Re: [Intel-gfx] [PATCH] drm/i915: Selectively enable self-reclaim

["M. Vefa Bicakci" <bicave@superonline.com>]

On 02/07/10 04:28 AM, Linus Torvalds wrote:
> On Thu, Jul 1, 2010 at 5:49 PM, Dave Airlie <airlied@gmail.com> wrote:
>>
>> RECLAIMABLE added also seems fine, of course you can't have
>> RECLAIMABLE and MOVABLE (I find this out when it oopses on boot).
> 
> Yes. They are both flags for the anti-fragmentation code, and I think
> I'll leave the decision as to whether the i915 driver should use
> __GFP_RECLAIMABLE to the people who work with and care about the
> fragmentation issues. I doubt it matters much in practice, at least
> not for the loads that the fragmentation people tend to care most
> about.
> 
>> So I suspect MOVABLE is the problem. but I don't know enough about gfp
>> flags to know what RECLAIMABLE buys us, and where it  might bite us so
>> I can test some more.
> 
> I think I'll just apply your previous tested patch - GFP_HIGHUSER
> should take care of all the flags that matter fundamentally, and then
> the reclaimable flag is really just a small detail for others to worry
> about.
> 

Dear Linus,

I have bad news regarding your fix for self-reclaim and i915.

Apparently, I haven't tried enough hibernate/thaw cycles while
initially testing your fix.

After applying your fix to 2.6.34.1 and using it for two weeks,
I noticed that every now and then I get a black screen or random
kernel errors after thawing. I thought maybe this might be the
same problem caused by d8e0902806c0bd2ccc4f6a267ff52565a3ec933b .
(It turns out that my guess was right.)

So I compiled two vanilla 2.6.34.1 kernels. One with
d8e0902806c0bd2ccc4f6a267ff52565a3ec933b reverted to get back
to pre 2.6.32.8 state, and another one with your fix applied.

Then I set up an automated process where the computer would
hibernate, and reboot at the end of the hibernation sequence
(by setting /sys/power/disk to reboot) and then thaw back.
I made this process loop at least 20 times.

The kernel with d8e0902806c0bd2ccc4f6a267ff52565a3ec933b reverted
was able to hibernate/thaw at least 40 times in one go, while
the one with your fix applied was able to hibernate/thaw at most
17 times (in two separate trials) after which it crashed during
the next thaw.

Is there anything I can do find out the correct flags to use
in addition to GFP_HIGHUSER ? Can I do something like a bisection
for the flags one by one starting from the pre 2.6.32.8 state?
If you could outline a procedure to do this, I would be glad to
follow it.

Sorry to bug you again about this problem because of incomplete
testing on my part.

Regards,

M. Vefa Bicakci