Linux NFS development
 help / color / mirror / Atom feed
From: Benjamin Coddington <ben.coddington@hammerspace.com>
To: Trond Myklebust <trondmy@kernel.org>,
	Anna Schumaker <anna@kernel.org>,
	Mike Snitzer <snitzer@kernel.org>
Cc: linux-nfs@vger.kernel.org
Subject: [PATCH v2 0/5] pNFS/flexfiles: I/O without a layout under FF_FLAGS_NO_IO_THRU_MDS
Date: Thu, 19 Nov 2026 15:27:23 -0500	[thread overview]
Message-ID: <cover.1795117974.git.bcodding@hammerspace.com> (raw)

From: Benjamin Coddington <bcodding@hammerspace.com>

When a file's flexfiles layouts carry FF_FLAGS_NO_IO_THRU_MDS, the client
keeps reads and writes off the MDS when it has no layout segment for
them.  Reads and writes handle that differently:

 - Reads test the inode's NFS_INO_NO_IO_THRU_MDS bit, wait a second
   between attempts, and honour io_maxretrans on soft mounts.

 - Writes test a bit on the layout header and return -EAGAIN.
   nfs_writepages() and the O_DIRECT reschedule path resend at once, so
   writers spin for as long as there is no layout.  nfs_wb_folio()
   returns the -EAGAIN, so a read(2) or write(2) that has to flush a
   folio another open context dirtied fails with EAGAIN.  Once the
   header is freed the bit is gone and writes go through the MDS, and
   ff_layout_pg_init_write() sends them there whenever
   pnfs_generic_pg_check_layout() drops an invalidated segment.

Neither direction uses the MDS when the server has said there is no
layout to give.  After NFS4ERR_LAYOUTUNAVAILABLE, RFC 8435 says the I/O
is to go to the MDS (section 5.1.1), and the flag is a hint (section
5.1.2).  Instead, reads wait and writes spin until the 120 second fail
bit expires, then send a LAYOUTGET that can fail the same way.

Patch 1 sends writeback of a single locked folio (nfs_wb_folio(),
reclaim, migration) through the MDS instead of refusing it.  Resending
with the folio locked could deadlock against a bulk recall, which can't
complete while nfs_writepages() holds a layout segment, and
nfs_writepages() would be waiting for that folio.  Patch 2 makes the
O_DIRECT write work a delayed_work.  Patch 3 has the remaining writes
test the inode bit.  After the folios and requests of a refused pass
are unlocked, data-integrity nfs_writepages() waits a second before it
resends, and the O_DIRECT reschedule queues its next pass a second
later; other writeback leaves the folios dirty.  Patch 4 removes the
header bit, which nothing tests after patch 3.  Patch 5 sends reads,
writes and the resend after a data server error through the MDS while
the fail bit for their iomode is set.

That last change is also the path for a data server the client can't
use, such as a version mismatch: the client reports the failure with
LAYOUTERROR or LAYOUTRETURN, and the server decides whether to answer
the next LAYOUTGET with NFS4ERR_LAYOUTUNAVAILABLE.  The missing-nfsv3
case is covered separately by Roberto's "NFSv4/flexfiles: couple back
with NFSv3".

Patch 3 replaces the header bit test that the first of them fixes.
These patches are independent of the series "pNFS: don't reply
NOMATCHING to a layout recall with I/O in flight", but both change
ff_layout_pg_get_mirror_count_write(), so whichever goes second will
need a small rebase.

Tested against a reffs MDS and DS with a six-way mirrored layout,
injecting LAYOUTGET errors after a file recall.  With
NFS4ERR_LAYOUTUNAVAILABLE, reads and writes go through the MDS at once,
and the data written through the MDS is on every mirror.  With a
retryable error (NFS4ERR_SERVERFAULT) held for 6 to 8 seconds, fsync
and the O_DIRECT reschedule retry once a second with no I/O to the MDS,
SIGKILL ends a writer blocked in fsync or an O_DIRECT write,
sync_file_range(SYNC_FILE_RANGE_WRITE) returns 0, and a soft mount with
io_maxretrans=3 returns EIO after about 4 seconds.  During the hold, a
read of a folio that another open context had dirtied flushes it
through the MDS and returns without EAGAIN.

