[PATCH 0/4] NFS/localio: various improvements

[Mike Snitzer <snitzer@kernel.org>]

Hi,

This series has been rebased on Trond's recent patchset:
https://lore.kernel.org/linux-nfs/cover.1767459435.git.trond.myklebust@hammerspace.com/

but the first 2 patches were developed while chasing a particularly
nasty deadlock that was reproducible when putting LOCALIO under heavy
load on systems with smaller amounts of memory (16G).  Ultimately that
deadlock happens to also be fixed by Trond's recent commit here:
https://git.linux-nfs.org/?p=trondmy/linux-nfs.git;a=commitdiff;h=cce0be6eb4971456b703aaeafd571650d314bcca

FYI, prior to that nfs_release_folio() deadlock fix, tasks would be
stuck like:

[92]            "khugepaged"
[<0>] wait_on_commit+0x71/0xb0 [nfs]
[<0>] __nfs_commit_inode+0x131/0x180 [nfs]
[<0>] nfs_wb_folio+0xa7/0x1b0 [nfs]
[<0>] nfs_release_folio+0x6b/0x120 [nfs]
[<0>] split_huge_page_to_list_to_order+0x367/0x840
[<0>] migrate_pages_batch+0x208/0x7b0
[<0>] migrate_pages+0x220/0x550
[<0>] compact_zone+0x300/0x6e0
[<0>] compact_zone_order+0xc7/0x110
[<0>] try_to_compact_pages+0xe6/0x2d0
[<0>] __alloc_pages_direct_compact+0x8a/0x210
[<0>] __alloc_pages_slowpath.constprop.0+0x218/0x9c0
[<0>] __alloc_pages_noprof+0x335/0x350
[<0>] __folio_alloc_noprof+0x14/0x30
[<0>] alloc_charge_folio+0xdd/0x1b0
[<0>] collapse_huge_page+0x63/0x700
[<0>] hpage_collapse_scan_pmd+0x627/0x750
[<0>] khugepaged_scan_mm_slot.constprop.0+0x3cf/0x5a0
[<0>] khugepaged+0x105/0x200
[<0>] kthread+0xef/0x230
[<0>] ret_from_fork+0x31/0x50
[<0>] ret_from_fork_asm+0x1a/0x30

and:

[12813]         "nfsd"
[<0>] wait_on_commit+0x71/0xb0 [nfs]
[<0>] __nfs_commit_inode+0x131/0x180 [nfs]
[<0>] nfs_wb_folio+0xa7/0x1b0 [nfs]
[<0>] nfs_release_folio+0x6b/0x120 [nfs]
[<0>] split_huge_page_to_list_to_order+0x367/0x840
[<0>] migrate_pages_batch+0x208/0x7b0
[<0>] migrate_pages+0x220/0x550
[<0>] compact_zone+0x300/0x6e0
[<0>] compact_zone_order+0xc7/0x110
[<0>] try_to_compact_pages+0xe6/0x2d0
[<0>] __alloc_pages_direct_compact+0x8a/0x210
[<0>] __alloc_pages_slowpath.constprop.0+0x218/0x9c0
[<0>] __alloc_pages_noprof+0x335/0x350
[<0>] alloc_pages_mpol_noprof+0x8f/0x1f0
[<0>] folio_alloc_noprof+0x5b/0xb0
[<0>] __filemap_get_folio+0x177/0x330
[<0>] nfs_write_begin+0x81/0x370 [nfs]
[<0>] generic_perform_write+0x91/0x290
[<0>] nfs_file_write+0x1ea/0x2f0 [nfs]
[<0>] vfs_iocb_iter_write+0xc0/0x210
[<0>] nfsd_vfs_write+0x26e/0x760 [nfsd]
[<0>] nfsd_write+0x17c/0x1c0 [nfsd]
[<0>] nfsd3_proc_write+0x10b/0x1a0 [nfsd]
[<0>] nfsd_dispatch+0xea/0x220 [nfsd]
[<0>] svc_process_common+0x4cd/0x6c0 [sunrpc]
[<0>] svc_process+0x145/0x210 [sunrpc]
[<0>] svc_handle_xprt+0x481/0x550 [sunrpc]
[<0>] svc_recv+0x170/0x2c0 [sunrpc]
[<0>] nfsd+0x8f/0xf0 [nfsd]
[<0>] kthread+0xef/0x230
[<0>] ret_from_fork+0x31/0x50
[<0>] ret_from_fork_asm+0x1a/0x30

Anyway, I think this LOCALIO series is worthwhile on its own and ready
for wider review and hopeful inclusion.

Thanks,
Mike

Mike Snitzer (4):
  NFS/localio: prevent direct reclaim recursion into NFS via nfs_writepages
  NFS/localio: use GFP_NOIO and non-memreclaim workqueue in nfs_local_commit
  NFS/localio: remove -EAGAIN handling in nfs_local_doio()
  NFS/localio: switch nfs_local_do_read and nfs_local_do_write to return void

 fs/nfs/localio.c | 60 +++++++++++++++++++++++++++++-------------------
 1 file changed, 36 insertions(+), 24 deletions(-)

-- 
2.44.0