the full count.

There's no addition to the test suite here, since you need to convince write() to fail in order to see the problem. The simplest reproduction recipe I came up with is to trigger ENOSPC (this only works on Linux, obviously):

  # make a limited-size filesystem
  dd if=/dev/zero of=small.disk bs=1M count=1
  mke2fs small.disk
  mkdir mnt
  sudo mount -o loop small.disk mnt
  cd mnt
  sudo chown $USER:$USER .

  # make a config file with some content
  git config --file=config one.key value
  git config --file=config two.key value

  # now fill up the disk
  dd if=/dev/zero of=fill

  # and try to delete a key, which requires copying the rest
  # of the file to config.lock, and will fail on write()
  git config --file=config --unset two.key

That final command should (and does after this patch) produce an error message due to the failed write, and leave the file intact. With the current code it silently ignores the failure and renames config.lock into place, leaving you with a totally empty config file!

Reported-by: demerphq <demerphq@gmail.com>
Signed-off-by: Jeff King <peff@peff.net>
---
 config.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)