On Thu, 2020-01-23 at 01:04 +0800, Yan, Zheng wrote: > On Wed, Jan 22, 2020 at 3:31 AM Jeff Layton <jlayton@xxxxxxxxxx> wrote: > > v3: > > - move some cephfs-specific code into ceph.ko > > - present and track inode numbers as u64 values > > - fix up check for dentry and cap eligibility checks > > - set O_CEPH_EXCL on async creates > > - attempt to handle errors better on async create (invalidate dentries > > and dir completeness). > > - ensure that fsync waits for async create to complete > > > > v2: > > - move cached layout to dedicated field in inode > > - protect cached layout with i_ceph_lock > > - wipe cached layout in __check_cap_issue > > - set max_size of file to layout.stripe_unit > > - set truncate_size to (u64)-1 > > - use dedicated CephFS feature bit instead of CEPHFS_FEATURE_OCTOPUS > > - set cap_id to 1 in async created inode > > - allocate inode number before submitting request > > - rework the prep for an async create to be more efficient > > - don't allow MDS or cap messages involving an inode until we get async > > create reply > > > > Still not quite ready for merge, but I've cleaned up a number of warts > > in the v2 set. Performance numbers still look about the same. > > > > There is definitely still a race of some sort that causes the client to > > try to asynchronously create a dentry that already exists. I'm still > > working on tracking that down. > > > > Jeff Layton (10): > > ceph: move net/ceph/ceph_fs.c to fs/ceph/util.c > > ceph: make ceph_fill_inode non-static > > ceph: make dentry_lease_is_valid non-static > > ceph: make __take_cap_refs non-static > > ceph: decode interval_sets for delegated inos > > ceph: add flag to designate that a request is asynchronous > > ceph: add infrastructure for waiting for async create to complete > > ceph: add new MDS req field to hold delegated inode number > > ceph: cache layout in parent dir on first sync create > > ceph: attempt to do async create when possible > > > > fs/ceph/Makefile | 2 +- > > fs/ceph/caps.c | 38 +++-- > > fs/ceph/dir.c | 13 +- > > fs/ceph/file.c | 240 +++++++++++++++++++++++++-- > > fs/ceph/inode.c | 50 +++--- > > fs/ceph/mds_client.c | 123 ++++++++++++-- > > fs/ceph/mds_client.h | 17 +- > > fs/ceph/super.h | 16 +- > > net/ceph/ceph_fs.c => fs/ceph/util.c | 4 - > > include/linux/ceph/ceph_fs.h | 8 +- > > net/ceph/Makefile | 2 +- > > 11 files changed, 443 insertions(+), 70 deletions(-) > > rename net/ceph/ceph_fs.c => fs/ceph/util.c (94%) > > > > -- > > 2.24.1 > > > > I realized that there still are two issues: > - we needs to clear delegated inos when mds failover I'm guessing we need to do this whenever any session is reconnected, right? I think we may have bigger problems here though: The issue is that with this set we assign out ino_t's prior to submitting the request. We could get down into it and find that we're reconnecting the session. At that point, that ino_t is no longer valid for the session. > - we needs to clear caps for async dir operations when reconnecting to > mds. (see last commit of https://github.com/ceph/ceph/pull/32576) > I guess we can use ceph_iterate_session_caps to do this. That said, looking at your patch: if (in->is_dir()) { // remove caps for async dir operations cap.implemented &= (CEPH_CAP_ANY_SHARED | CEPH_CAP_ANY_EXCL); } We remove all but Fsx here. That seems like quite a subtle difference from how FILE caps work. Given the way that we're doing the inode delegation and handling dir caps, I think we may need to rework things such that async requests never get queued to the s_waiting list, and instead return some sort of distinct error that cues the caller to fall back to a synchronous request. That would help prevent some of the potential races here. -- Jeff Layton <jlayton@xxxxxxxxxx>
