Re: [PATCH 1/2] MM: handle THP in swap_*page_fs()

[ying.huang@xxxxxxxxx]

On Thu, 2022-05-05 at 09:41 +1000, NeilBrown wrote:
> On Tue, 03 May 2022, Yang Shi wrote:
> > On Sun, May 1, 2022 at 9:23 PM NeilBrown <neilb@xxxxxxx> wrote:
> > > 
> > > On Sat, 30 Apr 2022, Yang Shi wrote:
> > > > On Thu, Apr 28, 2022 at 5:44 PM NeilBrown <neilb@xxxxxxx> wrote:
> > > > > 
> > > > > Pages passed to swap_readpage()/swap_writepage() are not necessarily all
> > > > > the same size - there may be transparent-huge-pages involves.
> > > > > 
> > > > > The BIO paths of swap_*page() handle this correctly, but the SWP_FS_OPS
> > > > > path does not.
> > > > > 
> > > > > So we need to use thp_size() to find the size, not just assume
> > > > > PAGE_SIZE, and we need to track the total length of the request, not
> > > > > just assume it is "page * PAGE_SIZE".
> > > > 
> > > > Swap-over-nfs doesn't support THP swap IIUC. So SWP_FS_OPS should not
> > > > see THP at all. But I agree to remove the assumption about page size
> > > > in this path.
> > > 
> > > Can you help me understand this please.  How would the swap code know
> > > that swap-over-NFS doesn't support THP swap?  There is no reason that
> > > NFS wouldn't be able to handle 2MB writes.  Even 1GB should work though
> > > NFS would have to split into several smaller WRITE requests.
> > 
> > AFAICT, THP swap is only supported on non-rotate block devices, for
> > example, SSD, PMEM, etc. IIRC, the swap device has to support the
> > cluster in order to swap THP. The cluster is only supported by
> > non-rotate block devices.
> > 
> > Looped Ying in, who is the author of THP swap.
> 
> I hunted around the code and found that THP swap only happens if a
> 'cluster_info' is allocated, and that only happens if 
> 	if (p->bdev && bdev_nonrot(p->bdev)) {
> in the swapon syscall.
> 

And in get_swap_pages(), the cluster is only allocated for block
devices.

		if (size == SWAPFILE_CLUSTER) {
			if (si->flags & SWP_BLKDEV)
				n_ret = swap_alloc_cluster(si, swp_entries);
		} else
			n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
						    n_goal, swp_entries);

We may remove this restriction in the future if someone can show the
benefit.

Best Regards,
Huang, Ying

> I guess "nonrot" is being use as a synonym for "low latency"...
> So even if NFS was low-latency it couldn't benefit from THP swap.
> 
> So as you say it is not currently possible for THP pages to be send to
> NFS for swapout.  It makes sense to prepare for it though I think - if
> only so that the code is more consistent and less confusing.
> 
> Thanks,
> NeilBrown