Re: SCSI bug

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Sun, 2016-02-21 at 13:43 -0500, John David Anglin wrote:
> I verified that commit 54efd50bfd873e2dbf784e0b21a8027ba4299a3e in
> linux-block fails to
> boot and commit 41609892701e26724b8617201f43254cadf2e7ae (blk-cgroup:
> Drop unlikely
> before IS_ERR(_OR_NULL)) does boot successfully.  Commit
> 41609892701e26724b8617201f43254cadf2e7ae
> is previous commit in tree.
> 
> I don't believe that the change can be reverted from Linus' tree as
> this commit allowed other
> stuff to be removed (see second paragraph of commit description).

OK, can you just verify you can boot 4.5-rc5 without the sata_sil24
driver?

My theory, based on what Helge produced is that this commit is building
a large transfer list > 65535 and then splitting it wrongly.  I think
the problem is that it's not respecting the DMA boundary, so Helge sees
a transfer size of 69632 which I think slops over on both sides,
requiring 3 segments to describe.  However, the merge code thinks we
only need two (because the length is < 2*65536).  The reason we only
see this with ATA drivers is because virtually no SCSI drivers set the
DMA boundary.  Most ATA drivers require a dma boundary of 65535.

James