Hi Igor I've sent you these directly as they're a bit chunky. Let me know if you haven't got them. Thx, Chris On 20/05/2020 14:43, Igor Fedotov wrote:

Short update on the issue: Finally we're able to reproduce the issue in master (not octopus), investigating further.. @Chris - to make sure you're facing the same issue could you please check the content of the broken file. To do so: 1) run "ceph-bluestore-tool --path <path-to-osd> --our-dir <target dir> --command bluefs-export This will export bluefs files to <target dir> 2) Check the content for file db.wal/002040.log at offset 0x470000 This will presumably contain 32K of zero bytes. Is this the case? No hurry as I'm just making sure symptoms in Octopus are the same... Thanks, Igor On 5/20/2020 5:24 PM, Igor Fedotov wrote: > Chris, > > got them, thanks! > > Investigating.... > > > Thanks, > > Igor > > On 5/20/2020 5:23 PM, Chris Palmer wrote: >> Hi Igor >> I've sent you these directly as they're a bit chunky. Let me know if >> you haven't got them. >> Thx, Chris >> >> On 20/05/2020 14:43, Igor Fedotov wrote: >>> Hi Cris, >>> >>> could you please share the full log prior to the first failure? >>> >>> Also if possible please set debug-bluestore/debug bluefs to 20 and >>> collect another one for failed OSD startup. >>> >>> >>> Thanks, >>> >>> Igor >>> >>> >>> On 5/20/2020 4:39 PM, Chris Palmer wrote: >>>> I'm getting similar errors after rebooting a node. Cluster was >>>> upgraded 15.2.1 -> 15.2.2 yesterday. No problems after rebooting >>>> during upgrade. >>>> >>>> On the node I just rebooted, 2/4 OSDs won't restart. Similar logs >>>> from both. Logs from one below. >>>> Neither OSDs have compression enabled, although there is a >>>> compression-related error in the log. >>>> Both are replicated x3. One has data on HDD & separate WAL/DB on >>>> NVMe partition, the other is everything on NVMe partition only. >>>> >>>> Feeling kinda nervous here - advice welcomed!! >>>> >>>> Thx, Chris >>>> >>>> >>>> >>>> 2020-05-20T13:14:00.837+0100 7f2e0d273700  3 rocksdb: >>>> [table/block_based_table_reader.cc:1117] Encountered error while >>>> reading data from compression dictionary block Corruption: block >>>> checksum mismatch: expected 0, got 3423870535  in db/000304.sst >>>> offset 18446744073709551615 size 18446744073709551615 >>>> 2020-05-20T13:14:00.841+0100 7f2e1957ee00  4 rocksdb: >>>> [db/version_set.cc:3757] Recovered from manifest >>>> file:db/MANIFEST-000312 succeeded,manifest_file_number is 312, >>>> next_file_number is 314, last_sequence is 22320582, log_number is >>>> 309,prev_log_number is 0,max_column_family is >>>> 0,min_log_number_to_keep is 0 >>>> >>>> 2020-05-20T13:14:00.841+0100 7f2e1957ee00  4 rocksdb: >>>> [db/version_set.cc:3766] Column family [default] (ID 0), log number >>>> is 309 >>>> >>>> 2020-05-20T13:14:00.841+0100 7f2e1957ee00  4 rocksdb: EVENT_LOG_v1 >>>> {"time_micros": 1589976840843199, "job": 1, "event": >>>> "recovery_started", "log_files": [313]} >>>> 2020-05-20T13:14:00.841+0100 7f2e1957ee00  4 rocksdb: >>>> [db/db_impl_open.cc:583] Recovering log #313 mode 0 >>>> 2020-05-20T13:14:00.937+0100 7f2e1957ee00  3 rocksdb: >>>> [db/db_impl_open.cc:518] db.wal/000313.log: dropping 9044 bytes; >>>> Corruption: error in middle of record >>>> 2020-05-20T13:14:00.937+0100 7f2e1957ee00  3 rocksdb: >>>> [db/db_impl_open.cc:518] db.wal/000313.log: dropping 86 bytes; >>>> Corruption: missing start of fragmented record(2) >>>> 2020-05-20T13:14:00.937+0100 7f2e1957ee00  4 rocksdb: >>>> [db/db_impl.cc:390] Shutdown: canceling all background work >>>> 2020-05-20T13:14:00.937+0100 7f2e1957ee00  4 rocksdb: >>>> [db/db_impl.cc:563] Shutdown complete >>>> 2020-05-20T13:14:00.937+0100 7f2e1957ee00 -1 rocksdb: Corruption: >>>> error in middle of record >>>> 2020-05-20T13:14:00.937+0100 7f2e1957ee00 -1 >>>> bluestore(/var/lib/ceph/osd/ceph-9) _open_db erroring opening db: >>>> 2020-05-20T13:14:00.937+0100 7f2e1957ee00  1 bluefs umount >>>> 2020-05-20T13:14:00.937+0100 7f2e1957ee00  1 fbmap_alloc >>>> 0x55daf2b3a900 shutdown >>>> 2020-05-20T13:14:00.937+0100 7f2e1957ee00  1 bdev(0x55daf3838700 >>>> /var/lib/ceph/osd/ceph-9/block) close >>>> 2020-05-20T13:14:01.093+0100 7f2e1957ee00  1 bdev(0x55daf3838000 >>>> /var/lib/ceph/osd/ceph-9/block) close >>>> 2020-05-20T13:14:01.341+0100 7f2e1957ee00 -1 osd.9 0 OSD:init: >>>> unable to mount object store >>>> 2020-05-20T13:14:01.341+0100 7f2e1957ee00 -1 ESC[0;31m ** ERROR: >>>> osd init failed: (5) Input/output errorESC[0m >>>> _______________________________________________ >>>> ceph-users mailing list -- ceph-users(a)ceph.io >>>> To unsubscribe send an email to ceph-users-leave(a)ceph.io >> > _______________________________________________ > ceph-users mailing list -- ceph-users(a)ceph.io > To unsubscribe send an email to ceph-users-leave(a)ceph.io