The series is based on v7.4-rc3 with three flexfiles fixes that were
posted separately for stable:

  pNFS/flexfiles: read the layout hdr under i_lock in pg_get_mirror_count_write
  https://lore.kernel.org/linux-nfs/23435d398481f51ce2680053cd3a151fe673f91c.1795108831.git.bcodding@hammerspace.com/
  pNFS/flexfiles: don't delay or restart an rpc_task that LOCALIO never ran
  https://lore.kernel.org/linux-nfs/9c46fda0f2cd6d3843b198dac304b4722514b8e3.1795108831.git.bcodding@hammerspace.com/
  pNFS/flexfiles: restart a write that the DS error handler retries
  https://lore.kernel.org/linux-nfs/8ee107c9d0c68b6d0ee7305605ff7edbe0ffa160.1795108832.git.bcodding@hammerspace.com/

Changes since v1:
https://lore.kernel.org/linux-nfs/cover.1795024981.git.bcodding@hammerspace.com/
 - Rebased onto the three fixes above.
 - Patch 1: shorter pg_allow_mds comment; the nfs_wb_folio() comment
   says the refused folio was redirtied (Anna).
 - New patch 2, so that patch 3 no longer sleeps on nfsiod.
 - Patch 3 (was 2): only data-integrity nfs_writepages() waits, and a
   fatal signal ends the wait; other writeback returns 0 instead of
   waiting a second per folio.  The O_DIRECT reschedule requeues after a
   second instead of sleeping.  The tracepoints that read the layout
   header without an lseg are emitted under i_lock.

Benjamin Coddington (5):
  pNFS/flexfiles: use the MDS for locked-folio writeback without a
    layout
  NFS: make the O_DIRECT write work a delayed_work
  pNFS/flexfiles: pace writes that have no layout and can't use the MDS
  pNFS/flexfiles: remove the layout-header NO_IO_THRU_MDS bit
  pNFS/flexfiles: use the MDS when the server has no layout to give

 fs/nfs/direct.c                        |  11 +-
 fs/nfs/flexfilelayout/flexfilelayout.c | 138 ++++++++++++++++---------
 fs/nfs/flexfilelayout/flexfilelayout.h |  16 ---
 fs/nfs/internal.h                      |  14 +--
 fs/nfs/nfs4trace.c                     |   1 +
 fs/nfs/nfs4trace.h                     |   1 +
 fs/nfs/pagelist.c                      |   2 +
 fs/nfs/pnfs.c                          |  22 ++++
 fs/nfs/pnfs.h                          |   1 +
 fs/nfs/write.c                         |  31 +++++-
 include/linux/nfs_page.h               |   2 +
 11 files changed, 162 insertions(+), 77 deletions(-)


base-commit: aebb7274476403040d582c61478df8f4a7bf9cd3
prerequisite-patch-id: 71097d575f7ba74684462ab64b028ff8815dff63
prerequisite-patch-id: 5a73a1fb9d47a31d4ec337ec0264eadbcaa9fd4e
prerequisite-patch-id: 2663ab8e1949d6a11d957cc2784161bf2816cac2
-- 
2.53.0


             reply	other threads:[~2026-11-19 20:27 UTC|newest]

Thread overview: 6+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-11-19 20:27 Benjamin Coddington [this message]
2026-11-19 20:27 ` [PATCH v2 1/5] pNFS/flexfiles: use the MDS for locked-folio writeback without a layout Benjamin Coddington
2026-11-19 20:27 ` [PATCH v2 2/5] NFS: make the O_DIRECT write work a delayed_work Benjamin Coddington
2026-11-19 20:27 ` [PATCH v2 3/5] pNFS/flexfiles: pace writes that have no layout and can't use the MDS Benjamin Coddington
2026-11-19 20:27 ` [PATCH v2 4/5] pNFS/flexfiles: remove the layout-header NO_IO_THRU_MDS bit Benjamin Coddington
2026-11-19 20:27 ` [PATCH v2 5/5] pNFS/flexfiles: use the MDS when the server has no layout to give Benjamin Coddington

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=cover.1795117974.git.bcodding@hammerspace.com \
    --to=ben.coddington@hammerspace.com \
    --cc=anna@kernel.org \
    --cc=linux-nfs@vger.kernel.org \
    --cc=snitzer@kernel.org \
    --cc=trondmy@kernel.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox