[PATCH] drm/ttm: replace dma_resv object on deleted BOs v3
Pan, Xinhui
Xinhui.Pan at amd.com
Thu Feb 13 11:15:11 UTC 2020
> 2020年2月13日 18:01，Koenig, Christian <Christian.Koenig at amd.com> 写道：
>
> Am 13.02.20 um 05:11 schrieb Pan, Xinhui:
>>
>>
>>> 2020年2月12日 19:53，Christian König <ckoenig.leichtzumerken at gmail.com> 写道：
>>>
>>> Am 12.02.20 um 07:23 schrieb Pan, Xinhui:
>>>>> 2020年2月11日 23:43，Christian König <ckoenig.leichtzumerken at gmail.com> 写道：
>>>>>
>>>>> When non-imported BOs are resurrected for delayed delete we replace
>>>>> the dma_resv object to allow for easy reclaiming of the resources.
>>>>>
>>>>> v2: move that to ttm_bo_individualize_resv
>>>>> v3: add a comment to explain what's going on
>>>>>
>>>>> Signed-off-by: Christian König <christian.koenig at amd.com>
>>>>> Reviewed-by: xinhui pan <xinhui.pan at amd.com>
>>>>> ---
>>>>> drivers/gpu/drm/ttm/ttm_bo.c | 14 +++++++++++++-
>>>>> 1 file changed, 13 insertions(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
>>>>> index bfc42a9e4fb4..8174603d390f 100644
>>>>> --- a/drivers/gpu/drm/ttm/ttm_bo.c
>>>>> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
>>>>> @@ -393,6 +393,18 @@ static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
>>>>>
>>>>> r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
>>>>> dma_resv_unlock(&bo->base._resv);
>>>>> + if (r)
>>>>> + return r;
>>>>> +
>>>>> + if (bo->type != ttm_bo_type_sg) {
>>>>> + /* This works because the BO is about to be destroyed and nobody
>>>>> + * reference it any more. The only tricky case is the trylock on
>>>>> + * the resv object while holding the lru_lock.
>>>>> + */
>>>>> + spin_lock(&ttm_bo_glob.lru_lock);
>>>>> + bo->base.resv = &bo->base._resv;
>>>>> + spin_unlock(&ttm_bo_glob.lru_lock);
>>>>> + }
>>>>>
>>>> how about something like that.
>>>> the basic idea is to do the bo cleanup work in bo release first and avoid any race with evict.
>>>> As in bo dieing progress, evict also just do bo cleanup work.
>>>>
>>>> If bo is busy, neither bo_release nor evict can do cleanupwork on it. For the bo release case, we just add bo back to lru list.
>>>> So we can clean it up both in workqueue and shrinker as the past way did.
>>>>
>>>> @@ -405,8 +405,9 @@ static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
>>>> if (bo->type != ttm_bo_type_sg) {
>>>> spin_lock(&ttm_bo_glob.lru_lock);
>>>> - bo->base.resv = &bo->base._resv;
>>>> + ttm_bo_del_from_lru(bo);
>>>> spin_unlock(&ttm_bo_glob.lru_lock);
>>>> + bo->base.resv = &bo->base._resv;
>>>> }
>>>> return r;
>>>> @@ -606,10 +607,9 @@ static void ttm_bo_release(struct kref *kref)
>>>> * shrinkers, now that they are queued for
>>>> * destruction.
>>>> */
>>>> - if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
>>>> + if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT)
>>>> bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
>>>> - ttm_bo_move_to_lru_tail(bo, NULL);
>>>> - }
>>>> + ttm_bo_add_mem_to_lru(bo, &bo->mem);
>>>> kref_init(&bo->kref);
>>>> list_add_tail(&bo->ddestroy, &bdev->ddestroy);
>>> Yeah, thought about that as well. But this has the major drawback that the deleted BO moves to the end of the LRU, which is something we don't want.
>> well, as the bo is busy, looks like it needs time to being idle. putting it to tail seems fair.
>
> No, see BOs should move to the tail of the LRU whenever they are used. Freeing up a BO is basically the opposite of using it.
