[dm-devel] regression: data corruption with ext4 on LUKS on nvme with torvalds master

[Mikulas Patocka]

On Thu, 13 May 2021, Bart Van Assche wrote:

> On 5/13/21 12:22 PM, Mikulas Patocka wrote:
> > We already had problems with too large bios in dm-crypt and we fixed it by 
> > adding this piece of code:
> > 
> >         /*
> >          * Check if bio is too large, split as needed.
> >          */
> >         if (unlikely(bio->bi_iter.bi_size > (BIO_MAX_VECS << PAGE_SHIFT)) &&
> >             (bio_data_dir(bio) == WRITE || cc->on_disk_tag_size))
> >                 dm_accept_partial_bio(bio, ((BIO_MAX_VECS << PAGE_SHIFT) >> SECTOR_SHIFT));
> > 
> > It will ask the device mapper to split the bio if it is too large. So, 
> > crypt_alloc_buffer can't receive a bio that is larger than BIO_MAX_VECS << 
> > PAGE_SHIFT.
> 
> Hi Mikulas,
> 
> Are you perhaps referring to commit 4e870e948fba ("dm crypt: fix error
> with too large bios")? Did that commit go upstream before multi-page
> bvec support?

Yes. It's from 2016.

> Can larger bios be supported in case of two or more
> contiguous pages now that multi-page bvec support is upstream?

No - we need to allocate a buffer for the written data. The buffer size is 
limited to PAGE_SIZE * BIO_MAX_VECS.

> Thanks,
> 
> Bart.

Mikulas