On Thu, Nov 25, 2021 at 3:10 AM Catalin Marinas [off-list ref] wrote:

For this specific btrfs case, if we want go with tuning the offset based
on the fault address, we'd need copy_to_user_nofault() (or a new
function) to be exact.

I really don't see why you harp on the exactness.

I really believe that the fix is to make the read/write probing just
be more aggressive.

Make the read/write probing require that AT LEAST <n> bytes be
readable/writable at the beginning, where 'n' is 'min(len,ALIGN)', and
ALIGN is whatever size that copy_from/to_user_xyz() might require just
because it might do multi-byte accesses.

In fact, make ALIGN be perhaps something reasonable like 512 bytes or
whatever, and then you know you can handle the btrfs "copy a whole
structure and reset if that fails" case too.

Don't require that the fundamental copying routines (and whatever
fixup the code might need) be some kind of byte-precise - it's the
error case that should instead be made stricter.

If the user gave you a range that triggered a pointer color mismatch,
then returning an error is fine, rather than say "we'll do as much as
we can and waste time and effort on being byte-exact too".

Your earlier argument was that it was too expensive to probe things.
That was based on looking at the whole range that migth be MB (or GB)
in size. So just make it check the first <n> bytes, and problem
solved.

                 Linus