Redis偶发连接失败案例分析
【作者】
张延俊:携程技术保障中心资深dba,对数据库架构和疑难问题分析排查有浓厚的兴趣。
寿向晨:携程技术保障中心高级dba,主要负责携程redis及db的运维工作,在自动化运维,流程化及监控排障等方面有较多的实践经验,喜欢深入分析问题,提高团队运维效率。
【问题描述】
生产环境有一个redis会偶尔发生连接失败的报错,报错的时间点、客户端ip并没有特别明显的规律,过一会儿,报错会自动恢复。
以下是客户端报错信息:
credis.client.rexceptions.excutecommandexception: unable to connect redis server: ---> credis.third.redis.redisexception: unable to connect redis server: 在 credis.third.redis.redisnativeclient.createconnectionerror() 在 credis.third.redis.redisnativeclient.sendexpectdata(byte[][] cmdwithbinaryargs) 在 credis.client.entities.redisserver.<>c__displayclassd`1.
从报错的信息来看,应该是连接不上redis所致。redis的版本是2.8.19。虽然版本有点老,但基本运行稳定。
线上环境只有这个集群有偶尔报错。这个集群的一个比较明显的特征是客户端服务器比较多,有上百台。
【问题分析】
从报错的信息来看,客户端连接不到服务端。常见的原因有以下几点:
- 一个常见的原因是由于端口耗尽,对网络连接进行排查,在出问题的点上,tcp连接数远没有达到端口耗尽的场景,因此这个不是redis连接不上的根本原因。
- 另外一种常见的场景是在服务端有慢查询,导致redis服务阻塞。我们在redis服务端,把运行超过10毫秒的语句进行抓取,也没有抓到运行慢的语句。
从服务端的部署的监控来看,出问题的点上,连接数有一个突然飙升,从3500个连接突然飙升至4100个连接。如下图显示:
同时间,服务器端显示redis服务端有丢包现象:345539 – 344683 = 856个包。
sat apr 7 10:41:40 cst 2018 1699 outgoing packets dropped 92 dropped because of missing route 344683 syns to listen sockets dropped 344683 times the listen queue of a socket overflowed
sat apr 7 10:41:41 cst 2018 1699 outgoing packets dropped 92 dropped because of missing route 345539 syns to listen sockets dropped 345539 times the listen queue of a socket overflowed
客户端报错的原因基本确定,是因为建连速度太快,导致服务端backlog队列溢出,连接被server端reset。
【关于backlog overflow】
在高并发的短连接服务中,这是一种很常见的tcp报错类型。一个正常的tcp建连过程如下:
1.client发送一个(syn)给server
2.server返回一个(syn,ack)给client
3.client返回一个(ack)
三次握手结束,对client来说建连成功,client可以继续发送数据包给server,但是这个时候server端未必ready,如下图所示 :
在bsd版本内核实现的tcp协议中,server端建连过程需要两个队列,一个是syn queue,一个是accept queue。前者叫半开连接(或者半连接)队列,在接收到client发送的syn时加入队列。(一种常见的网络攻击方式就是不断发送syn但是不发送ack从而导致server端的半开队列撑爆,server端拒绝服务。)后者叫全连接队列,server返回(syn,ack),在接收到client发送ack后(此时client会认为建连已经完成,会开始发送psh包),如果accept queue没有满,那么server从syn queue把连接信息移到accept queue;如果此时accept queue溢出的话,server的行为要看配置。如果tcp_abort_on_overflow为0(默认),那么直接drop掉client发送的psh包,此时client会进入重发过程,一段时间后server端重新发送syn,ack,重新从建连的第二步开始;如果tcp_abort_on_overflow为1,那么server端发现accept queue满之后直接发送reset。
通过wireshark搜索发现在一秒内有超过2000次对redis server端发起建连请求。我们尝试修改tcp backlog大小,从511调整到2048, 问题并没有得到解决。所以此类微调,并不能彻底的解决问题。
【网络包分析】
我们用wireshark来识别网络拥塞的准确时间点和原因。我们已经有了准确的报错时间点,先用editcap把超大的tcp包裁剪一下,裁成30秒间隔,并通过wireshark i/o 100ms间隔分析网络阻塞的准确时间点:
根据图标可以明显看到tcp的packets来往存在block。
对该block前后的网络包进行明细分析,网络包来往情况如下:
time | source | dest | description |
---|---|---|---|
12:01:54.6536050 | redis-server | clients | tcp:flags=…ap… |
12:01:54.6538580 | redis-server | clients | tcp:flags=…ap… |
12:01:54.6539770 | redis-server | clients | tcp:flags=…ap… |
12:01:54.6720580 | redis-server | clients | tcp:flags=…a..s.. |
12:01:54.6727200 | redis-server | clients | tcp:flags=…a…… |
12:01:54.6808480 | redis-server | clients | tcp:flags=…ap….. |
12:01:54.6910840 | redis-server | clients | tcp:flags=…a…s., |
12:01:54.6911950 | redis-server | clients | tcp:flags=…a…… |
… | … | … | … |
12:01:56.1181350 | redis-server | clients | tcp:flags=…ap…. |
