通过阅读Redis源码,配合GDB和抓包等调试手段,分析Redis发布订阅的实现原理,思考相关问题。
源码版本:Redis 6.0.10
思考问题发布订阅基本概念介绍订阅频道 —— SUBSCRIBE命令的实现退订频道 —— USUBSCRIBE命令的实现订阅模式 —— PSUBSCRIBE命令的实现退订模式 —— PUNSUBSCRIBE命令的实现发送消息 —— PUBLISH命令的实现
Redis发布订阅的特点为什么设计pubsub_patterns_dict字典? 一次完整的Redis订阅、发布流程分析
为什么redis-cli中执行了subscribe命令后无法再执行unsubscribe命令?TCP是没有消息边界的,订阅者是如何保证正确解析发布者的消息?订阅的客户端断线,服务端是怎么感知的?如何保证这个客户端的频道、模式订阅信息也被同步删除?服务器正常关闭场景,正在订阅的客户端是怎么感知到服务器离线的? 查看订阅消息 发布订阅基本概念介绍
Redis的发布订阅功能由SUBSCRIBE,PSUBSCRIBE,PUBLISH等命令组成,每个命令对应的功能和相关源码参考下表:
Redis发布订阅的示例可参考菜鸟教程 Redis 发布订阅
订阅频道 —— SUBSCRIBE命令的实现客户端向Redis服务器发送SUBSCRIBE命令,Redis服务器收到客户端的订阅请求,将频道的订阅关系保存在字典pubsub_channels中,数据结构如下:
struct redisServer server; struct redisServer { // server.h dict *pubsub_channels; };
pubsub_channels字典的键表示频道,为字符串类型,字典的值为一个链表,保存所有订阅了这个频道的客户端。
每当一个客户端执行SUBSCRIBE channel命令后,服务端将执行如下 *** 作:
如果pubsub_channel字典不包含channel这个键,说明此时这个频道还没有任何订阅者,此时服务端将首先在字典中创建一个叫channel的键,值初始成一个空链表,然后将这个客户端加入空链表。如果pubsub_channel字典包含channel这个键,说明此时这个频道已经有1个或多个订阅者,此时服务端在字典中找到这个键对应的链表,再将这个客户端加入链表末尾。
举例:有4个客户端(client1~client4),其中client1, client2订阅channel1, client3, client4订阅channel2,此时,pubsub_channels字典可以是下图所示:
订阅频道的源码实现参考pubsub.c/subscribeCommand
void subscribeCommand(client *c) { int j; for (j = 1; j < c->argc; j++) pubsubSubscribeChannel(c,c->argv[j]); c->flags |= CLIENT_PUBSUB; } int pubsubSubscribeChannel(client *c, robj *channel) { dictEntry *de; list *clients = NULL; int retval = 0; if (dictAdd(c->pubsub_channels,channel,NULL) == DICT_OK) { retval = 1; incrRefCount(channel); de = dictFind(server.pubsub_channels,channel); if (de == NULL) { // 说明channel还没有任何订阅者,需先在字典中新增channel键,再初始化1个链表,并将客户端加入链表 clients = listCreate(); dictAdd(server.pubsub_channels,channel,clients); incrRefCount(channel); } else { // 说明channel已有1个或多个订阅者,直接将这个客户端加入链表即可 clients = dictGetVal(de); } listAddNodeTail(clients,c); } addReplyPubsubSubscribed(c,channel); return retval; }退订频道 —— USUBSCRIBE命令的实现
Redis服务器收到客户端的退订频道请求后,通过修改字典pubsub_channels,解除客户端与频道之间的订阅关系,具体 *** 作如下:
如果pubsub_channel字典包含channel这个键,说明此时这个频道已经有1个或多个订阅者:
如果频道有多于1个的订阅者,Redis服务器会查询pubsub_channels字典,找到频道对应的客户端链表,然后从链表中删除这个客户端。如果频道只有1个订阅者,Redis服务器将对应的客户端链表变成空链表,并且从pubsub_channels字典中删除这个频道对应的键。 如果pubsub_channel字典不包含channel这个键,说明没有客户端订阅这个频道,此时Redis服务器不对pubsub_channels字典做任何修改,只是简单向客户端发送回复。(回复内容包括三部分:“unsubscribe”, “channel”, 这个客户端订阅的频道个数和模式个数之和)
举例:有4个客户端(client1~client4),其中client1, client2订阅channel1, client3, client4订阅channel2。
假设client2执行unsubscribe channel1,client3, client4分别执行unsubscribe channel2,那么退订前后的pubsub_channels字典变化如下所示:
退订频道的源码实现参考pubsub.c/unsubscribeCommand
void unsubscribeCommand(client *c) { if (c->argc == 1) { pubsubUnsubscribeAllChannels(c,1); // unsubscribe不加任何参数,表示解除所有频道的订阅 } else { int j; for (j = 1; j < c->argc; j++) pubsubUnsubscribeChannel(c,c->argv[j],1); // 解除这个客户端与某个频道的订阅关系 } if (clientSubscriptionsCount(c) == 0) c->flags &= ~CLIENT_PUBSUB; } int pubsubUnsubscribeChannel(client *c, robj *channel, int notify) { dictEntry *de; list *clients; listNode *ln; int retval = 0; incrRefCount(channel); if (dictDelete(c->pubsub_channels,channel) == DICT_OK) { retval = 1; de = dictFind(server.pubsub_channels,channel); serverAssertWithInfo(c,NULL,de != NULL); clients = dictGetVal(de); // 1、查询pubsub_channels字典,找到频道对应的客户端链表clients ln = listSearchKey(clients,c); // 2、在链表clients中查找客户端c serverAssertWithInfo(c,NULL,ln != NULL); listDelNode(clients,ln); // 3、然后从链表中删除这个客户端c if (listLength(clients) == 0) { // 4、如果频道只有1个订阅者,还需要从pubsub_channels字典中删除这个频道对应的键 dictDelete(server.pubsub_channels,channel); } } if (notify) addReplyPubsubUnsubscribed(c,channel); decrRefCount(channel); return retval; }订阅模式 —— PSUBSCRIBE命令的实现
Redis服务器将所有模式订阅信息记录在链表pubsub_patterns和字典pubsub_pattern_dict,相关数据结构定义如下:
struct redisServer server; struct redisServer { list *pubsub_patterns; dict *pubsub_patterns_dict; };
1、pubsub_patterns链表节点的类型为struct pubsubPattern,其中pattern成员表示订阅的模式串,client成员记录订阅该模式的客户端。
typedef struct pubsubPattern { client *client; // 记录订阅该模式的客户端 robj *pattern; // 表示订阅的模式串 } pubsubPattern;
2、pubsub_patterns_dict字典的键为被订阅的模式(robj *类型),字典的值为链表,这个链表记录所有订阅了这个模式的客户端 。
当客户端执行PSUBSCRIBE命令后,Redis服务器会收到这个模式订阅请求,并执行如下 *** 作:
创建一个struct pubsubPattern结构,其中pattern成员设置为被订阅的模式,client成员设置为订阅该模式的客户端将struct pubsubPattern结构加入到pubsub_patterns链表的末尾。然后在pubsub_patterns_dict字典中查找这个模式:
如果没查到,说明这个模式还没有任何客户端订阅。此时Redis在pubsub_patterns_dict字典中为这个模式创建一个键,将这个键的值设为空链表,再将客户端加入链表。如果能查到,说明这个模式已有客户端订阅,pubsub_patterns_dict字典中必有对应的客户端链表,此时Redis只需将客户端添加到这个客户端链表的末尾即可。
举例:有4个客户端(client1~client4),其中:
client1订阅模式pat1*client2订阅模式pat2*client3, client4订阅模式pat3*。
此时pubsub_patterns_dict字典和pubsub_patterns链表如下图所示:
GDB打印模式字典和模式链表的方法订阅模式的源码参考pubsub.c/pubsubSubscribePattern
int pubsubSubscribePattern(client *c, robj *pattern) { dictEntry *de; list *clients; int retval = 0; if (listSearchKey(c->pubsub_patterns,pattern) == NULL) { retval = 1; pubsubPattern *pat; listAddNodeTail(c->pubsub_patterns,pattern); incrRefCount(pattern); pat = zmalloc(sizeof(*pat)); // 1、构造一个pubsubPattern结构 pat->pattern = getDecodedObject(pattern); // 2、pattern成员设置为被订阅的模式 pat->client = c; // 3、client成员设置为订阅这个模式的客户端 listAddNodeTail(server.pubsub_patterns,pat); // 4、添加这个pubsubPatern结构到pubsub_patterns链表尾 de = dictFind(server.pubsub_patterns_dict,pattern); if (de == NULL) { // 如果没查到,说明这个模式还没有客户端订阅,此时先在字典中为这个模式创建一个键,键的值为空链表 clients = listCreate(); dictAdd(server.pubsub_patterns_dict,pattern,clients); incrRefCount(pattern); } else { clients = dictGetVal(de); } listAddNodeTail(clients,c); // 将客户端c加入链表clients } addReplyPubsubPatSubscribed(c,pattern); return retval; }退订模式 —— PUNSUBSCRIBE命令的实现
模式退订 *** 作和模式订阅的 *** 作正好相反,当客户端执行PUNSUBSCRIBE命令退订一个模式时,Redis服务器将发生如下 *** 作:
在pubsub_patterns链表中查找这个客户端并删除。在pubsub_patterns_dict字典中,找到这个模式对应的客户端链表,在链表中查找这个客户端并删除。
举例:有4个客户端(client1~client4),其中:
client1订阅了模式pat1*client2订阅了模式pat2*client3, client4订阅了模式pat3*。
假设此时client2执行unpsubscribe pat2*,client4执行unsubscribe pat3*,那么模式退订前后的pubsub_patterns链表和pubsub_patterns_dict字典变化如下图所示:
模式退订的源码参考pubsub.c/punsubscribeCommand:
