On Thu, Nov 22, 2018 at 06:32:09PM +0800, Ming Lei wrote:

On Thu, Nov 22, 2018 at 11:04:28AM +0100, Christoph Hellwig wrote:

quoted

On Thu, Nov 22, 2018 at 05:33:00PM +0800, Ming Lei wrote:

quoted

However, using virt boundary limit on non-cluster seems over-kill,
because the bio will be over-split(each small bvec may be split as one bio)
if it includes lots of small segment.

The combination of the virt boundary of PAGE_SIZE - 1 and a
max_segment_size of PAGE_SIZE will only split if the to me merged
segment is in a different page than the previous one, which is exactly
what we need here.  Multiple small bvec inside the same page (e.g.
512 byte buffer_heads) will still be merged.

quoted

What we want to do is just to avoid to merge bvecs to segment, which
should have been done by NO_SG_MERGE simply. However, after multi-page
is enabled, two adjacent bvecs won't be merged any more, I just forget
to remove the bvec merge code in V11.

So seems we can simply avoid to use virt boundary limit for non-cluster
after multipage bvec is enabled?

No, we can't just remove it.  As explained in the patch there is one very
visible difference of setting the flag amd that is no segment will span a
page boundary, and at least the iSCSI code seems to rely on that.

IMO, we should use queue_segment_boundary() to enhance the rule during splitting
segment after multi-page bvec is enabled.

Seems we miss the segment boundary limit in bvec_split_segs().

Yes, that looks like the right fix!