[bug report] block: integer overflow in __bvec_gap_to_prev()

[bug report] block: integer overflow in __bvec_gap_to_prev()

[Roman Smirnov]

Hello.

There is a case of integer overflow in __bvec_gap_to_prev():

	((bprv->bv_offset + bprv->bv_len) & lim->virt_boundary_mask);

bio_vec can cross multiple pages:

https://lore.kernel.org/lkml/20190215111324.30129-1-ming.lei@redhat.com/t/

So, in case bio has one bio_vec bv_len can have a maximum value of UINT_MAX.
The check happens in bio_full(). In the case when bv_len is equal to
UINT_MAX and bv_offset is greater than zero, an overflow may occur.

Found by Linux Verification Center (linuxtesting.org) with Svace.

Re: [bug report] block: integer overflow in __bvec_gap_to_prev()

[Keith Busch]

On Tue, Jun 11, 2024 at 02:23:48PM +0000, Roman Smirnov wrote:
> Hello.
> 
> There is a case of integer overflow in __bvec_gap_to_prev():
> 
> 	((bprv->bv_offset + bprv->bv_len) & lim->virt_boundary_mask);
> 
> bio_vec can cross multiple pages:
> 
> https://lore.kernel.org/lkml/20190215111324.30129-1-ming.lei@redhat.com/t/
> 
> So, in case bio has one bio_vec bv_len can have a maximum value of UINT_MAX.
> The check happens in bio_full(). In the case when bv_len is equal to
> UINT_MAX and bv_offset is greater than zero, an overflow may occur.

Does it matter? The lower bits checked against the mask should be the
same regardless of overflow.

Re: [bug report] block: integer overflow in __bvec_gap_to_prev()

[Roman Smirnov]

On Tue, 2024-06-11 at 09:15 -0600, Keith Busch wrote:
> On Tue, Jun 11, 2024 at 02:23:48PM +0000, Roman Smirnov wrote:
> > Hello.
> > 
> > There is a case of integer overflow in __bvec_gap_to_prev():
> > 
> >         ((bprv->bv_offset + bprv->bv_len) & lim->virt_boundary_mask);
> > 
> > bio_vec can cross multiple pages:
> > 
> > https://lore.kernel.org/lkml/20190215111324.30129-1-ming.lei@redhat.com/t/
> > 
> > So, in case bio has one bio_vec bv_len can have a maximum value of UINT_MAX.
> > The check happens in bio_full(). In the case when bv_len is equal to
> > UINT_MAX and bv_offset is greater than zero, an overflow may occur.
> 
> Does it matter? The lower bits checked against the mask should be the
> same regardless of overflow.


There are several other places where this kind of thing happens:

https://elixir.bootlin.com/linux/latest/source/block/blk-merge.c#L292
https://elixir.bootlin.com/linux/latest/source/block/blk.h#L331

I think in those cases overflow would make a difference.

I also found a comment before __bio_add_page(). It says that the
caller should watch out for free space in bio:

https://elixir.bootlin.com/linux/latest/source/block/bio.c#L1075

But what happens if it doesn't keep a check on it? Such code
won't get into the kernel?