void punsubscribeCommand(client *c) { if (c->argc == 1) { pubsubUnsubscribeAllPatterns(c,1); } else { int j; for (j = 1; j < c->argc; j++) pubsubUnsubscribePattern(c,c->argv[j],1); } if (clientSubscriptionsCount(c) == 0) c->flags &= ~CLIENT_PUBSUB; } int pubsubUnsubscribePattern(client *c, robj *pattern, int notify) { dictEntry *de; list *clients; listNode *ln; pubsubPattern pat; int retval = 0; incrRefCount(pattern); if ((ln = listSearchKey(c->pubsub_patterns,pattern)) != NULL) { retval = 1; listDelNode(c->pubsub_patterns,ln); pat.client = c; pat.pattern = pattern; ln = listSearchKey(server.pubsub_patterns,&pat); listDelNode(server.pubsub_patterns,ln); // 在pubsub_patterns链表中查找这个客户端并删除。 de = dictFind(server.pubsub_patterns_dict,pattern); serverAssertWithInfo(c,NULL,de != NULL); clients = dictGetVal(de); ln = listSearchKey(clients,c); // 查找pubsub_patterns_dict字典,找到客户端链表,在链表中删除这个客户端节点 serverAssertWithInfo(c,NULL,ln != NULL); listDelNode(clients,ln); if (listLength(clients) == 0) { // 如果删除节点后的链表长度为0,把键pattern也从字典中删除。 dictDelete(server.pubsub_patterns_dict,pattern); } } if (notify) addReplyPubsubPatUnsubscribed(c,pattern); decrRefCount(pattern); return retval; }发送消息 —— PUBLISH命令的实现
客户端执行PUBLISH channel message命令时,Redis服务器将执行如下两个重要 *** 作:
首先,查找pubsub_channels字典,将消息发送给这个频道的所有订阅者。其次,查找pubsub_patterns_dict字典,找出所有匹配这个频道的模式,再将消息发送给这些模式的订阅者。
发布消息的源码实现参考publishCommand
void publishCommand(client *c) { int receivers = pubsubPublishMessage(c->argv[1],c->argv[2]); if (server.cluster_enabled) clusterPropagatePublish(c->argv[1],c->argv[2]); else forceCommandPropagation(c,PROPAGATE_REPL); addReplyLongLong(c,receivers); } int pubsubPublishMessage(robj *channel, robj *message) { int receivers = 0; dictEntry *de; dictIterator *di; listNode *ln; listIter li; // 1、先查找频道字典`pubsub_channels` ,将消息发送给这个频道的所有订阅者 de = dictFind(server.pubsub_channels,channel); if (de) { list *list = dictGetVal(de); listNode *ln; listIter li; listRewind(list,&li); while ((ln = listNext(&li)) != NULL) { client *c = ln->value; addReplyPubsubMessage(c,channel,message); // 将消息发送给这个频道的所有订阅者 receivers++; } } // 2、再查找模式字典pubsub_patterns_dict, 找出所有匹配这个channel的模式,将消息发送给这些模式的所有订阅者 di = dictGetIterator(server.pubsub_patterns_dict); if (di) { channel = getDecodedObject(channel); while((de = dictNext(di)) != NULL) { robj *pattern = dictGetKey(de); list *clients = dictGetVal(de); if (!stringmatchlen((char*)pattern->ptr, // 判断模式是否与频道channel匹配 sdslen(pattern->ptr), (char*)channel->ptr, sdslen(channel->ptr),0)) continue; // 遍历这个模式的客户端链表,将消息发送到所有订阅这个模式的客户端。 listRewind(clients,&li); while ((ln = listNext(&li)) != NULL) { client *c = listNodevalue(ln); addReplyPubsubPatMessage(c,pattern,channel,message); receivers++; } } decrRefCount(channel); dictReleaseIterator(di); } return receivers; }
结合源码,可以看出Redis发布订阅的特点:
发布的消息不保留在内存。也就是说,必须先有订阅者,发布者发送的消息才有意义。
订阅的模式名称支持glob风格的匹配符,示例参考:Redis 发布与订阅模式匹配符示例
为什么设计pubsub_patterns_dict字典,有什么好处?假设有N个客户端, 每个客户端订阅了1个模式;如果只使用链表,判断所有模式是否与指定频道匹配的次数固定为N次,效率较低。
由于实际场景中,存在同一个模式被多个客户端订阅的场景,所以设计pubsub_patterns_dict字典的好处在于,将这个判断次数从N次降低为最多N次,提高效率。
