view src/event/modules/ngx_select_module.c @ 4207:4fc91bae6f83

Better recheck of dead upstream servers. Previously nginx used to mark backend again as live as soon as fail_timeout passes (10s by default) since last failure. On the other hand, detecting dead backend takes up to 60s (proxy_connect_timeout) in typical situation "backend is down and doesn't respond to any packets". This resulted in suboptimal behaviour in the above situation (up to 23% of requests were directed to dead backend with default settings). More detailed description of the problem may be found here (in Russian): http://mailman.nginx.org/pipermail/nginx-ru/2011-August/042172.html Fix is to only allow one request after fail_timeout passes, and mark backend as "live" only if this request succeeds. Note that with new code backend will not be marked "live" unless "check" request is completed, and this may take a while in some specific workloads (e.g. streaming). This is believed to be acceptable.
author Maxim Dounin <mdounin@mdounin.ru>
date Wed, 12 Oct 2011 14:22:48 +0000
parents ab353d7dc182
children d620f497c50f
line wrap: on
line source


/*
 * Copyright (C) Igor Sysoev
 */


#include <ngx_config.h>
#include <ngx_core.h>
#include <ngx_event.h>


static ngx_int_t ngx_select_init(ngx_cycle_t *cycle, ngx_msec_t timer);
static void ngx_select_done(ngx_cycle_t *cycle);
static ngx_int_t ngx_select_add_event(ngx_event_t *ev, ngx_int_t event,
    ngx_uint_t flags);
static ngx_int_t ngx_select_del_event(ngx_event_t *ev, ngx_int_t event,
    ngx_uint_t flags);
static ngx_int_t ngx_select_process_events(ngx_cycle_t *cycle, ngx_msec_t timer,
    ngx_uint_t flags);
static void ngx_select_repair_fd_sets(ngx_cycle_t *cycle);
static char *ngx_select_init_conf(ngx_cycle_t *cycle, void *conf);


static fd_set         master_read_fd_set;
static fd_set         master_write_fd_set;
static fd_set         work_read_fd_set;
static fd_set         work_write_fd_set;

static ngx_int_t      max_fd;
static ngx_uint_t     nevents;

static ngx_event_t  **event_index;


static ngx_str_t    select_name = ngx_string("select");

ngx_event_module_t  ngx_select_module_ctx = {
    &select_name,
    NULL,                                  /* create configuration */
    ngx_select_init_conf,                  /* init configuration */

    {
        ngx_select_add_event,              /* add an event */
        ngx_select_del_event,              /* delete an event */
        ngx_select_add_event,              /* enable an event */
        ngx_select_del_event,              /* disable an event */
        NULL,                              /* add an connection */
        NULL,                              /* delete an connection */
        NULL,                              /* process the changes */
        ngx_select_process_events,         /* process the events */
        ngx_select_init,                   /* init the events */
        ngx_select_done                    /* done the events */
    }

};

ngx_module_t  ngx_select_module = {
    NGX_MODULE_V1,
    &ngx_select_module_ctx,                /* module context */
    NULL,                                  /* module directives */
    NGX_EVENT_MODULE,                      /* module type */
    NULL,                                  /* init master */
    NULL,                                  /* init module */
    NULL,                                  /* init process */
    NULL,                                  /* init thread */
    NULL,                                  /* exit thread */
    NULL,                                  /* exit process */
    NULL,                                  /* exit master */
    NGX_MODULE_V1_PADDING
};


static ngx_int_t
ngx_select_init(ngx_cycle_t *cycle, ngx_msec_t timer)
{
    ngx_event_t  **index;

    if (event_index == NULL) {
        FD_ZERO(&master_read_fd_set);
        FD_ZERO(&master_write_fd_set);
        nevents = 0;
    }

    if (ngx_process >= NGX_PROCESS_WORKER
        || cycle->old_cycle == NULL
        || cycle->old_cycle->connection_n < cycle->connection_n)
    {
        index = ngx_alloc(sizeof(ngx_event_t *) * 2 * cycle->connection_n,
                          cycle->log);
        if (index == NULL) {
            return NGX_ERROR;
        }

        if (event_index) {
            ngx_memcpy(index, event_index, sizeof(ngx_event_t *) * nevents);
            ngx_free(event_index);
        }

        event_index = index;
    }

    ngx_io = ngx_os_io;

    ngx_event_actions = ngx_select_module_ctx.actions;

    ngx_event_flags = NGX_USE_LEVEL_EVENT;

    max_fd = -1;