>
> So what would happen on the next memory contention is that the MM would evict BOs which are still used and only after come to the delete BO which could have been removed long ago.
>
>>> I think the real solution to this problem is to go a completely different way and remove the delayed delete feature from TTM altogether. Instead this should be part of some DRM domain handler component.
>>>
>> yes, completely agree. As long as we can shrink bos when OOM, the workqueue is not necessary, The workqueue does not help in a heavy workload case.
>>
>> Pls see my patches below, I remove the workqueue, and what’s more, we can clearup the bo without lru lock hold.
>> That would reduce the lock contention. I run kfdtest and got a good performance result.
>
> No, that's an approach we had before as well and it also doesn't work that well.
>
> See the problem is that we can only remove the BO from the LRU after it has released the memory it references. Otherwise we run into the issue that some threads can't wait for the memory to be freed any more and run into an OOM situation.
>
ok, we can keep the workqueue at it is.
Now I come up with another idea that evict and swap can touch the destroy list first, then lru list.
Looks like putting a dieing bo in lru list is useless.
thanks
xinhui
> Regards,
> Christian.
>
>>
>>
>>> In other words it should not matter if a BO is evicted, moved or freed. Whenever a piece of memory becomes available again we keep around a fence which marks the end of using this piece of memory.
>>>
>>> When then somebody asks for new memory we work through the LRU and test if using a certain piece of memory makes sense or not. If we find that a BO needs to be evicted for this we return a reference to the BO in question to the upper level handling.
>>>
>>> If we find that we can do the allocation but only with recently freed up memory we gather the fences and say you can only use the newly allocated memory after waiting for those.
>>>
>>> HEY! Wait a second! Did I just outlined what a potential replacement to TTM would look like?
>> yes, that is a good picture. Looks like we could do more work hen. :)
>>
>> thanks
>> xinhui
>>
>>
>> --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
>> index e795d5b5f8af..ac826a09b4ec 100644
>> --- a/drivers/gpu/drm/ttm/ttm_bo.c
>> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
>> @@ -405,6 +405,9 @@ static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
>> if (bo->type != ttm_bo_type_sg) {
>> spin_lock(&ttm_bo_glob.lru_lock);
>> + /* it is very likely to release bo successfully.
>> + * if not, just adding it back.
>> + */
>> ttm_bo_del_from_lru(bo);
>> spin_unlock(&ttm_bo_glob.lru_lock);
>> bo->base.resv = &bo->base._resv;
>> @@ -466,18 +469,20 @@ static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
>> if (unlock_resv)
>> dma_resv_unlock(bo->base.resv);
>> - spin_unlock(&ttm_bo_glob.lru_lock);
>> lret = dma_resv_wait_timeout_rcu(resv, true, interruptible,
>> 30 * HZ);
>> if (lret < 0)
>> - return lret;
>> - else if (lret == 0)
>> - return -EBUSY;
>> + goto busy;
>> + else if (lret == 0) {
>> + ret = -EBUSY;
>> + goto busy;
>> + }
>> - spin_lock(&ttm_bo_glob.lru_lock);
>> if (unlock_resv && !dma_resv_trylock(bo->base.resv)) {
>> + /* no race should be on it now */
>> + BUG();
>> /*
>> * We raced, and lost, someone else holds the reservation now,
>> * and is probably busy in ttm_bo_cleanup_memtype_use.
>> @@ -486,20 +491,18 @@ static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
>> * delayed destruction would succeed, so just return success
>> * here.
>> */
>> - spin_unlock(&ttm_bo_glob.lru_lock);
>> return 0;
>> }
>> ret = 0;
>> }
>> - if (ret || unlikely(list_empty(&bo->ddestroy))) {
>> + if (ret) {
>> if (unlock_resv)
>> dma_resv_unlock(bo->base.resv);
>> - spin_unlock(&ttm_bo_glob.lru_lock);
>> - return ret;
>> + goto busy;
>> }
>> - ttm_bo_del_from_lru(bo);
>> + spin_lock(&ttm_bo_glob.lru_lock);
>> list_del_init(&bo->ddestroy);
>> spin_unlock(&ttm_bo_glob.lru_lock);
>> ttm_bo_cleanup_memtype_use(bo);
>> @@ -510,11 +513,20 @@ static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
>> ttm_bo_put(bo);
>> return 0;
>> +
>> +busy:
>> + spin_lock(&ttm_bo_glob.lru_lock);
>> + ttm_bo_add_mem_to_lru(bo, &bo->mem);
>> + spin_unlock(&ttm_bo_glob.lru_lock);
>> +
>> + return ret;
>> }
>> /**
>> * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
>> * encountered buffers.