一次完整的Redis订阅、发布流程分析假设有1个频道订阅者和1个消息发布者,一次完整的Redis订阅、发布流程如下图所示:
观察流程图中的第5步,订阅者成功建立长连接后,会进入while死循环,每次循环调用函数cliReadReply读取一条频道的消息,所以无法再执行unsubscribe命令。redis-cli的调用栈参考如下:
(gdb) bt #0 0x00007ffff7e5d5b0 in __libc_recv (fd=3, buf=0x7fffffffa200, len=16384, flags=flags@entry=0) at ../sysdeps/unix/sysv/linux/recv.c:28 #1 0x000055555558ec0b in recv (__flags=0, __n=TCP是没有消息边界的,订阅者是如何保证正确解析发布者的消息?, __buf= , __fd= ) at /usr/include/x86_64-linux-gnu/bits/socket2.h:44 #2 redisNetRead (c=0x555555841300, buf= , bufcap= ) at net.c:61 #3 0x0000555555587702 in redisBufferRead (c=0x555555841300) at hiredis.c:881 #4 0x0000555555587a92 in main (c=0x555555841300, reply=0x7fffffffe280) at hiredis.c:954 #5 0x000055555557fadd in cliReadReply (output_raw_strings=0) at redis-cli.c:1204 #6 0x0000555555581e0d in cliSendCommand (argc=2, argv=0x7ffff780a000, repeat=0) at redis-cli.c:1361 #7 0x0000555555582006 in issueCommandRepeat (argc=2, argv=0x7ffff780a000, repeat=1) at redis-cli.c:1858 #8 0x000055555556dd8d in issueCommand (argv=0x7ffff780a000, argc= ) at redis-cli.c:2090 #9 noninteractive (argv=0x7ffff780a000, argc= ) at redis-cli.c:2090 #10 main (argc= , argv= ) at redis-cli.c:8251
考虑如下场景,发布者依次发布两条消息到服务端,假设消息内容分别为"hello" “world”。由于TCP协议是没有消息边界的,如果客户端仅通过循环调用read()的方式读取消息,可能会发生如下不符合预期的情况:
read只调用了1次,返回的是"helloworld"read调用了2次,但返回的是"hell" “oworld”,或者"hellowor" “ld”,等等。read调用了3次,依次返回"hell", “owo”, “rld”…
针对以上的TCP无消息边界问题,一般有如下3种常见的解决策略:
每次只发送固定长度的消息把消息的大小和内容一并发送使用特殊标记划分消息边界
其中,Redis采用第三种策略,即通过自定义RESP协议的方式,使用CRLF(rn)作为消息边界。
比如服务器向客户端回复消息"subscribe channel1 1",对应的TCP报文内容如下:
15:36:47.498803 IP localhost.6379 > localhost.50208: Flags [P.], seq 1:38, ack 34, win 512, options [nop,nop,TS val 3705662770 ecr 3705662770], length 37: RESP "subscribe" "channel1" "1" 0x0000: 4500 0059 d241 4000 4006 6a5b 7f00 0001 E..Y.A@.@.j[.... 0x0010: 7f00 0001 18eb c420 f57a bd66 c0c1 f5d6 .........z.f.... 0x0020: 8018 0200 fe4d 0000 0101 080a dcdf ed32 .....M.........2 0x0030: dcdf ed32 2a33 0d0a 2439 0d0a 7375 6273 ...2*3..$9..subs 0x0040: 6372 6962 650d 0a24 380d 0a63 6861 6e6e cribe..$8..chann 0x0050: 656c 310d 0a3a 310d 0a el1..:1..附:一次订阅中的连接建立,请求消息的过程抓包结果
可以看出,"subscribe channel1 1"命令在TCP报文中存储的协议内容如下:
*3rnrnsubscribernrnchannel1rn:1
协议内容的解析参考Redis协议规范(RESP) , 以下只做简单的解释:
*3 表示数组长度为3, 数组元素依次为 [“subscribe” channel1” “1”]$9 表示字符串 “subscribe” 的长度$8表示字符串 “channel1” 的长度:表示类型为整数,:1表示第三个元素为整数,取值为1
另外,客户端程序redis-cli中,通过封装函数redisGetReply获取一次完整的消息,而不是简单地循环调用read(),从而确保正确地解析服务器发送的消息。
订阅的客户端断线,服务端是怎么感知的?如何保证这个客户端的频道、模式订阅信息也被同步删除?假设1个客户端通过redis-cli订阅频道channel1之后,按下了Ctrl+C退出,此时服务器将依次发生如下 *** 作:
1、首先通过I/O多路复用器(select/poll/epoll)监听这个客户端套接字的可读事件,触发命令请求处理的回调函数readQueryFromClient。调用链aeMain -> aeProcessEvents -> readQueryFromClient -> read,调用栈参考:
(gdb) bt // 1、首先通过I/O多路复用器监听这个客户端套接字的可读事件,触发命令请求处理的回调函数readQueryFromClient #0 connSocketRead (conn=0x7fcbeae150c0, buf=0x7fcbeae7cfc5, buf_len=16384) at connection.c:182 #1 0x000056041195df03 in connRead (buf_len=, buf= , conn= ) at connection.h:152 #2 readQueryFromClient (conn=0x7fcbeae150c0) at networking.c:2026 #3 0x00005604119e3e3c in callHandler (handler= , conn=0x7fcbeae150c0) at connhelpers.h:79 #4 connSocketEventHandler (el= , fd= , clientData=0x7fcbeae150c0, mask= ) at connection.c:296 #5 0x0000560411942723 in aeProcessEvents (eventLoop=eventLoop@entry=0x7fcbeae0b480, flags=flags@entry=27) at ae.c:479 #6 0x0000560411942a5d in aeMain (eventLoop=0x7fcbeae0b480) at ae.c:539 #7 0x000056041193eed8 in main (argc= , argv=0x7ffc77472d58) at server.c:5498 (gdb) n int ret = read(conn->fd, buf, buf_len); // 2、调用read()读套接字, 发现返回值ret为0,从而感知到对端已关闭连接。 (gdb) n if (!ret) { (gdb) p ret $2 = 0
2、readQueryFromClient中,调用read()读客户端套接字,发现返回值为0,从而感知到对端已关闭连接,再调用函数freeClientAsync,异步地关闭这个客户端。freeClientAsync源码参考:
void freeClientAsync(client *c) { if (c->flags & CLIENT_CLOSE_ASAP || c->flags & CLIENT_LUA) return; c->flags |= CLIENT_CLOSE_ASAP; if (server.io_threads_num == 1) { listAddNodeTail(server.clients_to_close,c); return; } static pthread_mutex_t async_free_queue_mutex = PTHREAD_MUTEX_INITIALIZER; pthread_mutex_lock(&async_free_queue_mutex); listAddNodeTail(server.clients_to_close,c); pthread_mutex_unlock(&async_free_queue_mutex); }
可以看出,freeClientAsync的实现很简单,只是把这个客户端加入服务器链表clients_to_close, 并不立即释放这个客户端,所以说是异步地关闭。
3、在下一次事件循环中,调用freeClientsInAsyncFreeQueue从服务器链表clients_to_close取出待释放的客户端;再调用freeClient真正地释放这个客户端,接着调用pubsubUnsubscribeAllChannels保证这个客户端的频道和模式订阅信息被同步删除。调用链:aeMain -> aeProcessEvents -> beforeSleep -> freeClientsInAsyncFreeQueue -> freeClient, 调用栈参考如下:
(gdb) bt #0 dictGenericDelete (d=0x7fcbeae0d2e0, key=key@entry=0x7fcbeae0e360, nofree=nofree@entry=0) at dict.c:393 #1 0x0000560411944bdf in dictDelete (ht=, key=key@entry=0x7fcbeae0e360) at dict.c:406 #2 0x000056041198e2bf in pubsubUnsubscribeChannel (c=c@entry=0x7fcbeaf4d700, channel=0x7fcbeae0e360, notify=notify@entry=0) at pubsub.c:198 #3 0x000056041198e5bd in pubsubUnsubscribeAllChannels (c=c@entry=0x7fcbeaf4d700, notify=notify@entry=0) at pubsub.c:284 #4 0x0000560411957cf7 in freeClient (c=0x7fcbeaf4d700) at networking.c:1251 #5 0x00005604119584fd in freeClientsInAsyncFreeQueue () at networking.c:1345 #6 0x000056041194634b in beforeSleep (eventLoop= ) at server.c:2204 #7 beforeSleep (eventLoop= ) at server.c:2117 #8 0x00005604119425e8 in aeProcessEvents (eventLoop=eventLoop@entry=0x7fcbeae0b480, flags=flags@entry=27) at ae.c:443 #9 0x0000560411942a5d in aeMain (eventLoop=0x7fcbeae0b480) at ae.c:539 #10 0x000056041193eed8 in main (argc= , argv=0x7ffc77472d58) at server.c:5498
freeClientsInAsyncFreeQueue的源码实现参考如下,逻辑很简单,就是从服务器链表client_to_close中依次取出并释放所有的客户端。
int freeClientsInAsyncFreeQueue(void) { int freed = 0; listIter li; listNode *ln; listRewind(server.clients_to_close,&li); // 从clients_to_close链表中依次取出所有客户端,并释放这个客户端。 while ((ln = listNext(&li)) != NULL) { client *c = listNodevalue(ln); if (c->flags & CLIENT_PROTECTED) continue; c->flags &= ~CLIENT_CLOSE_ASAP; freeClient(c); listDelNode(server.clients_to_close,ln); freed++; } return freed; // 返回释放的客户端总数 }服务器正常关闭场景,正在订阅的客户端是怎么感知到服务器离线的?
