I sent an initial RFC set for this around 10 months ago. Since then,
the requisite patches for the MDS have been merged for the octopus
release. This adds support to the kclient to take advantage of
asynchronous unlinks.
In earlier testing (with a vstart cluster backed by a rotating HDD), I
saw roughly a 2x speedup when doing an rmdir on a directory with 10000
files in it. When testing with a cluster backed by an NVMe SSD though,
I only saw about a 20% speedup.
I'd like to put this in the testing branch now, so that it's ready for
merge in the upcoming v5.6 merge window. Once this is in, asynchronous
create support will be the next step.
Jeff Layton (4):
ceph: close holes in struct ceph_mds_session
ceph: hold extra reference to r_parent over life of request
ceph: register MDS request with dir inode from the start
ceph: add refcounting for Fx caps
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/caps.c | 84 ++++++++++++++++++++++++++----------
fs/ceph/dir.c | 70 ++++++++++++++++++++++++++++--
fs/ceph/inode.c | 9 +++-
fs/ceph/mds_client.c | 27 ++++++------
fs/ceph/mds_client.h | 2 +-
fs/ceph/super.c | 4 ++
fs/ceph/super.h | 17 +++-----
include/linux/ceph/ceph_fs.h | 9 ++++
8 files changed, 169 insertions(+), 53 deletions(-)
--
2.24.1
