[PATCH] intel-gpu-tools: skip gem_mmap_offset_exhaustion on Android

[PATCH] intel-gpu-tools: skip gem_mmap_offset_exhaustion on Android

[tim.gore]

From: Tim Gore <tim.gore@intel.com>

gem_mmap_offset_exhaustion relies on purgeable memory
allocations getting swapped out, freeing up physical
memory for further allocations. On Android we have no
swap partition so this cannot happen and the test gets
killed by the low memory killer before mmap offset
exhaustion can happen, thus defeating the tests purpose.

Signed-off-by: Tim Gore <tim.gore@intel.com>
---
 tests/gem_mmap_offset_exhaustion.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/tests/gem_mmap_offset_exhaustion.c b/tests/gem_mmap_offset_exhaustion.c
index 914fe6e..016143d 100644
--- a/tests/gem_mmap_offset_exhaustion.c
+++ b/tests/gem_mmap_offset_exhaustion.c
@@ -77,6 +77,10 @@ igt_simple_main
 {
 	int fd, i;
 
+#ifdef ANDROID
+	igt_skip("Test not valid on Android\n");
+#endif
+
 	igt_skip_on_simulation();
 
 	fd = drm_open_any();
-- 
1.9.2

Re: [PATCH] intel-gpu-tools: skip gem_mmap_offset_exhaustion on Android

[Daniel Vetter]

On Fri, Jul 25, 2014 at 10:00:19AM +0100, tim.gore@intel.com wrote:
> From: Tim Gore <tim.gore@intel.com>
> 
> gem_mmap_offset_exhaustion relies on purgeable memory
> allocations getting swapped out, freeing up physical
> memory for further allocations. On Android we have no
> swap partition so this cannot happen and the test gets
> killed by the low memory killer before mmap offset
> exhaustion can happen, thus defeating the tests purpose.
> 
> Signed-off-by: Tim Gore <tim.gore@intel.com>



	/* we happily leak objects to exhaust mmap offset space, the kernel will
	 * reap backing storage. */
	gem_madvise(fd, handle, I915_MADV_DONTNEED);

There's really no way you should be able to run out of memory. I suspect
android kernel's will fall over even with swap.
-Daniel

> ---
>  tests/gem_mmap_offset_exhaustion.c | 4 ++++
>  1 file changed, 4 insertions(+)
> 
> diff --git a/tests/gem_mmap_offset_exhaustion.c b/tests/gem_mmap_offset_exhaustion.c
> index 914fe6e..016143d 100644
> --- a/tests/gem_mmap_offset_exhaustion.c
> +++ b/tests/gem_mmap_offset_exhaustion.c
> @@ -77,6 +77,10 @@ igt_simple_main
>  {
>  	int fd, i;
>  
> +#ifdef ANDROID
> +	igt_skip("Test not valid on Android\n");
> +#endif
> +
>  	igt_skip_on_simulation();
>  
>  	fd = drm_open_any();
> -- 
> 1.9.2
> 

-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

Re: [PATCH] intel-gpu-tools: skip gem_mmap_offset_exhaustion on Android

[Gore, Tim]

> -----Original Message-----
> From: Daniel Vetter [mailto:daniel.vetter@ffwll.ch] On Behalf Of Daniel
> Vetter
> Sent: Friday, July 25, 2014 10:07 AM
> To: Gore, Tim
> Cc: intel-gfx@lists.freedesktop.org; daniel.vetter@ffwll.ch
> Subject: Re: [PATCH] intel-gpu-tools: skip gem_mmap_offset_exhaustion on
> Android
> 
> On Fri, Jul 25, 2014 at 10:00:19AM +0100, tim.gore@intel.com wrote:
> > From: Tim Gore <tim.gore@intel.com>
> >
> > gem_mmap_offset_exhaustion relies on purgeable memory allocations
> > getting swapped out, freeing up physical memory for further
> > allocations. On Android we have no swap partition so this cannot
> > happen and the test gets killed by the low memory killer before mmap
> > offset exhaustion can happen, thus defeating the tests purpose.
> >
> > Signed-off-by: Tim Gore <tim.gore@intel.com>
> 
> 
> 
> 	/* we happily leak objects to exhaust mmap offset space, the kernel
> will
> 	 * reap backing storage. */
> 	gem_madvise(fd, handle, I915_MADV_DONTNEED);
> 
> There's really no way you should be able to run out of memory. I suspect
> android kernel's will fall over even with swap.
> -Daniel
> 
Well, not sure I fully understand how GEM works, but I can clearly see the free memory
Shrinking until the OOM killer steps in. Since the bo's are not destroyed, surely the only
Way for the physical memory to be reclaimed is if it gets swapped out, which Android
Wont do. Perhaps I misunderstand "purgeable". Should kswapd free such memory?
  Tim

