一次ceph recovery经历
背景
- 这是一个測试环境。
- 该环境中是cephfs
- 一共12个节点， 2个client、2个mds、8个osd
- mds： 2颗CPU，每一个4核。一共是8核。 128G内存， 单独的两个节点，仅仅作为mds
cpu型号： Intel(R) Xeon(R) CPU E5-1620 v3 @ 3.50GHz osd节点， 每一个24核， 8 × 4T SATA盘， 1 SSD：INTEL SSD SC2BB48 (480G) 64G内存
cpu型号: Intel(R) Xeon(R) CPU E5-2620 v3 @ 2.40GHz
当中。有两个osd各有3块nvme SSD，3个nvme SSD，每一个分4个分区，两个journal，两个osd，一共是6个osd来做为metadata pool
測试任务
- 10亿个小文件（2M-4M）。终于我们选择了256K的文件， 由于为了达到文件数量， 仅仅能选择小文件， 否则容量不够。
- 到出现故障的时候， ceph cluster有7亿多的文件。cluster正常。
意外
- 由于測试环境的物理条件限制。温度过高。跳闸了。
悲剧发生，我们的raid卡用的cache没带电池，物理机开启后，8台osd节点，86个osd。一共同拥有40块左右磁盘都故障了， 无法mount。
磁盘恢复
问题： 磁盘文件系统损坏, mount不成功(error message: log mount/recovery failed: error -117); 即使mount成功, 进入挂载文件夹, ls会显示input/output error.
解决: 最初想使用例如以下的步骤来修复disk:
1. xfs_check /dev/sdb1
2. xfs_metadump /dev/sdb1 sdb1.meta
3. xfs_mdrestore sdb1.meta sdb1.img
4. xfs_repair sdb1.img结果, 在第二步就报错过不去, 所以仅仅能强制修复了:
root@host2:~# xfs_repair /dev/sdb1
Phase 1 - find and verify superblock...
Phase 2 - using internal log
- zero log...
ERROR: The filesystem has valuable metadata changes in a log which needs to
be replayed. Mount the filesystem to replay the log, and unmount it before
re-running xfs_repair. If you are unable to mount the filesystem, then use
the -L option to destroy the log and attempt a repair.
Note that destroying the log may cause corruption -- please attempt a mount
of the filesystem before doing this.
// -L选项 表示强制修复, 假设遇到dirty log, 也要强制设为0, 非常危急, easy丢数据, 由于是測试环境, 没办法, 仅仅能这样.
# xfs_repair -L /dev/sdb1- 当时断电后, 有3台机器, 一直能够ping通, 可是就是不能ssh连接, log例如以下:
[Thu Mar 30 12:05:00 2017] INFO: task mount:1984 blocked for more than 120 seconds.
[Thu Mar 30 12:05:00 2017] Not tainted 3.19.0-25-generic #26~14.04.1-Ubuntu
[Thu Mar 30 12:05:00 2017] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[Thu Mar 30 12:05:00 2017] mount D ffff881059927c28 0 1984 1889 0x00000000
[Thu Mar 30 12:05:00 2017] ffff881059927c28 ffff88085af16bf0 0000000000013e80 ffff881059927fd8
[Thu Mar 30 12:05:00 2017] 0000000000013e80 ffff88105c265850 ffff88085af16bf0 ffff88084b57c000
[Thu Mar 30 12:05:00 2017] ffff88085b302900 ffff88084b57c000 ffff88085b302940 ffff88085b302968
[Thu Mar 30 12:05:00 2017] Call Trace:
[Thu Mar 30 12:05:00 2017] [] schedule+0x29/0x70
[Thu Mar 30 12:05:00 2017] [] xfs_ail_push_all_sync+0xa9/0xe0 [xfs]
[Thu Mar 30 12:05:00 2017] [] ? prepare_to_wait_event+0x110/0x110
[Thu Mar 30 12:05:00 2017] [] xfs_log_quiesce+0x37/0x70 [xfs]
[Thu Mar 30 12:05:00 2017] [] xfs_log_unmount+0x1a/0x70 [xfs]
[Thu Mar 30 12:05:00 2017] [] xfs_mountfs+0x5e5/0x760 [xfs]
[Thu Mar 30 12:05:00 2017] [] xfs_fs_fill_super+0x2c6/0x340 [xfs]
[Thu Mar 30 12:05:00 2017] [] mount_bdev+0x1b0/0x1f0
[Thu Mar 30 12:05:00 2017] [] ? xfs_parseargs+0xbe0/0xbe0 [xfs] [Thu Mar 30 12:05:00 2017] [
copy_mount_options+0x3a/0x160 [Thu Mar 30 12:05:00 2017] [
copy_mount_options+0x3a/0x160 [Thu Mar 30 12:07:01 2017] [
可能是磁盘损坏, 而ceph-osd进程一直在试图mount, mount不成功, 堵塞, 操作系统发现进程超时, 仅仅能调整hung_task_timeout_secs,
我们再深入的研究这个问题, 但基本上断定: 这个时候系统的CPU忙于这些, 连ssh都不处理. 我们觉得这是操作系统的问题, 兴许须要细致研究一下这个问题.
