Re: do_wait fix for 2.6.10-rc1

[Klaus Dittrich]

Sripathi Kodi wrote:

> Can you test this patch please?
>
> > --
> > Regards Klaus
>
>
> Hi Klaus,
>
> I have tested your patch against my testcase and it seems to be alright.
>
> My testcase is right here in my first post on LKML about this: 
> http://lkml.org/lkml/2004/11/5/34! The essence of the testcase is two 
> threads racing in waitpid to collect the exit status of a child. 
> Sometimes the thread that loses this race gets stuck in waitpid 
> forever unless manually interrupted. I had been able to recreate this 
> problem only on a multiprocessor machine. The essence of the fix was 
> to make the thread that loses this race not to set 'flag =1'. The 
> patch you sent me does not alter this behavior, so it works alright.
>
> Thanks and regards,
> Sripathi.

I encountered a bug in waitpid after Sripathi's patch
was applied.
See lkml
"kernel reports kill too late? Thu Nov 18 2004 - 07:36:48 EST"

Attached is a patch for do_wait() that may fix both problems.

--
Regards Klaus
--- kernel/exit.ORIG.c	2004-11-24 17:20:32.000000000 +0100
+++ kernel/exit.c	2004-11-26 17:38:36.000000000 +0100
@@ -1316,6 +1316,9 @@ static long do_wait(pid_t pid, int optio
 	DECLARE_WAITQUEUE(wait, current);
 	struct task_struct *tsk;
 	int flag, retval;
+	struct task_struct *p;
+	struct list_head *_p;
+	int ret;
 
 	add_wait_queue(&current->wait_chldexit,&wait);
 repeat:
@@ -1324,9 +1327,6 @@ repeat:
 	read_lock(&tasklist_lock);
 	tsk = current;
 	do {
-		struct task_struct *p;
-		struct list_head *_p;
-		int ret;
 
 		list_for_each(_p,&tsk->children) {
 			p = list_entry(_p,struct task_struct,sibling);
@@ -1377,10 +1377,11 @@ repeat:
 						goto end;
 					break;
 				}
-				flag = 1;
 check_continued:
-				if (!unlikely(options & WCONTINUED))
+				if (!unlikely(options & WCONTINUED)) {
+					flag = 1;
 					continue;
+				}
 				retval = wait_task_continued(
 					p, (options & WNOWAIT),
 					infop, stat_addr, ru);