归置组排障¶
归置组总不整洁¶
When you create a cluster and your cluster remains in active
,
active+remapped
or active+degraded
status and never achieve an
active+clean
status, you likely have a problem with your configuration.
You may need to review settings in the 存储池、归置组和 CRUSH 配置参考 and make appropriate adjustments.
As a general rule, you should run your cluster with more than one OSD and a pool size greater than 1 object replica.
单节点集群¶
Ceph no longer provides documentation for operating on a single node, because you would never deploy a system designed for distributed computing on a single node. Additionally, mounting client kernel modules on a single node containing a Ceph daemon may cause a deadlock due to issues with the Linux kernel itself (unless you use VMs for the clients). You can experiment with Ceph in a 1-node configuration, in spite of the limitations as described herein.
If you are trying to create a cluster on a single node, you must change the
default of the osd crush chooseleaf type
setting from 1
(meaning
host
or node
) to 0
(meaning osd
) in your Ceph configuration
file before you create your monitors and OSDs. This tells Ceph that an OSD
can peer with another OSD on the same host. If you are trying to set up a
1-node cluster and osd crush chooseleaf type
is greater than 0
,
Ceph will try to peer the PGs of one OSD with the PGs of another OSD on
another node, chassis, rack, row, or even datacenter depending on the setting.
Tip
DO NOT mount kernel clients directly on the same node as your Ceph Storage Cluster, because kernel conflicts can arise. However, you can mount kernel clients within virtual machines (VMs) on a single node.
If you are creating OSDs using a single disk, you must create directories for the data manually first. For example:
ceph-deploy osd create --data {disk} {host}
OSD 数量小于副本数¶
If you’ve brought up two OSDs to an up
and in
state, but you still
don’t see active + clean
placement groups, you may have an
osd pool default size
set to greater than 2
.
There are a few ways to address this situation. If you want to operate your
cluster in an active + degraded
state with two replicas, you can set the
osd pool default min size
to 2
so that you can write objects in
an active + degraded
state. You may also set the osd pool default size
setting to 2
so that you only have two stored replicas (the original and
one replica), in which case the cluster should achieve an active + clean
state.
Note
You can make the changes at runtime. If you make the changes in your Ceph configuration file, you may need to restart your cluster.
存储池副本数为 1¶
If you have the osd pool default size
set to 1
, you will only have
one copy of the object. OSDs rely on other OSDs to tell them which objects
they should have. If a first OSD has a copy of an object and there is no
second copy, then no second OSD can tell the first OSD that it should have
that copy. For each placement group mapped to the first OSD (see
ceph pg dump
), you can force the first OSD to notice the placement groups
it needs by running:
ceph osd force-create-pg <pgid>
CRUSH 图错误¶
Another candidate for placement groups remaining unclean involves errors in your CRUSH map.
卡住的归置组¶
有失败时归置组会进入“degraded”(降级)或“peering”(连接建立中)状态,这事时有发生,通常这些状态意味着正常的失败恢复正在进行。然而,如果一个归置组长时间处于某个这些状态就意味着有更大的问题,因此监视器在归置组卡 (stuck) 在非最优状态时会警告。我们具体检查:
inactive
(不活跃)——归置组长时间无活跃(即它不能提供读写服务了);unclean
(不干净)——归置组长时间不干净(例如它未能从前面的失败完全恢复);stale
(不新鲜)——归置组状态没有被ceph-osd
更新,表明存储这个归置组的所有节点可能都挂了。
你可以摆出卡住的归置组:
ceph pg dump_stuck stale
ceph pg dump_stuck inactive
ceph pg dump_stuck unclean
卡在 stale
状态的归置组通过修复 ceph-osd
进程通常可以修复;卡在 inactive
状态的归置组通常是互联问题(参见 归置组挂了——互联失败 );卡在 unclean
状态的归置组通常是由于某些原因阻止了恢复的完成,像未找到的对象(参见 未找到的对象 )。
归置组挂了——互联失败¶
在某些情况下, ceph-osd
连接建立进程会遇到问题,使 PG 不能活跃、可用,例如 ceph health
也许显示:
ceph health detail
HEALTH_ERR 7 pgs degraded; 12 pgs down; 12 pgs peering; 1 pgs recovering; 6 pgs stuck unclean; 114/3300 degraded (3.455%); 1/3 in osds are down
...