这里仅讨论服务器正常关闭的场景,即客户端发送SHUTDOWN命令关闭服务端。此时服务端将执行如下 *** 作:
1、接受客户端的SHUTDOWN命令,调用函数shutdownCommand处理这个关闭请求。调用链aeMain -> processInputBuffer -> shutdownCommand,调用栈如下:
(gdb) bt #0 closeListeningSockets (unlink_unix_socket=1) at server.c:3809 #1 0x0000562ee4b31992 in prepareForShutdown (flags=, flags@entry=0) at server.c:3916 #2 0x0000562ee4b4bbdb in shutdownCommand (c=0x7ff3c594dd80) at db.c:1061 #3 0x0000562ee4b30701 in call (c=0x7ff3c594dd80, flags=15) at server.c:3368 #4 0x0000562ee4b311c6 in processCommand (c=c@entry=0x7ff3c594dd80) at server.c:3797 #5 0x0000562ee4b3fca4 in processCommandAndResetClient (c=c@entry=0x7ff3c594dd80) at networking.c:1895 #6 0x0000562ee4b448fa in processInputBuffer (c=0x7ff3c594dd80) at networking.c:1978 #7 0x0000562ee4bcae3c in callHandler (handler= , conn=0x7ff3c58150c0) at connhelpers.h:79 #8 connSocketEventHandler (el= , fd= , clientData=0x7ff3c58150c0, mask= ) at connection.c:296 #9 0x0000562ee4b29723 in aeProcessEvents (eventLoop=eventLoop@entry=0x7ff3c580b480, flags=flags@entry=27) at ae.c:479 #10 0x0000562ee4b29a5d in aeMain (eventLoop=0x7ff3c580b480) at ae.c:539 #11 0x0000562ee4b25ed8 in main (argc= , argv=0x7ffe9a269bf8) at server.c:5498
2、接着调用closeListeningSockets关闭listen套接字,之后再调用exit退出服务器程序。源码参考如下:
void closeListeningSockets(int unlink_unix_socket) { int j; // 调用close(), 关闭服务端的listen套接字 for (j = 0; j < server.ipfd_count; j++) close(server.ipfd[j]); for (j = 0; j < server.tlsfd_count; j++) close(server.tlsfd[j]); if (server.sofd != -1) close(server.sofd); if (server.cluster_enabled) for (j = 0; j < server.cfd_count; j++) close(server.cfd[j]); if (unlink_unix_socket && server.unixsocket) { serverLog(LL_NOTICE,"Removing the unix socket file."); unlink(server.unixsocket); } }
3、客户端通过死循环调用redisGetReply获取订阅信息,而redisGetReply最终是通过系统调用recv读取套接字数据,当服务器调用exit退出后,客户端的recv调用会返回0,从而感知到服务器离线,最终调用exit退出客户端程序。调用栈参考:
(gdb) bt #0 0x00007ffff7e5d5b0 in __libc_recv (fd=3, buf=0x7fffffffa200, len=16384, flags=flags@entry=0) at ../sysdeps/unix/sysv/linux/recv.c:28 #1 0x000055555558ec0b in recv (__flags=0, __n=调试方法参考 1、频道订阅场景,用GDB打印字典pubsub_channels 的方法参考如下:, __buf= , __fd= ) at /usr/include/x86_64-linux-gnu/bits/socket2.h:44 #2 redisNetRead (c=0x555555841300, buf= , bufcap= ) at net.c:61 #3 0x0000555555587702 in redisBufferRead (c=0x555555841300) at hiredis.c:881 #4 0x0000555555587a92 in redisGetReply (c=0x555555841300, reply=0x7fffffffe280) at hiredis.c:954 #5 0x000055555557fadd in cliReadReply (output_raw_strings=0) at redis-cli.c:1204 #6 0x0000555555581e0d in cliSendCommand (argc=2, argv=0x7ffff780a000, repeat=0) at redis-cli.c:1361 #7 0x0000555555582006 in issueCommandRepeat (argc=2, argv=0x7ffff780a000, repeat=1) at redis-cli.c:1858 #8 0x000055555556dd8d in issueCommand (argv=0x7ffff780a000, argc= ) at redis-cli.c:2090 #9 noninteractive (argv=0x7ffff780a000, argc= ) at redis-cli.c:2090 #10 main (argc= , argv= ) at redis-cli.c:8251
