nginx动态修改upstream ngx_http_dyups_module源码分析 – 作者:stan1y

简介

nginx 动态修改upstream不reload nginx模块，ngx_http_dyups_module分析。

主要围绕https://github.com/yzprofile/ngx_http_dyups_module/blob/master/ngx_http_dyups_module.c进行分析记录下来。

开整……

在create_main_conf的时候初始化这个数组

static void * ngx_http_dyups_create_main_conf(ngx_conf_t *cf)

{

…

if (ngx_array_init(&dmcf->dy_upstreams, cf->pool, 1024, sizeof(ngx_http_dyups_srv_conf_t)) != NGX_OK)

{

return NULL;

}

…

}

ngx_http_dyups_init

在dyups init的时把upstream中的conf取出来放进去。

初始化dy_upstream链以及全局ngx_http_dyups_deleted_upstream。

static ngx_int_t ngx_http_dyups_init(ngx_conf_t *cf)

{

…

dmcf = ngx_http_conf_get_module_main_conf(cf, ngx_http_dyups_module);

umcf = ngx_http_conf_get_module_main_conf(cf, ngx_http_upstream_module);

uscfp = umcf->upstreams.elts;

for (i = 0; i < umcf->upstreams.nelts; i++) {

duscf = ngx_array_push(&dmcf->dy_upstreams);

// 清零

ngx_memzero(duscf, sizeof(ngx_http_dyups_srv_conf_t));

duscf->pool = NULL;

// 赋值

duscf->upstream = uscfp[i];

duscf->dynamic = (uscfp[i]->port == 0

&& uscfp[i]->srv_conf && uscfp[i]->servers

&& uscfp[i]->flags & NGX_HTTP_UPSTREAM_CREATE);

duscf->deleted = 0;

// 赋值index

duscf->idx = i;

}

…

}

dyups share memory同步机制

shm初始化是在ngx_http_dyups_init_main_conf函数中实现的，同时设置了read_mesg的超时时间，并且指定了大小。

static char *ngx_http_dyups_init_main_conf(ngx_conf_t *cf, void *conf)

{

…

if (dmcf->read_msg_timeout == NGX_CONF_UNSET_MSEC) {

// 一秒一次

dmcf->read_msg_timeout = 1000;

}

if (dmcf->shm_size == NGX_CONF_UNSET_UINT) {

dmcf->shm_size = 2 * 1024 * 1024;

}

return ngx_http_dyups_init_shm(cf, conf);

…

}

static char *ngx_http_dyups_init_shm(ngx_conf_t *cf, void *conf)

{

…

shm_zone = ngx_shared_memory_add(cf, &dmcf->shm_name, dmcf->shm_size,

&ngx_http_dyups_module);

shm_zone->data = cf->pool;

// 加进去的这个名头的共享内存块的init函数会在初始化的时候统一调用

shm_zone->init = ngx_http_dyups_init_shm_zone;

…

}

static ngx_int_t ngx_http_dyups_init_shm_zone(ngx_shm_zone_t *shm_zone, void *data)

{

…

shpool = (ngx_slab_pool_t *) shm_zone->shm.addr;

sh = ngx_slab_alloc(shpool, sizeof(ngx_dyups_shctx_t));

if (sh == NULL) {

return NGX_ERROR;

}

// 全局变量，sh和shpool

ngx_dyups_global_ctx.sh = sh;

ngx_dyups_global_ctx.shpool = shpool;

// 初始化msg->queue

ngx_queue_init(&sh->msg_queue);

sh->version = 0;

sh->status = NULL;

…

}

ngx_http_dyups_init_process

该函数在启动进程时候调用，设定了一些定时器。

初始化共享内存，判断如果是非正常退出的，那么重新加载upstream配置。

static ngx_int_t ngx_http_dyups_init_process(ngx_cycle_t *cycle)

{

…

// 设定定时器来定时read msg，同步信息

timer = &ngx_dyups_global_ctx.msg_timer;

timer->handler = ngx_http_dyups_read_msg;

ngx_add_timer(timer, dmcf->read_msg_timeout);

