On Mon, Apr 04, 2016 at 08:50:42PM +0100, Al Viro wrote:
> What happens if somebody issues SG_IO with 256-segment vector, each segment
> 1 byte long and page-aligned? Will the driver really be happy with the
> resulting request, as long as it hasn't claimed non-zero queue_virt_boundary?
> Because AFAICS we'll get a request with a pile of bvecs, each with
> ->bv_offset equal to 0 and ->bv_len equal to 1; can that really work?
OK, it really doesn't make sense. What happened, AFAICS, is that when
blk_rq_map_user_iov() has grown a "misaligned, need to copy" code, the
check had been mishandled - rather than checking both the base and the
length of segments, as blk_rq_map_{user,kern} used to do (and as ..._kern
is still doing) it checked only the base.
Then in "block: use blk_rq_map_user_iov to implement blk_rq_map_user" you've
missed that problem, which got us the current situaiton. Note that e.g.
PIO case of libata really wants copy in case of 500 bytes + 12 bytes
vector - it'll splat the last 12 bytes adjacent to the end of the first
segment, etc.
AFAICS, what we need there is simply
nr_pages = iov_iter_npages(iter);
alignment = iov_iter_alignment(iter);
if (alignment & (queue_dma_alignment(q) | q->dma_pad_mask))
copy = true;
and I really wonder if we care about special-casing the situation when the
ends are not aligned to queue_virt_boundary(q). If we don't, we might as
well add queue_virt_boundary(q) to the mask we are checking. If we do,
it's not hard to add a variant that would calculate both the alignment and
alignment for internal boundaries...
--
To unsubscribe from this list: send the line "unsubscribe linux-block" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
