Re: Memory allocations in .suspend became very unreliable

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

On Saturday 16 January 2010, Maxim Levitsky wrote:
> On Sat, 2010-01-16 at 01:57 +0100, Rafael J. Wysocki wrote: 
> > On Saturday 16 January 2010, Maxim Levitsky wrote:
> > > On Fri, 2010-01-15 at 23:03 +0100, Rafael J. Wysocki wrote: 
> > > > On Friday 15 January 2010, Maxim Levitsky wrote:
> > > > > Hi,
> > > > 
> > > > Hi,
> > > > 
> > > > > I know that this is very controversial, because here I want to describe
> > > > > a problem in a proprietary driver that happens now in 2.6.33-rc3
> > > > > I am taking about nvidia driver.
> > > > > 
> > > > > Some time ago I did very long hibernate test and found no errors after
> > > > > more that 200 cycles.
> > > > > 
> > > > > Now I update to 2.6.33 and notice that system will hand when nvidia
> > > > > driver allocates memory is their .suspend functions. 
> > > > 
> > > > They shouldn't do that, there's no guarantee that's going to work at all.
> > > > 
> > > > > This could fail in 2.6.32 if I would run many memory hungry
> > > > > applications, but now this happens with most of memory free.
> > > > 
> > > > This sounds a little strange.  What's the requested size of the image?
> > > Don't know, but system has to be very tight on memory.
> > 
> > Can you send full dmesg, please?
> 
> I deleted it, but for this case I think that hang was somewhere else.
> This task was hand on doing forking, which probably happened even before
> the freezer.
> 
> Anyway, the problem is clear. Now __get_free_pages blocks more often,
> and can block in .suspend even if there is plenty of memory free.
> 
> I now patched nvidia to use GFP_ATOMIC _always_, and problem disappear.
> It isn't such great solution when memory is tight though....
> 
> This is going to hit hard all nvidia users...

Well, generally speaking, no driver should ever allocate memory using
GFP_KERNEL in its .suspend() routine, because that's not going to work, as you
can readily see.  So this is a NVidia bug, hands down.

Now having said that, we've been considering a change that will turn all
GFP_KERNEL allocations into GFP_NOIO during suspend/resume, so perhaps I'll
prepare a patch to do that and let's see what people think.

Rafael