    return NGX_OK;
}


static void
ngx_select_done(ngx_cycle_t *cycle)
{
    ngx_free(event_index);

    event_index = NULL;
}


static ngx_int_t
ngx_select_add_event(ngx_event_t *ev, ngx_int_t event, ngx_uint_t flags)
{
    ngx_connection_t  *c;

    c = ev->data;

    ngx_log_debug2(NGX_LOG_DEBUG_EVENT, ev->log, 0,
                   "select add event fd:%d ev:%i", c->fd, event);

    if (ev->index != NGX_INVALID_INDEX) {
        ngx_log_error(NGX_LOG_ALERT, ev->log, 0,
                      "select event fd:%d ev:%i is already set", c->fd, event);
        return NGX_OK;
    }

    if ((event == NGX_READ_EVENT && ev->write)
        || (event == NGX_WRITE_EVENT && !ev->write))
    {
        ngx_log_error(NGX_LOG_ALERT, ev->log, 0,
                      "invalid select %s event fd:%d ev:%i",
                      ev->write ? "write" : "read", c->fd, event);
        return NGX_ERROR;
    }

    if (event == NGX_READ_EVENT) {
        FD_SET(c->fd, &master_read_fd_set);

    } else if (event == NGX_WRITE_EVENT) {
        FD_SET(c->fd, &master_write_fd_set);
    }

    if (max_fd != -1 && max_fd < c->fd) {
        max_fd = c->fd;
    }

    ev->active = 1;

    event_index[nevents] = ev;
    ev->index = nevents;
    nevents++;

    return NGX_OK;
}


static ngx_int_t
ngx_select_del_event(ngx_event_t *ev, ngx_int_t event, ngx_uint_t flags)
{
    ngx_event_t       *e;
    ngx_connection_t  *c;

    c = ev->data;

    ev->active = 0;

    if (ev->index == NGX_INVALID_INDEX) {
        return NGX_OK;
    }

    ngx_log_debug2(NGX_LOG_DEBUG_EVENT, ev->log, 0,
                   "select del event fd:%d ev:%i", c->fd, event);

    if (event == NGX_READ_EVENT) {
        FD_CLR(c->fd, &master_read_fd_set);

    } else if (event == NGX_WRITE_EVENT) {
        FD_CLR(c->fd, &master_write_fd_set);
    }

    if (max_fd == c->fd) {
        max_fd = -1;
    }

    if (ev->index < --nevents) {
        e = event_index[nevents];
        event_index[ev->index] = e;
        e->index = ev->index;
    }

    ev->index = NGX_INVALID_INDEX;

    return NGX_OK;
}


static ngx_int_t
ngx_select_process_events(ngx_cycle_t *cycle, ngx_msec_t timer,
    ngx_uint_t flags)
{
    int                ready, nready;
    ngx_err_t          err;
    ngx_uint_t         i, found;
    ngx_event_t       *ev, **queue;
    struct timeval     tv, *tp;
    ngx_connection_t  *c;

    if (max_fd == -1) {
        for (i = 0; i < nevents; i++) {
            c = event_index[i]->data;
            if (max_fd < c->fd) {
                max_fd = c->fd;
            }
        }

        ngx_log_debug1(NGX_LOG_DEBUG_EVENT, cycle->log, 0,
                       "change max_fd: %d", max_fd);
    }

#if (NGX_DEBUG)
    if (cycle->log->log_level & NGX_LOG_DEBUG_ALL) {
        for (i = 0; i < nevents; i++) {
            ev = event_index[i];
            c = ev->data;
            ngx_log_debug2(NGX_LOG_DEBUG_EVENT, cycle->log, 0,
                           "select event: fd:%d wr:%d", c->fd, ev->write);
        }

        ngx_log_debug1(NGX_LOG_DEBUG_EVENT, cycle->log, 0,
                       "max_fd: %d", max_fd);
    }
#endif

    if (timer == NGX_TIMER_INFINITE) {
        tp = NULL;

    } else {
        tv.tv_sec = (long) (timer / 1000);
        tv.tv_usec = (long) ((timer % 1000) * 1000);
        tp = &tv;
    }

    ngx_log_debug1(NGX_LOG_DEBUG_EVENT, cycle->log, 0,
                   "select timer: %M", timer);

    work_read_fd_set = master_read_fd_set;
    work_write_fd_set = master_write_fd_set;

    ready = select(max_fd + 1, &work_read_fd_set, &work_write_fd_set, NULL, tp);

    err = (ready == -1) ? ngx_errno : 0;

    if (flags & NGX_UPDATE_TIME || ngx_event_timer_alarm) {
        ngx_time_update();
    }

    ngx_log_debug1(NGX_LOG_DEBUG_EVENT, cycle->log, 0,
                   "select ready %d", ready);

    if (err) {
        ngx_uint_t  level;

        if (err == NGX_EINTR) {

            if (ngx_event_timer_alarm) {
                ngx_event_timer_alarm = 0;
                return NGX_OK;
            }

            level = NGX_LOG_INFO;

