Skip to contentSkip to conversation

ttm: wait mem space if user allow while gpu busy

PATCH In review Chunming Zhou 6 participants 17 messages Apr 22, 2019 – Apr 24, 2019

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ttm: wait mem space if user allow while gpu busy

heavy gpu job could occupy memory long time, which could lead to other user fail to get memory.

Change-Id I0b322d98cd76e5ac32b00462bbae8008d76c5e11
---
 drivers/gpu/drm/ttm/ttm_bo.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
drivers/gpu/drm/ttm/ttm_bo.c +4 -2
index 7c484729f9b2..6c596cc24bec 100644
@@ -830,8 +830,10 @@ static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
 		if (mem->mm_node)
 			break;
 		ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
-		if (unlikely(ret != 0))
-			return ret;
+		if (unlikely(ret != 0)) {
+			if (!ctx || ctx->no_wait_gpu || ret != -EBUSY)
+				return ret;
+		}
 	} while (1);
 	mem->mem_type = mem_type;
 	return ttm_bo_add_move_fence(bo, man, mem);
Signature · 6 lines
2.17.1

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Acked-by on v1 1/1
#1 lore raw

Thanks,
Prike
-----Original Message-----
From: Chunming Zhou <[email protected]>
Sent: Monday, April 22, 2019 6:39 PM
To: [email protected]
Cc: Liang, Prike <[email protected]>; Zhou, David(ChunMing) <[email protected]>
Subject: [PATCH] ttm: wait mem space if user allow while gpu busy

heavy gpu job could occupy memory long time, which could lead to other user fail to get memory.

Change-Id I0b322d98cd76e5ac32b00462bbae8008d76c5e11
---
 drivers/gpu/drm/ttm/ttm_bo.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
drivers/gpu/drm/ttm/ttm_bo.c +4 -2
@@ -830,8 +830,10 @@ static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
 		if (mem->mm_node)
 			break;
 		ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
-		if (unlikely(ret != 0))
-			return ret;
+		if (unlikely(ret != 0)) {
+			if (!ctx || ctx->no_wait_gpu || ret != -EBUSY)
+				return ret;
+		}
 	} while (1);
 	mem->mem_type = mem_type;
 	return ttm_bo_add_move_fence(bo, man, mem);

--
2.17.1

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#2 lore raw

Well that is certainly a NAK because it can lead to deadlock in the memory management.

You can't just busy wait with all those locks held.

Regards,
Christian.

Am 23.04.19 um 03:45 schrieb Liang, Prike:

Quoted 38 lines

Thanks,
Prike
-----Original Message-----
From: Chunming Zhou <[email protected]>
Sent: Monday, April 22, 2019 6:39 PM
To: [email protected]
Cc: Liang, Prike <[email protected]>; Zhou, David(ChunMing) <[email protected]>
Subject: [PATCH] ttm: wait mem space if user allow while gpu busy

heavy gpu job could occupy memory long time, which could lead to other user fail to get memory.

Change-Id I0b322d98cd76e5ac32b00462bbae8008d76c5e11
---
  drivers/gpu/drm/ttm/ttm_bo.c | 6 ++++--
  1 file changed, 4 insertions(+), 2 deletions(-)
drivers/gpu/drm/ttm/ttm_bo.c +4 -3
@@ -830,8 +830,10 @@ static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
  		if (mem->mm_node)
  			break;
  		ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
-		if (unlikely(ret != 0))
-			return ret;
+		if (unlikely(ret != 0)) {
+			if (!ctx || ctx->no_wait_gpu || ret != -EBUSY)
+				return ret;
+		}
  	} while (1);
  	mem->mem_type = mem_type;
  	return ttm_bo_add_move_fence(bo, man, mem);
--

2.17.1

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#3 lore raw

How about adding more condition ctx->resv inline to address your concern? As well as don't wait from same user, shouldn't lead to deadlock.

Otherwise, any other idea?

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Liang, Prike" ,"Zhou, David(ChunMing)" ,[email protected] CC:

Well that is certainly a NAK because it can lead to deadlock in the memory management.

You can't just busy wait with all those locks held.

Regards,
Christian.

Am 23.04.19 um 03:45 schrieb Liang, Prike:

Quoted 38 lines

Thanks,
Prike
-----Original Message-----
From: Chunming Zhou <[email protected]>
Sent: Monday, April 22, 2019 6:39 PM
To: [email protected]
Cc: Liang, Prike <[email protected]>; Zhou, David(ChunMing) <[email protected]>
Subject: [PATCH] ttm: wait mem space if user allow while gpu busy

heavy gpu job could occupy memory long time, which could lead to other user fail to get memory.

Change-Id I0b322d98cd76e5ac32b00462bbae8008d76c5e11
---
  drivers/gpu/drm/ttm/ttm_bo.c | 6 ++++--
  1 file changed, 4 insertions(+), 2 deletions(-)
drivers/gpu/drm/ttm/ttm_bo.c +4 -3
@@ -830,8 +830,10 @@ static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
               if (mem->mm_node)
                       break;
               ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
-             if (unlikely(ret != 0))
-                     return ret;
+             if (unlikely(ret != 0)) {
+                     if (!ctx || ctx->no_wait_gpu || ret != -EBUSY)
+                             return ret;
+             }
       } while (1);
       mem->mem_type = mem_type;
       return ttm_bo_add_move_fence(bo, man, mem);
--

2.17.1

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#4 lore raw

Well that's not so easy of hand.

The basic problem here is that when you busy wait at this place you can easily run into situations where application A busy waits for B while B busy waits for A -> deadlock.

So what we need here is the deadlock detection logic of the ww_mutex. To use this we at least need to do the following steps:

1. Reserve the BO in DC using a ww_mutex ticket (trivial).

2. If we then run into this EBUSY condition in TTM check if the BO we need memory for (or rather the ww_mutex of its reservation object) has a ticket assigned.

3. If we have a ticket we grab a reference to the first BO on the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

4. If getting the reservation lock with the ticket succeeded we check if the BO is still the first one on the LRU in question (the BO could have moved).

5. If the BO is still the first one on the LRU in question we try to evict it as we would evict any other BO.

6. If any of the "If's" above fail we just back off and return -EBUSY.

Steps 2-5 are certainly not trivial, but doable as far as I can see.

Have fun :)
Christian.

