Re: Re: [RFC][PATCH -mm 5/6] Freezer: Use freezing timeout more efficiently

[Oleg Nesterov <oleg@tv-sign.ru>]

On 07/10, Rafael J. Wysocki wrote:
>
> On Tuesday, 10 July 2007 19:17, Oleg Nesterov wrote:
> 
> > Just watch the "todo", it it doesn't decrease during BREAK_TIMEOUT, then retry.
> > No?
> 
> No, it is computed from the start in every interation and I compare the result
> from the previous iteration with the current result.  Still, I don't compare 'todo'
> with 'todo from the previous step' directly, because the lock up can only
> happen if 'todo' is equal to 'blocking'.  For this reason, it is sufficient to
> compare 'blocking' with 'prev_blocking' (ie. 'blocking' from the previous
> step) and 'todo' with 'blocking' (the test if 'todo' > 0 is not essential;
> well, I should remove it in accordance with my signature ;-)).

Yes, I see what the code does. My question was: why it is not good enough to just
compare 'todo' with 'todo from the previous step' ? If 'todo' does not decrease
during the BREAK_TIMEOUT period, probably we should retry. A fork() from user-space
should not succeed because we send a fake signal to all tasks.

> > > +	if (todo && freeze_user_space && !time_after(jiffies, end_time)) {
> > > +		/*
> > > +		 * Some tasks have not been able to freeze.  They might be stuck
> > > +		 * in TASK_UNINTERRUPTIBLE waiting for the frozen tasks.  Try to
> > > +		 * thaw the tasks that have frozen without clearing the freeze
> > > +		 * requests of the remaining tasks and repeat.
> > > +		 */
> > 
> > Another question, why do we check "freeze_user_space" here?
> 
> Well, the assumption is that the freezable kernel threads won't lock up the
> freezer in such a way (that may be overoptimistic, but I'd like to make it
> for now).

OK, understand.

Oleg.