lkml.org 
[lkml]   [2010]   [Dec]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: 2.6.35.5: hibernation broken... AGAIN
Date
On Thursday, December 02, 2010, KAMEZAWA Hiroyuki wrote:
> On Wed, 1 Dec 2010 23:23:31 +0100
> "Rafael J. Wysocki" <rjw@sisk.pl> wrote:
>
> > On Wednesday, December 01, 2010, Rafael J. Wysocki wrote:
> > > On Wednesday, December 01, 2010, KAMEZAWA Hiroyuki wrote:
> > > > On Wed, 1 Dec 2010 01:38:53 +0100
> > > > "Rafael J. Wysocki" <rjw@sisk.pl> wrote:
> > > >
> > > > > On Wednesday, December 01, 2010, Rafael J. Wysocki wrote:
> > > > > > On Tuesday, November 30, 2010, Hugh Dickins wrote:
> > > > > > > On Sat, 27 Nov 2010, Rafael J. Wysocki wrote:
> > > > > > > > On Saturday, November 27, 2010, Ondrej Zary wrote:
> > > > > ...
> > > > > > >
> > > > > > > Trivial point, I suppose, but it bothers me that PM is accumulating
> > > > > > > wrappers around wrappers around gfp_allowed_mask. Looks like
> > > > > > > clear_gfp_allowed_mask and set_gfp_allowed_mask (oddly asymmetrical)
> > > > > > > were not really what we need. How about scrapping them, and putting
> > > > > > > pm_restrict_gfp_mask() and pm_restore_gfp_mask() into page_alloc.c?
> > > > > >
> > > > > > Sure, that sounds like a good idea indeed.
> > > > >
> > > > > Below is an updated patch in which I tried to address your comments.
> > > > >
> > > > > I didn't find it very useful to couple pm_restore_gfp_mask() with the thawing
> > > > > of tasks, but nevertheless I think all of the spots where it's needed are
> > > > > covered now.
> > > > >
> > > > > The patch has only been compile-tested for now, so caveat emptor.
> > > > >
> > > >
> > > > Hmm, can't we have some error check as
> > > >
> > > > > +static gfp_t saved_gfp_mask;
> > > >
> > > > atomic_t gfp_mask_save_mode_counter;
> > > >
> > > > > +
> > > > > +void pm_restore_gfp_mask(void)
> > > > > {
> > > > > WARN_ON(!mutex_is_locked(&pm_mutex));
> > > > > - gfp_allowed_mask = mask;
> > > >
> > > > if (atomic_dec_return(&gfp_mask_save_mode_counter))
> > > > WARN_ONCE()
> > > >
> > > > > + if (saved_gfp_mask) {
> > > > > + gfp_allowed_mask = saved_gfp_mask;
> > > > > + saved_gfp_mask = 0;
> > > > > + }
> > > > > }
> > > >
> > > > > +void pm_restrict_gfp_mask(void)
> > > > > {
> > > > > - gfp_t ret = gfp_allowed_mask;
> > > > > -
> > > > > WARN_ON(!mutex_is_locked(&pm_mutex));
> > > > > - gfp_allowed_mask &= ~mask;
> > > > > - return ret;
> > > > > + saved_gfp_mask = gfp_allowed_mask;
> > > > > + gfp_allowed_mask &= ~GFP_IOFS;
> > > >
> > > > if (atomic_inc_return(&gfp_mask_save_mode_counter) > 1)
> > > > WARN_ONCE()
> > > >
> > > > or some ?
> > >
> > > What exactly would that be useful for?
> >
> > Please note that pm_restore_gfp_mask() can be legitimately called before
> > pm_restrict_gfp_mask() via the SNAPSHOT_CREATE_IMAGE hibernate ioctl, so the
> > test you're suggesting wouldn't really work.
> >
>
> Hm, I just wonder some tests not for breaking gfp_allowed_mask by
> - double call

That's easy. It's sufficient to do WARN_ON(saved_gfp_mask) at the beginning
of pm_restrict_gfp_mask(). This also takes care of the "forget to restore"
case to some extent, because pm_restrict_gfp_mask() will generally be only
called twice in a row if there's a missing pm_restore_gfp_mask() in between.

> - forget to restore

This is kind of difficult in general. You can't detect a missing
pm_restore_gfp_mask() other than by checking if pm_restrict_gfp_mask() is not
called twice in a row IMHO.

> That will be fatal.

Yes, it will.

> I'm not very interested in implementation detail

Well ...

Thanks,
Rafael


\
 
 \ /
  Last update: 2010-12-02 01:03    [from the cache]
©2003-2011 Jasper Spaans