Am 23.04.19 um 15:19 schrieb Zhou, David(ChunMing):

Quoted 67 lines

How about adding more condition ctx->resv inline to address your concern? As well as don't wait from same user, shouldn't lead to deadlock.

Otherwise, any other idea?

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Liang, Prike" ,"Zhou, David(ChunMing)" ,[email protected] CC:

Well that is certainly a NAK because it can lead to deadlock in the memory management.

You can't just busy wait with all those locks held.

Regards,
Christian.

Am 23.04.19 um 03:45 schrieb Liang, Prike:

Quoted 9 lines

Thanks,
Prike
-----Original Message-----
From: Chunming Zhou <[email protected]>
Sent: Monday, April 22, 2019 6:39 PM
To: [email protected]
Cc: Liang, Prike <[email protected]>; Zhou, David(ChunMing)

<[email protected]>

Quoted 3 lines

Subject: [PATCH] ttm: wait mem space if user allow while gpu busy

heavy gpu job could occupy memory long time, which could lead to

other user fail to get memory.

Review

Quoted 8 lines
Change-Id I0b322d98cd76e5ac32b00462bbae8008d76c5e11

---
   drivers/gpu/drm/ttm/ttm_bo.c | 6 ++++--

   1 file changed, 4 insertions(+), 2 deletions(-)
(quoted diff) +0 -0

b/drivers/gpu/drm/ttm/ttm_bo.c index 7c484729f9b2..6c596cc24bec 100644

drivers/gpu/drm/ttm/ttm_bo.c static int ttm_bo_mem_force_space(struct

Quoted 3 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -0
@@ -830,8 +830,10 @@ static int ttm_bo_mem_force_space(struct 

ttm_buffer_object *bo,

Quoted 15 lines
drivers/gpu/drm/ttm/ttm_bo.c +4 -3
                 if (mem->mm_node)
                         break;
                 ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
-             if (unlikely(ret != 0))
-                     return ret;
+             if (unlikely(ret != 0)) {
+                     if (!ctx || ctx->no_wait_gpu || ret != -EBUSY)
+                             return ret;
+             }
         } while (1);
         mem->mem_type = mem_type;
         return ttm_bo_add_move_fence(bo, man, mem);
--

2.17.1

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#5 lore raw
Quoted 1 line

3. If we have a ticket we grab a reference to the first BO on the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

The BO on LRU is already locked by cs user, can it be dropped here by DC user? and then DC user grab its lock with ticket, how does CS grab it again?

If you think waiting in ttm has this risk, how about just adding a wrapper for pin function as below? amdgpu_get_pin_bo_timeout()
{
do {
amdgpo_bo_reserve();
r=amdgpu_bo_pin();

if(!r)

        break;

amdgpu_bo_unreserve();
timeout--;

} while(timeout>0);

}

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Zhou, David(ChunMing)" ,"Koenig, Christian" ,"Liang, Prike" ,[email protected] CC:

Well that's not so easy of hand.

The basic problem here is that when you busy wait at this place you can easily run into situations where application A busy waits for B while B busy waits for A -> deadlock.

So what we need here is the deadlock detection logic of the ww_mutex. To use this we at least need to do the following steps:

1. Reserve the BO in DC using a ww_mutex ticket (trivial).

2. If we then run into this EBUSY condition in TTM check if the BO we need memory for (or rather the ww_mutex of its reservation object) has a ticket assigned.

3. If we have a ticket we grab a reference to the first BO on the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

4. If getting the reservation lock with the ticket succeeded we check if the BO is still the first one on the LRU in question (the BO could have moved).

5. If the BO is still the first one on the LRU in question we try to evict it as we would evict any other BO.

6. If any of the "If's" above fail we just back off and return -EBUSY.

Steps 2-5 are certainly not trivial, but doable as far as I can see.

Have fun :)
Christian.

Am 23.04.19 um 15:19 schrieb Zhou, David(ChunMing):
How about adding more condition ctx->resv inline to address your concern? As well as don't wait from same user, shouldn't lead to deadlock.

Otherwise, any other idea?

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Liang, Prike" ,"Zhou, David(ChunMing)" ,[email protected]<mailto:[email protected]> CC:

Well that is certainly a NAK because it can lead to deadlock in the memory management.

You can't just busy wait with all those locks held.

Regards,
Christian.

Am 23.04.19 um 03:45 schrieb Liang, Prike:

Quoted 38 lines

Thanks,
Prike
-----Original Message-----
From: Chunming Zhou <[email protected]><mailto:[email protected]>
Sent: Monday, April 22, 2019 6:39 PM
To: [email protected]<mailto:[email protected]>
Cc: Liang, Prike <[email protected]><mailto:[email protected]>; Zhou, David(ChunMing) <[email protected]><mailto:[email protected]>
Subject: [PATCH] ttm: wait mem space if user allow while gpu busy

heavy gpu job could occupy memory long time, which could lead to other user fail to get memory.

Change-Id I0b322d98cd76e5ac32b00462bbae8008d76c5e11
---
  drivers/gpu/drm/ttm/ttm_bo.c | 6 ++++--
  1 file changed, 4 insertions(+), 2 deletions(-)
drivers/gpu/drm/ttm/ttm_bo.c +4 -3
@@ -830,8 +830,10 @@ static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
               if (mem->mm_node)
                       break;
               ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
-             if (unlikely(ret != 0))
-                     return ret;
+             if (unlikely(ret != 0)) {
+                     if (!ctx || ctx->no_wait_gpu || ret != -EBUSY)
+                             return ret;
+             }
       } while (1);
       mem->mem_type = mem_type;
       return ttm_bo_add_move_fence(bo, man, mem);
--

2.17.1

Quoted 4 lines

_______________________________________________
dri-devel mailing list [email protected]<mailto:[email protected]>
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list [email protected]<mailto:[email protected]>
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#6 lore raw
Quoted 1 line

how about just adding a wrapper for pin function as below?

I considered this as well and don't think it will work reliable.

We could use it as a band aid for this specific problem, but in general we need to improve the handling in TTM to resolve those kind of resource conflicts.

Regards,
Christian.

Am 23.04.19 um 17:09 schrieb Zhou, David(ChunMing):

Quoted 124 lines
Quoted 1 line

3. If we have a ticket we grab a reference to the first BO on the

LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

The BO on LRU is already locked by cs user, can it be dropped here by DC user? and then DC user grab its lock with ticket, how does CS grab it again?