Thanks Igor, Do you have any idea on a e.t.a or plan for people that are running 15.2.2 to be able to patch / fix the issue. I had a read of the ticket and seems the corruption is happening but the WAL is not read till OSD restart, so I imagine we will need some form of fix / patch we can apply to a running OSD before we then restart the OSD, as a normal OSD upgrade will require the OSD to restart to apply the code resulting in a corrupt OSD. Thanks ---- On Sat, 23 May 2020 00:12:59 +0800 Igor Fedotov &lt;ifedotov(a)suse.de&gt; wrote ---- Status update: Finally we have the first patch to fix the issue in master: https://github.com/ceph/ceph/pull/35201 And ticket has been updated with root cause analysis:https://tracker.ceph.com/issues/45613On 5/21/2020 2:07 PM, Igor Fedotov wrote: @Chris - unfortunately it looks like the corruption is permanent since  valid WAL data are presumably overwritten with another stuff. Hence I don't know any way to recover - perhaps you can try cutting WAL file off which will allow OSD to start. With some latest ops lost. Once can use exported BlueFS as a drop in replacement for regular DB volume but I'm not aware of details. And the above are just speculations, can't say for sure if it helps... I can't explain why WAL doesn't have zero block in your case though. Little chances this is a different issue. Just in case - could you please search for 32K zero blocks over the whole file? And the same for another OSD? Thanks, Igor _______________________________________________ ceph-users mailing list -- mailto:ceph-users@ceph.io To unsubscribe send an email to mailto:ceph-users-leave@ceph.io

