2024-05-10 04:48:26

by Al Viro

[permalink] [raw]
Subject: Re: [PATCH] libfs: fix accidental overflow in offset calculation

On Fri, May 10, 2024 at 03:26:08AM +0000, Justin Stitt wrote:

> This feels like a case of accidental correctness. You demonstrated that
> even with overflow we end up going down a control path that returns an
> error code so all is good.

No. It's about a very simple arithmetical fact: the smallest number that
wraps to 0 is 2^N, which is more than twice the maximal signed N-bit
value. So wraparound on adding a signed N-bit to non-negative signed N-bit
will always end up with negative result. That's *NOT* a hard math. Really.

As for the rest... SEEK_CUR semantics is "seek to current position + offset";
just about any ->llseek() instance will have that shape - calculate the
position we want to get to, then forget about the difference between
SEEK_SET and SEEK_CUR. So noticing that wraparound ends with negative
is enough - we reject straight SEEK_SET to negatives anyway, so no
extra logics is needed.

> However, I think finding the solution
> shouldn't require as much mental gymnastics. We clearly don't want our
> file offsets to wraparound and a plain-and-simple check for that lets
> readers of the code understand this.

No comments that would be suitable for any kind of polite company.


2024-05-10 11:22:04

by Al Viro

[permalink] [raw]
Subject: Re: [PATCH] libfs: fix accidental overflow in offset calculation

On Fri, May 10, 2024 at 05:48:05AM +0100, Al Viro wrote:
> On Fri, May 10, 2024 at 03:26:08AM +0000, Justin Stitt wrote:
>
> > This feels like a case of accidental correctness. You demonstrated that
> > even with overflow we end up going down a control path that returns an
> > error code so all is good.
>
> No. It's about a very simple arithmetical fact: the smallest number that
> wraps to 0 is 2^N, which is more than twice the maximal signed N-bit
> value. So wraparound on adding a signed N-bit to non-negative signed N-bit
> will always end up with negative result. That's *NOT* a hard math. Really.
>
> As for the rest... SEEK_CUR semantics is "seek to current position + offset";
> just about any ->llseek() instance will have that shape - calculate the
> position we want to get to, then forget about the difference between
> SEEK_SET and SEEK_CUR. So noticing that wraparound ends with negative
> is enough - we reject straight SEEK_SET to negatives anyway, so no
> extra logics is needed.
>
> > However, I think finding the solution
> > shouldn't require as much mental gymnastics. We clearly don't want our
> > file offsets to wraparound and a plain-and-simple check for that lets
> > readers of the code understand this.
>
> No comments that would be suitable for any kind of polite company.

FWIW, exchange of nasty cracks aside, I believe that this kind of
whack-a-mole in ->llseek() instances is just plain wrong. We have
80-odd instances in the tree.

Sure, a lot of them a wrappers for standard helpers, but that's
still way too many places to spill that stuff over. And just
about every instance that supports SEEK_CUR has exact same kind
of logics.

As the matter of fact, it would be interesting to find out
which instances, if any, do *not* have that relationship
between SEEK_CUR and SEEK_SET. If such are rare, it might
make sense to mark them as such in file_operations and
have vfs_llseek() check that - it would've killed a whole
lot of boilerplate. And there it a careful handling of
overflow checks (or a clear comment explaining what's
going on) would make a lot more sense.

IF we know that an instance deals with SEEK_CUR as SEEK_SET to
offset + ->f_pos, we can translate SEEK_CUR into SEEK_SET
in the caller.