If you think waiting in ttm has this risk, how about just adding a wrapper for pin function as below? amdgpu_get_pin_bo_timeout()
{
do {
amdgpo_bo_reserve();
r=amdgpu_bo_pin();

if(!r)
        break;
amdgpu_bo_unreserve();
timeout--;

} while(timeout>0);

}

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Zhou, David(ChunMing)" ,"Koenig, Christian" ,"Liang, Prike" ,[email protected] CC:

Well that's not so easy of hand.

The basic problem here is that when you busy wait at this place you can easily run into situations where application A busy waits for B while B busy waits for A -> deadlock.

So what we need here is the deadlock detection logic of the ww_mutex. To use this we at least need to do the following steps:

1. Reserve the BO in DC using a ww_mutex ticket (trivial).

2. If we then run into this EBUSY condition in TTM check if the BO we need memory for (or rather the ww_mutex of its reservation object) has a ticket assigned.

3. If we have a ticket we grab a reference to the first BO on the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

4. If getting the reservation lock with the ticket succeeded we check if the BO is still the first one on the LRU in question (the BO could have moved).

5. If the BO is still the first one on the LRU in question we try to evict it as we would evict any other BO.

6. If any of the "If's" above fail we just back off and return -EBUSY.

Steps 2-5 are certainly not trivial, but doable as far as I can see.

Have fun :)
Christian.

Am 23.04.19 um 15:19 schrieb Zhou, David(ChunMing):

Quoted 67 lines

How about adding more condition ctx->resv inline to address your concern? As well as don't wait from same user, shouldn't lead to deadlock.

Otherwise, any other idea?

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Liang, Prike" ,"Zhou, David(ChunMing)" ,[email protected] CC:

Well that is certainly a NAK because it can lead to deadlock in the memory management.

You can't just busy wait with all those locks held.

Regards,
Christian.

Am 23.04.19 um 03:45 schrieb Liang, Prike:

Quoted 9 lines

Thanks,
Prike
-----Original Message-----
From: Chunming Zhou <[email protected]>
Sent: Monday, April 22, 2019 6:39 PM
To: [email protected]
Cc: Liang, Prike <[email protected]>; Zhou, David(ChunMing)

<[email protected]>

Quoted 3 lines

Subject: [PATCH] ttm: wait mem space if user allow while gpu busy

heavy gpu job could occupy memory long time, which could lead to

other user fail to get memory.

Review

Quoted 8 lines
Change-Id I0b322d98cd76e5ac32b00462bbae8008d76c5e11

---
   drivers/gpu/drm/ttm/ttm_bo.c | 6 ++++--

   1 file changed, 4 insertions(+), 2 deletions(-)
(quoted diff) +0 -0

b/drivers/gpu/drm/ttm/ttm_bo.c index 7c484729f9b2..6c596cc24bec 100644

drivers/gpu/drm/ttm/ttm_bo.c static int ttm_bo_mem_force_space(struct

Quoted 3 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -0
@@ -830,8 +830,10 @@ static int ttm_bo_mem_force_space(struct 

ttm_buffer_object *bo,

Quoted 15 lines
drivers/gpu/drm/ttm/ttm_bo.c +4 -3
                 if (mem->mm_node)
                         break;
                 ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
-             if (unlikely(ret != 0))
-                     return ret;
+             if (unlikely(ret != 0)) {
+                     if (!ctx || ctx->no_wait_gpu || ret != -EBUSY)
+                             return ret;
+             }
         } while (1);
         mem->mem_type = mem_type;
         return ttm_bo_add_move_fence(bo, man, mem);
--

2.17.1

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#7 lore raw

I made a patch as attached.

I'm not sure how to make patch as your proposal, Could you make a patch for that if mine isn't enough?

-David

Quoted 133 lines · Christian König
Quoted 1 line

how about just adding a wrapper for pin function as below?

I considered this as well and don't think it will work reliable.

We could use it as a band aid for this specific problem, but in general we need to improve the handling in TTM to resolve those kind of resource conflicts.

Regards,
Christian.

Am 23.04.19 um 17:09 schrieb Zhou, David(ChunMing):

Quoted 124 lines
Quoted 1 line

3. If we have a ticket we grab a reference to the first BO on the

LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

The BO on LRU is already locked by cs user, can it be dropped here by DC user? and then DC user grab its lock with ticket, how does CS grab it again?

If you think waiting in ttm has this risk, how about just adding a wrapper for pin function as below? amdgpu_get_pin_bo_timeout()
{
do {
amdgpo_bo_reserve();
r=amdgpu_bo_pin();

if(!r)
        break;
amdgpu_bo_unreserve();
timeout--;

} while(timeout>0);

}

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Zhou, David(ChunMing)" ,"Koenig, Christian" ,"Liang, Prike" ,[email protected] CC:

Well that's not so easy of hand.

The basic problem here is that when you busy wait at this place you can easily run into situations where application A busy waits for B while B busy waits for A -> deadlock.

So what we need here is the deadlock detection logic of the ww_mutex. To use this we at least need to do the following steps:

1. Reserve the BO in DC using a ww_mutex ticket (trivial).

2. If we then run into this EBUSY condition in TTM check if the BO we need memory for (or rather the ww_mutex of its reservation object) has a ticket assigned.

3. If we have a ticket we grab a reference to the first BO on the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

4. If getting the reservation lock with the ticket succeeded we check if the BO is still the first one on the LRU in question (the BO could have moved).

5. If the BO is still the first one on the LRU in question we try to evict it as we would evict any other BO.

6. If any of the "If's" above fail we just back off and return -EBUSY.

Steps 2-5 are certainly not trivial, but doable as far as I can see.

Have fun :)
Christian.

Am 23.04.19 um 15:19 schrieb Zhou, David(ChunMing):

Quoted 71 lines

How about adding more condition ctx->resv inline to address your concern? As well as don't wait from same user, shouldn't lead to deadlock.

Otherwise, any other idea?

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Liang, Prike" ,"Zhou, David(ChunMing)" ,[email protected] CC:

Well that is certainly a NAK because it can lead to deadlock in the memory management.

You can't just busy wait with all those locks held.

Regards,
Christian.

Am 23.04.19 um 03:45 schrieb Liang, Prike:

Quoted 9 lines

Thanks,
Prike
-----Original Message-----
From: Chunming Zhou <[email protected]>
Sent: Monday, April 22, 2019 6:39 PM
To: [email protected]
Cc: Liang, Prike <[email protected]>; Zhou, David(ChunMing)

<[email protected]>

Quoted 3 lines

Subject: [PATCH] ttm: wait mem space if user allow while gpu busy

heavy gpu job could occupy memory long time, which could lead to

other user fail to get memory.