pg 0.5 is down+peering
pg 1.4 is down+peering
...
osd.1 is down since epoch 69, last address 192.168.106.220:6801/8651
可以查询到 PG 为何被标记为 down
:
ceph pg 0.5 query
{ "state": "down+peering",
...
"recovery_state": [
{ "name": "Started\/Primary\/Peering\/GetInfo",
"enter_time": "2012-03-06 14:40:16.169679",
"requested_info_from": []},
{ "name": "Started\/Primary\/Peering",
"enter_time": "2012-03-06 14:40:16.169659",
"probing_osds": [
0,
1],
"blocked": "peering is blocked due to down osds",
"down_osds_we_would_probe": [
1],
"peering_blocked_by": [
{ "osd": 1,
"current_lost_at": 0,
"comment": "starting or marking this osd lost may let us proceed"}]},
{ "name": "Started",
"enter_time": "2012-03-06 14:40:16.169513"}
]
}
recovery_state
段告诉我们连接建立因 ceph-osd
进程挂了而被阻塞,本例是 osd.1
挂了,启动这个进程应该就可以恢复。
另外,如果 osd.1
是灾难性的失败(如硬盘损坏),我们可以告诉集群它丢失( lost
)了,让集群尽力完成副本拷贝。
Important
集群不能保证其它数据副本是一致且最新就危险了!
让 Ceph 无论如何都继续:
ceph osd lost 1
恢复将继续。
未找到的对象¶
某几种失败相组合可能导致 Ceph 抱怨有找不到( unfound
)的对象:
ceph health detail
HEALTH_WARN 1 pgs degraded; 78/3778 unfound (2.065%)
pg 2.4 is active+degraded, 78 unfound
这意味着存储集群知道一些对象(或者存在对象的较新副本)存在,却没有找到它们的副本。下例展示了这种情况是如何发生的,一个 PG 的数据存储在 ceph-osd 1 和 2 上:
1 挂了;
2 独自处理一些写动作;
1 起来了;
1 和 2 重新互联, 1 上面丢失的对象加入队列准备恢复;
新对象还未拷贝完, 2 挂了。
这时, 1 知道这些对象存在,但是活着的 ceph-osd
都没有副本,这种情况下,读写这些对象的 IO 就会被阻塞,集群只能指望节点早点恢复。这时我们假设用户希望先得到一个 IO 错误。
首先,你应该确认哪些对象找不到了:
ceph pg 2.4 list_unfound [starting offset, in json]
{ "offset": { "oid": "",
"key": "",
"snapid": 0,
"hash": 0,
"max": 0},
"num_missing": 0,
"num_unfound": 0,
"objects": [
{ "oid": "object 1",
"key": "",
"hash": 0,
"max": 0 },
...