// 1、打印pubsub_channels字典 (gdb) p *(server.pubsub_channels) $1 = { type = 0x557a48da11602、模式订阅场景,用GDB打印模式链表pubsub_patterns和模式字典pubsub_patterns_dict的方法参考如下:, privdata = 0x0, ht = {{ table = 0x7f4f0bc0e3a0, size = 4, sizemask = 3, used = 2 // channel频道总数为2 }, { table = 0x0, size = 0, sizemask = 0, used = 0 }}, rehashidx = -1, iterators = 0 } (gdb) p server.pubsub_channels.ht[0].table[0] $2 = (dictEntry *) 0x0 (gdb) p server.pubsub_channels.ht[0].table[1] $3 = (dictEntry *) 0x7f4f0bc4b728 (gdb) p server.pubsub_channels.ht[0].table[2] $4 = (dictEntry *) 0x7f4f0bc4b7d0 (gdb) p server.pubsub_channels.ht[0].table[3] $5 = (dictEntry *) 0x0 // 1.1 打印所有的channel,此例中channel个数为2,分别是channel1, channel2 (gdb) p (char *)((robj *)(server.pubsub_channels.ht[0].table[1].key)).ptr $6 = 0x7f4f0bc0e373 "channel1" (gdb) p (char *)((robj *)(server.pubsub_channels.ht[0].table[2].key)).ptr $7 = 0x7f4f0bc0e433 "channel2" // 2、打印每个channel对应的客户端链表 (gdb) p (list *)server.pubsub_channels.ht[0].table[1].v.val $8 = (list *) 0x7f4f0bc0fa20 // 2.1 链表长度为2,表示有2个客户端正在订阅channel1频道 (gdb) p *(list *)server.pubsub_channels.ht[0].table[1].v.val $9 = { head = 0x7f4f0bc4b740, tail = 0x7f4f0bc4b7a0, dup = 0x0, free = 0x0, match = 0x0, len = 2 } // 2.2 链表长度为2,表示有2个客户端正在订阅channel2频道 (gdb) p *(list *)server.pubsub_channels.ht[0].table[2].v.val $10 = { head = 0x7f4f0bc4b7e8, tail = 0x7f4f0bc4b830, dup = 0x0, free = 0x0, match = 0x0, len = 2 }
// 打印模式链表pubsub_patterns p *(server.pubsub_patterns) $3 = {head = 0x7f499144b740, tail = 0x7f499144b7d0, dup = 0x0, free = 0x55c68a57c1103、通过tcpdump抓包,分析一次订阅中的连接建立、请求消息的过程, match = 0x55c68a57c130 , len = 4} // 链表长度为4 (gdb) p (sds)(robj *)(((struct pubsubPattern *)server.pubsub_patterns->head->value)->pattern)->ptr $12 = (sds) 0x7f499140e353 "pat1*" (gdb) p (sds)(robj *)(((struct pubsubPattern *)server.pubsub_patterns->head->next->value)->pattern)->ptr $13 = (sds) 0x7f499140e3b3 "pat2*" (gdb) p (sds)(robj *)(((struct pubsubPattern *)server.pubsub_patterns->head->next->next->value)->pattern)->ptr $14 = (sds) 0x7f499140e3d3 "pat3*" (gdb) p (sds)(robj *)(((struct pubsubPattern *)server.pubsub_patterns->head->next->next->next->value)->pattern)->ptr $15 = (sds) 0x7f499140e413 "pat3*" // 打印模式字典 (gdb) p *server.pubsub_patterns_dict $16 = {type = 0x55c68a6ce160 , privdata = 0x0, ht = {{table = 0x7f499140e380, size = 4, sizemask = 3, used = 3}, {table = 0x0, size = 0, sizemask = 0, used = 0}}, rehashidx = -1, iterators = 0} // 字典有3个key p (sds)((robj *)server.pubsub_patterns_dict->ht[0].table[1]->key)->ptr $27 = (sds) 0x7f499140e353 "pat1*" (gdb) p (sds)((robj *)server.pubsub_patterns_dict->ht[0].table[3]->key)->ptr $28 = (sds) 0x7f499140e3d3 "pat3*" (gdb) p (sds)((robj *)server.pubsub_patterns_dict->ht[0].table[3]->next->key)->ptr $29 = (sds) 0x7f499140e3b3 "pat2*"
