[PATCH v2 09/67] fscache: Implement volume registration

Linus Torvalds torvalds at linux-foundation.org
Thu Dec 9 09:32:34 PST 2021


On Thu, Dec 9, 2021 at 8:55 AM David Howells <dhowells at redhat.com> wrote:
>
> +static long fscache_compare_volume(const struct fscache_volume *a,
> +                                  const struct fscache_volume *b)
> +{
> +       size_t klen;
> +
> +       if (a->key_hash != b->key_hash)
> +               return (long)a->key_hash - (long)b->key_hash;
> +       if (a->cache != b->cache)
> +               return (long)a->cache    - (long)b->cache;
> +       if (a->key[0] != b->key[0])
> +               return (long)a->key[0]   - (long)b->key[0];
> +
> +       klen = round_up(a->key[0] + 1, sizeof(unsigned int));
> +       return memcmp(a->key, b->key, klen);

None of this is endianness-independent except for the final memcmp()
(and that one assumes the data is just a "stream of bytes")

In fact, even if everybody is little-endian, the above gives different
results on 32-bit and 64-bit architectures, since you're doing math in
(possibly) 64 bits but using a 32-bit "key_hash". So sign bits will
differ, afaik.

And once again, that key_hash isn't actually endianness-independent anyway:

> +       volume->key_hash = fscache_hash(0, (unsigned int *)key,
> +                                       hlen / sizeof(unsigned int));

Yeah, for the same key data, this will give entirely different results
on LE vs BE, unless you've made sure to always convert whatever keys
from soem on-disk fixed-32-bit-endianness format to a in-memory host
endianness.

Which is a fundamental design mistake in itself. That kind of "one
endianness on disk, another in memory" is garbage.

I'm not sure any of these matter - maybe all these hashes are entirely
for in-memory stuff and never haev any longer lifetimes, so the fact
that they get calculated and compared differently depending on
endianness and depending on word size may not matter at all. You may
only care about "stable on the native architecture".

But then you shouldn't have your own hash function that you claim is
somehow endianness-safe.

If you really want to be endianness safe, *ALL* the data you work on
needs to be a proper fixed endianness format. All throught the code.
Make all key pointers always be "__le32 *", and never randomly cast
the pointer from some other data like I see in every use I actually
looked at.

                  Linus



More information about the linux-afs mailing list