I sent a preliminary patchset for this back in April, which relied
on a totally hacked-up MDS patchset. Since then, Zheng has modified
the approach somewhat to make the MDS grant the client explicit
capabilities for asynchronous directory operations.
This patchset is an updated version of the earlier set. With this,
and the companion MDS set in play, removing a directory with a large
number of files in it is roughly twice as fast as doing it
synchronously.
In addition this set includes some new tracepoints that allow the
admin to better view what's happening with caps. They're mostly
limited to unlink and cap handling here, but I expect we'll add
more of those as time goes on.
I don't think we'll want to merge this just yet, until the MDS
support is merged. Once that goes in, and assuming we don't have
any changes to the client/MDS interface, we should clear to do
so.
Jeff Layton (7):
ceph: make several helper accessors take const pointers
ceph: hold extra reference to r_parent over life of request
ceph: register MDS request with dir inode from the get-go
ceph: add refcounting for Fx caps
ceph: wait for async dir ops to complete before doing synchronous dir
ops
ceph: new tracepoints when adding and removing caps
ceph: add tracepoints for async and sync unlink
Yan, Zheng (2):
ceph: check inode type for CEPH_CAP_FILE_{CACHE,RD,REXTEND,LAZYIO}
ceph: perform asynchronous unlink if we have sufficient caps
fs/ceph/Makefile | 3 +-
fs/ceph/caps.c | 88 +++++++++++++++++------
fs/ceph/dir.c | 121 ++++++++++++++++++++++++++++++--
fs/ceph/file.c | 4 ++
fs/ceph/inode.c | 9 ++-
fs/ceph/mds_client.c | 27 +++----
fs/ceph/super.h | 28 ++++----
fs/ceph/trace.c | 76 ++++++++++++++++++++
fs/ceph/trace.h | 86 +++++++++++++++++++++++
include/linux/ceph/ceph_debug.h | 1 +
include/linux/ceph/ceph_fs.h | 9 +++
11 files changed, 393 insertions(+), 59 deletions(-)
create mode 100644 fs/ceph/trace.c
create mode 100644 fs/ceph/trace.h
--
2.21.0