12:01:54.6808480, redis server端向客户端发送了一个push包,也就是对于查询请求的一个结果返回。后面的包都是在做连接处理,包括ack包,ack确认包,以及重置的rst包,紧接着下面一个push包是在12:01:56.1181350发出的。中间的间隔是1.4372870秒。也就是说,在这1.4372870秒期间,redis的服务器端,除了做一个查询,其他的操作都是在做建连,或拒绝连接。
客户端报错的前后逻辑已经清楚了,redis-server卡了1.43秒,client的connection pool被打满,疯狂新建连接,server的accept queue满,直接拒绝服务,client报错。开始怀疑client发送了特殊命令,这时需要确认一下client的最后几个命令是什么,找到redis-server卡死前的第一个包,装一个wireshark的redis插件,看到最后几个命令是简单的get,并且key-value都很小,不至于需要耗费1.43秒才能完成。服务端也没有slow log,此时排障再次陷入僵局。
【进一步分析】
为了了解这1.43秒之内,redis server在做什么事情,我们用pstack来抓取信息。pstack本质上是gdb attach. 高频率的抓取会影响redis的吞吐。死循环0.5秒一次无脑抓,在redis-server卡死的时候抓到堆栈如下(过滤了没用的栈信息):
thu may 31 11:29:18 cst 2018 thread 1 (thread 0x7ff2db6de720 (lwp 8378)): #0 0x000000000048cec4 in ?? () #1 0x00000000004914a4 in je_arena_ralloc () #2 0x00000000004836a1 in je_realloc () #3 0x0000000000422cc5 in zrealloc () #4 0x00000000004213d7 in sdsremovefreespace () #5 0x000000000041ef3c in clientscronresizequerybuffer () #6 0x00000000004205de in clientscron () #7 0x0000000000420784 in servercron () #8 0x0000000000418542 in aeprocessevents () #9 0x000000000041873b in aemain () #10 0x0000000000420fce in main () thu may 31 11:29:19 cst 2018 thread 1 (thread 0x7ff2db6de720 (lwp 8378)): #0 0x0000003729ee5407 in madvise () from /lib64/libc.so.6 #1 0x0000000000493a4e in je_pages_purge () #2 0x000000000048cf70 in ?? () #3 0x00000000004914a4 in je_arena_ralloc () #4 0x00000000004836a1 in je_realloc () #5 0x0000000000422cc5 in zrealloc () #6 0x00000000004213d7 in sdsremovefreespace () #7 0x000000000041ef3c in clientscronresizequerybuffer () #8 0x00000000004205de in clientscron () #9 0x0000000000420784 in servercron () #10 0x0000000000418542 in aeprocessevents () #11 0x000000000041873b in aemain () #12 0x0000000000420fce in main () thu may 31 11:29:19 cst 2018 thread 1 (thread 0x7ff2db6de720 (lwp 8378)): #0 0x000000000048108c in je_malloc_usable_size () #1 0x0000000000422be6 in zmalloc () #2 0x00000000004220bc in sdsnewlen () #3 0x000000000042c409 in createstringobject () #4 0x000000000042918e in processmultibulkbuffer () #5 0x0000000000429662 in processinputbuffer () #6 0x0000000000429762 in readqueryfromclient () #7 0x000000000041847c in aeprocessevents () #8 0x000000000041873b in aemain () #9 0x0000000000420fce in main () thu may 31 11:29:20 cst 2018 thread 1 (thread 0x7ff2db6de720 (lwp 8378)): #0 0x000000372a60e7cd in write () from /lib64/libpthread.so.0 #1 0x0000000000428833 in sendreplytoclient () #2 0x0000000000418435 in aeprocessevents () #3 0x000000000041873b in aemain () #4 0x0000000000420fce in main ()
重复多次抓取后,从堆栈中发现可疑堆栈clientscronresizequerybuffer位置,属于servercron()函数下,这个redis-server内部的定时调度,并不在用户线程下,这个解释了为什么卡死的时候没有出现慢查询。
查看redis源码,确认到底redis-server在做什么:
clientscron(server.h): #define clients_cron_min_iterations 5 void clientscron(void) { /* make sure to process at least numclients/server.hz of clients * per call. since this function is called server.hz times per second * we are sure that in the worst case we process all the clients in 1 * second. */ int numclients = listlength(server.clients); int iterations = numclients/server.hz; mstime_t now = mstime(); /* process at least a few clients while we are at it, even if we need * to process less