v2: fix a hang that would occasionally occur under fsstress This is the second version of this patchset, and is a bit more of an overhaul. I revised the set because I was seeing fsstress hangs. Zheng mentioned that he had gotten a report that testing with vdbench on a v4.8 kernel showed a large increase in the number of GETATTR calls being made. Since we had changed the d_revalidate code to use a GETATTR call and not set r_locked_dir in the request, the dentry lease wasn't getting updated from the traces, even when the MDS had granted one. We need to use a LOOKUP when revalidating dentries so we can update the dentry lease if things look ok. That currently depends however on having r_locked_dir set to the parent directory. In d_revalidate though, we may very well not hold the parent's i_rwsem. Manipulating the dcache is not allowed without it, but we can update the inode and the dentry leases. The basic approach here is to separate r_locked_dir into an inode pointer representing the parent, and a flag indicating whether we know that it's locked (note that it _may_ still be locked even if that flag isn't set). In ceph_fill_trace, we can then do all of the parts that don't require a locked parent whenever the parent pointer is set, but avoid the dcache manipulation when it isn't. With this, dentry leases are again being updated as a result of d_revalidate lookups. The switch to using a flags fields instead of bools also shrinks the request size slightly. If everyone hates that, I can just plop in another bool, but I think this gives us more room to expand with other flags as needed. Jeff Layton (13): ceph: add new r_req_flags field to ceph_mds_request ceph: move r_aborted flag into r_req_flags ceph: move r_got_unsafe flag into r_req_flags ceph: move r_got_safe flag into r_req_flags ceph: move r_got_result into r_req_flags ceph: move r_did_prepopulate into r_req_flags ceph: remove "Debugging hook" from ceph_fill_trace ceph: don't need session pointer to ceph_fill_trace ceph: add a new flag to indicate whether parent is locked ceph: don't update_dentry_lease unless we actually got one ceph: vet the parent inode before updating dentry lease ceph: call update_dentry_lease even when r_locked dir is not set ceph: do a LOOKUP in d_revalidate instead of GETATTR fs/ceph/debugfs.c | 2 +- fs/ceph/dir.c | 30 +++++++++------ fs/ceph/export.c | 3 +- fs/ceph/file.c | 3 +- fs/ceph/inode.c | 104 ++++++++++++++++++++++----------------------------- fs/ceph/mds_client.c | 73 +++++++++++++++++++----------------- fs/ceph/mds_client.h | 15 +++++--- fs/ceph/super.h | 3 +- 8 files changed, 118 insertions(+), 115 deletions(-) -- 2.9.3 -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