Review

Quoted 8 lines
Change-Id I0b322d98cd76e5ac32b00462bbae8008d76c5e11

---
   drivers/gpu/drm/ttm/ttm_bo.c | 6 ++++--

   1 file changed, 4 insertions(+), 2 deletions(-)
(quoted diff) +0 -0

b/drivers/gpu/drm/ttm/ttm_bo.c index 7c484729f9b2..6c596cc24bec 100644

drivers/gpu/drm/ttm/ttm_bo.c static int ttm_bo_mem_force_space(struct

Quoted 3 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -0
@@ -830,8 +830,10 @@ static int ttm_bo_mem_force_space(struct 

ttm_buffer_object *bo,

Quoted 15 lines
drivers/gpu/drm/ttm/ttm_bo.c +4 -3
                 if (mem->mm_node)
                         break;
                 ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
-             if (unlikely(ret != 0))
-                     return ret;
+             if (unlikely(ret != 0)) {
+                     if (!ctx || ctx->no_wait_gpu || ret != -EBUSY)
+                             return ret;
+             }
         } while (1);
         mem->mem_type = mem_type;
         return ttm_bo_add_move_fence(bo, man, mem);
--

2.17.1

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 7 lines
Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

From 184941165665d80ad1992179e7652d7e4d739b2e Mon Sep 17 00:00:00 2001

From: Chunming Zhou <[email protected]>
Date: Wed, 24 Apr 2019 11:35:29 +0800
Subject: [PATCH] ttm: return EBUSY to user

EBUSY means there is no idle BO on LRU which can be evicted, the error is meaningful to user.

Change-Id I3857c4b99859c9d2f354cf2c486dbacb8520d0ed
---
 drivers/gpu/drm/ttm/ttm_bo.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/ttm/ttm_bo.c +1 -1
index 7c484729f9b2..dce90053f29e 100644
@@ -1000,7 +1000,7 @@ int ttm_bo_mem_space(struct ttm_buffer_object *bo,
 		return -EINVAL;
 	}
 
-	return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
+	return ret == -EBUSY ? ret : (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
 }
 EXPORT_SYMBOL(ttm_bo_mem_space);
 
Signature · 1 line
2.17.1

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#8 lore raw

That would change the semantics of ttm_bo_mem_space() and so could change the return code in an IOCTL as well. Not a good idea, cause that could have a lot of side effects.

Instead in the calling DC code you could check if you get an -ENOMEM and then call schedule().

If after the schedule() we see that we have now BOs on the LRU we can try again and see if pinning the frame buffer now succeeds.

Christian.

Am 24.04.19 um 09:17 schrieb zhoucm1:

Quoted 148 lines

I made a patch as attached.

I'm not sure how to make patch as your proposal, Could you make a patch for that if mine isn't enough?

-David

Quoted 138 lines · Christian König
Quoted 1 line

how about just adding a wrapper for pin function as below?

I considered this as well and don't think it will work reliable.

We could use it as a band aid for this specific problem, but in general we need to improve the handling in TTM to resolve those kind of resource conflicts.

Regards,
Christian.

Am 23.04.19 um 17:09 schrieb Zhou, David(ChunMing):

Quoted 128 lines
Quoted 1 line

3. If we have a ticket we grab a reference to the first BO on the

LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

The BO on LRU is already locked by cs user, can it be dropped here by DC user? and then DC user grab its lock with ticket, how does CS grab it again?

If you think waiting in ttm has this risk, how about just adding a wrapper for pin function as below? amdgpu_get_pin_bo_timeout()
{
do {
amdgpo_bo_reserve();
r=amdgpu_bo_pin();

if(!r)
        break;
amdgpu_bo_unreserve();
timeout--;

} while(timeout>0);

}

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Zhou, David(ChunMing)" ,"Koenig, Christian" ,"Liang, Prike" ,[email protected] CC:

Well that's not so easy of hand.

The basic problem here is that when you busy wait at this place you can easily run into situations where application A busy waits for B while B busy waits for A -> deadlock.

So what we need here is the deadlock detection logic of the ww_mutex. To use this we at least need to do the following steps:

1. Reserve the BO in DC using a ww_mutex ticket (trivial).

2. If we then run into this EBUSY condition in TTM check if the BO we need memory for (or rather the ww_mutex of its reservation object) has a ticket assigned.

3. If we have a ticket we grab a reference to the first BO on the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

4. If getting the reservation lock with the ticket succeeded we check if the BO is still the first one on the LRU in question (the BO could have moved).

5. If the BO is still the first one on the LRU in question we try to evict it as we would evict any other BO.

6. If any of the "If's" above fail we just back off and return -EBUSY.

Steps 2-5 are certainly not trivial, but doable as far as I can see.

Have fun :)
Christian.

Am 23.04.19 um 15:19 schrieb Zhou, David(ChunMing):

Quoted 67 lines

How about adding more condition ctx->resv inline to address your concern? As well as don't wait from same user, shouldn't lead to deadlock.

Otherwise, any other idea?

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Liang, Prike" ,"Zhou, David(ChunMing)" ,[email protected] CC:

Well that is certainly a NAK because it can lead to deadlock in the memory management.

You can't just busy wait with all those locks held.

Regards,
Christian.

Am 23.04.19 um 03:45 schrieb Liang, Prike:

Quoted 9 lines

Thanks,
Prike
-----Original Message-----
From: Chunming Zhou <[email protected]>
Sent: Monday, April 22, 2019 6:39 PM
To: [email protected]
Cc: Liang, Prike <[email protected]>; Zhou, David(ChunMing)

<[email protected]>

Quoted 3 lines

Subject: [PATCH] ttm: wait mem space if user allow while gpu busy

heavy gpu job could occupy memory long time, which could lead to

other user fail to get memory.