Hello Chris, Great to hear, few questions. Once you have injected the bluefs_preextend_wal_files to false, are you just rebuilding the OSD's that failed? Or are you going through and rebuilding every OSD even the working one's? Or does setting the bluefs_preextend_wal_files value to false and leaving the OSD running fix the WAL automatically? Thanks ---- On Sat, 23 May 2020 15:53:42 +0800 Chris Palmer &lt;chris.palmer(a)pobox.com&gt; wrote ---- Hi Ashley Igor has done a great job of tracking down the problem, and we have finally shown evidence of the type of corruption it would produce in one of my WALs. Our feeling at the moment is that the problem can be detoured by setting bluefs_preextend_wal_files to false on affected OSDs while they are running (but see below), although Igor does note that there is a small risk in doing this. I've agreed a plan of action based on this route, recreating the failed OSDs, and then cycling through the others until all are healthy. I've started this now, and so far it looks promising, although of course I have to wait for recovery/rebalancing. This is the fastest route to recovery, although there other options. I'll post as it progresses. The good news seems to be that there shouldn't be any actual data corruption or loss, providing that this can be done before OSDs are taken down (other than as part of this process). My understanding is that there will some degree of performance penalty until the root cause is fixed in the next release and preextending can be turned back on. However it does seem like I can get back to a stable/safe position without waiting for a software release. I'm just working through this at the moment though, so please don't take the above as any form of recommendation. It is important not to try to restart OSDs though in the meantime. I'm sure Igor will publish some more expert recommendations in due course... Regards, Chris On 23/05/2020 06:54, Ashley Merrick wrote: Thanks Igor, Do you have any idea on a e.t.a or plan for people that are running 15.2.2 to be able to patch / fix the issue. I had a read of the ticket and seems the corruption is happening but the WAL is not read till OSD restart, so I imagine we will need some form of fix / patch we can apply to a running OSD before we then restart the OSD, as a normal OSD upgrade will require the OSD to restart to apply the code resulting in a corrupt OSD. Thanks ---- On Sat, 23 May 2020 00:12:59 +0800 Igor Fedotov mailto:ifedotov@suse.de wrote ---- Status update: Finally we have the first patch to fix the issue in master: https://github.com/ceph/ceph/pull/35201 And ticket has been updated with root cause analysis:https://tracker.ceph.com/issues/45613On 5/21/2020 2:07 PM, Igor Fedotov wrote: @Chris - unfortunately it looks like the corruption is permanent since  valid WAL data are presumably overwritten with another stuff. Hence I don't know any way to recover - perhaps you can try cutting WAL file off which will allow OSD to start. With some latest ops lost. Once can use exported BlueFS as a drop in replacement for regular DB volume but I'm not aware of details. And the above are just speculations, can't say for sure if it helps... I can't explain why WAL doesn't have zero block in your case though. Little chances this is a different issue. Just in case - could you please search for 32K zero blocks over the whole file? And the same for another OSD? Thanks, Igor (not octopus), could you please <path-to-osd> --our-dir <target offset 0x470000 this the case? are the same... chunky. Let me know if to the first failure? debug-bluestore/debug bluefs to 20 and startup. rebooting a node. Cluster was yesterday. No problems after rebooting won't restart. Similar logs enabled, although there is a on HDD & separate WAL/DB on everything on NVMe partition only. welcomed!! 7f2e0d273700  3 rocksdb: [table/block_based_table_reader.cc:1117] Encountered error while dictionary block Corruption: block 3423870535  in db/000304.sst 18446744073709551615 7f2e1957ee00  4 rocksdb: from manifest last_sequence is 22320582, log_number is 0,max_column_family is >>>>> 0,min_log_number_to_keep is 0 >>>>> 7f2e1957ee00  4 rocksdb: family [default] (ID 0), log >>>>> number is 309 >>>>> 7f2e1957ee00  4 rocksdb: EVENT_LOG_v1 "job": 1, "event": [313]} 7f2e1957ee00  4 rocksdb: log #313 mode 0 7f2e1957ee00  3 rocksdb: db.wal/000313.log: dropping 9044 bytes; >>>>> Corruption: error in middle of record 7f2e1957ee00  3 rocksdb: db.wal/000313.log: dropping 86 bytes; fragmented record(2) 7f2e1957ee00  4 rocksdb: canceling all background work 7f2e1957ee00  4 rocksdb: >>>>> [db/db_impl.cc:563] Shutdown complete 7f2e1957ee00 -1 rocksdb: Corruption: >>>>> error in middle of record 7f2e1957ee00 -1 _open_db erroring opening db: 7f2e1957ee00  1 bluefs umount 7f2e1957ee00  1 fbmap_alloc >>>>> 0x55daf2b3a900 shutdown 7f2e1957ee00  1 bdev(0x55daf3838700 7f2e1957ee00  1 bdev(0x55daf3838000 7f2e1957ee00 -1 osd.9 0 OSD:init: 7f2e1957ee00 -1 ESC[0;31m ** ERROR: errorESC[0m _______________________________________________ _______________________________________________ ceph-users mailing list -- mailto:ceph-users@ceph.io To unsubscribe send an email to mailto:ceph-users-leave@ceph.io

Status date: We seem to have success. I followed the steps below. Only one more OSD (on node3) failed to restart, showing the same WAL corruption messages. After replacing that & backfilling I could then restart it. So we have a healthy cluster with restartable OSDs again, with bluefs_preextend_wal_files=false until its deemed safe to re-enable it. Many thanks Igor! Regards, Chris On 23/05/2020 11:06, Chris Palmer wrote:

Hi Ashley The command to reset the flag for ALL OSDs is       ceph config set osd bluefs_preextend_wal_files false And for just an individual OSD:       ceph config set osd.5 bluefs_preextend_wal_files false And to remove it from an individual one (so you just have the global one left):       ceph config rm osd.5 bluefs_preextend_wal_files BUT: I can't stress enough how important it is to only take down ONE OSD AT A TIME. And not to take any others down until that one is properly back up (replaced and backfilled if necessary). *Rebooting nodes without doing this may very well cause irretrievable data loss, no matter how long it has been since you reset that parameter.* This all seems to have worked for me but you should get expert advice. Regards, Chris On 23/05/2020 16:32, Ashley Merrick wrote: