[PATCH 3/3] drm/ttm: make up to 90% of system memory available

[Christian König]

Am 17.11.20 um 18:19 schrieb Daniel Vetter:
> On Tue, Nov 17, 2020 at 03:06:15PM +0100, Christian König wrote:
>> Increase the ammount of system memory drivers can use to about 90% of
>> the total available.
>>
>> Signed-off-by: Christian König <christian.koenig at amd.com>
>> ---
>>   drivers/gpu/drm/ttm/ttm_bo.c | 2 +-
>>   1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
>> index a958135cb3fe..0a93df93dba4 100644
>> --- a/drivers/gpu/drm/ttm/ttm_bo.c
>> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
>> @@ -1267,7 +1267,7 @@ static int ttm_bo_global_init(void)
>>   	 * the available system memory.
>>   	 */
>>   	num_pages = (u64)si.totalram * si.mem_unit;
>> -	num_pages = (num_pages * 50 / 100) >> PAGE_SHIFT;
>> +	num_pages = (num_pages * 90 / 100) >> PAGE_SHIFT;
> I don't think this is the design we want. As long as it was set at "half
> of system memory" it was clear that a) it's a hack b) precision didn't
> matter.

Yeah, I'm also wondering what to do here exactly. In generally I would 
completely nuke that limit if possible.

> But if you go to the limit and still want to keep the "we make sure
> there's no OOM", then precision starts to matter:
> - memory hotplug and hotunplug is a thing
> - userspace can mlock, and it's configureable
> - drivers can pin_user_pages for IO and random other stuff. Some of it is
>    accounted as some subsystem specific mlock (like rdma does iirc), some
>    is just yolo or short term enough (like)
> - none of what we do here takes into considerations any interactions with
>    core mm tracking (like cgroups or numa or anything like that)

OOM is perfectly fine with me, we should just not run into an OOM killer 
situation because we call shmem_read_mapping_page_gfp() in the shrinker 
callback.

Any idea how we could guarantee that?

> If we want to drop the "half of system ram" limit (and yes that makes
> sense) I think the right design is:
>
> - we give up on the "no OOM" guarantee.
>
> - This means if you want real isolation of tasks, we need cgroups, and we
>    need to integrate ttm cgroups with system memory cgroups somehow. Unlike
>    randomly picked hardcoded limits this should work a lot more reliably
>    and be a lot more useful in practical use in the field.
>
> - This also means that drivers start to fail in interesting ways. I think
>    locking headaches are covered with all the lockdep annotations I've
>    pushed, plus some of the things I still have in-flight (I have a
>    might_alloc() annotations somewhere). That leaves validation of error
>    paths for when allocations fail. Ime a very effective way we used in
>    i915 is (ab)using EINTR restarting, which per drmIoctl uapi spec is
>    requried. We could put a debug mode into ttm_tt which randomly fails
>    with -EINTR to make sure it's all working correctly (plus anything else
>    that allocates memory), and unlike real out-of-memory injection piglit
>    and any other cts will complete without failure. Which gives us an
>    excellent metric for "does it work". Actualy OOM, even injected one,
>    tends to make stuff blow up in a way that's very hard to track and make
>    sure you've got good coverage, since all your usual tests die pretty
>    quickly.
>
> - ttm_tt needs to play fair with every other system memory user. We need
>    to register a shrinker for each ttm_tt (so usually one per device I
>    guess), which walks the lru (in shrink_count) and uses dma_resv_trylock
>    for actual shrinking. We probably want to move it to SYSTEM first for
>    that shrinker to pick up, so that there's some global fairness across
>    all ttm_tt.

I already have patches for this.

What's still missing is teaching the OOM killer which task is using the 
memory since memory referenced through the file descriptors are 
currently not accounted towards any process resources.

> - for GFP_DMA32 that means zone aware shrinkers. We've never used those,
>    because thus far i915 didn't have any big need for low memory, so we
>    haven't used this in practice. But it's supposed to be a thing.

I think we can mostly forget about GFP_DMA32, this should only be used 
for AGP and other ancient hardware.

Christian.

> It's a bit more code than the oneliner above, but I also think it's a lot
> more solid. Plus it would resolve the last big issue where i915 gem is
> fairly fundamentally different compared to ttm. For that question I think
> once Maarten's locking rework for i915 has landed and all the other ttm
> rework from you and Dave is in, we've resolved them all.
>
>
>>   	/* But for DMA32 we limit ourself to only use 2GiB maximum. */
>>   	num_dma32_pages = (u64)(si.totalram - si.totalhigh) * si.mem_unit;
>> -- 
>> 2.25.1
>>