Review

Quoted 8 lines
Change-Id I0b322d98cd76e5ac32b00462bbae8008d76c5e11

---
   drivers/gpu/drm/ttm/ttm_bo.c | 6 ++++--

   1 file changed, 4 insertions(+), 2 deletions(-)
(quoted diff) +0 -0

b/drivers/gpu/drm/ttm/ttm_bo.c index 7c484729f9b2..6c596cc24bec 100644

drivers/gpu/drm/ttm/ttm_bo.c static int ttm_bo_mem_force_space(struct

Quoted 3 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -0
@@ -830,8 +830,10 @@ static int ttm_bo_mem_force_space(struct 

ttm_buffer_object *bo,

Quoted 15 lines
drivers/gpu/drm/ttm/ttm_bo.c +4 -3
                 if (mem->mm_node)
                         break;
                 ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
-             if (unlikely(ret != 0))
-                     return ret;
+             if (unlikely(ret != 0)) {
+                     if (!ctx || ctx->no_wait_gpu || ret != -EBUSY)
+                             return ret;
+             }
         } while (1);
         mem->mem_type = mem_type;
         return ttm_bo_add_move_fence(bo, man, mem);
--

2.17.1

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#9 lore raw

how about new attached?

-David

Quoted 158 lines · Christian König

That would change the semantics of ttm_bo_mem_space() and so could change the return code in an IOCTL as well. Not a good idea, cause that could have a lot of side effects.

Instead in the calling DC code you could check if you get an -ENOMEM and then call schedule().

If after the schedule() we see that we have now BOs on the LRU we can try again and see if pinning the frame buffer now succeeds.

Christian.

Am 24.04.19 um 09:17 schrieb zhoucm1:

Quoted 149 lines

I made a patch as attached.

I'm not sure how to make patch as your proposal, Could you make a patch for that if mine isn't enough?

-David

Quoted 139 lines · Christian König
Quoted 1 line

how about just adding a wrapper for pin function as below?

I considered this as well and don't think it will work reliable.

We could use it as a band aid for this specific problem, but in general we need to improve the handling in TTM to resolve those kind of resource conflicts.

Regards,
Christian.

Am 23.04.19 um 17:09 schrieb Zhou, David(ChunMing):

Quoted 129 lines
Quoted 1 line

3. If we have a ticket we grab a reference to the first BO on the

LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

The BO on LRU is already locked by cs user, can it be dropped here by DC user? and then DC user grab its lock with ticket, how does CS grab it again?

If you think waiting in ttm has this risk, how about just adding a wrapper for pin function as below? amdgpu_get_pin_bo_timeout()
{
do {
amdgpo_bo_reserve();
r=amdgpu_bo_pin();

if(!r)
        break;
amdgpu_bo_unreserve();
timeout--;

} while(timeout>0);

}

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Zhou, David(ChunMing)" ,"Koenig, Christian" ,"Liang, Prike" ,[email protected] CC:

Well that's not so easy of hand.

The basic problem here is that when you busy wait at this place you can easily run into situations where application A busy waits for B while B busy waits for A -> deadlock.

So what we need here is the deadlock detection logic of the ww_mutex. To use this we at least need to do the following steps:

1. Reserve the BO in DC using a ww_mutex ticket (trivial).

2. If we then run into this EBUSY condition in TTM check if the BO we need memory for (or rather the ww_mutex of its reservation object) has a ticket assigned.

3. If we have a ticket we grab a reference to the first BO on the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

4. If getting the reservation lock with the ticket succeeded we check if the BO is still the first one on the LRU in question (the BO could have moved).

5. If the BO is still the first one on the LRU in question we try to evict it as we would evict any other BO.

6. If any of the "If's" above fail we just back off and return -EBUSY.

Steps 2-5 are certainly not trivial, but doable as far as I can see.

Have fun :)
Christian.

Am 23.04.19 um 15:19 schrieb Zhou, David(ChunMing):

Quoted 68 lines

How about adding more condition ctx->resv inline to address your concern? As well as don't wait from same user, shouldn't lead to deadlock.

Otherwise, any other idea?

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Liang, Prike" ,"Zhou, David(ChunMing)" ,[email protected] CC:

Well that is certainly a NAK because it can lead to deadlock in the memory management.

You can't just busy wait with all those locks held.

Regards,
Christian.

Am 23.04.19 um 03:45 schrieb Liang, Prike:

Quoted 9 lines

Thanks,
Prike
-----Original Message-----
From: Chunming Zhou <[email protected]>
Sent: Monday, April 22, 2019 6:39 PM
To: [email protected]
Cc: Liang, Prike <[email protected]>; Zhou, David(ChunMing)

<[email protected]>

Quoted 3 lines

Subject: [PATCH] ttm: wait mem space if user allow while gpu busy

heavy gpu job could occupy memory long time, which could lead to

other user fail to get memory.

Review

Quoted 8 lines
Change-Id I0b322d98cd76e5ac32b00462bbae8008d76c5e11

---
   drivers/gpu/drm/ttm/ttm_bo.c | 6 ++++--

   1 file changed, 4 insertions(+), 2 deletions(-)
(quoted diff) +0 -0

b/drivers/gpu/drm/ttm/ttm_bo.c index 7c484729f9b2..6c596cc24bec 100644

drivers/gpu/drm/ttm/ttm_bo.c static int ttm_bo_mem_force_space(struct

Quoted 3 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -0
@@ -830,8 +830,10 @@ static int ttm_bo_mem_force_space(struct 

ttm_buffer_object *bo,

drivers/gpu/drm/ttm/ttm_bo.c static int ttm_bo_mem_force_space(struct · line 832

Quoted 3 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -0
                 if (mem->mm_node)
                         break;
                 ret = ttm_mem_evict_first(bdev, mem_type, place, 

ctx);

Quoted 12 lines
drivers/gpu/drm/ttm/ttm_bo.c +4 -3
-             if (unlikely(ret != 0))
-                     return ret;
+             if (unlikely(ret != 0)) {
+                     if (!ctx || ctx->no_wait_gpu || ret != -EBUSY)
+                             return ret;
+             }
         } while (1);
         mem->mem_type = mem_type;
         return ttm_bo_add_move_fence(bo, man, mem);
--

2.17.1

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 7 lines
Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

From 61174dd44618bbfcb035484a44435b7ed4df86a5 Mon Sep 17 00:00:00 2001

From: Chunming Zhou <[email protected]>
Date: Wed, 24 Apr 2019 11:35:29 +0800
Subject: [PATCH] drm/amdgpu: wait memory idle when fb is pinned fail

heavy gpu job execution could occupy memory long time, which could lead to other user fail to get memory.