> > ---
> >  tests/gem_mmap_offset_exhaustion.c | 4 ++++
> >  1 file changed, 4 insertions(+)
> >
> > diff --git a/tests/gem_mmap_offset_exhaustion.c
> > b/tests/gem_mmap_offset_exhaustion.c
> > index 914fe6e..016143d 100644
> > --- a/tests/gem_mmap_offset_exhaustion.c
> > +++ b/tests/gem_mmap_offset_exhaustion.c
> > @@ -77,6 +77,10 @@ igt_simple_main
> >  {
> >  	int fd, i;
> >
> > +#ifdef ANDROID
> > +	igt_skip("Test not valid on Android\n"); #endif
> > +
> >  	igt_skip_on_simulation();
> >
> >  	fd = drm_open_any();
> > --
> > 1.9.2
> >
> 
> --
> Daniel Vetter
> Software Engineer, Intel Corporation
> +41 (0) 79 365 57 48 - http://blog.ffwll.ch

Re: [PATCH] intel-gpu-tools: skip gem_mmap_offset_exhaustion on Android

[Daniel Vetter]

On Fri, Jul 25, 2014 at 09:14:51AM +0000, Gore, Tim wrote:
> 
> 
> > -----Original Message-----
> > From: Daniel Vetter [mailto:daniel.vetter@ffwll.ch] On Behalf Of Daniel
> > Vetter
> > Sent: Friday, July 25, 2014 10:07 AM
> > To: Gore, Tim
> > Cc: intel-gfx@lists.freedesktop.org; daniel.vetter@ffwll.ch
> > Subject: Re: [PATCH] intel-gpu-tools: skip gem_mmap_offset_exhaustion on
> > Android
> > 
> > On Fri, Jul 25, 2014 at 10:00:19AM +0100, tim.gore@intel.com wrote:
> > > From: Tim Gore <tim.gore@intel.com>
> > >
> > > gem_mmap_offset_exhaustion relies on purgeable memory allocations
> > > getting swapped out, freeing up physical memory for further
> > > allocations. On Android we have no swap partition so this cannot
> > > happen and the test gets killed by the low memory killer before mmap
> > > offset exhaustion can happen, thus defeating the tests purpose.
> > >
> > > Signed-off-by: Tim Gore <tim.gore@intel.com>
> > 
> > 
> > 
> > 	/* we happily leak objects to exhaust mmap offset space, the kernel
> > will
> > 	 * reap backing storage. */
> > 	gem_madvise(fd, handle, I915_MADV_DONTNEED);
> > 
> > There's really no way you should be able to run out of memory. I suspect
> > android kernel's will fall over even with swap.
> > -Daniel
> > 
> Well, not sure I fully understand how GEM works, but I can clearly see the free memory
> Shrinking until the OOM killer steps in. Since the bo's are not destroyed, surely the only
> Way for the physical memory to be reclaimed is if it gets swapped out, which Android
> Wont do. Perhaps I misunderstand "purgeable". Should kswapd free such memory?

Well that's how i915 is designed. The shrinker is officially the right
interface for the core vm to tell various other pieces in the kernel when
they should tighten up. If the lowmemkiller in android is a bit too
enthusiastic about this and starts shooting down random process even
before we run out of memory for real then that's a design-screw up.

I've never looked at it, but a possible fix might be to remove the
lowmemkiller from the shrinker and instead wire it up as a true OOM
callback.

