Re: [Bisected Regression in 2.6.35] A full tmpfs filesystem causeshibernation to hang

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

On Thursday, September 02, 2010, Rafael J. Wysocki wrote:
> On Thursday, September 02, 2010, KOSAKI Motohiro wrote:
> > > On Wednesday, September 01, 2010, KOSAKI Motohiro wrote:
> > > > > === 8< ===
> > > > > PM: Marking nosave pages: ...0009f000 - ...000100000
> > > > > PM: basic memory bitmaps created
> > > > > PM: Syncing filesystems ... done
> > > > > Freezing user space processes ... (elapsed 0.01 seconds) done.
> > > > > Freezing remaining freezable tasks ... (elapsed 0.01 seconds) done.
> > > > > PM: Preallocating image memory...
> > > > > shrink_all_memory start
> > > > > PM: shrink memory: pass=1, req:310171 reclaimed:15492 free:360936
> > > > > PM: shrink memory: pass=2, req:294679 reclaimed:28864 free:373981
> > > > > PM: shrink memory: pass=3, req:265815 reclaimed:60311 free:405374
> > > > > PM: shrink memory: pass=4, req:205504 reclaimed:97870 free:443024
> > > > > PM: shrink memory: pass=5, req:107634 reclaimed:146948 free:492141
> > > > > shrink_all_memory: req:107634 reclaimed:146948 free:492141
> > > > > PM: preallocate_image_highmem 556658 278329
> > > > > PM: preallocate_image_memory 103139 103139
> > > > > PM: preallocate_highmem_fraction 183908 556658 760831 -> 183908
> > > > > === >8 ===
> > > > 
> > > > Rafael, this log mean hibernate_preallocate_memory() has a bug.
> > > 
> > > Well, it works as designed ...
> > > 
> > > > It allocate memory as following order.
> > > >  1. preallocate_image_highmem()  (i.e. __GFP_HIGHMEM)
> > > >  2. preallocate_image_memory()   (i.e. GFP_KERNEL)
> > > >  3. preallocate_highmem_fraction (i.e. __GFP_HIGHMEM)
> > > >  4. preallocate_image_memory()   (i.e. GFP_KERNEL)
> > > > 
> > > > But, please imazine following scenario (as Vefa's scenario).
> > > >  - system has 3GB memory. 1GB is normal. 2GB is highmem.
> > > >  - all normal memory is free
> > > >  - 1.5GB memory of highmem are used for tmpfs. rest 500MB is free.
> > > 
> > > Indeed, that's a memory allocation pattern I didn't anticipate.
> > > 
> > > > At that time, hibernate_preallocate_memory() works as following.
> > > > 
> > > > 1. call preallocate_image_highmem(1GB)
> > > > 2. call preallocate_image_memory(500M)		total 1.5GB allocated
> > > > 3. call preallocate_highmem_fraction(660M)	total 2.2GB allocated
> > > > 
> > > > then, all of normal zone memory was exhaust. next preallocate_image_memory()
> > > > makes OOM, and oom_killer_disabled makes infinite loop.
> > > > (oom_killer_disabled careless is vmscan bug. I'll fix it soon)
> > > 
> > > So, it looks like the problem will go away if we check if there are any normal
> > > pages to allocate from before calling the last preallocate_image_memory()?
> > > 
> > > Like in the patch below, perhaps?
> > 
> > Looks like fine. but I have one question. hibernate_preallocate_memory() call
> > preallocate_image_memory() two times. Why do you only care latter one?
> > former one seems similar risk.
> 
> The first one is mandatory, ie. if we can't allocate the requested number of
> pages at this point, we fail the entire hibernation.  In that case the
> performance hit doesn't matter.

IOW, your patch at http://lkml.org/lkml/2010/9/2/262 is still necessary to
protect against the infinite loop in that case.

Thanks,
Rafael