#362 closed defect (fixed)
deadlock on win32 with accept_mutex and multiple worker_processes
| Reported by: | Ogurets Aggressive | Owned by: | |
|---|---|---|---|
| Priority: | minor | Milestone: | |
| Component: | nginx-module | Version: | |
| Keywords: | win32 worker_processes accept_mutex deadlock | Cc: | |
| uname -a: | Microsoft Windows [Version 5.2.3790] (Windows 2003 Server SP2 Enterprise x64) | ||
| nginx -V: |
nginx version: nginx/1.4.1
TLS SNI support enabled configure arguments: --with-cc=cl --builddir=objs.msvc8 --with-debug --prefix= --conf-path=conf/nginx.conf --pid-path=logs/nginx.pid --http-log-path=logs/access.log --error-log-path=logs/error.log --sbin-path=nginx.exe --http-client-body-temp-path=temp/client_body_temp --http-proxy-temp-path=temp/proxy_temp --http-fastcgi-temp-path=temp/fastcgi_temp --http-scgi-temp-path=temp/scgi_temp --http-uwsgi-temp-path=temp/uwsgi_temp --with-cc-opt=-DFD_SETSIZE=1024 --with-pcre=objs.msvc8/lib/pcre-8.32 --with-zlib=objs.msvc8/lib/zlib-1.2.7 --with-select_module --with-http_realip_module --with-http_addition_module --with-http_sub_module --with-http_dav_module --with-http_stub_status_module --with-http_flv_module --with-http_mp4_module --with-http_gunzip_module --with-http_gzip_static_module --with-http_random_index_module --with-http_secure_link_module --with-mail --with-openssl=objs.msvc8/lib/openssl-1.0.1e --with-openssl-opt=enable-tlsext --with-http_ssl_module --with-mail_ssl_module --with-ipv6 |
||
Description
To reproduce an error following things are required (all of them):
- worker_processes > 1
- accept_mutex on
- server record like this:
server {
listen 80;
server_name localhost;
location / {
proxy_pass http://SOME_IP;
proxy_buffering off;
}
}
Where "SOME_IP" is an address of another server (not sure if it matters though).
Accessing this location from a browser stucks in "waiting for answer" for a very long time.
Following workarounds seem to solve the problem:
- Set "worker_processes" to 1
or
- Set "accept_mutex" to "off"
Also, non-proxy locations (without "proxy-pass") seem to work fine under the same conditions, only "proxies" are affected.
Change History (3)
comment:1 by , 12 years ago
| Keywords: | win32 worker_processes accept_mutex added; proxy-pass worker mutex removed |
|---|---|
| Status: | new → accepted |
| Summary: | Infinite "waiting for answer" for servers with "proxy-pass" → deadlock on win32 with accept_mutex and multiple worker_processes |
comment:2 by , 12 years ago
| Resolution: | → fixed |
|---|---|
| Status: | accepted → closed |
The patch is committed as [c9fe549b127b].
Note:
See TracTickets
for help on using tickets.
This is a known issue with windows version, thanks for filing the ticket. See http://nginx.org/en/docs/windows.html:
With accept_mutex switched on this results in a deadlock if accept_mutex happens to be grabbed by a worker process which can't accept connections. The issue isn't related to a proxy_pass, but using it might increase probability of accept_mutex being obtained by a wrong worker process.
The following trivial patch completely disables accept_mutex on win32:
--- a/src/event/ngx_event.c +++ b/src/event/ngx_event.c @@ -607,6 +607,17 @@ ngx_event_process_init(ngx_cycle_t *cycl ngx_use_accept_mutex = 0; } +#if (NGX_WIN32) + + /* + * disable accept mutex on win32 as it may cause deadlock if + * grabbed by a process which can't accept connections + */ + + ngx_use_accept_mutex = 0; + +#endif + #if (NGX_THREADS) ngx_posted_events_mutex = ngx_mutex_init(cycle->log, 0); if (ngx_posted_events_mutex == NULL) {Probably this is a right way to go, at least till the underlying limitation is resolved.