Re: [PATCH 1/2] [WIP]drm/ttm: add waiter list to prevent allocation not in order

[Christian König]

Am 05.02.2018 um 09:22 schrieb Chunming Zhou:

On 2018年01月31日 18:54, Christian König wrote:

So I think preventing validation on same place is a simpler way:
process B bo's place is fpfn~lpfn, it will only try to evict LRU BOs 
in that range, while eviction, we just prevent those validation to 
this range(fpfn~lpfn), if out of this range, the 
allocation/validation still can be go on.


Any negative?
That won't work either. The most common use of fpfn~lpfn range is to 
limit a BO to visible VRAM, the other use cases are to fullfil 
hardware limitations.


So blocking this would result in blocking all normal GTT and VRAM 
allocations, adding a mutex to validate would have the same effect.
No, different effect, mutex will make the every allocation serialized. 
My approach only effects busy case, that is said, only when space is 
used up, the allocation is serialized in that range, otherwise still 
in parallel.


That would still not allow for concurrent evictions, not as worse as 
completely blocking concurrent validation but still not a good idea as 
far as I can see.


Going to give it a try today to use the ww_mutex owner to detect 
eviction of already reserved BOs.


Maybe that can be used instead,
Christian.



Regards,
David Zhou




Regards,
Christian.

Am 31.01.2018 um 11:30 schrieb Chunming Zhou:




On 2018年01月26日 22:35, Christian König wrote:
I just realized that a change I'm thinking about for a while would 
solve your problem as well, but keep concurrent allocation possible.


See ttm_mem_evict_first() unlocks the BO after evicting it:

    ttm_bo_del_from_lru(bo);
    spin_unlock(&glob->lru_lock);

    ret = ttm_bo_evict(bo, ctx);
    if (locked) {
    ttm_bo_unreserve(bo); < here
    } else {
    spin_lock(&glob->lru_lock);
    ttm_bo_add_to_lru(bo);
    spin_unlock(&glob->lru_lock);
    }

    kref_put(&bo->list_kref, ttm_bo_release_list);


The effect is that in your example process C can not only beat 
process B once, but many many times because we run into a ping/pong 
situation where B evicts resources while C moves them back in.
For ping/pong case, I want to disable busy placement for allocation 
period, only enable it for cs bo validation.




For a while now I'm thinking about dropping those reservations only 
after the original allocation succeeded.


The effect would be that process C can still beat process B 
initially, but sooner or process B would evict some resources from 
process C as well and then it can succeed with its allocation.
If it is from process C cs validation, process B still need evict 
the resource only after process C command submission completion.




The problem is for this approach to work we need to core change to 
the ww_mutexes to be able to handle this efficiently.
Yes, ww_mutex doesn't support this net lock, which easily deadlock 
without ticket and class.


So I think preventing validation on same place is a simpler way:
process B bo's place is fpfn~lpfn, it will only try to evict LRU BOs 
in that range, while eviction, we just prevent those validation to 
this range(fpfn~lpfn), if out of this range, the 
allocation/validation still can be go on.


Any negative?

Regards,
David Zhou


Regards,
Christian.

Am 26.01.2018 um 14:59 schrieb Christian König:
I know, but this has the same effect. You prevent concurrent 
allocation from happening.


What we could do is to pipeline reusing of deleted memory as well, 
this makes it less likely to cause the problem you are seeing 
because the evicting processes doesn't need to block for deleted 
BOs any more.


But that other processes can grab memory during eviction is 
intentional. Otherwise greedy processes would completely dominate 
command submission.


Regards,
Christian.

Am 26.01.2018 um 14:50 schrieb Zhou, David(ChunMing):
I don't want to prevent all, my new approach is to prevent the 
later allocation is trying and ahead of front to get the memory 
space that the front made from eviction.



发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午9:24写道：


Yes, exactly that's the problem.

See when you want to prevent a process B from allocating the 
memory process A has evicted, you need to prevent all concurrent 
allocation.


And we don't do that because it causes a major performance drop.

Regards,
Christian.

Am 26.01.2018 um 14:21 schrieb Zhou, David(ChunMing):
You patch will prevent concurrent allocation, and will result in 
allocation performance drop much.


发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午9:04写道：


Attached is what you actually want to do cleanly implemented. 
But as I said this is a NO-GO.


Regards,
Christian.

Am 26.01.2018 um 13:43 schrieb Christian König:
After my investigation, this issue should be detect of TTM 
design self, which breaks scheduling balance.

Yeah, but again. This is indented design we can't change easily.

Regards,
Christian.

Am 26.01.2018 um 13:36 schrieb Zhou, David(ChunMing):
I am off work

Re: [PATCH 1/2] [WIP]drm/ttm: add waiter list to prevent allocation not in order

[Chunming Zhou]

On 2018年01月31日 18:54, Christian König wrote:

So I think preventing validation on same place is a simpler way:
process B bo's place is fpfn~lpfn, it will only try to evict LRU BOs 
in that range, while eviction, we just prevent those validation to 
this range(fpfn~lpfn), if out of this range, the 
allocation/validation still can be go on.


Any negative?
That won't work either. The most common use of fpfn~lpfn range is to 
limit a BO to visible VRAM, the other use cases are to fullfil 
hardware limitations.


So blocking this would result in blocking all normal GTT and VRAM 
allocations, adding a mutex to validate would have the same effect.
No, different effect, mutex will make the every allocation serialized. 
My approach only effects busy case, that is said, only when space is 
used up, the allocation is serialized in that range, otherwise still in 
parallel.


Regards,
David Zhou




Regards,
Christian.

Am 31.01.2018 um 11:30 schrieb Chunming Zhou:




On 2018年01月26日 22:35, Christian König wrote:
I just realized that a change I'm thinking about for a while would 
solve your problem as well, but keep concurrent allocation possible.


See ttm_mem_evict_first() unlocks the BO after evicting it:

    ttm_bo_del_from_lru(bo);
    spin_unlock(&glob->lru_lock);

    ret = ttm_bo_evict(bo, ctx);
    if (locked) {
    ttm_bo_unreserve(bo); < here
    } else {
    spin_lock(&glob->lru_lock);
    ttm_bo_add_to_lru(bo);
    spin_unlock(&glob->lru_lock);
    }

    kref_put(&bo->list_kref, ttm_bo_release_list);


The effect is that in your example process C can not only beat 
process B once, but many many times because we run into a ping/pong 
situation where B evicts resources while C moves them back in.
For ping/pong case, I want to disable busy placement for allocation 
period, only enable it for cs bo validation.




For a while now I'm thinking about dropping those reservations only 
after the original allocation succeeded.


The effect would be that process C can still beat process B 
initially, but sooner or process B would evict some resources from 
process C as well and then it can succeed with its allocation.
If it is from process C cs validation, process B still need evict the 
resource only after process C command submission completion.




The problem is for this approach to work we need to core change to 
the ww_mutexes to be able to handle this efficiently.
Yes, ww_mutex doesn't support this net lock, which easily deadlock 
without ticket and class.


So I think preventing validation on same place is a simpler way:
process B bo's place is fpfn~lpfn, it will only try to evict LRU BOs 
in that range, while eviction, we just prevent those validation to 
this range(fpfn~lpfn), if out of this range, the 
allocation/validation still can be go on.


Any negative?

Regards,
David Zhou


Regards,
Christian.

Am 26.01.2018 um 14:59 schrieb Christian König:
I know, but this has the same effect. You prevent concurrent 
allocation from happening.


What we could do is to pipeline reusing of deleted memory as well, 
this makes it less likely to cause the problem you are seeing 
because the evicting processes doesn't need to block for deleted 
BOs any more.


But that other processes can grab memory during eviction is 
intentional. Otherwise greedy processes would completely dominate 
command submission.


Regards,
Christian.

Am 26.01.2018 um 14:50 schrieb Zhou, David(ChunMing):
I don't want to prevent all, my new approach is to prevent the 
later allocation is trying and ahead of front to get the memory 
space that the front made from eviction.



发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午9:24写道：


Yes, exactly that's the problem.

See when you want to prevent a process B from allocating the 
memory process A has evicted, you need to prevent all concurrent 
allocation.


And we don't do that because it causes a major performance drop.

Regards,
Christian.

Am 26.01.2018 um 14:21 schrieb Zhou, David(ChunMing):
You patch will prevent concurrent allocation, and will result in 
allocation performance drop much.


发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午9:04写道：


Attached is what you actually want to do cleanly implemented. But 
as I said this is a NO-GO.


Regards,
Christian.

Am 26.01.2018 um 13:43 schrieb Christian König:
After my investigation, this issue should be detect of TTM 
design self, which breaks scheduling balance.

Yeah, but again. This is indented design we can't change easily.

Regards,
Christian.

Am 26.01.2018 um 13:36 schrieb Zhou, David(ChunMing):
I am off work, so reply mail by phone, the format could not be 
text.


back to topic itself:
the problem indeed happen on amdgpu driver, someone reports me 
that application runs with two instances, the performance are 
different.
I also reproduced the issue with unit test(bo_eviction_test). 
They always think our scheduler isn't working as expected.


After my inv

Re: [PATCH 1/2] [WIP]drm/ttm: add waiter list to prevent allocation not in order

[Christian König]

So I think preventing validation on same place is a simpler way:
process B bo's place is fpfn~lpfn, it will only try to evict LRU BOs 
in that range, while eviction, we just prevent those validation to 
this range(fpfn~lpfn), if out of this range, the allocation/validation 
still can be go on.


Any negative?
That won't work either. The most common use of fpfn~lpfn range is to 
limit a BO to visible VRAM, the other use cases are to fullfil hardware 
limitations.


So blocking this would result in blocking all normal GTT and VRAM 
allocations, adding a mutex to validate would have the same effect.


Regards,
Christian.

Am 31.01.2018 um 11:30 schrieb Chunming Zhou:




On 2018年01月26日 22:35, Christian König wrote:
I just realized that a change I'm thinking about for a while would 
solve your problem as well, but keep concurrent allocation possible.


See ttm_mem_evict_first() unlocks the BO after evicting it:

    ttm_bo_del_from_lru(bo);
    spin_unlock(&glob->lru_lock);

    ret = ttm_bo_evict(bo, ctx);
    if (locked) {
    ttm_bo_unreserve(bo); < here
    } else {
    spin_lock(&glob->lru_lock);
    ttm_bo_add_to_lru(bo);
    spin_unlock(&glob->lru_lock);
    }

    kref_put(&bo->list_kref, ttm_bo_release_list);


The effect is that in your example process C can not only beat 
process B once, but many many times because we run into a ping/pong 
situation where B evicts resources while C moves them back in.
For ping/pong case, I want to disable busy placement for allocation 
period, only enable it for cs bo validation.




For a while now I'm thinking about dropping those reservations only 
after the original allocation succeeded.


The effect would be that process C can still beat process B 
initially, but sooner or process B would evict some resources from 
process C as well and then it can succeed with its allocation.
If it is from process C cs validation, process B still need evict the 
resource only after process C command submission completion.




The problem is for this approach to work we need to core change to 
the ww_mutexes to be able to handle this efficiently.
Yes, ww_mutex doesn't support this net lock, which easily deadlock 
without ticket and class.


So I think preventing validation on same place is a simpler way:
process B bo's place is fpfn~lpfn, it will only try to evict LRU BOs 
in that range, while eviction, we just prevent those validation to 
this range(fpfn~lpfn), if out of this range, the allocation/validation 
still can be go on.


Any negative?

Regards,
David Zhou


Regards,
Christian.

Am 26.01.2018 um 14:59 schrieb Christian König:
I know, but this has the same effect. You prevent concurrent 
allocation from happening.


What we could do is to pipeline reusing of deleted memory as well, 
this makes it less likely to cause the problem you are seeing 
because the evicting processes doesn't need to block for deleted BOs 
any more.


But that other processes can grab memory during eviction is 
intentional. Otherwise greedy processes would completely dominate 
command submission.


Regards,
Christian.

Am 26.01.2018 um 14:50 schrieb Zhou, David(ChunMing):
I don't want to prevent all, my new approach is to prevent the 
later allocation is trying and ahead of front to get the memory 
space that the front made from eviction.



发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午9:24写道：


Yes, exactly that's the problem.

See when you want to prevent a process B from allocating the memory 
process A has evicted, you need to prevent all concurrent allocation.


And we don't do that because it causes a major performance drop.

Regards,
Christian.

Am 26.01.2018 um 14:21 schrieb Zhou, David(ChunMing):
You patch will prevent concurrent allocation, and will result in 
allocation performance drop much.


发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午9:04写道：


Attached is what you actually want to do cleanly implemented. But 
as I said this is a NO-GO.


Regards,
Christian.

Am 26.01.2018 um 13:43 schrieb Christian König:
After my investigation, this issue should be detect of TTM 
design self, which breaks scheduling balance.

Yeah, but again. This is indented design we can't change easily.

Regards,
Christian.

Am 26.01.2018 um 13:36 schrieb Zhou, David(ChunMing):

I am off work, so reply mail by phone, the format could not be text.

back to topic itself:
the problem indeed happen on amdgpu driver, someone reports me 
that application runs with two instances, the performance are 
different.
I also reproduced the issue with unit test(bo_eviction_test). 
They always think our scheduler isn't working as expected.


After my investigation, this issue should be detect of TTM 
design self, which breaks scheduling balance.


Further, if we run containers for our gpu, container A could run 
high score, container B runs low score with same benchmark.


So this is bug that we need fix.

Regards,
David Zhou

发自坚果 Pro

Christian 

Re: [PATCH 1/2] [WIP]drm/ttm: add waiter list to prevent allocation not in order

[Chunming Zhou]

On 2018年01月26日 22:35, Christian König wrote:
I just realized that a change I'm thinking about for a while would 
solve your problem as well, but keep concurrent allocation possible.


See ttm_mem_evict_first() unlocks the BO after evicting it:

    ttm_bo_del_from_lru(bo);
    spin_unlock(&glob->lru_lock);

    ret = ttm_bo_evict(bo, ctx);
    if (locked) {
    ttm_bo_unreserve(bo); < here
    } else {
    spin_lock(&glob->lru_lock);
    ttm_bo_add_to_lru(bo);
    spin_unlock(&glob->lru_lock);
    }

    kref_put(&bo->list_kref, ttm_bo_release_list);


The effect is that in your example process C can not only beat process 
B once, but many many times because we run into a ping/pong situation 
where B evicts resources while C moves them back in.
For ping/pong case, I want to disable busy placement for allocation 
period, only enable it for cs bo validation.




For a while now I'm thinking about dropping those reservations only 
after the original allocation succeeded.


The effect would be that process C can still beat process B initially, 
but sooner or process B would evict some resources from process C as 
well and then it can succeed with its allocation.
If it is from process C cs validation, process B still need evict the 
resource only after process C command submission completion.




The problem is for this approach to work we need to core change to the 
ww_mutexes to be able to handle this efficiently.
Yes, ww_mutex doesn't support this net lock, which easily deadlock 
without ticket and class.


So I think preventing validation on same place is a simpler way:
process B bo's place is fpfn~lpfn, it will only try to evict LRU BOs in 
that range, while eviction, we just prevent those validation to this 
range(fpfn~lpfn), if out of this range, the allocation/validation still 
can be go on.


Any negative?

Regards,
David Zhou


Regards,
Christian.

Am 26.01.2018 um 14:59 schrieb Christian König:
I know, but this has the same effect. You prevent concurrent 
allocation from happening.


What we could do is to pipeline reusing of deleted memory as well, 
this makes it less likely to cause the problem you are seeing because 
the evicting processes doesn't need to block for deleted BOs any more.


But that other processes can grab memory during eviction is 
intentional. Otherwise greedy processes would completely dominate 
command submission.


Regards,
Christian.

Am 26.01.2018 um 14:50 schrieb Zhou, David(ChunMing):
I don't want to prevent all, my new approach is to prevent the later 
allocation is trying and ahead of front to get the memory space that 
the front made from eviction.



发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午9:24写道：


Yes, exactly that's the problem.

See when you want to prevent a process B from allocating the memory 
process A has evicted, you need to prevent all concurrent allocation.


And we don't do that because it causes a major performance drop.

Regards,
Christian.

Am 26.01.2018 um 14:21 schrieb Zhou, David(ChunMing):
You patch will prevent concurrent allocation, and will result in 
allocation performance drop much.


发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午9:04写道：


Attached is what you actually want to do cleanly implemented. But 
as I said this is a NO-GO.


Regards,
Christian.

Am 26.01.2018 um 13:43 schrieb Christian König:
After my investigation, this issue should be detect of TTM design 
self, which breaks scheduling balance.

Yeah, but again. This is indented design we can't change easily.

Regards,
Christian.

Am 26.01.2018 um 13:36 schrieb Zhou, David(ChunMing):

I am off work, so reply mail by phone, the format could not be text.

back to topic itself:
the problem indeed happen on amdgpu driver, someone reports me 
that application runs with two instances, the performance are 
different.
I also reproduced the issue with unit test(bo_eviction_test). 
They always think our scheduler isn't working as expected.


After my investigation, this issue should be detect of TTM design 
self, which breaks scheduling balance.


Further, if we run containers for our gpu, container A could run 
high score, container B runs low score with same benchmark.


So this is bug that we need fix.

Regards,
David Zhou

发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午6:31写道：


Am 26.01.2018 um 11:22 schrieb Chunming Zhou:
> there is a scheduling balance issue about get node like:
> a. process A allocates full memory and use it for submission.
> b. process B tries to allocates memory, will wait for process A 
BO idle in eviction.
> c. process A completes the job, process B eviction will put 
process A BO node,
> but in the meantime, process C is comming to allocate BO, whill 
directly get node successfully, and do submission,

> process B will again wait for process C BO idle.
> d. repeat the above setps, process B could be delayed much more.
>
> later allocation must not be ahead of front in same place.

Agai

Re: [PATCH 1/2] [WIP]drm/ttm: add waiter list to prevent allocation not in order

[Christian König]

I just realized that a change I'm thinking about for a while would solve 
your problem as well, but keep concurrent allocation possible.


See ttm_mem_evict_first() unlocks the BO after evicting it:

    ttm_bo_del_from_lru(bo);
    spin_unlock(&glob->lru_lock);

    ret = ttm_bo_evict(bo, ctx);
    if (locked) {
    ttm_bo_unreserve(bo); < here
    } else {
    spin_lock(&glob->lru_lock);
    ttm_bo_add_to_lru(bo);
    spin_unlock(&glob->lru_lock);
    }

    kref_put(&bo->list_kref, ttm_bo_release_list);


The effect is that in your example process C can not only beat process B 
once, but many many times because we run into a ping/pong situation 
where B evicts resources while C moves them back in.


For a while now I'm thinking about dropping those reservations only 
after the original allocation succeeded.


The effect would be that process C can still beat process B initially, 
but sooner or process B would evict some resources from process C as 
well and then it can succeed with its allocation.


The problem is for this approach to work we need to core change to the 
ww_mutexes to be able to handle this efficiently.


Regards,
Christian.

Am 26.01.2018 um 14:59 schrieb Christian König:
I know, but this has the same effect. You prevent concurrent 
allocation from happening.


What we could do is to pipeline reusing of deleted memory as well, 
this makes it less likely to cause the problem you are seeing because 
the evicting processes doesn't need to block for deleted BOs any more.


But that other processes can grab memory during eviction is 
intentional. Otherwise greedy processes would completely dominate 
command submission.


Regards,
Christian.

Am 26.01.2018 um 14:50 schrieb Zhou, David(ChunMing):
I don't want to prevent all, my new approach is to prevent the later 
allocation is trying and ahead of front to get the memory space that 
the front made from eviction.



发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午9:24写道：


Yes, exactly that's the problem.

See when you want to prevent a process B from allocating the memory 
process A has evicted, you need to prevent all concurrent allocation.


And we don't do that because it causes a major performance drop.

Regards,
Christian.

Am 26.01.2018 um 14:21 schrieb Zhou, David(ChunMing):
You patch will prevent concurrent allocation, and will result in 
allocation performance drop much.


发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午9:04写道：


Attached is what you actually want to do cleanly implemented. But as 
I said this is a NO-GO.


Regards,
Christian.

Am 26.01.2018 um 13:43 schrieb Christian König:
After my investigation, this issue should be detect of TTM design 
self, which breaks scheduling balance.

Yeah, but again. This is indented design we can't change easily.

Regards,
Christian.

Am 26.01.2018 um 13:36 schrieb Zhou, David(ChunMing):

I am off work, so reply mail by phone, the format could not be text.

back to topic itself:
the problem indeed happen on amdgpu driver, someone reports me 
that application runs with two instances, the performance are 
different.
I also reproduced the issue with unit test(bo_eviction_test). They 
always think our scheduler isn't working as expected.


After my investigation, this issue should be detect of TTM design 
self, which breaks scheduling balance.


Further, if we run containers for our gpu, container A could run 
high score, container B runs low score with same benchmark.


So this is bug that we need fix.

Regards,
David Zhou

发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午6:31写道：


Am 26.01.2018 um 11:22 schrieb Chunming Zhou:
> there is a scheduling balance issue about get node like:
> a. process A allocates full memory and use it for submission.
> b. process B tries to allocates memory, will wait for process A 
BO idle in eviction.
> c. process A completes the job, process B eviction will put 
process A BO node,
> but in the meantime, process C is comming to allocate BO, whill 
directly get node successfully, and do submission,

> process B will again wait for process C BO idle.
> d. repeat the above setps, process B could be delayed much more.
>
> later allocation must not be ahead of front in same place.

Again NAK to the whole approach.

At least with amdgpu the problem you described above never occurs
because evictions are pipelined operations. We could only block for
deleted regions to become free.

But independent of that incoming memory requests while we make 
room for

eviction are intended to be served first.

Changing that is certainly a no-go cause that would favor memory 
hungry

applications over small clients.

Regards,
Christian.

>
> Change-Id: I3daa892e50f82226c552cc008a29e55894a98f18
> Signed-off-by: Chunming Zhou 
> ---
>   drivers/gpu/drm/ttm/ttm_bo.c    | 69 
+++--

>   include/drm/ttm/ttm_bo_api.h    |  7 +
>   include/drm/ttm/ttm_bo_driver.h |  7 +
>   3 fil

Re: [PATCH 1/2] [WIP]drm/ttm: add waiter list to prevent allocation not in order

[Christian König]

I know, but this has the same effect. You prevent concurrent allocation 
from happening.


What we could do is to pipeline reusing of deleted memory as well, this 
makes it less likely to cause the problem you are seeing because the 
evicting processes doesn't need to block for deleted BOs any more.


But that other processes can grab memory during eviction is intentional. 
Otherwise greedy processes would completely dominate command submission.


Regards,
Christian.

Am 26.01.2018 um 14:50 schrieb Zhou, David(ChunMing):
I don't want to prevent all, my new approach is to prevent the later 
allocation is trying and ahead of front to get the memory space that 
the front made from eviction.



发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午9:24写道：


Yes, exactly that's the problem.

See when you want to prevent a process B from allocating the memory 
process A has evicted, you need to prevent all concurrent allocation.


And we don't do that because it causes a major performance drop.

Regards,
Christian.

Am 26.01.2018 um 14:21 schrieb Zhou, David(ChunMing):
You patch will prevent concurrent allocation, and will result in 
allocation performance drop much.


发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午9:04写道：


Attached is what you actually want to do cleanly implemented. But as 
I said this is a NO-GO.


Regards,
Christian.

Am 26.01.2018 um 13:43 schrieb Christian König:
After my investigation, this issue should be detect of TTM design 
self, which breaks scheduling balance.

Yeah, but again. This is indented design we can't change easily.

Regards,
Christian.

Am 26.01.2018 um 13:36 schrieb Zhou, David(ChunMing):

I am off work, so reply mail by phone, the format could not be text.

back to topic itself:
the problem indeed happen on amdgpu driver, someone reports me that 
application runs with two instances, the performance are different.
I also reproduced the issue with unit test(bo_eviction_test). They 
always think our scheduler isn't working as expected.


After my investigation, this issue should be detect of TTM design 
self, which breaks scheduling balance.


Further, if we run containers for our gpu, container A could run 
high score, container B runs low score with same benchmark.


So this is bug that we need fix.

Regards,
David Zhou

发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午6:31写道：


Am 26.01.2018 um 11:22 schrieb Chunming Zhou:
> there is a scheduling balance issue about get node like:
> a. process A allocates full memory and use it for submission.
> b. process B tries to allocates memory, will wait for process A 
BO idle in eviction.
> c. process A completes the job, process B eviction will put 
process A BO node,
> but in the meantime, process C is comming to allocate BO, whill 
directly get node successfully, and do submission,

> process B will again wait for process C BO idle.
> d. repeat the above setps, process B could be delayed much more.
>
> later allocation must not be ahead of front in same place.

Again NAK to the whole approach.

At least with amdgpu the problem you described above never occurs
because evictions are pipelined operations. We could only block for
deleted regions to become free.

But independent of that incoming memory requests while we make room 
for

eviction are intended to be served first.

Changing that is certainly a no-go cause that would favor memory 
hungry

applications over small clients.

Regards,
Christian.

>
> Change-Id: I3daa892e50f82226c552cc008a29e55894a98f18
> Signed-off-by: Chunming Zhou 
> ---
>   drivers/gpu/drm/ttm/ttm_bo.c    | 69 
+++--

>   include/drm/ttm/ttm_bo_api.h    |  7 +
>   include/drm/ttm/ttm_bo_driver.h |  7 +
>   3 files changed, 80 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c 
b/drivers/gpu/drm/ttm/ttm_bo.c

> index d33a6bb742a1..558ec2cf465d 100644
> --- a/drivers/gpu/drm/ttm/ttm_bo.c
> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
> @@ -841,6 +841,58 @@ static int ttm_bo_add_move_fence(struct 
ttm_buffer_object *bo,

>    return 0;
>   }
>
> +static void ttm_man_init_waiter(struct ttm_bo_waiter *waiter,
> + struct ttm_buffer_object *bo,
> + const struct ttm_place *place)
> +{
> + waiter->tbo = bo;
> + memcpy((void *)&waiter->place, (void *)place, sizeof(*place));
> + INIT_LIST_HEAD(&waiter->list);
> +}
> +
> +static void ttm_man_add_waiter(struct ttm_mem_type_manager *man,
> +    struct ttm_bo_waiter *waiter)
> +{
> + if (!waiter)
> + return;
> + spin_lock(&man->wait_lock);
> + list_add_tail(&waiter->list, &man->waiter_list);
> + spin_unlock(&man->wait_lock);
> +}
> +
> +static void ttm_man_del_waiter(struct ttm_mem_type_manager *man,
> +    struct ttm_bo_waiter *waiter)
> +{
> + if (!waiter)
> + return;
> + spin_lock(&man->wait_lock);
> + if (!list_empty(&waiter->list))
> + list_del(&wa

Re: [PATCH 1/2] [WIP]drm/ttm: add waiter list to prevent allocation not in order

[Zhou, David(ChunMing)]

I don't want to prevent all, my new approach is to prevent the later allocation 
is trying and ahead of front to get the memory space that the front made from 
eviction.



发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 下午9:24写道：

Yes, exactly that's the problem.

See when you want to prevent a process B from allocating the memory process A 
has evicted, you need to prevent all concurrent allocation.

And we don't do that because it causes a major performance drop.

Regards,
Christian.

Am 26.01.2018 um 14:21 schrieb Zhou, David(ChunMing):
You patch will prevent concurrent allocation, and will result in allocation 
performance drop much.


发自坚果 Pro

Christian K鰊ig 
 于 
2018年1月26日 下午9:04写道：

Attached is what you actually want to do cleanly implemented. But as I said 
this is a NO-GO.

Regards,
Christian.

Am 26.01.2018 um 13:43 schrieb Christian König:
After my investigation, this issue should be detect of TTM design self, which 
breaks scheduling balance.
Yeah, but again. This is indented design we can't change easily.

Regards,
Christian.

Am 26.01.2018 um 13:36 schrieb Zhou, David(ChunMing):
I am off work, so reply mail by phone, the format could not be text.

back to topic itself:
the problem indeed happen on amdgpu driver, someone reports me that application 
runs with two instances, the performance are different.
I also reproduced the issue with unit test(bo_eviction_test). They always think 
our scheduler isn't working as expected.

After my investigation, this issue should be detect of TTM design self, which 
breaks scheduling balance.

Further, if we run containers for our gpu, container A could run high score, 
container B runs low score with same benchmark.

So this is bug that we need fix.

Regards,
David Zhou


发自坚果 Pro

Christian K鰊ig 
 于 
2018年1月26日 下午6:31写道：

Am 26.01.2018 um 11:22 schrieb Chunming Zhou:
> there is a scheduling balance issue about get node like:
> a. process A allocates full memory and use it for submission.
> b. process B tries to allocates memory, will wait for process A BO idle in 
> eviction.
> c. process A completes the job, process B eviction will put process A BO node,
> but in the meantime, process C is comming to allocate BO, whill directly get 
> node successfully, and do submission,
> process B will again wait for process C BO idle.
> d. repeat the above setps, process B could be delayed much more.
>
> later allocation must not be ahead of front in same place.

Again NAK to the whole approach.

At least with amdgpu the problem you described above never occurs
because evictions are pipelined operations. We could only block for
deleted regions to become free.

But independent of that incoming memory requests while we make room for
eviction are intended to be served first.

Changing that is certainly a no-go cause that would favor memory hungry
applications over small clients.

Regards,
Christian.

>
> Change-Id: I3daa892e50f82226c552cc008a29e55894a98f18
> Signed-off-by: Chunming Zhou 
> ---
>   drivers/gpu/drm/ttm/ttm_bo.c| 69 
> +++--
>   include/drm/ttm/ttm_bo_api.h|  7 +
>   include/drm/ttm/ttm_bo_driver.h |  7 +
>   3 files changed, 80 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
> index d33a6bb742a1..558ec2cf465d 100644
> --- a/drivers/gpu/drm/ttm/ttm_bo.c
> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
> @@ -841,6 +841,58 @@ static int ttm_bo_add_move_fence(struct 
> ttm_buffer_object *bo,
>return 0;
>   }
>
> +static void ttm_man_init_waiter(struct ttm_bo_waiter *waiter,
> + struct ttm_buffer_object *bo,
> + const struct ttm_place *place)
> +{
> + waiter->tbo = bo;
> + memcpy((void *)&waiter->place, (void *)place, sizeof(*place));
> + INIT_LIST_HEAD(&waiter->list);
> +}
> +
> +static void ttm_man_add_waiter(struct ttm_mem_type_manager *man,
> +struct ttm_bo_waiter *waiter)
> +{
> + if (!waiter)
> + return;
> + spin_lock(&man->wait_lock);
> + list_add_tail(&waiter->list, &man->waiter_list);
> + spin_unlock(&man->wait_lock);
> +}
> +
> +static void ttm_man_del_waiter(struct ttm_mem_type_manager *man,
> +struct ttm_bo_waiter *waiter)
> +{
> + if (!waiter)
> + return;
> + spin_lock(&man->wait_lock);
> + if (!list_empty(&waiter->list))
> + list_del(&waiter->list);
> + spin_unlock(&man->wait_lock);
> + kfree(waiter);
> +}
> +
> +int ttm_man_check_bo(struct ttm_mem_type_manager *man,
> +  struct ttm_buffer_object *bo,
> +  const struct ttm_place *place)
> +{
> + struct ttm_bo_waiter *waiter, *tmp;
> +
> + spin_lock(&man->wait_lock);
> + list_for_each_entry_safe(waiter, tmp, &man->waiter_list, list) {
> + if ((

Re: [PATCH 1/2] [WIP]drm/ttm: add waiter list to prevent allocation not in order

[Christian König]

Yes, exactly that's the problem.

See when you want to prevent a process B from allocating the memory 
process A has evicted, you need to prevent all concurrent allocation.


And we don't do that because it causes a major performance drop.

Regards,
Christian.

Am 26.01.2018 um 14:21 schrieb Zhou, David(ChunMing):
You patch will prevent concurrent allocation, and will result in 
allocation performance drop much.


发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午9:04写道：


Attached is what you actually want to do cleanly implemented. But as I 
said this is a NO-GO.


Regards,
Christian.

Am 26.01.2018 um 13:43 schrieb Christian König:
After my investigation, this issue should be detect of TTM design 
self, which breaks scheduling balance.

Yeah, but again. This is indented design we can't change easily.

Regards,
Christian.

Am 26.01.2018 um 13:36 schrieb Zhou, David(ChunMing):

I am off work, so reply mail by phone, the format could not be text.

back to topic itself:
the problem indeed happen on amdgpu driver, someone reports me that 
application runs with two instances, the performance are different.
I also reproduced the issue with unit test(bo_eviction_test). They 
always think our scheduler isn't working as expected.


After my investigation, this issue should be detect of TTM design 
self, which breaks scheduling balance.


Further, if we run containers for our gpu, container A could run 
high score, container B runs low score with same benchmark.


So this is bug that we need fix.

Regards,
David Zhou

发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午6:31写道：


Am 26.01.2018 um 11:22 schrieb Chunming Zhou:
> there is a scheduling balance issue about get node like:
> a. process A allocates full memory and use it for submission.
> b. process B tries to allocates memory, will wait for process A BO 
idle in eviction.
> c. process A completes the job, process B eviction will put 
process A BO node,
> but in the meantime, process C is comming to allocate BO, whill 
directly get node successfully, and do submission,

> process B will again wait for process C BO idle.
> d. repeat the above setps, process B could be delayed much more.
>
> later allocation must not be ahead of front in same place.

Again NAK to the whole approach.

At least with amdgpu the problem you described above never occurs
because evictions are pipelined operations. We could only block for
deleted regions to become free.

But independent of that incoming memory requests while we make room for
eviction are intended to be served first.

Changing that is certainly a no-go cause that would favor memory hungry
applications over small clients.

Regards,
Christian.

>
> Change-Id: I3daa892e50f82226c552cc008a29e55894a98f18
> Signed-off-by: Chunming Zhou 
> ---
>   drivers/gpu/drm/ttm/ttm_bo.c    | 69 
+++--

>   include/drm/ttm/ttm_bo_api.h    |  7 +
>   include/drm/ttm/ttm_bo_driver.h |  7 +
>   3 files changed, 80 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c 
b/drivers/gpu/drm/ttm/ttm_bo.c

> index d33a6bb742a1..558ec2cf465d 100644
> --- a/drivers/gpu/drm/ttm/ttm_bo.c
> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
> @@ -841,6 +841,58 @@ static int ttm_bo_add_move_fence(struct 
ttm_buffer_object *bo,

>    return 0;
>   }
>
> +static void ttm_man_init_waiter(struct ttm_bo_waiter *waiter,
> + struct ttm_buffer_object *bo,
> + const struct ttm_place *place)
> +{
> + waiter->tbo = bo;
> + memcpy((void *)&waiter->place, (void *)place, sizeof(*place));
> + INIT_LIST_HEAD(&waiter->list);
> +}
> +
> +static void ttm_man_add_waiter(struct ttm_mem_type_manager *man,
> +    struct ttm_bo_waiter *waiter)
> +{
> + if (!waiter)
> + return;
> + spin_lock(&man->wait_lock);
> + list_add_tail(&waiter->list, &man->waiter_list);
> + spin_unlock(&man->wait_lock);
> +}
> +
> +static void ttm_man_del_waiter(struct ttm_mem_type_manager *man,
> +    struct ttm_bo_waiter *waiter)
> +{
> + if (!waiter)
> + return;
> + spin_lock(&man->wait_lock);
> + if (!list_empty(&waiter->list))
> + list_del(&waiter->list);
> + spin_unlock(&man->wait_lock);
> + kfree(waiter);
> +}
> +
> +int ttm_man_check_bo(struct ttm_mem_type_manager *man,
> +  struct ttm_buffer_object *bo,
> +  const struct ttm_place *place)
> +{
> + struct ttm_bo_waiter *waiter, *tmp;
> +
> + spin_lock(&man->wait_lock);
> + list_for_each_entry_safe(waiter, tmp, &man->waiter_list, list) {
> + if ((bo != waiter->tbo) &&
> + ((place->fpfn >= waiter->place.fpfn &&
> +   place->fpfn <= waiter->place.lpfn) ||
> +  (place->lpfn <= waiter->place.lpfn && 
place->lpfn >=

> +   waiter->place.fpfn)))
> + goto later_bo;
> + }
> + spin_unlock(&m

Re: [PATCH 1/2] [WIP]drm/ttm: add waiter list to prevent allocation not in order

[Zhou, David(ChunMing)]

You patch will prevent concurrent allocation, and will result in allocation 
performance drop much.


发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 下午9:04写道：

Attached is what you actually want to do cleanly implemented. But as I said 
this is a NO-GO.

Regards,
Christian.

Am 26.01.2018 um 13:43 schrieb Christian König:
After my investigation, this issue should be detect of TTM design self, which 
breaks scheduling balance.
Yeah, but again. This is indented design we can't change easily.

Regards,
Christian.

Am 26.01.2018 um 13:36 schrieb Zhou, David(ChunMing):
I am off work, so reply mail by phone, the format could not be text.

back to topic itself:
the problem indeed happen on amdgpu driver, someone reports me that application 
runs with two instances, the performance are different.
I also reproduced the issue with unit test(bo_eviction_test). They always think 
our scheduler isn't working as expected.

After my investigation, this issue should be detect of TTM design self, which 
breaks scheduling balance.

Further, if we run containers for our gpu, container A could run high score, 
container B runs low score with same benchmark.

So this is bug that we need fix.

Regards,
David Zhou


发自坚果 Pro

Christian K鰊ig 
 于 
2018年1月26日 下午6:31写道：

Am 26.01.2018 um 11:22 schrieb Chunming Zhou:
> there is a scheduling balance issue about get node like:
> a. process A allocates full memory and use it for submission.
> b. process B tries to allocates memory, will wait for process A BO idle in 
> eviction.
> c. process A completes the job, process B eviction will put process A BO node,
> but in the meantime, process C is comming to allocate BO, whill directly get 
> node successfully, and do submission,
> process B will again wait for process C BO idle.
> d. repeat the above setps, process B could be delayed much more.
>
> later allocation must not be ahead of front in same place.

Again NAK to the whole approach.

At least with amdgpu the problem you described above never occurs
because evictions are pipelined operations. We could only block for
deleted regions to become free.

But independent of that incoming memory requests while we make room for
eviction are intended to be served first.

Changing that is certainly a no-go cause that would favor memory hungry
applications over small clients.

Regards,
Christian.

>
> Change-Id: I3daa892e50f82226c552cc008a29e55894a98f18
> Signed-off-by: Chunming Zhou 
> ---
>   drivers/gpu/drm/ttm/ttm_bo.c| 69 
> +++--
>   include/drm/ttm/ttm_bo_api.h|  7 +
>   include/drm/ttm/ttm_bo_driver.h |  7 +
>   3 files changed, 80 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
> index d33a6bb742a1..558ec2cf465d 100644
> --- a/drivers/gpu/drm/ttm/ttm_bo.c
> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
> @@ -841,6 +841,58 @@ static int ttm_bo_add_move_fence(struct 
> ttm_buffer_object *bo,
>return 0;
>   }
>
> +static void ttm_man_init_waiter(struct ttm_bo_waiter *waiter,
> + struct ttm_buffer_object *bo,
> + const struct ttm_place *place)
> +{
> + waiter->tbo = bo;
> + memcpy((void *)&waiter->place, (void *)place, sizeof(*place));
> + INIT_LIST_HEAD(&waiter->list);
> +}
> +
> +static void ttm_man_add_waiter(struct ttm_mem_type_manager *man,
> +struct ttm_bo_waiter *waiter)
> +{
> + if (!waiter)
> + return;
> + spin_lock(&man->wait_lock);
> + list_add_tail(&waiter->list, &man->waiter_list);
> + spin_unlock(&man->wait_lock);
> +}
> +
> +static void ttm_man_del_waiter(struct ttm_mem_type_manager *man,
> +struct ttm_bo_waiter *waiter)
> +{
> + if (!waiter)
> + return;
> + spin_lock(&man->wait_lock);
> + if (!list_empty(&waiter->list))
> + list_del(&waiter->list);
> + spin_unlock(&man->wait_lock);
> + kfree(waiter);
> +}
> +
> +int ttm_man_check_bo(struct ttm_mem_type_manager *man,
> +  struct ttm_buffer_object *bo,
> +  const struct ttm_place *place)
> +{
> + struct ttm_bo_waiter *waiter, *tmp;
> +
> + spin_lock(&man->wait_lock);
> + list_for_each_entry_safe(waiter, tmp, &man->waiter_list, list) {
> + if ((bo != waiter->tbo) &&
> + ((place->fpfn >= waiter->place.fpfn &&
> +   place->fpfn <= waiter->place.lpfn) ||
> +  (place->lpfn <= waiter->place.lpfn && place->lpfn >=
> +   waiter->place.fpfn)))
> + goto later_bo;
> + }
> + spin_unlock(&man->wait_lock);
> + return true;
> +later_bo:
> + spin_unlock(&man->wait_lock);
> + return false;
> +}
>   /**
>* Repeatedly evict memory from the LRU for @mem_type until we create enough
>* space, or we've evicted everything and there

Re: [PATCH 1/2] [WIP]drm/ttm: add waiter list to prevent allocation not in order

[Christian König]

Attached is what you actually want to do cleanly implemented. But as I 
said this is a NO-GO.


Regards,
Christian.

Am 26.01.2018 um 13:43 schrieb Christian König:
After my investigation, this issue should be detect of TTM design 
self, which breaks scheduling balance.

Yeah, but again. This is indented design we can't change easily.

Regards,
Christian.

Am 26.01.2018 um 13:36 schrieb Zhou, David(ChunMing):

I am off work, so reply mail by phone, the format could not be text.

back to topic itself:
the problem indeed happen on amdgpu driver, someone reports me that 
application runs with two instances, the performance are different.
I also reproduced the issue with unit test(bo_eviction_test). They 
always think our scheduler isn't working as expected.


After my investigation, this issue should be detect of TTM design 
self, which breaks scheduling balance.


Further, if we run containers for our gpu, container A could run high 
score, container B runs low score with same benchmark.


So this is bug that we need fix.

Regards,
David Zhou

发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午6:31写道：


Am 26.01.2018 um 11:22 schrieb Chunming Zhou:
> there is a scheduling balance issue about get node like:
> a. process A allocates full memory and use it for submission.
> b. process B tries to allocates memory, will wait for process A BO 
idle in eviction.
> c. process A completes the job, process B eviction will put process 
A BO node,
> but in the meantime, process C is comming to allocate BO, whill 
directly get node successfully, and do submission,

> process B will again wait for process C BO idle.
> d. repeat the above setps, process B could be delayed much more.
>
> later allocation must not be ahead of front in same place.

Again NAK to the whole approach.

At least with amdgpu the problem you described above never occurs
because evictions are pipelined operations. We could only block for
deleted regions to become free.

But independent of that incoming memory requests while we make room for
eviction are intended to be served first.

Changing that is certainly a no-go cause that would favor memory hungry
applications over small clients.

Regards,
Christian.

>
> Change-Id: I3daa892e50f82226c552cc008a29e55894a98f18
> Signed-off-by: Chunming Zhou 
> ---
>   drivers/gpu/drm/ttm/ttm_bo.c    | 69 
+++--

>   include/drm/ttm/ttm_bo_api.h    |  7 +
>   include/drm/ttm/ttm_bo_driver.h |  7 +
>   3 files changed, 80 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c 
b/drivers/gpu/drm/ttm/ttm_bo.c

> index d33a6bb742a1..558ec2cf465d 100644
> --- a/drivers/gpu/drm/ttm/ttm_bo.c
> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
> @@ -841,6 +841,58 @@ static int ttm_bo_add_move_fence(struct 
ttm_buffer_object *bo,

>    return 0;
>   }
>
> +static void ttm_man_init_waiter(struct ttm_bo_waiter *waiter,
> + struct ttm_buffer_object *bo,
> + const struct ttm_place *place)
> +{
> + waiter->tbo = bo;
> + memcpy((void *)&waiter->place, (void *)place, sizeof(*place));
> + INIT_LIST_HEAD(&waiter->list);
> +}
> +
> +static void ttm_man_add_waiter(struct ttm_mem_type_manager *man,
> +    struct ttm_bo_waiter *waiter)
> +{
> + if (!waiter)
> + return;
> + spin_lock(&man->wait_lock);
> + list_add_tail(&waiter->list, &man->waiter_list);
> + spin_unlock(&man->wait_lock);
> +}
> +
> +static void ttm_man_del_waiter(struct ttm_mem_type_manager *man,
> +    struct ttm_bo_waiter *waiter)
> +{
> + if (!waiter)
> + return;
> + spin_lock(&man->wait_lock);
> + if (!list_empty(&waiter->list))
> + list_del(&waiter->list);
> + spin_unlock(&man->wait_lock);
> + kfree(waiter);
> +}
> +
> +int ttm_man_check_bo(struct ttm_mem_type_manager *man,
> +  struct ttm_buffer_object *bo,
> +  const struct ttm_place *place)
> +{
> + struct ttm_bo_waiter *waiter, *tmp;
> +
> + spin_lock(&man->wait_lock);
> + list_for_each_entry_safe(waiter, tmp, &man->waiter_list, list) {
> + if ((bo != waiter->tbo) &&
> + ((place->fpfn >= waiter->place.fpfn &&
> +   place->fpfn <= waiter->place.lpfn) ||
> +  (place->lpfn <= waiter->place.lpfn && place->lpfn >=
> +   waiter->place.fpfn)))
> + goto later_bo;
> + }
> + spin_unlock(&man->wait_lock);
> + return true;
> +later_bo:
> + spin_unlock(&man->wait_lock);
> + return false;
> +}
>   /**
>    * Repeatedly evict memory from the LRU for @mem_type until we 
create enough

>    * space, or we've evicted everything and there isn't enough space.
> @@ -853,17 +905,26 @@ static int ttm_bo_mem_force_space(struct 
ttm_buffer_object *bo,

>   {
>    struct ttm_bo_device *bdev = bo->bdev;
>    struct ttm_mem_type_manager *man = &bdev->

Re: [PATCH 1/2] [WIP]drm/ttm: add waiter list to prevent allocation not in order

[Christian König]

After my investigation, this issue should be detect of TTM design 
self, which breaks scheduling balance.

Yeah, but again. This is indented design we can't change easily.

Regards,
Christian.

Am 26.01.2018 um 13:36 schrieb Zhou, David(ChunMing):

I am off work, so reply mail by phone, the format could not be text.

back to topic itself:
the problem indeed happen on amdgpu driver, someone reports me that 
application runs with two instances, the performance are different.
I also reproduced the issue with unit test(bo_eviction_test). They 
always think our scheduler isn't working as expected.


After my investigation, this issue should be detect of TTM design 
self, which breaks scheduling balance.


Further, if we run containers for our gpu, container A could run high 
score, container B runs low score with same benchmark.


So this is bug that we need fix.

Regards,
David Zhou

发自坚果 Pro

Christian K鰊ig  于 2018年1月26日 
下午6:31写道：


Am 26.01.2018 um 11:22 schrieb Chunming Zhou:
> there is a scheduling balance issue about get node like:
> a. process A allocates full memory and use it for submission.
> b. process B tries to allocates memory, will wait for process A BO 
idle in eviction.
> c. process A completes the job, process B eviction will put process 
A BO node,
> but in the meantime, process C is comming to allocate BO, whill 
directly get node successfully, and do submission,

> process B will again wait for process C BO idle.
> d. repeat the above setps, process B could be delayed much more.
>
> later allocation must not be ahead of front in same place.

Again NAK to the whole approach.

At least with amdgpu the problem you described above never occurs
because evictions are pipelined operations. We could only block for
deleted regions to become free.

But independent of that incoming memory requests while we make room for
eviction are intended to be served first.

Changing that is certainly a no-go cause that would favor memory hungry
applications over small clients.

Regards,
Christian.

>
> Change-Id: I3daa892e50f82226c552cc008a29e55894a98f18
> Signed-off-by: Chunming Zhou 
> ---
>   drivers/gpu/drm/ttm/ttm_bo.c    | 69 
+++--

>   include/drm/ttm/ttm_bo_api.h    |  7 +
>   include/drm/ttm/ttm_bo_driver.h |  7 +
>   3 files changed, 80 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
> index d33a6bb742a1..558ec2cf465d 100644
> --- a/drivers/gpu/drm/ttm/ttm_bo.c
> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
> @@ -841,6 +841,58 @@ static int ttm_bo_add_move_fence(struct 
ttm_buffer_object *bo,

>    return 0;
>   }
>
> +static void ttm_man_init_waiter(struct ttm_bo_waiter *waiter,
> + struct ttm_buffer_object *bo,
> + const struct ttm_place *place)
> +{
> + waiter->tbo = bo;
> + memcpy((void *)&waiter->place, (void *)place, sizeof(*place));
> + INIT_LIST_HEAD(&waiter->list);
> +}
> +
> +static void ttm_man_add_waiter(struct ttm_mem_type_manager *man,
> +    struct ttm_bo_waiter *waiter)
> +{
> + if (!waiter)
> + return;
> + spin_lock(&man->wait_lock);
> + list_add_tail(&waiter->list, &man->waiter_list);
> + spin_unlock(&man->wait_lock);
> +}
> +
> +static void ttm_man_del_waiter(struct ttm_mem_type_manager *man,
> +    struct ttm_bo_waiter *waiter)
> +{
> + if (!waiter)
> + return;
> + spin_lock(&man->wait_lock);
> + if (!list_empty(&waiter->list))
> + list_del(&waiter->list);
> + spin_unlock(&man->wait_lock);
> + kfree(waiter);
> +}
> +
> +int ttm_man_check_bo(struct ttm_mem_type_manager *man,
> +  struct ttm_buffer_object *bo,
> +  const struct ttm_place *place)
> +{
> + struct ttm_bo_waiter *waiter, *tmp;
> +
> + spin_lock(&man->wait_lock);
> + list_for_each_entry_safe(waiter, tmp, &man->waiter_list, list) {
> + if ((bo != waiter->tbo) &&
> + ((place->fpfn >= waiter->place.fpfn &&
> +   place->fpfn <= waiter->place.lpfn) ||
> +  (place->lpfn <= waiter->place.lpfn && place->lpfn >=
> +   waiter->place.fpfn)))
> + goto later_bo;
> + }
> + spin_unlock(&man->wait_lock);
> + return true;
> +later_bo:
> + spin_unlock(&man->wait_lock);
> + return false;
> +}
>   /**
>    * Repeatedly evict memory from the LRU for @mem_type until we 
create enough

>    * space, or we've evicted everything and there isn't enough space.
> @@ -853,17 +905,26 @@ static int ttm_bo_mem_force_space(struct 
ttm_buffer_object *bo,

>   {
>    struct ttm_bo_device *bdev = bo->bdev;
>    struct ttm_mem_type_manager *man = &bdev->man[mem_type];
> + struct ttm_bo_waiter waiter;
>    int ret;
>
> + ttm_man_init_waiter(&waiter, bo, place);
> + ttm_man_add_waiter(man, &waiter);
>   

Re: [PATCH 1/2] [WIP]drm/ttm: add waiter list to prevent allocation not in order

[Zhou, David(ChunMing)]

I am off work, so reply mail by phone, the format could not be text.

back to topic itself:
the problem indeed happen on amdgpu driver, someone reports me that application 
runs with two instances, the performance are different.
I also reproduced the issue with unit test(bo_eviction_test). They always think 
our scheduler isn't working as expected.

After my investigation, this issue should be detect of TTM design self, which 
breaks scheduling balance.

Further, if we run containers for our gpu, container A could run high score, 
container B runs low score with same benchmark.

So this is bug that we need fix.

Regards,
David Zhou


发自坚果 Pro

Christian K�nig  于 2018年1月26日 下午6:31写道：

Am 26.01.2018 um 11:22 schrieb Chunming Zhou:
> there is a scheduling balance issue about get node like:
> a. process A allocates full memory and use it for submission.
> b. process B tries to allocates memory, will wait for process A BO idle in 
> eviction.
> c. process A completes the job, process B eviction will put process A BO node,
> but in the meantime, process C is comming to allocate BO, whill directly get 
> node successfully, and do submission,
> process B will again wait for process C BO idle.
> d. repeat the above setps, process B could be delayed much more.
>
> later allocation must not be ahead of front in same place.

Again NAK to the whole approach.

At least with amdgpu the problem you described above never occurs
because evictions are pipelined operations. We could only block for
deleted regions to become free.

But independent of that incoming memory requests while we make room for
eviction are intended to be served first.

Changing that is certainly a no-go cause that would favor memory hungry
applications over small clients.

Regards,
Christian.

>
> Change-Id: I3daa892e50f82226c552cc008a29e55894a98f18
> Signed-off-by: Chunming Zhou 
> ---
>   drivers/gpu/drm/ttm/ttm_bo.c| 69 
> +++--
>   include/drm/ttm/ttm_bo_api.h|  7 +
>   include/drm/ttm/ttm_bo_driver.h |  7 +
>   3 files changed, 80 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
> index d33a6bb742a1..558ec2cf465d 100644
> --- a/drivers/gpu/drm/ttm/ttm_bo.c
> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
> @@ -841,6 +841,58 @@ static int ttm_bo_add_move_fence(struct 
> ttm_buffer_object *bo,
>return 0;
>   }
>
> +static void ttm_man_init_waiter(struct ttm_bo_waiter *waiter,
> + struct ttm_buffer_object *bo,
> + const struct ttm_place *place)
> +{
> + waiter->tbo = bo;
> + memcpy((void *)&waiter->place, (void *)place, sizeof(*place));
> + INIT_LIST_HEAD(&waiter->list);
> +}
> +
> +static void ttm_man_add_waiter(struct ttm_mem_type_manager *man,
> +struct ttm_bo_waiter *waiter)
> +{
> + if (!waiter)
> + return;
> + spin_lock(&man->wait_lock);
> + list_add_tail(&waiter->list, &man->waiter_list);
> + spin_unlock(&man->wait_lock);
> +}
> +
> +static void ttm_man_del_waiter(struct ttm_mem_type_manager *man,
> +struct ttm_bo_waiter *waiter)
> +{
> + if (!waiter)
> + return;
> + spin_lock(&man->wait_lock);
> + if (!list_empty(&waiter->list))
> + list_del(&waiter->list);
> + spin_unlock(&man->wait_lock);
> + kfree(waiter);
> +}
> +
> +int ttm_man_check_bo(struct ttm_mem_type_manager *man,
> +  struct ttm_buffer_object *bo,
> +  const struct ttm_place *place)
> +{
> + struct ttm_bo_waiter *waiter, *tmp;
> +
> + spin_lock(&man->wait_lock);
> + list_for_each_entry_safe(waiter, tmp, &man->waiter_list, list) {
> + if ((bo != waiter->tbo) &&
> + ((place->fpfn >= waiter->place.fpfn &&
> +   place->fpfn <= waiter->place.lpfn) ||
> +  (place->lpfn <= waiter->place.lpfn && place->lpfn >=
> +   waiter->place.fpfn)))
> + goto later_bo;
> + }
> + spin_unlock(&man->wait_lock);
> + return true;
> +later_bo:
> + spin_unlock(&man->wait_lock);
> + return false;
> +}
>   /**
>* Repeatedly evict memory from the LRU for @mem_type until we create enough
>* space, or we've evicted everything and there isn't enough space.
> @@ -853,17 +905,26 @@ static int ttm_bo_mem_force_space(struct 
> ttm_buffer_object *bo,
>   {
>struct ttm_bo_device *bdev = bo->bdev;
>struct ttm_mem_type_manager *man = &bdev->man[mem_type];
> + struct ttm_bo_waiter waiter;
>int ret;
>
> + ttm_man_init_waiter(&waiter, bo, place);
> + ttm_man_add_waiter(man, &waiter);
>do {
>ret = (*man->func->get_node)(man, bo, place, mem);
> - if (unlikely(ret != 0))
> + if (unlikely(ret != 0)) {
> + ttm_man_del_waiter(man, &waiter);
>return

Re: [PATCH 1/2] [WIP]drm/ttm: add waiter list to prevent allocation not in order

[Christian König]

Am 26.01.2018 um 11:22 schrieb Chunming Zhou:

there is a scheduling balance issue about get node like:
a. process A allocates full memory and use it for submission.
b. process B tries to allocates memory, will wait for process A BO idle in 
eviction.
c. process A completes the job, process B eviction will put process A BO node,
but in the meantime, process C is comming to allocate BO, whill directly get 
node successfully, and do submission,
process B will again wait for process C BO idle.
d. repeat the above setps, process B could be delayed much more.

later allocation must not be ahead of front in same place.


Again NAK to the whole approach.

At least with amdgpu the problem you described above never occurs 
because evictions are pipelined operations. We could only block for 
deleted regions to become free.


But independent of that incoming memory requests while we make room for 
eviction are intended to be served first.


Changing that is certainly a no-go cause that would favor memory hungry 
applications over small clients.


Regards,
Christian.



Change-Id: I3daa892e50f82226c552cc008a29e55894a98f18
Signed-off-by: Chunming Zhou 
---
  drivers/gpu/drm/ttm/ttm_bo.c| 69 +++--
  include/drm/ttm/ttm_bo_api.h|  7 +
  include/drm/ttm/ttm_bo_driver.h |  7 +
  3 files changed, 80 insertions(+), 3 deletions(-)

diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index d33a6bb742a1..558ec2cf465d 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -841,6 +841,58 @@ static int ttm_bo_add_move_fence(struct ttm_buffer_object 
*bo,
return 0;
  }
  
+static void ttm_man_init_waiter(struct ttm_bo_waiter *waiter,

+   struct ttm_buffer_object *bo,
+   const struct ttm_place *place)
+{
+   waiter->tbo = bo;
+   memcpy((void *)&waiter->place, (void *)place, sizeof(*place));
+   INIT_LIST_HEAD(&waiter->list);
+}
+
+static void ttm_man_add_waiter(struct ttm_mem_type_manager *man,
+  struct ttm_bo_waiter *waiter)
+{
+   if (!waiter)
+   return;
+   spin_lock(&man->wait_lock);
+   list_add_tail(&waiter->list, &man->waiter_list);
+   spin_unlock(&man->wait_lock);
+}
+
+static void ttm_man_del_waiter(struct ttm_mem_type_manager *man,
+  struct ttm_bo_waiter *waiter)
+{
+   if (!waiter)
+   return;
+   spin_lock(&man->wait_lock);
+   if (!list_empty(&waiter->list))
+   list_del(&waiter->list);
+   spin_unlock(&man->wait_lock);
+   kfree(waiter);
+}
+
+int ttm_man_check_bo(struct ttm_mem_type_manager *man,
+struct ttm_buffer_object *bo,
+const struct ttm_place *place)
+{
+   struct ttm_bo_waiter *waiter, *tmp;
+
+   spin_lock(&man->wait_lock);
+   list_for_each_entry_safe(waiter, tmp, &man->waiter_list, list) {
+   if ((bo != waiter->tbo) &&
+   ((place->fpfn >= waiter->place.fpfn &&
+ place->fpfn <= waiter->place.lpfn) ||
+(place->lpfn <= waiter->place.lpfn && place->lpfn >=
+ waiter->place.fpfn)))
+   goto later_bo;
+   }
+   spin_unlock(&man->wait_lock);
+   return true;
+later_bo:
+   spin_unlock(&man->wait_lock);
+   return false;
+}
  /**
   * Repeatedly evict memory from the LRU for @mem_type until we create enough
   * space, or we've evicted everything and there isn't enough space.
@@ -853,17 +905,26 @@ static int ttm_bo_mem_force_space(struct 
ttm_buffer_object *bo,
  {
struct ttm_bo_device *bdev = bo->bdev;
struct ttm_mem_type_manager *man = &bdev->man[mem_type];
+   struct ttm_bo_waiter waiter;
int ret;
  
+	ttm_man_init_waiter(&waiter, bo, place);

+   ttm_man_add_waiter(man, &waiter);
do {
ret = (*man->func->get_node)(man, bo, place, mem);
-   if (unlikely(ret != 0))
+   if (unlikely(ret != 0)) {
+   ttm_man_del_waiter(man, &waiter);
return ret;
-   if (mem->mm_node)
+   }
+   if (mem->mm_node) {
+   ttm_man_del_waiter(man, &waiter);
break;
+   }
ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
-   if (unlikely(ret != 0))
+   if (unlikely(ret != 0)) {
+   ttm_man_del_waiter(man, &waiter);
return ret;
+   }
} while (1);
mem->mem_type = mem_type;
return ttm_bo_add_move_fence(bo, man, mem);
@@ -1450,6 +1511,8 @@ int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned 
type,
man->use_io_reserve_lru = false;
mutex_init(&man->io_reserve_mutex);
spin_lock_init(&man->move_lock);
+  

[PATCH 1/2] [WIP]drm/ttm: add waiter list to prevent allocation not in order

[Chunming Zhou]

there is a scheduling balance issue about get node like:
a. process A allocates full memory and use it for submission.
b. process B tries to allocates memory, will wait for process A BO idle in 
eviction.
c. process A completes the job, process B eviction will put process A BO node,
but in the meantime, process C is comming to allocate BO, whill directly get 
node successfully, and do submission,
process B will again wait for process C BO idle.
d. repeat the above setps, process B could be delayed much more.

later allocation must not be ahead of front in same place.

Change-Id: I3daa892e50f82226c552cc008a29e55894a98f18
Signed-off-by: Chunming Zhou 
---
 drivers/gpu/drm/ttm/ttm_bo.c| 69 +++--
 include/drm/ttm/ttm_bo_api.h|  7 +
 include/drm/ttm/ttm_bo_driver.h |  7 +
 3 files changed, 80 insertions(+), 3 deletions(-)

diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index d33a6bb742a1..558ec2cf465d 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -841,6 +841,58 @@ static int ttm_bo_add_move_fence(struct ttm_buffer_object 
*bo,
return 0;
 }
 
+static void ttm_man_init_waiter(struct ttm_bo_waiter *waiter,
+   struct ttm_buffer_object *bo,
+   const struct ttm_place *place)
+{
+   waiter->tbo = bo;
+   memcpy((void *)&waiter->place, (void *)place, sizeof(*place));
+   INIT_LIST_HEAD(&waiter->list);
+}
+
+static void ttm_man_add_waiter(struct ttm_mem_type_manager *man,
+  struct ttm_bo_waiter *waiter)
+{
+   if (!waiter)
+   return;
+   spin_lock(&man->wait_lock);
+   list_add_tail(&waiter->list, &man->waiter_list);
+   spin_unlock(&man->wait_lock);
+}
+
+static void ttm_man_del_waiter(struct ttm_mem_type_manager *man,
+  struct ttm_bo_waiter *waiter)
+{
+   if (!waiter)
+   return;
+   spin_lock(&man->wait_lock);
+   if (!list_empty(&waiter->list))
+   list_del(&waiter->list);
+   spin_unlock(&man->wait_lock);
+   kfree(waiter);
+}
+
+int ttm_man_check_bo(struct ttm_mem_type_manager *man,
+struct ttm_buffer_object *bo,
+const struct ttm_place *place)
+{
+   struct ttm_bo_waiter *waiter, *tmp;
+
+   spin_lock(&man->wait_lock);
+   list_for_each_entry_safe(waiter, tmp, &man->waiter_list, list) {
+   if ((bo != waiter->tbo) &&
+   ((place->fpfn >= waiter->place.fpfn &&
+ place->fpfn <= waiter->place.lpfn) ||
+(place->lpfn <= waiter->place.lpfn && place->lpfn >=
+ waiter->place.fpfn)))
+   goto later_bo;
+   }
+   spin_unlock(&man->wait_lock);
+   return true;
+later_bo:
+   spin_unlock(&man->wait_lock);
+   return false;
+}
 /**
  * Repeatedly evict memory from the LRU for @mem_type until we create enough
  * space, or we've evicted everything and there isn't enough space.
@@ -853,17 +905,26 @@ static int ttm_bo_mem_force_space(struct 
ttm_buffer_object *bo,
 {
struct ttm_bo_device *bdev = bo->bdev;
struct ttm_mem_type_manager *man = &bdev->man[mem_type];
+   struct ttm_bo_waiter waiter;
int ret;
 
+   ttm_man_init_waiter(&waiter, bo, place);
+   ttm_man_add_waiter(man, &waiter);
do {
ret = (*man->func->get_node)(man, bo, place, mem);
-   if (unlikely(ret != 0))
+   if (unlikely(ret != 0)) {
+   ttm_man_del_waiter(man, &waiter);
return ret;
-   if (mem->mm_node)
+   }
+   if (mem->mm_node) {
+   ttm_man_del_waiter(man, &waiter);
break;
+   }
ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
-   if (unlikely(ret != 0))
+   if (unlikely(ret != 0)) {
+   ttm_man_del_waiter(man, &waiter);
return ret;
+   }
} while (1);
mem->mem_type = mem_type;
return ttm_bo_add_move_fence(bo, man, mem);
@@ -1450,6 +1511,8 @@ int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned 
type,
man->use_io_reserve_lru = false;
mutex_init(&man->io_reserve_mutex);
spin_lock_init(&man->move_lock);
+   spin_lock_init(&man->wait_lock);
+   INIT_LIST_HEAD(&man->waiter_list);
INIT_LIST_HEAD(&man->io_reserve_lru);
 
ret = bdev->driver->init_mem_type(bdev, type, man);
diff --git a/include/drm/ttm/ttm_bo_api.h b/include/drm/ttm/ttm_bo_api.h
index 2cd025c2abe7..0fce4dbd02e7 100644
--- a/include/drm/ttm/ttm_bo_api.h
+++ b/include/drm/ttm/ttm_bo_api.h
@@ -40,6 +40,7 @@
 #include 
 #include 
 #include 
+#include 
 
 struct ttm_bo_device;
 
@@ -232,6 +233,12 @@ struct ttm_buffer_object {