>> + *
>> + * only called bo device release
>> */
>> static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
>> {
>> @@ -533,17 +545,17 @@ static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
>> list_move_tail(&bo->ddestroy, &removed);
>> if (!ttm_bo_get_unless_zero(bo))
>> continue;
>> + ttm_bo_del_from_lru(bo);
>> + spin_unlock(&glob->lru_lock);
>> if (remove_all || bo->base.resv != &bo->base._resv) {
>> - spin_unlock(&glob->lru_lock);
>> dma_resv_lock(bo->base.resv, NULL);
>> -
>> - spin_lock(&glob->lru_lock);
>> ttm_bo_cleanup_refs(bo, false, !remove_all, true);
>> -
>> } else if (dma_resv_trylock(bo->base.resv)) {
>> ttm_bo_cleanup_refs(bo, false, !remove_all, true);
>> } else {
>> + spin_lock(&glob->lru_lock);
>> + ttm_bo_add_mem_to_lru(bo, &bo->mem);
>> spin_unlock(&glob->lru_lock);
>> }
>> @@ -559,12 +571,8 @@ static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
>> static void ttm_bo_delayed_workqueue(struct work_struct *work)
>> {
>> - struct ttm_bo_device *bdev =
>> - container_of(work, struct ttm_bo_device, wq.work);
>> -
>> - if (!ttm_bo_delayed_delete(bdev, false))
>> - schedule_delayed_work(&bdev->wq,
>> - ((HZ / 100) < 1) ? 1 : HZ / 100);
>> + WARN(true, "do not schedule it");
>> + return;
>> }
>> static void ttm_bo_release(struct kref *kref)
>> @@ -595,6 +603,7 @@ static void ttm_bo_release(struct kref *kref)
>> ttm_mem_io_unlock(man);
>> }
>> + /* if bo is busy, spend a little more time to add bo to lru and ddestroy list*/
>> if (!dma_resv_test_signaled_rcu(bo->base.resv, true)) {
>> /* The BO is not idle, resurrect it for delayed destroy */
>> ttm_bo_flush_all_fences(bo);
>> @@ -615,8 +624,6 @@ static void ttm_bo_release(struct kref *kref)
>> list_add_tail(&bo->ddestroy, &bdev->ddestroy);
>> spin_unlock(&ttm_bo_glob.lru_lock);
>> - schedule_delayed_work(&bdev->wq,
>> - ((HZ / 100) < 1) ? 1 : HZ / 100);
>> return;
>> }
>> @@ -842,6 +849,9 @@ static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
>> return ret;
>> }
>> + ttm_bo_del_from_lru(bo);
>> + spin_unlock(&ttm_bo_glob.lru_lock);
>> +
>> if (bo->deleted) {
>> ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
>> ctx->no_wait_gpu, locked);
>> @@ -849,8 +859,6 @@ static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
>> return ret;
>> }
>> - spin_unlock(&ttm_bo_glob.lru_lock);
>> -
>> ret = ttm_bo_evict(bo, ctx);
>> if (locked)
>> ttm_bo_unreserve(bo);
>> @@ -1809,14 +1817,15 @@ int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx)
>> return ret;
>> }
>> + ttm_bo_del_from_lru(bo);
>> + spin_unlock(&glob->lru_lock);
>> +
>> if (bo->deleted) {
>> ret = ttm_bo_cleanup_refs(bo, false, false, locked);
>> ttm_bo_put(bo);
>> return ret;
>> }
>> - ttm_bo_del_from_lru(bo);
>> - spin_unlock(&glob->lru_lock);
>> /**
>> * Move to system cached
>
More information about the amd-gfx
mailing list