[PATCH 0/2] ceph: lazy I/O write and LAZYIO revoke fixes

[Xiubo Li via B4 Relay]

Two lazy I/O changes on top of the lazy I/O series already in testing.

Patch 1 stops a buffered lazy I/O write into part of a folio that is
not uptodate from reading the folio in first.  That read is done
synchronously under i_rwsem and serialises every writer of a shared
file on the client; with lazy I/O only the dirty range of the folio
goes out, so the read buys nothing.  16 writers on 4 clients writing
47008-byte records into one shared file go from 0.27 to 0.71 GiB/s.

Patch 2 makes ceph_adjust_caps_used_for_lazyio() move a use to LAZYIO
only when that use is in the used set and its cap is not issued.
Otherwise a file with only clean cached pages holds a LAZYIO revoke
until the delayed cap check, up to 60 s, and a getattr from another
client waits that long.

Signed-off-by: Xiubo Li <xiubo.li@xxxxxxxxx>
---
Xiubo Li (2):
      ceph: don't read a folio in for a partial lazy I/O write
      ceph: do not hold a LAZYIO revoke for pages that CACHE covers

 fs/ceph/addr.c | 163 ++++++++++++++++++++++++++++++++++++++++++++++++++++-----
 fs/ceph/caps.c |  11 +++-
 2 files changed, 159 insertions(+), 15 deletions(-)
---
base-commit: 5dcec86cc31f8144857c0d9bf2a7fa13b6e6f531
change-id: 20261107-ceph-lazyio-noread-087d6b8ccc9d

Best regards,
--  
Xiubo Li <xiubo.li@xxxxxxxxx>