Mercurial > hg > nginx-quic
view src/core/ngx_shmtx.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 | 0daede16d68b |
children | 3f6040cd731e 4919fb357a5d |
line wrap: on
line source
/* * Copyright (C) Igor Sysoev */ #include <ngx_config.h> #include <ngx_core.h> #if (NGX_HAVE_ATOMIC_OPS) ngx_int_t ngx_shmtx_create(ngx_shmtx_t *mtx, void *addr, u_char *name) { mtx->lock = addr; if (mtx->spin == (ngx_uint_t) -1) { return NGX_OK; } mtx->spin = 2048; #if (NGX_HAVE_POSIX_SEM) if (sem_init(&mtx->sem, 1, 0) == -1) { ngx_log_error(NGX_LOG_ALERT, ngx_cycle->log, ngx_errno, "sem_init() failed"); } else { mtx->semaphore = 1; } #endif return NGX_OK; } void ngx_shmtx_destory(ngx_shmtx_t *mtx) { #if (NGX_HAVE_POSIX_SEM) if (mtx->semaphore) { if (sem_destroy(&mtx->sem) == -1) { ngx_log_error(NGX_LOG_ALERT, ngx_cycle->log, ngx_errno, "sem_destroy() failed"); } } #endif } ngx_uint_t ngx_shmtx_trylock(ngx_shmtx_t *mtx) { ngx_atomic_uint_t val; val = *mtx->lock; return ((val & 0x80000000) == 0 && ngx_atomic_cmp_set(mtx->lock, val, val | 0x80000000)); } void ngx_shmtx_lock(ngx_shmtx_t *mtx) { ngx_uint_t i, n; ngx_atomic_uint_t val; ngx_log_debug0(NGX_LOG_DEBUG_CORE, ngx_cycle->log, 0, "shmtx lock"); for ( ;; ) { val = *mtx->lock; if ((val & 0x80000000) == 0 && ngx_atomic_cmp_set(mtx->lock, val, val | 0x80000000)) { return; } if (ngx_ncpu > 1) { for (n = 1; n < mtx->spin; n <<= 1) { for (i = 0; i < n; i++) { ngx_cpu_pause(); } val = *mtx->lock; if ((val & 0x80000000) == 0 && ngx_atomic_cmp_set(mtx->lock, val, val | 0x80000000)) { return; } } } #if (NGX_HAVE_POSIX_SEM) if (mtx->semaphore) { val = *mtx->lock; if ((val & 0x80000000) && ngx_atomic_cmp_set(mtx->lock, val, val + 1)) { ngx_log_debug1(NGX_LOG_DEBUG_CORE, ngx_cycle->log, 0, "shmtx wait %XA", val); while (sem_wait(&mtx->sem) == -1) { ngx_err_t err; err = ngx_errno; if (err != NGX_EINTR) { ngx_log_error(NGX_LOG_ALERT, ngx_cycle->log, err, "sem_wait() failed while waiting on shmtx"); break; } } ngx_log_debug0(NGX_LOG_DEBUG_CORE, ngx_cycle->log, 0, "shmtx awoke"); } continue; } #endif ngx_sched_yield(); } } void ngx_shmtx_unlock(ngx_shmtx_t *mtx) { ngx_atomic_uint_t val, old, wait; if (mtx->spin != (ngx_uint_t) -1) { ngx_log_debug0(NGX_LOG_DEBUG_CORE, ngx_cycle->log, 0, "shmtx unlock"); } for ( ;; ) { old = *mtx->lock; wait = old & 0x7fffffff; val = wait ? wait - 1 : 0; if (ngx_atomic_cmp_set(mtx->lock, old, val)) { break; } } #if (NGX_HAVE_POSIX_SEM) if (wait == 0 || !mtx->semaphore) { return; } ngx_log_debug1(NGX_LOG_DEBUG_CORE, ngx_cycle->log, 0, "shmtx wake %XA", old); if (sem_post(&mtx->sem) == -1) { ngx_log_error(NGX_LOG_ALERT, ngx_cycle->log, ngx_errno, "sem_post() failed while wake shmtx"); } #endif } #else ngx_int_t ngx_shmtx_create(ngx_shmtx_t *mtx, void *addr, u_char *name) { if (mtx->name) { if (ngx_strcmp(name, mtx->name) == 0) { mtx->name = name; return NGX_OK; } ngx_shmtx_destory(mtx); } mtx->fd = ngx_open_file(name, NGX_FILE_RDWR, NGX_FILE_CREATE_OR_OPEN, NGX_FILE_DEFAULT_ACCESS); if (mtx->fd == NGX_INVALID_FILE) { ngx_log_error(NGX_LOG_EMERG, ngx_cycle->log, ngx_errno, ngx_open_file_n " \"%s\" failed", name); return NGX_ERROR; } if (ngx_delete_file(name) == NGX_FILE_ERROR) { ngx_log_error(NGX_LOG_ALERT, ngx_cycle->log, ngx_errno, ngx_delete_file_n " \"%s\" failed", name); } mtx->name = name; return NGX_OK; } void ngx_shmtx_destory(ngx_shmtx_t *mtx) { if (ngx_close_file(mtx->fd) == NGX_FILE_ERROR) { ngx_log_error(NGX_LOG_ALERT, ngx_cycle->log, ngx_errno, ngx_close_file_n " \"%s\" failed", mtx->name); } } ngx_uint_t ngx_shmtx_trylock(ngx_shmtx_t *mtx) { ngx_err_t err; err = ngx_trylock_fd(mtx->fd); if (err == 0) { return 1; } if (err == NGX_EAGAIN) { return 0; } #if __osf__ /* Tru64 UNIX */ if (err == NGX_EACCESS) { return 0; } #endif ngx_log_abort(err, ngx_trylock_fd_n " %s failed", mtx->name); return 0; } void ngx_shmtx_lock(ngx_shmtx_t *mtx) { ngx_err_t err; err = ngx_lock_fd(mtx->fd); if (err == 0) { return; } ngx_log_abort(err, ngx_lock_fd_n " %s failed", mtx->name); } void ngx_shmtx_unlock(ngx_shmtx_t *mtx) { ngx_err_t err; err = ngx_unlock_fd(mtx->fd); if (err == 0) { return; } ngx_log_abort(err, ngx_unlock_fd_n " %s failed", mtx->name); } #endif