Re: [PATCH] pack-bitmap-write: use hashwrite_be32() in write_hash_cache()

From: Jeff King <peff@peff.net>
Date: Sep 7, 2020, 02:23 UTC
Message-ID: <20200907022340.GA1208024@coredump.intra.peff.net>
In-Reply-To: <20200906190235.GA6146@nand.local>

On Sun, Sep 06, 2020 at 03:02:35PM -0400, Taylor Blau wrote:

Show 13 quoted lines

> On Sun, Sep 06, 2020 at 10:59:06AM +0200, René Scharfe wrote:
> > -		uint32_t hash_value = htonl(entry->hash);
> > -		hashwrite(f, &hash_value, sizeof(hash_value));
> > +		hashwrite_be32(f, entry->hash);
> 
> This is an obviously correct translation of what's already written, and
> indeed it is shorter and easier to read.
> 
> Unfortunately, I think there is some more subtlety here since the hash
> cache isn't guarenteed to be aligned, and so blindly calling htonl()
> (either directly in write_hash_cache(), or indirectly in
> hashwrite_be32()) might cause tools like ASan to complain when loading
> data on architectures that don't support fast unaligned reads.

I think the alignment here is fine. We're just writing out an individual value. So in the original entry->hash and our local hash_value are both properly aligned, since they're declared as uint32_t. We pass the pointer to hashwrite(), but it doesn't expect any particular alignment. After the patch, the situation is the same, except that we're working with the uint32_t parameter to hashwrite_be32(), which is also properly aligned.

Show 5 quoted lines

> So, I think that we could do one of three things, depending on how much
> you care about improving this case ;-).
> 
>   - leave your patch alone, accepting that this case which was broken
>     before will remain broken, and leave it as #leftoverbits

So I think this is what we should do. :)

>   - discard your patch as-is, and replace the 'htonl' with 'get_be32()'
>     before handing it off to 'hashwrite()', or

No need here; get_be32() is for when we're reading in from an unaligned mmap.

>   - change the 'hashwrite_beXX()' implementations to use the correct
>     'get_beXX' wrappers which behave like htonl() on architectures with
>     fast unaligned loads, and fall back to byte reads and shifts on
>     architectures that don't.

Likewise, I don't think there's any reason to do this. hashwrite_be32() gets its parameter as a value, not a pointer. So even if it were coming from an unaligned mmap, it's actually the _caller_ who would have to use get_be32() when passing it.

> Credit goes to Peff for finding this issue in GitHub's fork. For what
> it's worth, we were planning on sending those patches to the list soon,
> but they are tied up with a longer series in the meantime.

There is a bug in our fork, but I don't think it's upstream. The relevant spot for the name-hash cache is in show_objects_for_type(), which reads from the bitmap->hashes pointer (that points into our unaligned mmap). But it does:

        if (bitmap_git->hashes)
                hash = get_be32(bitmap_git->hashes + entry->nr);

which is correct (using htonl() would not be). The bug is that in our fork, we have a custom bit-cache extension[1] which does use htonl(), and should be get_be32(). That's something we'll need to clean up when we send those patches upstream.

-Peff

[1] For the curious, the point is it keep a cache of the bit position of
    each object, which lets us ask "is this object's bit set" without
    having to load the revindex. It's helpful for bitmap-optimizing some
    algorithms like "branch --contains", though I think we should
    re-evaluate how much it helps now that we have commit-graphs with
    generation numbers.