Ceph » CephFS

Although actually based on how long fsstress is taking on this disk maybe nothing was blocked and it was just going slow. Ugh.

I managed to reproduce this on my mds_rename branch.

Well, it's a nested auth pin.

gregf@kai:~/logs/fsstress/freeze_not_finishing2$ grep -o "dir(10000000002) adjust_nested_auth_pins [-0-9]*/[-0-9]*" out/mds.as | sort | uniq -c
    658 dir(10000000002) adjust_nested_auth_pins -1/-1
    200 dir(10000000002) adjust_nested_auth_pins -1/0
    201 dir(10000000002) adjust_nested_auth_pins 1/0
    658 dir(10000000002) adjust_nested_auth_pins 1/1

Further analysis forthcoming, when I figure out a good way to do it...

If you can reproduce, you can enable the auth pin set define in mdstypes.h, which tracks who the pinners are.

//#define MDS_AUTHPIN_SET // define me for debugging auth pin leaks

Unfortunately, adjust_nested_auth_pins never sees the person who actually grabbed it. The others print out the grabbing pointer so it's easy enough to grep | sort | uniq to figure out who's taking and not dropping even without the set being kept in-memory (assuming you've got debugging cranked up).