// 拿到全局的pool和sh

shpool = ngx_dyups_global_ctx.shpool;

sh = ngx_dyups_global_ctx.sh;

ngx_shmtx_lock(&shpool->mutex);

// 初始化的时候肯定是NULL，，申请对应数量进程数的内存

if (sh->status == NULL) {

sh->status = ngx_slab_alloc_locked(shpool,

sizeof(ngx_dyups_status_t) * ccf->worker_processes);

if (sh->status == NULL) {

ngx_shmtx_unlock(&shpool->mutex);

return NGX_ERROR;

}

ngx_memzero(sh->status,

sizeof(ngx_dyups_status_t) * ccf->worker_processes);

ngx_shmtx_unlock(&shpool->mutex);

return NGX_OK;

}

ngx_shmtx_unlock(&shpool->mutex);

// 判断version，如果不是0的话，说明version已经在同步中被++了，所以是进程挂掉再被拉起来

if (sh->version != 0) {

//…

}

最核心的是ngx_http_dyups_read_msg函数，里面的是ngx_http_dyups_read_msg_locked函数

static void ngx_http_dyups_read_msg_locked(ngx_event_t *ev)

{

…

sh = ngx_dyups_global_ctx.sh;

shpool = ngx_dyups_global_ctx.shpool;

for (i = 0; i < ccf->worker_processes; i++) {

status = &sh->status[i];

if (status->pid == 0 || status->pid == ngx_pid) {

ngx_log_debug2(NGX_LOG_DEBUG_HTTP, ev->log, 0,

“[dyups] process %P update time %ui”,

status->pid, status->time);

// 遍历全部进程，将对应的pid赋值

status->pid = ngx_pid;

status->time = now;

break;

}

}

// 遍历消息队列

for (q = ngx_queue_last(&sh->msg_queue);

q != ngx_queue_sentinel(&sh->msg_queue);

q = ngx_queue_prev(q))

{

// 如果该msg的count和进程数一致，就是大家都同步过了，把这个msg删掉

if (msg->count == ccf->worker_processes) {

t = ngx_queue_next(q); ngx_queue_remove(q); q = t;

ngx_log_debug2(NGX_LOG_DEBUG_HTTP, ev->log, 0, “[dyups] destroy msg %V:%V”,&msg->name, &msg->content);

ngx_dyups_destroy_msg(shpool, msg);

continue;

}

found = 0;

for (i = 0; i < msg->count; i++) {

ngx_log_debug1(NGX_LOG_DEBUG_HTTP, ev->log, 0,”[dyups] msg pids [%P]”, msg->pid[i]);

if (msg->pid[i] == ngx_pid) {

found = 1;

break;

}

}

// 如果发现该进程了，就说明已经同步过了。

if (found) {

ngx_log_debug2(NGX_LOG_DEBUG_HTTP, ev->log, 0, “[dyups] msg %V count %ui found”, &msg->name, msg->count);

continue;

}

// 如果没发现的话，count++，pid更新

msg->pid[i] = ngx_pid;

msg->count++;

ngx_log_debug2(NGX_LOG_DEBUG_HTTP, ev->log, 0, “[dyups] msg %V count %ui”, &msg->name, msg->count);

// 取出来name和content

name = msg->name;

content = msg->content;

// 执行同步

rc = ngx_dyups_sync_cmd(pool, &name, &content, msg->flag);

if (rc != NGX_OK) {

ngx_log_error(NGX_LOG_ALERT, ev->log, 0, “[dyups] read msg error, may cause the ” “config inaccuracy,name:%V, content:%V”,&name, &content);

}

}

…

}

static ngx_int_t ngx_dyups_sync_cmd(ngx_pool_t *pool, ngx_str_t *name, ngx_str_t *content, ngx_uint_t flag)

{

…

} else if (flag == NGX_DYUPS_ADD) {

body.start = body.pos = content->data;

body.end = body.last = content->data + content->len;

body.temporary = 1;

rc = ngx_dyups_do_update(name, &body, &rv);

ngx_log_error(NGX_LOG_INFO, ngx_cycle->log, 0, “[dyups] sync add: %V rv: %V rc: %i”, name, &rv, rc);

if (rc != NGX_HTTP_OK) {

return NGX_ERROR;

