Re: [PATCH -mm] swsusp: freeze user space processes first

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Sunday 05 February 2006 12:39, Rafael J. Wysocki wrote:
> On Sunday 05 February 2006 12:18, Pavel Machek wrote:
> > On Ne 05-02-06 12:11:06, Rafael J. Wysocki wrote:
> > > On Sunday 05 February 2006 11:50, Ingo Molnar wrote:
> > > > 
> > > > * Rafael J. Wysocki <rjw@sisk.pl> wrote:
> > > > 
> > > > > > The logic in that loop makes my brain burst.
> > > > > > 
> > > > > > What happens if a process does vfork();sleep(100000000)?
> > > > > 
> > > > > The freezing of processes will fail due to the timeout.
> > > > > 
> > > > > Without the if (!p->vfork_done) it would fail too, because the child 
> > > > > would be frozen and the parent would wait for the vfork completion in 
> > > > > the TASK_UNINTERRUPTIBLE state (ie. unfreezeable).  But in that case 
> > > > > we have a race between the "freezer" and the child process (ie. if the 
> > > > > child gets frozen before it completes the vfork completion, the paret 
> > > > > will be unfreezeable) which sometimes leads to a failure when it 
> > > > > should not.  [We have a test case showing this.]
> > > > 
> > > > then i'd suggest to change the vfork implementation to make this code 
> > > > freezable.
> > > 
> > > I think you are right, but I don't know how to do this.
> > > 
> > > > Nothing that userspace does should cause freezing to fail.   If it does,
> > > > we've designed things incorrectly on the kernel side. 
> > > 
> > > I tend to agree.
> > > 
> > > Generally, the problem is due to the use of completions where userland
> > > processes are waited for.  The two places I know of are the vfork
> > > implementation and the usermode helper code.
> > 
> > Can you produce userland testcase? If we have uninterruptible process for
> > days... that's a bug in kernel, suspend or not.
> 
> Sure, no problem.  [Pretty scary, no?]

Actually it's not that bad, because the parent will be killable when the child
exit()s (or gets killed).

Greetings,
Rafael