Change-Id I3857c4b99859c9d2f354cf2c486dbacb8520d0ed
---
 drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 15 +++++++++++++++
 1 file changed, 15 insertions(+)
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +15 -0
index a5cacf846e1b..8577b629a640 100644
@@ -4093,6 +4093,8 @@ static const struct drm_plane_funcs dm_plane_funcs = {
 	.atomic_destroy_state = dm_drm_plane_destroy_state,
 };
 
+/* 100s default to wait for pin fb */
+#define DM_PIN_MEM_TIMEOUT 100000000000ULL
 static int dm_plane_helper_prepare_fb(struct drm_plane *plane,
 				      struct drm_plane_state *new_state)
 {
@@ -4103,6 +4105,7 @@ static int dm_plane_helper_prepare_fb(struct drm_plane *plane,
 	struct dm_plane_state *dm_plane_state_new, *dm_plane_state_old;
 	uint64_t tiling_flags;
 	uint32_t domain;
+	u64 timeout = nsecs_to_jiffies64(DM_PIN_MEM_TIMEOUT);
 	int r;
 
 	dm_plane_state_old = to_dm_plane_state(plane->state);
@@ -4117,6 +4120,8 @@ static int dm_plane_helper_prepare_fb(struct drm_plane *plane,
 	obj = new_state->fb->obj[0];
 	rbo = gem_to_amdgpu_bo(obj);
 	adev = amdgpu_ttm_adev(rbo->tbo.bdev);
+
+retry:
 	r = amdgpu_bo_reserve(rbo, false);
 	if (unlikely(r != 0))
 		return r;
@@ -4131,8 +4136,18 @@ static int dm_plane_helper_prepare_fb(struct drm_plane *plane,
 		if (r != -ERESTARTSYS)
 			DRM_ERROR("Failed to pin framebuffer with error %d\n", r);
 		amdgpu_bo_unreserve(rbo);
+		if (r == -ENOMEM) {
+			set_current_state(TASK_INTERRUPTIBLE);
+			if (timeout == 0)
+				return r;
+			if (signal_pending(current))
+				return r;
+			timeout = schedule_timeout(timeout);
+			goto retry;
+		}
 		return r;
 	}
+        __set_current_state(TASK_RUNNING);
 
 	r = amdgpu_ttm_alloc_gart(&rbo->tbo);
 	if (unlikely(r != 0)) {
Signature · 1 line
2.17.1

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#10 lore raw

This is used in a work item, so you don't need to check for signals.

And checking if the LRU is populated is mandatory here or otherwise you can run into an endless loop.

Christian.

Am 24.04.19 um 09:59 schrieb zhoucm1:

Quoted 174 lines

how about new attached?

-David

Quoted 164 lines · Christian König

That would change the semantics of ttm_bo_mem_space() and so could change the return code in an IOCTL as well. Not a good idea, cause that could have a lot of side effects.

Instead in the calling DC code you could check if you get an -ENOMEM and then call schedule().

If after the schedule() we see that we have now BOs on the LRU we can try again and see if pinning the frame buffer now succeeds.

Christian.

Am 24.04.19 um 09:17 schrieb zhoucm1:

Quoted 154 lines

I made a patch as attached.

I'm not sure how to make patch as your proposal, Could you make a patch for that if mine isn't enough?

-David

Quoted 140 lines · Christian König
Quoted 1 line

how about just adding a wrapper for pin function as below?

I considered this as well and don't think it will work reliable.

We could use it as a band aid for this specific problem, but in general we need to improve the handling in TTM to resolve those kind of resource conflicts.

Regards,
Christian.

Am 23.04.19 um 17:09 schrieb Zhou, David(ChunMing):

Quoted 130 lines
Quoted 1 line

3. If we have a ticket we grab a reference to the first BO on the

LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

The BO on LRU is already locked by cs user, can it be dropped here by DC user? and then DC user grab its lock with ticket, how does CS grab it again?

If you think waiting in ttm has this risk, how about just adding a wrapper for pin function as below? amdgpu_get_pin_bo_timeout()
{
do {
amdgpo_bo_reserve();
r=amdgpu_bo_pin();

if(!r)
        break;
amdgpu_bo_unreserve();
timeout--;

} while(timeout>0);

}

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Zhou, David(ChunMing)" ,"Koenig, Christian" ,"Liang, Prike" ,[email protected] CC:

Well that's not so easy of hand.

The basic problem here is that when you busy wait at this place you can easily run into situations where application A busy waits for B while B busy waits for A -> deadlock.

So what we need here is the deadlock detection logic of the ww_mutex. To use this we at least need to do the following steps:

1. Reserve the BO in DC using a ww_mutex ticket (trivial).

2. If we then run into this EBUSY condition in TTM check if the BO we need memory for (or rather the ww_mutex of its reservation object) has a ticket assigned.

3. If we have a ticket we grab a reference to the first BO on the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

4. If getting the reservation lock with the ticket succeeded we check if the BO is still the first one on the LRU in question (the BO could have moved).

5. If the BO is still the first one on the LRU in question we try to evict it as we would evict any other BO.

6. If any of the "If's" above fail we just back off and return -EBUSY.

Steps 2-5 are certainly not trivial, but doable as far as I can see.

Have fun :)
Christian.

Am 23.04.19 um 15:19 schrieb Zhou, David(ChunMing):

Quoted 69 lines

How about adding more condition ctx->resv inline to address your concern? As well as don't wait from same user, shouldn't lead to deadlock.

Otherwise, any other idea?

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Liang, Prike" ,"Zhou, David(ChunMing)" ,[email protected] CC:

Well that is certainly a NAK because it can lead to deadlock in the memory management.

You can't just busy wait with all those locks held.

Regards,
Christian.

Am 23.04.19 um 03:45 schrieb Liang, Prike:

Quoted 9 lines

Thanks,
Prike
-----Original Message-----
From: Chunming Zhou <[email protected]>
Sent: Monday, April 22, 2019 6:39 PM
To: [email protected]
Cc: Liang, Prike <[email protected]>; Zhou, David(ChunMing)

<[email protected]>

Quoted 3 lines

Subject: [PATCH] ttm: wait mem space if user allow while gpu busy

heavy gpu job could occupy memory long time, which could lead

to other user fail to get memory.