}

return NGX_OK;

}

…

}

同步其他进程接受的信息，如果是当前进程处理的就要把信息添加到消息队列中。

ngx_dyups_update_upstream

ngx_int_t ngx_dyups_update_upstream(ngx_str_t *name, ngx_buf_t *buf, ngx_str_t *rv)

{

…

ngx_http_dyups_read_msg_locked(timer);

// 沙箱测试配置

status = ngx_dyups_sandbox_update(buf, rv);

if (status != NGX_HTTP_OK) {

goto finish;

}

status = ngx_dyups_do_update(name, buf, rv);

if (status == NGX_HTTP_OK) {

//把操作发到队列中去

if (ngx_http_dyups_send_msg(name, buf, NGX_DYUPS_ADD)) {

ngx_str_set(rv, “alert: update success “

“but not sync to other process”);

status = NGX_HTTP_INTERNAL_SERVER_ERROR;

}

}

…

}

ngx_http_dyups_send_msg函数分析

static ngx_int_t ngx_http_dyups_send_msg(ngx_str_t *name, ngx_buf_t *body, ngx_uint_t flag)

{

…

// 初始化整个msg，将name和body填充进去

sh->version++;

ngx_queue_insert_head(&sh->msg_queue, &msg->queue);

…

}

ngx_dyups_do_update

在update之前先find寻找对应的upstream。

static ngx_http_dyups_srv_conf_t * ngx_dyups_find_upstream(ngx_str_t *name, ngx_int_t *idx)

{

…

duscfs = dumcf->dy_upstreams.elts;

for (i = 0; i < dumcf->dy_upstreams.nelts; i++) {

duscf = &duscfs[i];

uscf = duscf->upstream;

if (uscf->host.len != name->len

|| ngx_strncasecmp(uscf->host.data, name->data, uscf->host.len)

!= 0)

{

continue;

}

*idx = i;

return duscf;

}

…

}

如果寻找到了idx赋值。

一旦发现寻找到了对应name的dy_upstream就先判断。

然后调用的是ngx_dyups_mark_upstream_delete函数

static void ngx_dyups_mark_upstream_delete(ngx_http_dyups_srv_conf_t *duscf)

{

…

// 获取umcf和uscf

uscf = duscf->upstream;

umcf = ngx_http_cycle_get_module_main_conf(ngx_cycle,

ngx_http_upstream_module);

// us获取这个dynamic upstream下的servers

us = uscf->servers->elts;

for (i = 0; i < uscf->servers->nelts; i++) {

// 标志位置1

us[i].down = 1;

#if (NGX_HTTP_UPSTREAM_CHECK)

if (us[i].addrs) {

// 关闭peer，看宏定义主要关闭健康检查的peer

ngx_http_upstream_check_delete_dynamic_peer(&uscf->host,

us[i].addrs);

}

#endif

}

// 将upstream对应的index的配置变成一个dummy配置

uscfp[duscf->idx] = &ngx_http_dyups_deleted_upstream;

#if (NGX_HTTP_UPSTREAM_RBTREE)

ngx_rbtree_delete(&umcf->rbtree, &uscf->node);

#endif

duscf->deleted = NGX_DYUPS_DELETING;

…

}

其中最重要的是check_delete_dynamic_peer

void ngx_http_upstream_check_delete_dynamic_peer(ngx_str_t *name,ngx_addr_t *peer_addr)

{

…

/* 一堆比较 找到choosen*/

chosen = &peer[i];

chosen->shm->ref–;

if (chosen->shm->ref <= 0 && chosen->shm->delete != PEER_DELETED) {

ngx_http_upstream_check_clear_dynamic_peer_shm(chosen->shm);

chosen->shm->delete = PEER_DELETED;

}

ngx_shmtx_unlock(&chosen->shm->mutex);

ngx_http_upstream_check_clear_peer(chosen);

…

}

这样子删完一次之后，再find一次，idx大概率就变成-1了，就可以进行创建

static ngx_int_t ngx_dyups_do_update(ngx_str_t *name, ngx_buf_t *buf, ngx_str_t *rv)