# tcpdump tcp port 6379 -X -i lo tcpdump: verbose output suppressed, use -v or -vv for full protocol decode listening on lo, link-type EN10MB (Ethernet), capture size 262144 bytes 三次握手 # SYN 15:36:47.497450 IP localhost.50208 > localhost.6379: Flags [S], seq 3233936820, win 65495, options [mss 65495,sackOK,TS val 3705662768 ecr 0,nop,wscale 7], length 0 0x0000: 4500 003c 84bd 4000 4006 b7fc 7f00 0001 E..<..@.@....... 0x0010: 7f00 0001 c420 18eb c0c1 f5b4 0000 0000 ................ 0x0020: a002 ffd7 fe30 0000 0204 ffd7 0402 080a .....0.......... 0x0030: dcdf ed30 0000 0000 0103 0307 ...0........ # SYN + ACK 15:36:47.497456 IP localhost.6379 > localhost.50208: Flags [S.], seq 4118461797, ack 3233936821, win 65483, options [mss 65495,sackOK,TS val 3705662768 ecr 3705662768,nop,wscale 7], length 0 0x0000: 4500 003c 0000 4000 4006 3cba 7f00 0001 E..<..@.@.<..... 0x0010: 7f00 0001 18eb c420 f57a bd65 c0c1 f5b5 .........z.e.... 0x0020: a012 ffcb fe30 0000 0204 ffd7 0402 080a .....0.......... 0x0030: dcdf ed30 dcdf ed30 0103 0307 ...0...0.... # ACK 15:36:47.497462 IP localhost.50208 > localhost.6379: Flags [.], ack 1, win 512, options [nop,nop,TS val 3705662768 ecr 3705662768], length 0 0x0000: 4500 0034 84be 4000 4006 b803 7f00 0001 E..4..@.@....... 0x0010: 7f00 0001 c420 18eb c0c1 f5b5 f57a bd66 .............z.f 0x0020: 8010 0200 fe28 0000 0101 080a dcdf ed30 .....(.........0 0x0030: dcdf ed30 ...0 # 三次握手结束,订阅的长连接建立 # 客户端发送请求:"subscribe channel1",采用RESP协议 15:36:47.498719 IP localhost.50208 > localhost.6379: Flags [P.], seq 1:34, ack 1, win 512, options [nop,nop,TS val 3705662770 ecr 3705662768], length 33: RESP "subscribe" "channel1" 0x0000: 4500 0055 84bf 4000 4006 b7e1 7f00 0001 E..U..@.@....... 0x0010: 7f00 0001 c420 18eb c0c1 f5b5 f57a bd66 .............z.f 0x0020: 8018 0200 fe49 0000 0101 080a dcdf ed32 .....I.........2 0x0030: dcdf ed30 2a32 0d0a 2439 0d0a 7375 6273 ...0*2..$9..subs 0x0040: 6372 6962 650d 0a24 380d 0a63 6861 6e6e cribe..$8..chann 0x0050: 656c 310d 0a el1.. # 服务端的ACK 15:36:47.498725 IP localhost.6379 > localhost.50208: Flags [.], ack 34, win 512, options [nop,nop,TS val 3705662770 ecr 3705662770], length 0 0x0000: 4500 0034 d240 4000 4006 6a81 7f00 0001 E..4.@@.@.j..... 0x0010: 7f00 0001 18eb c420 f57a bd66 c0c1 f5d6 .........z.f.... 0x0020: 8010 0200 fe28 0000 0101 080a dcdf ed32 .....(.........2 0x0030: dcdf ed32 ...2 # 服务端回复请求:"subscribe channel1 1",采用RESP协议 15:36:47.498803 IP localhost.6379 > localhost.50208: Flags [P.], seq 1:38, ack 34, win 512, options [nop,nop,TS val 3705662770 ecr 3705662770], length 37: RESP "subscribe" "channel1" "1" 0x0000: 4500 0059 d241 4000 4006 6a5b 7f00 0001 E..Y.A@.@.j[.... 0x0010: 7f00 0001 18eb c420 f57a bd66 c0c1 f5d6 .........z.f.... 0x0020: 8018 0200 fe4d 0000 0101 080a dcdf ed32 .....M.........2 0x0030: dcdf ed32 2a33 0d0a 2439 0d0a 7375 6273 ...2*3..$9..subs 0x0040: 6372 6962 650d 0a24 380d 0a63 6861 6e6e cribe..$8..chann 0x0050: 656c 310d 0a3a 310d 0a el1..:1.. # 客户端的ACK 15:36:47.499058 IP localhost.50208 > localhost.6379: Flags [.], ack 38, win 512, options [nop,nop,TS val 3705662770 ecr 3705662770], length 0 0x0000: 4500 0034 84c0 4000 4006 b801 7f00 0001 E..4..@.@....... 0x0010: 7f00 0001 c420 18eb c0c1 f5d6 f57a bd8b .............z.. 0x0020: 8010 0200 fe28 0000 0101 080a dcdf ed32 .....(.........2 0x0030: dcdf ed32 # 以下省略客户端循环接受频道消息,四次挥手释放连接的过程......参考资料
【1】《Redis设计与实现》第18章 发布与订阅 —— 黄健宏
【2】Redis(十):pub/sub 发布订阅源码解析
【3】Redis 发布与订阅模式匹配符示例
【4】菜鸟教程 Redis 发布订阅
【5】Redis协议规范(RESP)
欢迎分享,转载请注明来源:内存溢出
评论列表(0条)