最后我们仅仅能到现场或通过IPMI, 强制重新启动机器, 进入安全模式, 来修复损坏的磁盘.
之所以要进入安全模式, 是由于正常模式进入后, 还是上面的问题, 我们在其它能够ssh进入的机器上, 先禁止了ceph服务的自启动:
root@host2:~# update-rc.d ceph disable
Disabling system startup links for /etc/init.d/ceph ...
Removing any system startup links for /etc/init.d/ceph ...
/etc/rc0.d/K20ceph
/etc/rc1.d/K20ceph
/etc/rc2.d/S20ceph
/etc/rc3.d/S20ceph
/etc/rc4.d/S20ceph
/etc/rc5.d/S20ceph
/etc/rc6.d/K20ceph
Adding system startup for /etc/init.d/ceph ...
/etc/rc0.d/K20ceph -> ../init.d/ceph
/etc/rc1.d/K20ceph -> ../init.d/ceph
/etc/rc6.d/K20ceph -> ../init.d/ceph
/etc/rc2.d/K80ceph -> ../init.d/ceph
/etc/rc3.d/K80ceph -> ../init.d/ceph
/etc/rc4.d/K80ceph -> ../init.d/ceph
/etc/rc5.d/K80ceph -> ../init.d/ceph日志盘修复
日志盘的修复事实上就是更换日志盘, 见还有一篇文章: links
osd重新启动
启动osd服务, 可是有10个osd起不来, 接下来就逐个分析原因, 个个击破.
osd恢复
- 有两个osd的leveldb出错, 详细log没保留下来, 非常遗憾.
解决:
download leveldb python 模块, 能够使用pip, 能够下载源代码安装, 链接: https://pypi.python.org/pypi/leveldb
修复leveldb: RepairDB的參数是leveldb的详细存储路径, 对于osd, 就是current/omap
root@host3:~/leveldb-0.20# python
Python 2.7.6 (default, Oct 26 2016, 20:30:19)
[GCC 4.8.4] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import leveldb
>>>
>>>
>>> leveldb.RepairDB("/var/lib/ceph/osd/ceph-44/current/omap/")
>>>
接下来的几个问题一直没有解决, 最重要的是osd直接crash, 无从谈及兴许的pg recovery, 尽管尝试了非常多方法, 包含把问题从正常的sod中导出, 然后增加进来, 都不起作用. 好在这是一个測试环境, 时间紧张, 我们仅仅能先放下, 转而使用多cephfs的特性来继续測试cephfs, 兴许还要跟代码来看看问题所在.
2. osd log例如以下:
-7> 2017-03-31 12:27:34.025244 7fdaeb424800 15 filestore(/var/lib/ceph/osd/ceph-47) omap_get_values 1.1daa_head/#1:55b80000::::head#
-6> 2017-03-31 12:27:34.025350 7fdaeb424800 15 filestore(/var/lib/ceph/osd/ceph-47) omap_get_values 1.1daa_head/#1:55b80000::::head# = -1
-5> 2017-03-31 12:27:34.025356 7fdaeb424800 15 filestore(/var/lib/ceph/osd/ceph-47) collection_getattr /var/lib/ceph/osd/ceph-47/current/1.1daa_head 'remove' len 1
-4> 2017-03-31 12:27:34.025378 7fdaeb424800 10 filestore(/var/lib/ceph/osd/ceph-47) collection_getattr /var/lib/ceph/osd/ceph-47/current/1.1daa_head 'remove' len 1 = -61
-3> 2017-03-31 12:27:34.025381 7fdaeb424800 10 osd.47 8625 pgid 1.1daa coll 1.1daa_head
-2> 2017-03-31 12:27:34.025385 7fdaeb424800 15 filestore(/var/lib/ceph/osd/ceph-47) omap_get_values 1.1daa_head/#1:55b80000::::head#
-1> 2017-03-31 12:27:34.025447 7fdaeb424800 15 filestore(/var/lib/ceph/osd/ceph-47) omap_get_values 1.1daa_head/#1:55b80000::::head# = -1
0> 2017-03-31 12:27:34.027079 7fdaeb424800 -1 osd/PG.cc: In function 'static int PG::peek_map_epoch(ObjectStore*, spg_t, epoch_t*, ceph::bufferlist*)' thread 7fdaeb424800 time 2017-03-31 12:27:34.025452
osd/PG.cc: 2947: FAILED assert(0 == "unable to open pg metadata")
ceph version 10.2.6 (656b5b63ed7c43bd014bcafd81b001959d5f089f)
1: (ceph::__ceph_assert_fail(char const*, char const*, int, char const*)+0x8b) [0x7fdaeaeba26b]
2: (PG::peek_map_epoch(ObjectStore*, spg_t, unsigned int*, ceph::buffer::list*)+0x727) [0x7fdaea906c97]
3: (OSD::load_pgs()+0x8bd) [0x7fdaea86840d]
4: (OSD::init()+0x1f74) [0x7fdaea87a734]
5: (main()+0x29d1) [0x7fdaea7e1d71]
6: (__libc_start_main()+0xf5) [0x7fdae72a9ec5]
7: (()+0x36a957) [0x7fdaea82a957]
NOTE: a copy of the executable, or `objdump -rdS ` is needed to interpret this. 未解决……
- osd log例如以下:
-3> 2017-04-01 10:24:08.243324 7f10393eb700 0 log_channel(cluster) log [INF] : 1.1fb5 starting backfill to osd.51 from (0'0,0'0] MAX to 8625'202097
-2> 2017-04-01 10:24:08.244198 7f10393eb700 5 osd.65 pg_epoch: 14087 pg[1.1fb5( v 8625'202097 (8612'199066,8625'202097] local-les=14087 n=87794 ec=350 les/c/f 12937/12937/0 14082/14086/14086) [65,51]/[65,57] r=0 lpr=14086 pi=9325-14085/17 bft=51 crt=8625'202097 lcod 0'0 mlcod 0'0 activating+remapped] enter Started/Primary/Active/Activating
-1> 2017-04-01 10:24:08.244211 7f1036be6700 0 log_channel(cluster) log [INF] : 1.14 starting backfill to osd.51 from (0'0,0'0] MAX to 8625'201676
0> 2017-04-01 10:24:08.246532 7f1037be8700 -1 osd/PGLog.h: In function 'void PGLog::IndexedLog::claim_log_and_clear_rollback_info(const pg_log_t&)' thread 7f1037be8700 time 2017-04-01 10:24:08.243637
osd/PGLog.h: 111: FAILED assert(rollback_info_trimmed_to_riter == log.rbegin())
ceph version 10.2.6 (656b5b63ed7c43bd014bcafd81b001959d5f089f)
1: (ceph::__ceph_assert_fail(char const*, char const*, int, char const*)+0x8b) [0x7f10824d026b]
2: (PG::RecoveryState::Stray::react(PG::MLogRec const&)+0x63e) [0x7f1081f497ae]
3: (boost::statechart::simple_state::RecoveryState::Stray, PG::RecoveryState::Started, boost::mpl::list::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na>, (boost::statechart::history_mode)0>::react_impl(boost::statechart::event_base const&, void const*)+0x1f4) [0x7f1081f84044]
4: (boost::statechart::state_machine::RecoveryState::RecoveryMachine, PG::RecoveryState::Initial, std::allocator::CephPeeringEvt>, PG::RecoveryCtx*)+0x1d5) [0x7f1081f367b5]
6: (OSD::process_peering_events(std::list, std::allocator, std::allocator` is needed to interpret this. 未解决......