],
"more": 0}
如果在一次查询里列出的对象太多, more
这个字段将为
true
,因此你可以查询更多。(命令行工具可能隐藏了,但这里没有)
其次,你可以找出哪些 OSD 上探测到、或可能包含数据:
ceph pg 2.4 query
"recovery_state": [
{ "name": "Started\/Primary\/Active",
"enter_time": "2012-03-06 15:15:46.713212",
"might_have_unfound": [
{ "osd": 1,
"status": "osd is down"}]},
本例中,集群知道 osd.1
可能有数据,但它挂了( down
)。所有可能的状态有:
已经探测到了
在查询
OSD 挂了
尚未查询
有时候集群要花一些时间来查询可能的位置。
还有一种可能性,对象存在于其它位置却未被列出,例如,集群里的一个 ceph-osd
停止且被剔出,然后完全恢复了;后来的失败、恢复后仍有未找到的对象,它也不会觉得早已死亡的 ceph-osd
上仍可能包含这些对象。(这种情况几乎不太可能发生)。
如果所有位置都查询过了仍有对象丢失,那就得放弃丢失的对象了。这仍可能是罕见的失败组合导致的,集群在写入完成前,未能得知写入是否已执行。以下命令把未找到的( unfound )对象标记为丢失( lost )。
ceph pg 2.5 mark_unfound_lost revert|delete
上述最后一个参数告诉集群应如何处理丢失的对象。
delete 选项将导致完全删除它们。
revert 选项(纠删码存储池不可用)会回滚到前一个版本或者(如果它是新对象的话)删除它。要慎用,它可能迷惑那些期望对象存在的应用程序。
无根归置组¶
拥有归置组拷贝的 OSD 都可以失败,在这种情况下,那一部分的对象存储不可用,监视器就不会收到那些归置组的状态更新了。为检测这种情况,监视器把任何主 OSD 失败的归置组标记为 stale
(不新鲜),例如:
ceph health
HEALTH_WARN 24 pgs stale; 3/300 in osds are down
你能找出哪些归置组 stale
、和存储这些归置组的最新 OSD ,命令如下:
ceph health detail
HEALTH_WARN 24 pgs stale; 3/300 in osds are down
...
pg 2.5 is stuck stale+active+remapped, last acting [2,0]
...
osd.10 is down since epoch 23, last address 192.168.106.220:6800/11080
osd.11 is down since epoch 13, last address 192.168.106.220:6803/11539
osd.12 is down since epoch 24, last address 192.168.106.220:6806/11861
如果想使归置组 2.5 重新在线,例如,上面的输出告诉我们它最后由 osd.0
和 osd.2
处理,重启这些 ceph-osd
将恢复之(还有其它的很多 PG )。
只有几个 OSD 接收数据¶
如果你的集群有很多节点,但只有其中几个接收数据,检查下存储池里的归置组数量。因为归置组是映射到多个 OSD 的,这样少量的归置组将不能分布于整个集群。试着创建个新存储池,其归置组数量是 OSD 数量的若干倍。详情见归置组,存储池的默认归置组数量没多大用,你可以参考这里更改它。
不能写入数据¶
如果你的集群已启动,但一些 OSD 没起来,导致不能写入数据,确认下运行的 OSD 数量满足归置组要求的最低 OSD 数。如果不能满足, Ceph 就不会允许你写入数据,因为 Ceph 不能保证复制能如愿进行。详情参见存储池、归置组和 CRUSH 配置参考里的 osd pool default min size
。
归置组不一致¶
如果你看到状态变成了 active + clean + inconsistent
,可能是洗刷时遇到了错误。与往常一样,我们可以这样找出不一致的归置组:
$ ceph health detail
HEALTH_ERR 1 pgs inconsistent; 2 scrub errors
pg 0.6 is active+clean+inconsistent, acting [0,1,2]
2 scrub errors
或者这样,如果你喜欢程序化的输出:
$ rados list-inconsistent-pg rbd
["0.6"]
一致的状态只有一种,然而在最坏的情况下,我们可能会遇到多个对象产生了各种各样的不一致。假设在 PG 0.6
里的一个名为 foo
的对象被截断了,我们可以这样查看:
$ rados list-inconsistent-obj 0.6 --format=json-pretty
{
"epoch": 14,
"inconsistents": [
{
"object": {
"name": "foo",
"nspace": "",
"locator": "",
"snap": "head",
"version": 1
},
"errors": [
"data_digest_mismatch",
"size_mismatch"
],
"union_shard_errors": [