{

…

if (idx == -1) {

duscf = ngx_array_push(&dumcf->dy_upstreams);

uscfp = ngx_array_push(&umcf->upstreams);

ngx_memzero(duscf, sizeof(ngx_http_dyups_srv_conf_t));

// 这里为了获取在umcf中的新upstream的index值。

idx = umcf->upstreams.nelts – 1;

}

duscf->idx = idx;

rc = ngx_dyups_init_upstream(duscf, name, idx);

rc = ngx_dyups_add_server(duscf, buf);

…

}

最重要的就是init_upstream和add_server。

首先是init ipstream，他的传参其实是dy_srv_conf_t。upstream的name，以及upstream链表中对应的index

static ngx_int_t ngx_dyups_init_upstream(ngx_http_dyups_srv_conf_t *duscf, ngx_str_t *name, ngx_uint_t index)

{

…

umcf = ngx_http_cycle_get_module_main_conf(ngx_cycle,

ngx_http_upstream_module);

uscfp = umcf->upstreams.elts;

/*初始化uscf 也就是upstream的各个结构体*/

uscfp[index] = uscf; // 赋值

duscf->dynamic = 1;

duscf->upstream = uscf;

ctx = ngx_pcalloc(duscf->pool, sizeof(ngx_http_conf_ctx_t));

// 存放ctx

duscf->ctx = ctx;

// insert进去uscf

uscf->node.key = ngx_crc32_short(uscf->host.data, uscf->host.len);

ngx_rbtree_insert(&umcf->rbtree, &uscf->node);

…

}

static ngx_int_t ngx_dyups_add_server(ngx_http_dyups_srv_conf_t *duscf, ngx_buf_t *buf)

{

…

ngx_dyups_parse_upstream(&cf, buf)

…

}

static char * ngx_dyups_parse_upstream(ngx_conf_t *cf, ngx_buf_t *buf)

{

…

b = *buf;

ngx_memzero(&conf_file, sizeof(ngx_conf_file_t));

conf_file.file.fd = NGX_INVALID_FILE;

conf_file.buffer = &b;

cf->conf_file = &conf_file;

return ngx_conf_parse(cf, NULL);

…

}

ngx_dyups_do_delete

static ngx_int_t ngx_dyups_do_delete(ngx_str_t *name, ngx_str_t *rv)

{

…

duscf = ngx_dyups_find_upstream(name, &dumy);

// 如果查出来的是NULL，或者是一个已经被标记删除，或者彻底删除的，就说明要删这个有异常

if (duscf == NULL || duscf->deleted) {

ngx_log_error(NGX_LOG_DEBUG, ngx_cycle->log, 0, “[dyups] not find upstream %V %p”, name, duscf);

ngx_str_set(rv, “not found uptream”);

return NGX_HTTP_NOT_FOUND;

}

// 没问题的话就执行正常删除

ngx_dyups_mark_upstream_delete(duscf);

…

}

ngx_dyups_find_upstream

find upstream做了很多事还做了一部分的删除操作。

static ngx_http_dyups_srv_conf_t * ngx_dyups_find_upstream(ngx_str_t *name, ngx_int_t *idx)

{

…

dumcf = ngx_http_cycle_get_module_main_conf(ngx_cycle, ngx_http_dyups_module);

duscfs = dumcf->dy_upstreams.elts;

for (i = 0; i < dumcf->dy_upstreams.nelts; i++) {

// 这里是在mark_upstream中被标记的

if (duscf->deleted == NGX_DYUPS_DELETING) {

// 确认可以删除，主要看这个ref的引用计数

if (*(duscf->ref) == 0) {

ngx_log_error(NGX_LOG_INFO, ngx_cycle->log, 0, “[dyups] free dynamic upstream in find upstream” ” %ui”, duscf->idx);

duscf->deleted = NGX_DYUPS_DELETED;

if (duscf->pool) {

ngx_destroy_pool(duscf->pool);

duscf->pool = NULL;

}

}

}

// 如果是deleted或者是deleting，算是没有找到，除非遍历完没找到返回一个deleted。

if (duscf->deleted == NGX_DYUPS_DELETING) {

continue;

