[LU-17916] Enable folio allocation support on the buffed io read/write path (BIO)

• Type: Improvement
• Resolution: Unresolved
• Priority: Minor
• Fix Version/s: Lustre 2.17.0
• Severity: 3
• Assignee: Shaun Tancheff
• Reporter: Shaun Tancheff

Description

When support for folios are added to generic_perform_write() we can enable folio based allocations without re-writing the lustre i/o path to use iomap:

https://www.spinics.net/lists/linux-nfs/msg104002.html
Updated patch:
https://lore.kernel.org/linux-fsdevel/[e-mail]/
Depends on (landed for v6.9):
https://patchwork.kernel.org/project/linux-fsdevel/patch/[e-mail]/

This is tested for NFS
https://www.spinics.net/lists/linux-nfs/msg104003.html

suggesting this alone is enough for a 2x improvement in buffered i/o.

Update lustre to use the folio allocation ie: __filemap_get_folio() instead of the grab_cache_page* family.

Provide a __filemap_get_folio() implementation for compatibility with older kernels.

Assuming no large regressions this should be available in kernel v6.11:
Linux kernel commit: v6.10-rc7-2-g9aac777aaf945
filemap: Convert generic_perform_write() to support large folios

Comments

[Shaun Tancheff]

Attached kernel-5.14.0-503.33.1.ldiskfs.el9_5.spec which adds attached patches 0001-filemap-Convert-generic_perform_write-to-support-lar.patch and 0001-iov_iter-Add-copy_folio_from_iter_atomic.patch to enable large folio support in el9.5 (these patches can also be applied to el9.6)

[Shaun Tancheff]

This is pretty young stuff (fs/netfs) (introduced: v5.12-rc4-40-g3a5829fefd3b) so means that we would have to backport the whole thing for older kernels?

netfs maps functionally to cl_page et. al. from my brief reading so that is not a small change at first glance

[Tim Day]

netfs would still require the folio conversion and most (if not all) of Shaun's patches. But using netfs could save us some amount of effort. Something worth looking at: https://docs.kernel.org/filesystems/netfs_library.html.

[James A Simmons]

Shaun last week Tim Day and I went to the Linux summit involving file systems and we had a discussion about instead of doing an folio abstract which is a large work load to use instead the netfs abstract that David Howells is working on. David would be greatly willing to help out with this work.

[Gerrit Updater]

"Shaun Tancheff <[e-mail]>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/58459
Subject: LU-17916 llite: large folio allocation option
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 61d851a4b4f4a95e200fc6a521d44a9ccc413565

[Gerrit Updater]

"Shaun Tancheff <[e-mail]>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/58458
Subject: LU-17916 osd: switch to struct folio
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 9f00105d639bea3d895ec72a6881484e723b37c2

[Gerrit Updater]

"Shaun Tancheff <[e-mail]>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/58457
Subject: LU-17916 osp: switch to struct folio
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: ac80d564e9ad4ce2ab3d9f15f6ebb2ee29edd2d6

[Gerrit Updater]

"Shaun Tancheff <[e-mail]>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/58456
Subject: LU-17916 osd-ldiskfs: switch to struct folio
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: d881fe29134c40990b92f22fc1d3696297a8e7fa

[Gerrit Updater]

"Shaun Tancheff <[e-mail]>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/58275
Subject: LU-17916 mdt: switch to struct folio
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 540da1549856a0d7fe86b3a14321af7f42830c4d

[Gerrit Updater]

"Shaun Tancheff <[e-mail]>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/58274
Subject: LU-17916 quota: switch to struct folio
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 96b4e4d76b7abd7610230a7c61365c907c2344c5