        } else {
            level = NGX_LOG_ALERT;
        }

        ngx_log_error(level, cycle->log, err, "select() failed");

        if (err == EBADF) {
            ngx_select_repair_fd_sets(cycle);
        }

        return NGX_ERROR;
    }

    if (ready == 0) {
        if (timer != NGX_TIMER_INFINITE) {
            return NGX_OK;
        }

        ngx_log_error(NGX_LOG_ALERT, cycle->log, 0,
                      "select() returned no events without timeout");
        return NGX_ERROR;
    }

    ngx_mutex_lock(ngx_posted_events_mutex);

    nready = 0;

    for (i = 0; i < nevents; i++) {
        ev = event_index[i];
        c = ev->data;
        found = 0;

        if (ev->write) {
            if (FD_ISSET(c->fd, &work_write_fd_set)) {
                found = 1;
                ngx_log_debug1(NGX_LOG_DEBUG_EVENT, cycle->log, 0,
                               "select write %d", c->fd);
            }

        } else {
            if (FD_ISSET(c->fd, &work_read_fd_set)) {
                found = 1;
                ngx_log_debug1(NGX_LOG_DEBUG_EVENT, cycle->log, 0,
                               "select read %d", c->fd);
            }
        }

        if (found) {
            ev->ready = 1;

            queue = (ngx_event_t **) (ev->accept ? &ngx_posted_accept_events:
                                                   &ngx_posted_events);
            ngx_locked_post_event(ev, queue);

            nready++;
        }
    }

    ngx_mutex_unlock(ngx_posted_events_mutex);

    if (ready != nready) {
        ngx_log_error(NGX_LOG_ALERT, cycle->log, 0,
                      "select ready != events: %d:%d", ready, nready);

        ngx_select_repair_fd_sets(cycle);
    }

    return NGX_OK;
}


static void
ngx_select_repair_fd_sets(ngx_cycle_t *cycle)
{
    int           n;
    socklen_t     len;
    ngx_err_t     err;
    ngx_socket_t  s;

    for (s = 0; s <= max_fd; s++) {

        if (FD_ISSET(s, &master_read_fd_set) == 0) {
            continue;
        }

        len = sizeof(int);

        if (getsockopt(s, SOL_SOCKET, SO_TYPE, &n, &len) == -1) {
            err = ngx_socket_errno;

            ngx_log_error(NGX_LOG_ALERT, cycle->log, err,
                          "invalid descriptor #%d in read fd_set", s);

            FD_CLR(s, &master_read_fd_set);
        }
    }

    for (s = 0; s <= max_fd; s++) {

        if (FD_ISSET(s, &master_write_fd_set) == 0) {
            continue;
        }

        len = sizeof(int);

        if (getsockopt(s, SOL_SOCKET, SO_TYPE, &n, &len) == -1) {
            err = ngx_socket_errno;

            ngx_log_error(NGX_LOG_ALERT, cycle->log, err,
                          "invalid descriptor #%d in write fd_set", s);

            FD_CLR(s, &master_write_fd_set);
        }
    }

    max_fd = -1;
}


static char *
ngx_select_init_conf(ngx_cycle_t *cycle, void *conf)
{
    ngx_event_conf_t  *ecf;

    ecf = ngx_event_get_conf(cycle->conf_ctx, ngx_event_core_module);

    if (ecf->use != ngx_select_module.ctx_index) {
        return NGX_CONF_OK;
    }

    /* disable warning: the default FD_SETSIZE is 1024U in FreeBSD 5.x */

    if (cycle->connection_n > FD_SETSIZE) {
        ngx_log_error(NGX_LOG_EMERG, cycle->log, 0,
                      "the maximum number of files "
                      "supported by select() is %ud", FD_SETSIZE);
        return NGX_CONF_ERROR;
    }

#if (NGX_THREADS)

    ngx_log_error(NGX_LOG_EMERG, cycle->log, 0,
                  "select() is not supported in the threaded mode");
    return NGX_CONF_ERROR;

#else

    return NGX_CONF_OK;

#endif
}