Date: Fri, 04 Jan 2002 20:14:47 +0300 From: "Vladimir B.Grebenschikov" <vova@sw.ru> To: FreeBSD-gnats-submit@freebsd.org Subject: kern/33542: a few of swapping completly freeze system with large number of processes Message-ID: <E16MXvb-0001s1-00@vbook.express.ru>
next in thread | raw e-mail | index | archive | help
>Number: 33542
>Category: kern
>Synopsis: a few of swapping completly freeze system with large number of processes
>Confidential: no
>Severity: critical
>Priority: medium
>Responsible: freebsd-bugs
>State: open
>Quarter:
>Keywords:
>Date-Required:
>Class: sw-bug
>Submitter-Id: current-users
>Arrival-Date: Fri Jan 04 09:20:01 PST 2002
>Closed-Date:
>Last-Modified:
>Originator: Vladimir B. Grebenschikov
>Release: FreeBSD 4.4-RELEASE i386
>Organization:
SWsoft
>Environment:
System: FreeBSD vbook.express.ru 4.4-RELEASE FreeBSD
>Description:
vmdaemon process (function vm_daemon() in kernel) runs through list of
processes and looks at all nested memory objects.
Kernel spends too long time ( 20-30 secs per cycle on system with 1000-2000 processes )
running through processes list in vm_daemon() (pageouting processes)
and system seems to be completly freezed (only network interrupts is servering) when
vm_daemon does not asleep.
>How-To-Repeat:
run large number of processes and reach end of real memory (until swapping begins)
>Fix:
This patch is far of ideal. But system responds on commands and seems
workable while pageout with it.
There are two sysctl variables to control pageout cycle:
vm.pageout_run_ticks - how long vmdaemon runs at once (in ticks)
vm.pageout_skip_ticks - how long vmdaemon sleeps before continue run through processes list (in ticks)
but vmdaemon will be awaken if new wakeup() called.
patch related to RELENG_4
--- sys/vm/vm_pageout.c.orig Wed Dec 26 04:28:12 2001
+++ sys/vm/vm_pageout.c Fri Jan 4 19:49:04 2002
@@ -143,6 +143,9 @@
static int defer_swap_pageouts=0;
static int disable_swap_pageouts=0;
+static int pageout_run_ticks=50;
+static int pageout_skip_ticks=50;
+
#if defined(NO_SWAPPING)
static int vm_swap_enabled=0;
static int vm_swap_idle_enabled=0;
@@ -191,6 +194,12 @@
SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
+SYSCTL_INT(_vm, OID_AUTO, pageout_run_ticks,
+ CTLFLAG_RW, &pageout_run_ticks, 0, "Pageout maximail run on single awake in ticks");
+
+SYSCTL_INT(_vm, OID_AUTO, pageout_skip_ticks,
+ CTLFLAG_RW, &pageout_skip_ticks, 0, "Pageout skip ticks after single run");
+
#define VM_PAGEOUT_PAGE_COUNT 16
int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
@@ -1433,10 +1442,13 @@
static void
vm_daemon()
{
+ static int pid = 0;
struct proc *p;
+ int timo = 0;
+ int cycleticks;
while (TRUE) {
- tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0);
+ tsleep(&vm_daemon_needed, PPAUSE, "psleep", timo);
if (vm_pageout_req_swapout) {
swapout_procs(vm_pageout_req_swapout);
vm_pageout_req_swapout = 0;
@@ -1446,7 +1458,28 @@
* process is swapped out -- deactivate pages
*/
- for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
+ cycleticks = ticks;
+
+ if (pid == 0)
+ p = allproc.lh_first;
+ else {
+ p = pfind(pid);
+ if (p) {
+ p = p->p_list.le_next;
+ if (!p) p = allproc.lh_first;
+ }
+ }
+
+ if (!p)
+ for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
+ if (p->p_pid < pid)
+ break;
+ }
+
+ if (!p)
+ p = allproc.lh_first;
+
+ for (; p != 0; p = p->p_list.le_next) {
vm_pindex_t limit, size;
/*
@@ -1479,11 +1512,26 @@
limit = 0; /* XXX */
size = vmspace_resident_count(p->p_vmspace);
+
+ if (ticks - cycleticks > pageout_run_ticks)
+ break;
+
if (limit >= 0 && size >= limit) {
vm_pageout_map_deactivate_pages(
&p->p_vmspace->vm_map, limit);
}
}
+
+ if (vm_page_count_target()) {
+ pid = (p) ? p->p_pid : 0;
+ /* sleep a bit - to give system possibility to do somet
+ timo = pageout_skip_ticks;
+ } else {
+ pid = 0;
+ /* sleep until awake */
+ timo = 0;
+ }
+
}
}
#endif
>Release-Note:
>Audit-Trail:
>Unformatted:
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-bugs" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?E16MXvb-0001s1-00>