On Thu, Nov 22, 2018 at 11:41:50AM +0100, Christoph Hellwig wrote: > 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: > > > On Thu, Nov 22, 2018 at 05:33:00PM +0800, Ming Lei wrote: > > > > 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. > > > > > > > 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! Then your patch should work by just replacing virt boundary with segment boudary limit. I will do that change in V12 if you don't object. Thanks, Ming
