On Jan 24, 2023, at 11:07, Olga Kornievskaia [off-list ref] wrote:

On Fri, Jan 20, 2023 at 12:10 AM [off-list ref] wrote:

quoted

From: Trond Myklebust <redacted>

This set of patches represents an initial effort to convert the page
cache I/O in the NFS client to use native folio functionality. It should
allow the nfs_page structs and their helpers to carry folios (including
folios of order > 0) and to pass their data contents through to the RPC
layer.
Note that because O_DIRECT uses pages, we still need to support the
traditional page based I/O, and so the new struct nfs_page will carry
both types.
I did not touch the fscache code, but I expect that to be able to
continue to work with order 0 folios.

The plan is to merge this functionality with order 0 folios first, in
order to catch any regressions in existing functionality. Then we can
enable order n > 0 once we're happy about the stability (at least for
the non-fscache case).

At this point, the xfstests are all passing without any regressions on
my setup, so I'm throwing the patches over the fence to allow for wider
testing.
Please make sure, in particular to test pNFS if your server supports it.
I didn't have to make any changes to the pNFS code, and I don't expect
any trouble, but it would be good to have validation of that assumption.

Here's my experience with running with these patches running thru
xfstest's quick group:
Against a linux server: takes about a couple of minutes longer to run
with folio patches (48m with folio, 45 without) but that's just from 1
run with and and without.

Against an ontap server (so pnfs case):  total time is higher with
patches: I have a difference of 47m with folio and 38m without.

While I don't believe this is related to folio patches but I need to
increase my vm's size to 4g to have a successful run of xfstest. Tests
generic/531 and generic/460 are problematic with 2G. generic/531 leads
to oom-killer and generic/460 hangs.

Hmm… Does the performance climb back to normal if you revert the patch 21f5ae90169b ("NFS: fix up nfs_release_folio() to try to release the page”)? That’s the only one I can think of that does more than just convert from struct page -> struct folio.

_________________________________
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@hammerspace.com