- osd log例如以下:
-3> 2017-04-01 09:07:03.256009 7f087f7e8700 1 osd.10 pg_epoch: 13637 pg[1.498( v 8625'201557 (8612'198458,8625'201557] local-les=13512 n=87583 ec=350 les/c/f 13512/13512/0 13636/13637/9027) [39,10] r=1 lpr=13637 pi=981-13636/70 crt=8625'201557 lcod 0'0 inactive NOTIFY] state: transitioning to Stray
-2> 2017-04-01 09:07:03.256570 7f087f7e8700 1 osd.10 pg_epoch: 13637 pg[1.ae4( v 8625' 201280 (8612'198271,8625'201280] local-les=13531 n=87307 ec=350 les/c/f 13531/13532/0 13636/13637/9026) [15,10] r=1 lpr=13637 pi=951-13636/70 crt=8625'201280 lcod 0'0 inactive NOTIFY] state<Start>: transitioning to Stray
-1> 2017-04-01 09:07:03.257173 7f087efe7700 1 osd.10 pg_epoch: 13637 pg[1.4e7( v 8625'200805 (8612'197750,8625'200805] local-les=13631 n=87570 ec=350 les/c/f 13512/9635/0 13636/13637/13636) [10,73]/[10] r=0 lpr=13637 pi=9634-13636/42 crt=8625'200805 lcod 0'0 mlcod 0'0 remapped] state<Start>: transitioning to Primary
0> 2017-04-01 09:07:03.267047 7f0874fff700 -1 osd/ReplicatedPG.cc: In function 'virtual void ReplicatedPG::on_local_recover(const hobject_t&, const ObjectRecoveryInfo&, ObjectContextRef, ObjectStore::Transaction*)' thread 7f0874fff700 time 2017-04-01 09:07:03.263494
osd/ReplicatedPG.cc: 209: FAILED assert(is_primary())
ceph version 10.2.6 (656b5b63ed7c43bd014bcafd81b001959d5f089f)
1: (ceph::__ceph_assert_fail(char const*, char const*, int, char const*)+0x8b) [0x7f08cc6d726b]
2: (ReplicatedPG::on_local_recover(hobject_t const&, ObjectRecoveryInfo const&, std::shared_ptr& , ThreadPool::TPHandle&)+0xed) [0x7f08cc1bd28d]
7: (OSD::dequeue_op(boost::intrusive_ptr& )+0x5d) [0x7f08cc07bead]
9: (OSD::ShardedOpWQ::_process(unsigned int, ceph::heartbeat_handle_d*)+0x869) [0x7f08cc0808c9]
10: (ShardedThreadPool::shardedthreadpool_worker(unsigned int)+0x877) [0x7f08cc6c7767]
11: (ShardedThreadPool::WorkThreadSharded::entry()+0x10) [0x7f08cc6c9690]
12: (()+0x8184) [0x7f08cabc5184]
13: (clone()+0x6d) [0x7f08c8b9fbed]
NOTE: a copy of the executable, or `objdump -rdS ` is needed to interpret this. 未解决......
尝试解决
- 找到问题pg, 从它的所在正常的osd中来导出pg, 再导入有问题的osd.
- 直接force_create问题pg