}

if (duscf->deleted == NGX_DYUPS_DELETED) {

*idx = i;

duscf_del = duscf;

continue;

}

// 如果找到了就正常返回

if (uscf->host.len != name->len

|| ngx_strncasecmp(uscf->host.data, name->data, uscf->host.len)

!= 0)

{

continue;

}

*idx = i;

return duscf;

}

…

}

引用计数ref

dyups的peer init get和free函数。

static ngx_int_t ngx_http_dyups_init_peer(ngx_http_request_t *r,

ngx_http_upstream_srv_conf_t *us)

{

…

// 设置上下文

ctx = ngx_pcalloc(r->pool, sizeof(ngx_http_dyups_ctx_t));

if (ctx == NULL) {

return NGX_ERROR;

}

// scf指向对应的dscf，ctx的data指向了自己

ctx->scf = dscf;

ctx->data = r->upstream->peer.data;

ctx->get = r->upstream->peer.get;

ctx->free = r->upstream->peer.free;

r->upstream->peer.data = ctx;

r->upstream->peer.get = ngx_http_dyups_get_peer;

r->upstream->peer.free = ngx_http_dyups_free_peer;

// 与客户端的连接的这个pool注册一个销毁函数

cln = ngx_pool_cleanup_add(r->pool, 0);

if (cln == NULL) {

return NGX_ERROR;

}

// 引用计数加一

dscf->ref++;

// 等调用这个就会将ref–。

cln->handler = ngx_http_dyups_clean_request;

cln->data = &dscf->ref;

…

}

// 这个函数是在ngx_dyups_add_server初始化upstream中被赋值的

uscf->peer.init = ngx_http_dyups_init_peer;

// 调用的位置是在这里

static void ngx_http_upstream_init_request(ngx_http_request_t *r)

{

…

if (uscf->peer.init(r, uscf) != NGX_OK) {

ngx_http_upstream_finalize_request(r, u,

NGX_HTTP_INTERNAL_SERVER_ERROR);

return;

}

ngx_http_upstream_connect(r, u);

…

}

static ngx_int_t ngx_http_dyups_get_peer(ngx_peer_connection_t *pc, void *data)

{

ngx_http_dyups_ctx_t  *ctx = data;

// 就用之前的peer get来。

return ctx->get(pc, ctx->data);

}

static void ngx_http_dyups_free_peer(ngx_peer_connection_t *pc, void *data,

ngx_uint_t state)

{

ngx_http_dyups_ctx_t  *ctx = data;

ngx_pool_cleanup_t  *cln;

/* upstream connect failed */

if (pc->connection == NULL) {

goto done;

}

if (pc->cached) {

goto done;

}

// free的时候先给ref++，等调用handler后ref–

ctx->scf->ref++;

// pool设置一个销毁pool的数据结构，赋值给pool->cleanup

cln = ngx_pool_cleanup_add(pc->connection->pool, 0);

if (cln == NULL) {

ngx_log_error(NGX_LOG_ERR, ngx_cycle->log, 0, “[dyups] dynamic upstream free peer may cause memleak %i”,ctx->scf->ref);

goto done;

}

// 销毁的递归函数

cln->handler = ngx_http_dyups_clean_request;

cln->data = &ctx->scf->ref;

done:

// 结束后调用之前保存的free

ctx->free(pc, ctx->data, state);

}

static void ngx_http_dyups_clean_request(void *data)

{

ngx_uint_t  *ref = data;

(*ref)–;

ngx_log_debug1(NGX_LOG_DEBUG_HTTP, ngx_cycle->log, 0, “[dyups] http clean request count %i”, *ref);

}

pool在被destroy的时候，会调用这个handler将引用计数减掉。

void ngx_destroy_pool(ngx_pool_t *pool)

{

ngx_pool_t          *p, *n;

ngx_pool_large_t    *l;

ngx_pool_cleanup_t  *c;

for (c = pool->cleanup; c; c = c->next) {

if (c->handler) {

ngx_log_debug1(NGX_LOG_DEBUG_ALLOC, pool->log, 0,

“run cleanup: %p”, c);

c->handler(c->data);

}

}

…

}

END

来源：freebuf.com 2020-12-01 00:46:40 by: stan1y