Luis Henriques lhenriques@suse.com writes:
"Yan, Zheng" ukernel@gmail.com writes:
On Tue, Mar 19, 2019 at 12:22 AM Luis Henriques lhenriques@suse.com wrote:
"Yan, Zheng" ukernel@gmail.com writes:
On Mon, Mar 18, 2019 at 6:33 PM Luis Henriques lhenriques@suse.com wrote:
"Yan, Zheng" ukernel@gmail.com writes:
On Fri, Mar 15, 2019 at 7:13 PM Luis Henriques lhenriques@suse.com wrote: > > I'm occasionally seeing a kmemleak warning in xfstest generic/013: > > unreferenced object 0xffff8881fccca940 (size 32): > comm "kworker/0:1", pid 12, jiffies 4295005883 (age 130.648s) > hex dump (first 32 bytes): > 01 00 00 00 00 00 00 00 01 00 00 00 00 00 00 00 ................ > 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ > backtrace: > [<00000000d741a1ea>] build_snap_context+0x5b/0x2a0 > [<0000000021a00533>] rebuild_snap_realms+0x27/0x90 > [<00000000ac538600>] rebuild_snap_realms+0x42/0x90 > [<000000000e955fac>] ceph_update_snap_trace+0x2ee/0x610 > [<00000000a9550416>] ceph_handle_snap+0x317/0x5f3 > [<00000000fc287b83>] dispatch+0x362/0x176c > [<00000000a312c741>] ceph_con_workfn+0x9ce/0x2cf0 > [<000000004168e3a9>] process_one_work+0x1d4/0x400 > [<000000002188e9e7>] worker_thread+0x2d/0x3c0 > [<00000000b593e4b3>] kthread+0x112/0x130 > [<00000000a8587dca>] ret_from_fork+0x35/0x40 > [<00000000ba1c9c1d>] 0xffffffffffffffff > > It looks like it is possible that we miss a flush_ack from the MDS when, > for example, umounting the filesystem. In that case, we can simply drop > the reference to the ceph_snap_context obtained in ceph_queue_cap_snap(). > > Link: https://tracker.ceph.com/issues/38224 > Cc: stable@vger.kernel.org > Signed-off-by: Luis Henriques lhenriques@suse.com > --- > fs/ceph/caps.c | 7 +++++++ > 1 file changed, 7 insertions(+) > > diff --git a/fs/ceph/caps.c b/fs/ceph/caps.c > index 36a8dc699448..208f4dc6f574 100644 > --- a/fs/ceph/caps.c > +++ b/fs/ceph/caps.c > @@ -1054,6 +1054,7 @@ int ceph_is_any_caps(struct inode *inode) > static void drop_inode_snap_realm(struct ceph_inode_info *ci) > { > struct ceph_snap_realm *realm = ci->i_snap_realm; > + > spin_lock(&realm->inodes_with_caps_lock); > list_del_init(&ci->i_snap_realm_item); > ci->i_snap_realm_counter++; > @@ -1063,6 +1064,12 @@ static void drop_inode_snap_realm(struct ceph_inode_info *ci) > spin_unlock(&realm->inodes_with_caps_lock); > ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc, > realm); > + /* > + * ci->i_head_snapc should be NULL, but we may still be waiting for a > + * flush_ack from the MDS. In that case, we still hold a ref for the > + * ceph_snap_context and we need to drop it. > + */ > + ceph_put_snap_context(ci->i_head_snapc); > } > > /*
This does not seem right. i_head_snapc is cleared when (ci->i_wrbuffer_ref_head == 0 && ci->i_dirty_caps == 0 && ci->i_flushing_caps == 0) . Nothing do with dropping ci->i_snap_realm. Did you see 'reconnect denied' during the test? If you did, the fix should be in iterate_session_caps()
No, I didn't saw any 'reconnect denied' in the test. The test actually seems to execute fine, except from the memory leak.
It's very difficult to reproduce this issue, but last time I managed to get this memory leak to trigger I actually had some debugging code in drop_inode_snap_realm, something like:
if (ci->i_head_snapc) printk("i_head_snapc: 0x%px\n", ci->i_head_snapc);
please add code that prints i_wrbuffer_ref_head, i_dirty_caps, i_flushing_caps. and try reproducing it again.
Ok, it took me a few hours, but I managed to reproduce the bug, with those extra printks. All those values are set to 0 when the bug triggers (and i_head_snapc != NULL).
Thanks, which test triggers this bug?
That's generic/013. It usually triggers after a few hours of running it in a loop (I'm using a vstart cluster for that).
I searched that code, found we may fail to cleanup i_head_snap in two places. One is in ceph_queue_cap_snap, Another is in remove_session_caps_cb().
Ah, great! I spent a lot of time looking but I couldn't really find it. My bet was that ceph_queue_cap_snap was doing the ceph_get_snap_context and that the corresponding ceph_put_snap_context would be done in handle_cap_flush_ack. That's why I mentioned the missing flush_ack from MDS.
Something that I didn't said explicitly is that I *know* that the unbalanced ceph_get_snap_context() is the one in function ceph_queue_cap_snap(). I know this for sure because I've managed to reproduce the issue several times with an instrumented kernel (with tracepoints) that allowed me to keep track of all ceph_snap_context operations (create, get, put, delete).
Cheers,