Review

Quoted 8 lines
Change-Id I0b322d98cd76e5ac32b00462bbae8008d76c5e11

---
   drivers/gpu/drm/ttm/ttm_bo.c | 6 ++++--

   1 file changed, 4 insertions(+), 2 deletions(-)
(quoted diff) +0 -0

b/drivers/gpu/drm/ttm/ttm_bo.c index 7c484729f9b2..6c596cc24bec 100644

drivers/gpu/drm/ttm/ttm_bo.c static int ttm_bo_mem_force_space(struct

Quoted 3 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -0
@@ -830,8 +830,10 @@ static int ttm_bo_mem_force_space(struct 

ttm_buffer_object *bo,

drivers/gpu/drm/ttm/ttm_bo.c static int ttm_bo_mem_force_space(struct · line 832

Quoted 3 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -0
                 if (mem->mm_node)
                         break;
                 ret = ttm_mem_evict_first(bdev, mem_type, place, 

ctx);

drivers/gpu/drm/ttm/ttm_bo.c static int ttm_bo_mem_force_space(struct · line 834

Quoted 4 lines
drivers/gpu/drm/ttm/ttm_bo.c +2 -2
-             if (unlikely(ret != 0))
-                     return ret;
+             if (unlikely(ret != 0)) {
+                     if (!ctx || ctx->no_wait_gpu || ret != 

-EBUSY)

Quoted 8 lines
drivers/gpu/drm/ttm/ttm_bo.c +2 -1
+                             return ret;
+             }
         } while (1);
         mem->mem_type = mem_type;
         return ttm_bo_add_move_fence(bo, man, mem);
--

2.17.1

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#11 lore raw
Quoted 1 line · Christian König

This is used in a work item, so you don't need to check for signals.

will remove.

Quoted 2 lines

And checking if the LRU is populated is mandatory here

How to check it outside of TTM? because the code is in dm.

Quoted 1 line

or otherwise you can run into an endless loop.

I already add a timeout for that.

-David

Quoted 184 lines

Christian.

Am 24.04.19 um 09:59 schrieb zhoucm1:

Quoted 179 lines

how about new attached?

-David

Quoted 165 lines · Christian König

That would change the semantics of ttm_bo_mem_space() and so could change the return code in an IOCTL as well. Not a good idea, cause that could have a lot of side effects.

Instead in the calling DC code you could check if you get an -ENOMEM and then call schedule().

If after the schedule() we see that we have now BOs on the LRU we can try again and see if pinning the frame buffer now succeeds.

Christian.

Am 24.04.19 um 09:17 schrieb zhoucm1:

Quoted 155 lines

I made a patch as attached.

I'm not sure how to make patch as your proposal, Could you make a patch for that if mine isn't enough?

-David

Quoted 141 lines · Christian König
Quoted 1 line

how about just adding a wrapper for pin function as below?

I considered this as well and don't think it will work reliable.

We could use it as a band aid for this specific problem, but in general we need to improve the handling in TTM to resolve those kind of resource conflicts.

Regards,
Christian.

Am 23.04.19 um 17:09 schrieb Zhou, David(ChunMing):

Quoted 131 lines
Quoted 1 line

3. If we have a ticket we grab a reference to the first BO on

the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

The BO on LRU is already locked by cs user, can it be dropped here by DC user? and then DC user grab its lock with ticket, how does CS grab it again?

If you think waiting in ttm has this risk, how about just adding a wrapper for pin function as below? amdgpu_get_pin_bo_timeout()
{
do {
amdgpo_bo_reserve();
r=amdgpu_bo_pin();

if(!r)
        break;
amdgpu_bo_unreserve();
timeout--;

} while(timeout>0);

}

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Zhou, David(ChunMing)" ,"Koenig, Christian" ,"Liang, Prike" ,[email protected] CC:

Well that's not so easy of hand.

The basic problem here is that when you busy wait at this place you can easily run into situations where application A busy waits for B while B busy waits for A -> deadlock.

So what we need here is the deadlock detection logic of the ww_mutex. To use this we at least need to do the following steps:

1. Reserve the BO in DC using a ww_mutex ticket (trivial).

2. If we then run into this EBUSY condition in TTM check if the BO we need memory for (or rather the ww_mutex of its reservation object) has a ticket assigned.

3. If we have a ticket we grab a reference to the first BO on the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

4. If getting the reservation lock with the ticket succeeded we check if the BO is still the first one on the LRU in question (the BO could have moved).

5. If the BO is still the first one on the LRU in question we try to evict it as we would evict any other BO.

6. If any of the "If's" above fail we just back off and return -EBUSY.

Steps 2-5 are certainly not trivial, but doable as far as I can see.

Have fun :)
Christian.

Am 23.04.19 um 15:19 schrieb Zhou, David(ChunMing):

Quoted 69 lines

How about adding more condition ctx->resv inline to address your concern? As well as don't wait from same user, shouldn't lead to deadlock.

Otherwise, any other idea?

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Liang, Prike" ,"Zhou, David(ChunMing)" ,[email protected] CC:

Well that is certainly a NAK because it can lead to deadlock in the memory management.

You can't just busy wait with all those locks held.

Regards,
Christian.

Am 23.04.19 um 03:45 schrieb Liang, Prike:

Quoted 9 lines

Thanks,
Prike
-----Original Message-----
From: Chunming Zhou <[email protected]>
Sent: Monday, April 22, 2019 6:39 PM
To: [email protected]
Cc: Liang, Prike <[email protected]>; Zhou, David(ChunMing)

<[email protected]>

Quoted 3 lines

Subject: [PATCH] ttm: wait mem space if user allow while gpu busy

heavy gpu job could occupy memory long time, which could lead

to other user fail to get memory.

Review

Quoted 8 lines
Change-Id I0b322d98cd76e5ac32b00462bbae8008d76c5e11

---
   drivers/gpu/drm/ttm/ttm_bo.c | 6 ++++--

   1 file changed, 4 insertions(+), 2 deletions(-)
(quoted diff) +0 -0

b/drivers/gpu/drm/ttm/ttm_bo.c index 7c484729f9b2..6c596cc24bec 100644

drivers/gpu/drm/ttm/ttm_bo.c static int ttm_bo_mem_force_space(struct

Quoted 3 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -0
@@ -830,8 +830,10 @@ static int ttm_bo_mem_force_space(struct 

ttm_buffer_object *bo,

drivers/gpu/drm/ttm/ttm_bo.c static int ttm_bo_mem_force_space(struct · line 832

Quoted 3 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -0
                 if (mem->mm_node)
                         break;
                 ret = ttm_mem_evict_first(bdev, mem_type, 

place, ctx);

drivers/gpu/drm/ttm/ttm_bo.c static int ttm_bo_mem_force_space(struct · line 834

Quoted 4 lines
drivers/gpu/drm/ttm/ttm_bo.c +2 -2
-             if (unlikely(ret != 0))
-                     return ret;
+             if (unlikely(ret != 0)) {
+                     if (!ctx || ctx->no_wait_gpu || ret != 

-EBUSY)

Quoted 8 lines
drivers/gpu/drm/ttm/ttm_bo.c +2 -1
+                             return ret;
+             }
         } while (1);
         mem->mem_type = mem_type;
         return ttm_bo_add_move_fence(bo, man, mem);
--

2.17.1

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#12 lore raw

Am 24.04.19 um 10:11 schrieb zhoucm1:

On 2019年04月24日 16:07, Christian König wrote:
This is used in a work item, so you don't need to check for signals. will remove.