Disabling the test because memory pressure handling on android is busted
is certainly not the right fix.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

Re: [PATCH] intel-gpu-tools: skip gem_mmap_offset_exhaustion on Android

[Chris Wilson]

On Fri, Jul 25, 2014 at 11:06:38AM +0200, Daniel Vetter wrote:
> On Fri, Jul 25, 2014 at 10:00:19AM +0100, tim.gore@intel.com wrote:
> > From: Tim Gore <tim.gore@intel.com>
> > 
> > gem_mmap_offset_exhaustion relies on purgeable memory
> > allocations getting swapped out, freeing up physical
> > memory for further allocations. On Android we have no
> > swap partition so this cannot happen and the test gets
> > killed by the low memory killer before mmap offset
> > exhaustion can happen, thus defeating the tests purpose.
> > 
> > Signed-off-by: Tim Gore <tim.gore@intel.com>
> 
> 
> 
> 	/* we happily leak objects to exhaust mmap offset space, the kernel will
> 	 * reap backing storage. */
> 	gem_madvise(fd, handle, I915_MADV_DONTNEED);
> 
> There's really no way you should be able to run out of memory. I suspect
> android kernel's will fall over even with swap.

No, it's just the android lowmemkiller hates i915 by design. The two are
more or less incompatible.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre

Re: [PATCH] intel-gpu-tools: skip gem_mmap_offset_exhaustion on Android

[Daniel Vetter]

On Fri, Jul 25, 2014 at 10:17:26AM +0100, Chris Wilson wrote:
> On Fri, Jul 25, 2014 at 11:06:38AM +0200, Daniel Vetter wrote:
> > On Fri, Jul 25, 2014 at 10:00:19AM +0100, tim.gore@intel.com wrote:
> > > From: Tim Gore <tim.gore@intel.com>
> > > 
> > > gem_mmap_offset_exhaustion relies on purgeable memory
> > > allocations getting swapped out, freeing up physical
> > > memory for further allocations. On Android we have no
> > > swap partition so this cannot happen and the test gets
> > > killed by the low memory killer before mmap offset
> > > exhaustion can happen, thus defeating the tests purpose.
> > > 
> > > Signed-off-by: Tim Gore <tim.gore@intel.com>
> > 
> > 
> > 
> > 	/* we happily leak objects to exhaust mmap offset space, the kernel will
> > 	 * reap backing storage. */
> > 	gem_madvise(fd, handle, I915_MADV_DONTNEED);
> > 
> > There's really no way you should be able to run out of memory. I suspect
> > android kernel's will fall over even with swap.
> 
> No, it's just the android lowmemkiller hates i915 by design. The two are
> more or less incompatible.

Well someone should fix up the lowmemkiller then. Disabling the test is
not really fixing it.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

Re: [PATCH] intel-gpu-tools: skip gem_mmap_offset_exhaustion on Android

[Tvrtko Ursulin]

On 07/25/2014 10:37 AM, Daniel Vetter wrote:
> On Fri, Jul 25, 2014 at 10:17:26AM +0100, Chris Wilson wrote:
>> On Fri, Jul 25, 2014 at 11:06:38AM +0200, Daniel Vetter wrote:
>>> On Fri, Jul 25, 2014 at 10:00:19AM +0100, tim.gore@intel.com wrote:
>>>> From: Tim Gore <tim.gore@intel.com>
>>>>
>>>> gem_mmap_offset_exhaustion relies on purgeable memory
>>>> allocations getting swapped out, freeing up physical
>>>> memory for further allocations. On Android we have no
>>>> swap partition so this cannot happen and the test gets
>>>> killed by the low memory killer before mmap offset
>>>> exhaustion can happen, thus defeating the tests purpose.
>>>>
>>>> Signed-off-by: Tim Gore <tim.gore@intel.com>
>>>
>>>
>>>
>>> 	/* we happily leak objects to exhaust mmap offset space, the kernel will
>>> 	 * reap backing storage. */
>>> 	gem_madvise(fd, handle, I915_MADV_DONTNEED);
>>>
>>> There's really no way you should be able to run out of memory. I suspect
>>> android kernel's will fall over even with swap.
>>
>> No, it's just the android lowmemkiller hates i915 by design. The two are
>> more or less incompatible.
>
> Well someone should fix up the lowmemkiller then. Disabling the test is
> not really fixing it.

AFAIR lowmemorykiller is a weird thing which tries to prevent pagecache 
dropping below a configurable threshold by killing processes. If free 
pages are below a threshold _and_ page cache is below the same threshold 
it will try to kill something. I suppose this is to prevent kernel in 
purging the page cache on its own. Interactions must be quite complex 
here given that thresholds (in num pages) and oom scores triggers per 
each threshold are multiple and configurable.

I was just disabling it when running IGT (echo 0 
 >/sys/module/lowmemorykiller/parameters/minfree). It is probably 
debatable whether this is OK or not, depending how you look at IGT in 
the overall pool of testing.

Tvrtko