merge r3294, r3305:

Fix a bug introduced in r2032: After a child process has read a terminate
message from a channel, the process tries to read the channel again.
The kernel (at least FreeBSD) may preempt the process and sends a SIGIO
signal to a master process. The master process sends a new terminate message,
the kernel switches again to the the child process, and the child process
reads the messages instead of an EAGAIN error. And this may repeat over
and over. Being that the child process can not exit the cycle and test
the termination flag set by the message handler.

The fix disallow the master process to send a new terminate message on
SIGIO signal reception. It may send the message only on SIGALARM signal.
This commit is contained in:
Igor Sysoev 2010-02-01 15:49:36 +00:00
parent 95a5ecb1ac
commit c419414db4

View File

@ -86,7 +86,7 @@ ngx_master_process_cycle(ngx_cycle_t *cycle)
u_char *p;
size_t size;
ngx_int_t i;
ngx_uint_t n;
ngx_uint_t n, sigio;
sigset_t set;
struct itimerval itv;
ngx_uint_t live;
@ -139,11 +139,13 @@ ngx_master_process_cycle(ngx_cycle_t *cycle)
ngx_new_binary = 0;
delay = 0;
sigio = 0;
live = 1;
for ( ;; ) {
if (delay) {
if (ngx_sigalrm) {
sigio = 0;
delay *= 2;
ngx_sigalrm = 0;
}
@ -168,7 +170,8 @@ ngx_master_process_cycle(ngx_cycle_t *cycle)
ngx_time_update(0, 0);
ngx_log_debug0(NGX_LOG_DEBUG_EVENT, cycle->log, 0, "wake up");
ngx_log_debug1(NGX_LOG_DEBUG_EVENT, cycle->log, 0,
"wake up, sigio %i", sigio);
if (ngx_reap) {
ngx_reap = 0;
@ -186,6 +189,13 @@ ngx_master_process_cycle(ngx_cycle_t *cycle)
delay = 50;
}
if (sigio) {
sigio--;
continue;
}
sigio = ccf->worker_processes + 2 /* cache processes */;
if (delay > 1000) {
ngx_signal_worker_processes(cycle, SIGKILL);
} else {