"data_digest_mismatch_info",
"size_mismatch_info"
],
"selected_object_info": "0:602f83fe:::foo:head(16'1 client.4110.0:1 dirty|data_digest|omap_digest s 968 uv 1 dd e978e67f od ffffffff alloc_hint [0 0 0])",
"shards": [
{
"osd": 0,
"errors": [],
"size": 968,
"omap_digest": "0xffffffff",
"data_digest": "0xe978e67f"
},
{
"osd": 1,
"errors": [],
"size": 968,
"omap_digest": "0xffffffff",
"data_digest": "0xe978e67f"
},
{
"osd": 2,
"errors": [
"data_digest_mismatch_info",
"size_mismatch_info"
],
"size": 0,
"omap_digest": "0xffffffff",
"data_digest": "0xffffffff"
}
}
]
}
此时,我们可以从输出里看到:
唯一不一致的对象名为
foo
,并且它的 head 不一致。不一致分为两类:
errors
: 这些错误表明不一致性出现在分片之间,但是没说明哪个(或哪些)分片有问题。如果 shards 阵列中有errors
字段,且不为空,它会指出问题所在。data_digest_mismatch
: OSD.2 内读取到的副本的数字摘要与 OSD.0 和 OSD.1 的不一样。size_mismatch
: OSD.2 内读取到的副本的尺寸是 0 ,而 OSD.0 和 OSD.1 说是 968 。
union_shard_errors
:shards
阵列中、所有与分片相关的错误errors
的并集。errors
是个错误原因集合,汇集了相关分片的这类问题,如read_error
。以oi
结尾的errors
表明它是与selected_object_info
的对照结果。从shards
阵列里可以查到哪个分片有什么样的错误。data_digest_mismatch_info
: 存储在 object-info ( 对象信息)里的数字签名不是0xffffffff
(这个是根据 OSD.2 上的分片计算出来的)。size_mismatch_info
: object-info 内存储的尺寸与 OSD.2 上的对象尺寸 0 不同。
你可以用下列命令修复不一致的归置组:
ceph pg repair {placement-group-ID}
此命令会用权威的副本覆盖有问题的。根据既定规则,多数情况下 Ceph 都能从若干副本中选择正确的,但是也会有例外。比如,存储的数字签名可能正好丢了,选择权威副本时又忽略了计算出的数字签名,总之,用此命令时小心为好。
如果一个分片的 errors
里出现了 read_error
,很可能是磁盘错误引起的不一致,你最好先查验那个 OSD 所用的磁盘。
如果你时不时遇到时钟偏移引起的 active + clean + inconsistent
状态,最好在监视器主机上配置 peer 角色的 NTP 服务。配置细节可参考网络时间协议和 Ceph 时钟选项。
纠删编码的归置组不是 active+clean¶
CRUSH 找不到足够多的 OSD 映射到某个 PG 时,它会显示为
2147483647
,意思是 ITEM_NONE 或 no OSD found
,例如:
[2,1,6,0,5,8,2147483647,7,4]
OSD 不够多¶
如果 Ceph 集群仅有 8 个 OSD ,但是纠删码存储池需要 9 个,就会显示上面的错误。这时候,你仍然可以另外创建需要较少 OSD 的纠删码存储池:
ceph osd erasure-code-profile set myprofile k=5 m=3
ceph osd pool create erasurepool 16 16 erasure myprofile
或者新增一个 OSD ,这个 PG 会自动用上的。
CRUSH 条件不能满足¶
即使集群拥有足够多的 OSD , CRUSH 规则的强制要求仍有可能无法满足。假如有 10 个 OSD 分布于两个主机上,且 CRUSH 规则要求相同归置组不得使用位于同一主机的两个 OSD ,这样映射就会失败,因为只能找到两个 OSD ,你可以从规则里查看必要条件:
$ ceph osd crush rule ls
[
"replicated_rule",
"erasurepool"]
$ ceph osd crush rule dump erasurepool
{ "rule_id": 1,
"rule_name": "erasurepool",
"ruleset": 1,
"type": 3,
"min_size": 3,
"max_size": 20,
"steps": [
{ "op": "take",
"item": -1,
"item_name": "default"},
{ "op": "chooseleaf_indep",
"num": 0,
"type": "host"},
{ "op": "emit"}]}