than clients_cron_min_iterations to meet our contract * of processing each client once per second. */ if (iterations < clients_cron_min_iterations) iterations = (numclients < clients_cron_min_iterations) ? numclients : clients_cron_min_iterations; while(listlength(server.clients) && iterations--) { client *c; listnode *head; /* rotate the list, take the current head, process. * this way if the client must be removed from the list it's the * first element and we don't incur into o(n) computation. */ listrotate(server.clients); head = listfirst(server.clients); c = listnodevalue(head); /* the following functions do different service checks on the client. * the protocol is that they return non-zero if the client was * terminated. */ if (clientscronhandletimeout(c,now)) continue; if (clientscronresizequerybuffer(c)) continue; } }
clientscron首先判断当前client的数量,用于控制一次清理连接的数量,生产服务器单实例的连接数量在5000不到,也就是一次清理的连接数是50个。
clientscronresizequerybuffer(server.h): /* the client query buffer is an sds.c string that can end with a lot of * free space not used, this function reclaims space if needed. * * the function always returns 0 as it never terminates the client. */ int clientscronresizequerybuffer(client *c) { size_t querybuf_size = sdsallocsize(c->querybuf); time_t idletime = server.unixtime - c->lastinteraction; /* 只在以下两种情况下会resize query buffer: * 1) query buffer > big_arg(在server.h 中定义#define proto_mbulk_big_arg (1024*32)) 且这个buffer的小于一段时间的客户端使用的峰值. * 2) 客户端空闲超过2s且buffer size大于1k. */ if (((querybuf_size > proto_mbulk_big_arg) && (querybuf_size/(c->querybuf_peak+1)) > 2) || (querybuf_size > 1024 && idletime > 2)) { /* only resize the query buffer if it is actually wasting space. */ if (sdsavail(c->querybuf) > 1024) { c->querybuf = sdsremovefreespace(c->querybuf); } } /* reset the peak again to capture the peak memory usage in the next * cycle. */ c->querybuf_peak = 0; return 0; }
如果redisclient对象的query buffer满足条件,那么就直接resize掉。满足条件的连接分成两种,一种是真的很大的,比该客户端一段时间内使用的峰值还大;还有一种是很闲(idle>2)的,这两种都要满足一个条件,就是buffer free的部分超过1k。那么redis-server卡住的原因就是正好有那么50个很大的或者空闲的并且free size超过了1k大小连接的同时循环做了resize,由于redis都属于单线程工作的程序,所以block了client。那么解决这个问题办法就很明朗了,让resize 的频率变低或者resize的执行速度变快。
既然问题出在query buffer上,我们先看一下这个东西被修改的位置:
readqueryfromclient(networking.c): redisclient *createclient(int fd) { redisclient *c = zmalloc(sizeof(redisclient)); /* passing -1 as fd it is possible to create a non connected client. * this is useful since all the redis commands needs to be executed * in the context of a client. when commands are executed in other * contexts (for instance a lua script) we need a non connected client. */ if (fd != -1) { anetnonblock(null,fd); anetenabletcpnodelay(null,fd); if (server.tcpkeepalive) anetkeepalive(null,fd,server.tcpkeepalive); if (aecreatefileevent(server.el,fd,ae_readable, readqueryfromclient, c) == ae_err) { close(fd); zfree(c); return null; } } selectdb(c,0); c->id = server.next_client_id++; c->fd = fd; c->name = null; c->bufpos = 0; c->querybuf = sdsempty(); 初始化是0 readqueryfromclient(networking.c): void readqueryfromclient(aeeventloop *el, int fd, void *privdata, int mask) { redisclient *c = (redisclient*) privdata; int nread, readlen; size_t qblen; redis_notused(el); redis_notused(mask); server.current_client = c; readlen = redis_iobuf_len; /* if this is a multi bulk request, and we are processing a bulk reply * that is large enough, try to maximize the probability that the query * buffer contains exactly the sds string representing the object, even * at the risk of requiring more read(2) calls. this way the function * processmultibulkbuffer() can avoid copying buffers to create the * redis object representing the argument. */ if (c->reqtype == redis_req_multibulk && c->multibulklen && c->bulklen != -1 && c->bulklen >= redis_mbulk_big_arg) { int remaining = (unsigned)(c->bulklen+2)-sdslen(c->querybuf); if (remaining < readlen) readlen = remaining; } qblen = sdslen(c->querybuf); if (c->querybuf_peak < qblen) c->querybuf_peak = qblen; c->querybuf = sdsmakeroomfor(c->querybuf, readlen); 在这里会被扩大
由此可见c->querybuf在连接第一次读取命令后的大小就会被分配至少1024*32,所以回过头再去看resize的清理逻辑就明显存在问题,每个被使用到的query buffer的大小至少就是1024*32,但是清理的时候判断条件是>1024,也就是说,所有的idle>2的被使用过的连接都会被resize掉,下次接收到请求的时候再重新分配到1024*32,这个其实是没有必要的,在访问比较频繁的群集,内存会被频繁得回收重分配,所以我们尝试将清理的判断条件改造为如下,就可以避免大部分没有必要的resize操作:
if (((querybuf_size > redis_mbulk_big_arg) && (querybuf_size/(c->querybuf_peak+1)) > 2) || (querybuf_size > 1024*32 && idletime > 2)) { /* only resize the query buffer if it is actually wasting space. */ if (sdsavail(c->querybuf) > 1024*32) { c->querybuf = sdsremovefreespace(c->querybuf); } }
这个改造的副作用是内存的开销,按照一个实例5k连接计算,5000*1024*32=160m,这点内存消耗对于上百g内存的服务器完全可以接受。
【问题重现】
在使用修改过源码的redis server后,问题仍然重现了,客户端还是会报同类型的错误,且报错的时候,服务器内存依然会出现抖动。抓取内存堆栈信息如下:
thu jun 14 21:56:54 cst 2018 #3 0x0000003729ee893d in clone () from /lib64/libc.so.6 thread 1 (thread 0x7f2dc108d720 (lwp 27851)): #0 0x0000003729ee5400 in madvise () from /lib64/libc.so.6 #1 0x0000000000493a1e in je_pages_purge () #2 0x000000000048cf40 in arena_purge () #3 0x00000000004a7dad in je_tcache_bin_flush_large () #4 0x00000000004a85e9 in je_tcache_event_hard () #5 0x000000000042c0b5 in decrrefcount () #6 0x000000000042744d in resetclient () #7 0x000000000042963b in processinputbuffer () #8 0x0000000000429762 in readqueryfromclient () #9 0x000000000041847c in aeprocessevents () #10 0x000000000041873b in aemain () #11 0x0000000000420fce in main () thu jun 14 21:56:54 cst 2018 thread 1 (thread 0x7f2dc108d720 (lwp 27851)): #0 0x0000003729ee5400 in madvise () from /lib64/libc.so.6 #1 0x0000000000493a1e in je_pages_purge () #2 0x000000000048cf40 in arena_purge () #3 0x00000000004a7dad in je_tcache_bin_flush_large () #4 0x00000000004a85e9 in je_tcache_event_hard () #5 0x000000000042c0b5 in decrrefcount () #6 0x000000000042744d in resetclient () #7 0x000000000042963b in processinputbuffer () #8 0x0000000000429762 in readqueryfromclient () #9 0x000000000041847c in aeprocessevents () #10 0x000000000041873b in aemain () #11 0x0000000000420fce in main ()
显然,querybuffer被频繁resize的问题已经得到了优化,但是还是会出现客户端报错。这就又陷入了僵局。难道还有其他因素导致query buffer resize变慢?我们再次抓取pstack。但这时,jemalloc引起了我们的注意。此时回想redis的内存分配机制,redis为避免libc内存不被释放导致大量内存碎片的问题,默认使用的是jemalloc用作内存分配管理,这次报错的堆栈信息中都是je_pages_purge () redis在调用jemalloc回收脏页。我们看下jemalloc做了些什么:
