On Thu, 25 May 2017, Łukasz Chrustek wrote: > Cześć, > > > On Thu, 25 May 2017, Łukasz Chrustek wrote: > >> Cześć, > >> > >> > On Wed, 24 May 2017, Łukasz Chrustek wrote: > >> >> Hello, > >> >> > >> >> >> > >> >> >> > This > >> >> >> > >> >> >> osd 6 - isn't startable > >> >> > >> >> > Disk completely 100% dead, or just borken enough that ceph-osd won't > >> >> > start? ceph-objectstore-tool can be used to extract a copy of the 2 pgs > >> >> > from this osd to recover any important writes on that osd. > >> >> > >> >> >> osd 10, 37, 72 are startable > >> >> > >> >> > With those started, I'd repeat the original sequence and get a fresh pg > >> >> > query to confirm that it still wants just osd.6. > >> >> > >> >> > use ceph-objectstore-tool to export the pg from osd.6, stop some other > >> >> > ranodm osd (not one of these ones), import the pg into that osd, and start > >> >> > again. once it is up, 'ceph osd lost 6'. the pg *should* peer at that > >> >> > point. repeat with the same basic process with the other pg. > >> >> > >> >> Here is output from ceph-objectstore-tool - also didn't success: > >> >> > >> >> https://pastebin.com/7XGAHdKH > >> > >> > Hmm, btrfs: > >> > >> > 2017-05-24 23:28:58.547456 7f500948e940 -1 > >> > filestore(/var/lib/ceph/osd/ceph-84) ERROR: > >> > /var/lib/ceph/osd/ceph-84/current/nosnap exists, not rolling back to avoid > >> > losing new data > >> > >> > You could try setting --osd-use-stale-snap as suggested. > >> > >> Yes... tried... and I simply get rided of 39GB data... > > > What does "get rided" mean? > > according to this pastebin: https://pastebin.com/QPcpkjg4 > > ls -R /var/lib/ceph/osd/ceph-33/current/ > > /var/lib/ceph/osd/ceph-33/current/: > > commit_op_seq omap > > > > /var/lib/ceph/osd/ceph-33/current/omap: > > 000003.log CURRENT LOCK MANIFEST-000002 > > earlier there were same data files. Yeah, looks like all the data was deleted from the device. :( > >> > >> > Is it the same error with the other one? > >> > >> Yes: https://pastebin.com/7XGAHdKH > >> > >> > >> > >> > >> > in particular, osd 37 38 48 67 all have incomplete copies of the PG (they > >> > are mid-backfill) and 68 has nothing. Some data is lost unless you can > >> > recovery another OSD with that PG. > >> > >> > The set of OSDs that might have data are: 6,10,33,72,84 > >> > >> > If that bears no fruit, then you can force last_backfill to report > >> complete on one of those OSDs and it'll think it has all the data even > >> though some of it is likely gone. (We can pick one that is farther > >> along... 38 48 and 67 seem to all match. > > Can You explain what do You mean by 'force last_backfill to report > complete' ? The current value for PG 1.60 is MAX and for 1.165 is > 1\/db616165\/rbd_data.ed9979641a9d82.000000000001dcee\/head ceph-objectstore-tool has a mark-complete operation. Do that one one of the OSDs that has the more advanced last_backfill (like the one above). After you restart the PG should recover. Good luck! sage