可以这样解决此问题,创建新存储池,其内的 PG 允许多个 OSD 位于同一主机,命令如下:
ceph osd erasure-code-profile set myprofile crush-failure-domain=osd
ceph osd pool create erasurepool 16 16 erasure myprofile
CRUSH 过早中止¶
假设集群拥有的 OSD 足以映射到 PG (比如有 9 个 OSD 和一个纠删码存储池的集群,每个 PG 需要 9 个 OSD ), CRUSH 仍然有可能在找到映射前就中止了。可以这样解决:
降低纠删存储池内 PG 的要求,让它使用较少的 OSD (需创建另一个存储池,因为纠删码配置不支持动态修改)。
向集群添加更多 OSD (无需修改纠删存储池,它会自动回到清洁状态)。
通过手工打造的 CRUSH 规则,让它多试几次以找到合适的映射。把
set_choose_tries
设置得高于默认值即可。
你从集群中提取出 crushmap 之后,应该先用 crushtool
校验一下是否有问题,这样你的试验就无需触及 Ceph 集群,只要在一个本地文件上测试即可:
$ ceph osd crush rule dump erasurepool
{ "rule_name": "erasurepool",
"ruleset": 1,
"type": 3,
"min_size": 3,
"max_size": 20,
"steps": [
{ "op": "take",
"item": -1,
"item_name": "default"},
{ "op": "chooseleaf_indep",
"num": 0,
"type": "host"},
{ "op": "emit"}]}
$ ceph osd getcrushmap > crush.map
got crush map from osdmap epoch 13
$ crushtool -i crush.map --test --show-bad-mappings \
--rule 1 \
--num-rep 9 \
--min-x 1 --max-x $((1024 * 1024))
bad mapping rule 8 x 43 num_rep 9 result [3,2,7,1,2147483647,8,5,6,0]
bad mapping rule 8 x 79 num_rep 9 result [6,0,2,1,4,7,2147483647,5,8]
bad mapping rule 8 x 173 num_rep 9 result [0,4,6,8,2,1,3,7,2147483647]
其中 --num-rep
是纠删码 CRUSH 规则所需的 OSD 数量,
--rule
是 ceph osd crush rule dump
命令结果中
ruleset
字段的值。此测试会尝试映射一百万个值(即
[--min-x,--max-x]
所指定的范围),且必须至少显示一个坏映射;如果它没有任何输出,说明所有映射都成功了,你可以就此打住:问题的根源不在这里。
反编译 crush 图后,你可以手动编辑其 CRUSH 规则:
$ crushtool --decompile crush.map > crush.txt
并把下面这行加进规则:
step set_choose_tries 100
然后 crush.txt
文件内的这部分大致如此:
rule erasurepool {
ruleset 1
type erasure
min_size 3
max_size 20
step set_chooseleaf_tries 5
step set_choose_tries 100
step take default
step chooseleaf indep 0 type host
step emit
}
然后编译、并再次测试:
$ crushtool --compile crush.txt -o better-crush.map
所有映射都成功时,用 crushtool
的 --show-choose-tries
选项能看到成功映射的尝试次数直方图:
$ crushtool -i better-crush.map --test --show-bad-mappings \
--show-choose-tries \
--rule 1 \
--num-rep 9 \
--min-x 1 --max-x $((1024 * 1024))
...
11: 42
12: 44
13: 54
14: 45
15: 35
16: 34
17: 30
18: 25
19: 19
20: 22
21: 20
22: 17
23: 13
24: 16
25: 13
26: 11
27: 11
28: 13
29: 11
30: 10
31: 6
32: 5
33: 10
34: 3
35: 7
36: 5
37: 2
38: 5
39: 5
40: 2
41: 5
42: 4
43: 1
44: 2
45: 2
46: 3
47: 1
48: 0
...
102: 0
103: 1
104: 0
...
有 42 个归置组需 11 次重试、 44 个归置组需 12 次重试,以此类推。这样,重试的最高次数就是防止坏映射的最低值,也就是
set_choose_tries
的取值(即上面输出中的 103 ,因为任意归置组成功映射的重试次数都没有超过 103 )。