[PATCH 0/2] NFSv4/pnfs: Keep data server connections across deviceid eviction

[Benjamin Coddington <ben.coddington@hammerspace.com>]

This applies on top of the flexfiles striping series:

  [PATCH v4 00/24] NFS: flexfiles device notifications and caching for
  wide striped layouts
  https://lore.kernel.org/linux-nfs/cover.1789474702.git.bcodding@hammerspace.com

and uses the hashed, address-keyed data server cache that series adds.

A metadata server that does not support deviceid notifications makes the
client set NFS_DEVICEID_NOCACHE, and a NOCACHE deviceid node is freed as
soon as the last layout referring to it is returned.  That takes the data
server's nfs_client and every transport underneath it along with it:

  nfs4_put_deviceid_node()
    ff_layout_free_deviceid_node()
      nfs4_pnfs_ds_put()
        destroy_ds()
          nfs_put_client()
            rpc_shutdown_client()

The next I/O rebuilds all of it: a synchronous connect and NULL ping
inline in pg_init(), plus the remaining nconnect-1 transports
reconnecting lazily as RPCs land on them.  I have a report of a wide
stripe at nconnect=16 spending upwards of ten seconds on this before I/O
can proceed, and paying it again on every layout cycle.

Keep the nfs4_pnfs_ds hashed once its last user goes away, and reap it later
if nothing claims it.

Benjamin Coddington (2):
  NFSv4/pnfs: Give the data server cache its own reference
  NFSv4/pnfs: Let unused data server connections linger

 fs/nfs/pnfs.c     |   4 +-
 fs/nfs/pnfs.h     |   4 ++
 fs/nfs/pnfs_nfs.c | 110 +++++++++++++++++++++++++++++++++++++++++++---
 3 files changed, 112 insertions(+), 6 deletions(-)

-- 
2.53.0