arena_purge(arena.c) static void arena_purge(arena_t *arena, bool all) { arena_chunk_t *chunk; size_t npurgatory; if (config_debug) { size_t ndirty = 0; arena_chunk_dirty_iter(&arena->chunks_dirty, null, chunks_dirty_iter_cb, (void *)&ndirty); assert(ndirty == arena->ndirty); } assert(arena->ndirty > arena->npurgatory || all); assert((arena->nactive >> opt_lg_dirty_mult) < (arena->ndirty - arena->npurgatory) || all); if (config_stats) arena->stats.npurge++; npurgatory = arena_compute_npurgatory(arena, all); arena->npurgatory += npurgatory; while (npurgatory > 0) { size_t npurgeable, npurged, nunpurged; /* get next chunk with dirty pages. */ chunk = arena_chunk_dirty_first(&arena->chunks_dirty); if (chunk == null) { arena->npurgatory -= npurgatory; return; } npurgeable = chunk->ndirty; assert(npurgeable != 0); if (npurgeable > npurgatory && chunk->nruns_adjac == 0) { arena->npurgatory += npurgeable - npurgatory; npurgatory = npurgeable; } arena->npurgatory -= npurgeable; npurgatory -= npurgeable; npurged = arena_chunk_purge(arena, chunk, all); nunpurged = npurgeable - npurged; arena->npurgatory += nunpurged; npurgatory += nunpurged; } }
jemalloc每次回收都会判断所有实际应该清理的chunck并对清理做count,这个操作对于高响应要求的系统是很奢侈的,所以我们考虑通过升级jemalloc的版本来优化purge的性能。redis 4.0版本发布后,性能有很大的改进,并可以通过命令回收内存,我们线上也正准备进行升级,跟随4.0发布的jemalloc版本为4.1,jemalloc的版本使用的在jemalloc的4.0之后版本的arena_purge()做了很多优化,去掉了计数器的调用,简化了很多判断逻辑,增加了arena_stash_dirty()方法合并了之前的计算和判断逻辑,增加了purge_runs_sentinel,用保持脏块在每个arena lru中的方式替代之前的保持脏块在arena树的dirty-run-containing chunck中的方式,大幅度减少了脏块purge的体积,并且在内存回收过程中不再移动内存块。代码如下:
arena_purge(arena.c) static void arena_purge(arena_t *arena, bool all) { chunk_hooks_t chunk_hooks = chunk_hooks_get(arena); size_t npurge, npurgeable, npurged; arena_runs_dirty_link_t purge_runs_sentinel; extent_node_t purge_chunks_sentinel; arena->purging = true; /* * calls to arena_dirty_count() are disabled even for debug builds * because overhead grows nonlinearly as memory usage increases. */ if (false && config_debug) { size_t ndirty = arena_dirty_count(arena); assert(ndirty == arena->ndirty); } assert((arena->nactive >> arena->lg_dirty_mult) < arena->ndirty || all); if (config_stats) arena->stats.npurge++; npurge = arena_compute_npurge(arena, all); qr_new(&purge_runs_sentinel, rd_link); extent_node_dirty_linkage_init(&purge_chunks_sentinel); npurgeable = arena_stash_dirty(arena, &chunk_hooks, all, npurge, &purge_runs_sentinel, &purge_chunks_sentinel); assert(npurgeable >= npurge); npurged = arena_purge_stashed(arena, &chunk_hooks, &purge_runs_sentinel, &purge_chunks_sentinel); assert(npurged == npurgeable); arena_unstash_purged(arena, &chunk_hooks, &purge_runs_sentinel, &purge_chunks_sentinel); arena->purging = false; }
【解决问题】
实际上我们有多个选项。可以使用google的tcmalloc来代替jemalloc,可以升级jemalloc的版本等等。我们根据上面的分析,尝试通过升级jemalloc版本,实际操作为升级redis版本来解决。我们将redis的版本升级到4.0.9之后观察,线上客户端连接超时这个棘手的问题得到了解决。
【问题总结】
redis在生产环境中因其支持高并发,响应快,易操作被广泛使用,对于运维人员而言,其响应时间的要求带来了各种各样的问题,redis的连接超时问题是其中比较典型的一种,从发现问题,客户端连接超时,到通过抓取客户端与服务端的网络包,内存堆栈定位问题,也被其中一些假象所迷惑,最终通过升级jemalloc(redis)的版本解决问题,这次最值得总结和借鉴的是整个分析的思路。
上一篇: 昨晚有国足的比赛
下一篇: html-路径的介绍