And checking if the LRU is populated is mandatory here How to check it outside of TTM? because the code is in dm.

Well just use a static cast on the first entry of the LRU.

We can't upstream that solution anyway, so just a hack should do for now.

or otherwise you can run into an endless loop.
I already add a timeout for that.

Thinking more about it we most likely actually need to grab the lock on the first BO entry from the LRU.

-David

Christian.

Am 24.04.19 um 09:59 schrieb zhoucm1:

how about new attached?

-David

On 2019年04月24日 15:30, Christian König wrote:
That would change the semantics of ttm_bo_mem_space() and so could change the return code in an IOCTL as well. Not a good idea, cause that could have a lot of side effects.

Instead in the calling DC code you could check if you get an -ENOMEM and then call schedule().

If after the schedule() we see that we have now BOs on the LRU we can try again and see if pinning the frame buffer now succeeds.

Christian.

Am 24.04.19 um 09:17 schrieb zhoucm1:

I made a patch as attached.

I'm not sure how to make patch as your proposal, Could you make a patch for that if mine isn't enough?

-David

On 2019年04月24日 15:12, Christian König wrote:
how about just adding a wrapper for pin function as below? I considered this as well and don't think it will work reliable.

We could use it as a band aid for this specific problem, but in general we need to improve the handling in TTM to resolve those kind of resource conflicts.

Regards,
Christian.

Am 23.04.19 um 17:09 schrieb Zhou, David(ChunMing):

Quoted 1 line

3. If we have a ticket we grab a reference to the first BO on the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

The BO on LRU is already locked by cs user, can it be dropped here by DC user? and then DC user grab its lock with ticket, how does CS grab it again?

If you think waiting in ttm has this risk, how about just adding a wrapper for pin function as below? amdgpu_get_pin_bo_timeout()
{
do {
amdgpo_bo_reserve();
r=amdgpu_bo_pin();

if(!r)

        break;

amdgpu_bo_unreserve();
timeout--;

} while(timeout>0);

}

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Zhou, David(ChunMing)" ,"Koenig, Christian" ,"Liang, Prike" ,[email protected]<mailto:[email protected]> CC:

Well that's not so easy of hand.

The basic problem here is that when you busy wait at this place you can easily run into situations where application A busy waits for B while B busy waits for A -> deadlock.

So what we need here is the deadlock detection logic of the ww_mutex. To use this we at least need to do the following steps:

1. Reserve the BO in DC using a ww_mutex ticket (trivial).

2. If we then run into this EBUSY condition in TTM check if the BO we need memory for (or rather the ww_mutex of its reservation object) has a ticket assigned.

3. If we have a ticket we grab a reference to the first BO on the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

4. If getting the reservation lock with the ticket succeeded we check if the BO is still the first one on the LRU in question (the BO could have moved).

5. If the BO is still the first one on the LRU in question we try to evict it as we would evict any other BO.

6. If any of the "If's" above fail we just back off and return -EBUSY.

Steps 2-5 are certainly not trivial, but doable as far as I can see.

Have fun :)
Christian.

Am 23.04.19 um 15:19 schrieb Zhou, David(ChunMing):
How about adding more condition ctx->resv inline to address your concern? As well as don't wait from same user, shouldn't lead to deadlock.

Otherwise, any other idea?

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Liang, Prike" ,"Zhou, David(ChunMing)" ,[email protected]<mailto:[email protected]> CC:

Well that is certainly a NAK because it can lead to deadlock in the memory management.

You can't just busy wait with all those locks held.

Regards,
Christian.

Am 23.04.19 um 03:45 schrieb Liang, Prike:

Quoted 38 lines

Thanks,
Prike
-----Original Message-----
From: Chunming Zhou <[email protected]><mailto:[email protected]>
Sent: Monday, April 22, 2019 6:39 PM
To: [email protected]<mailto:[email protected]>
Cc: Liang, Prike <[email protected]><mailto:[email protected]>; Zhou, David(ChunMing) <[email protected]><mailto:[email protected]>
Subject: [PATCH] ttm: wait mem space if user allow while gpu busy

heavy gpu job could occupy memory long time, which could lead to other user fail to get memory.

Change-Id I0b322d98cd76e5ac32b00462bbae8008d76c5e11
---
  drivers/gpu/drm/ttm/ttm_bo.c | 6 ++++--
  1 file changed, 4 insertions(+), 2 deletions(-)
drivers/gpu/drm/ttm/ttm_bo.c +4 -3
@@ -830,8 +830,10 @@ static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
               if (mem->mm_node)
                       break;
               ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
-             if (unlikely(ret != 0))
-                     return ret;
+             if (unlikely(ret != 0)) {
+                     if (!ctx || ctx->no_wait_gpu || ret != -EBUSY)
+                             return ret;
+             }
       } while (1);
       mem->mem_type = mem_type;
       return ttm_bo_add_move_fence(bo, man, mem);
--

2.17.1

Quoted 4 lines

_______________________________________________
dri-devel mailing list [email protected]<mailto:[email protected]>
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list [email protected]<mailto:[email protected]>
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list [email protected]<mailto:[email protected]>
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list [email protected]<mailto:[email protected]>
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list [email protected]<mailto:[email protected]>
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#13 lore raw

OK, Let's drop mine, then Could you draft a solution for that?

-David

Quoted 13 lines · Christian

Am 24.04.19 um 10:11 schrieb zhoucm1:

Quoted 9 lines · Christian König
Quoted 1 line

This is used in a work item, so you don't need to check for signals.

will remove.

Quoted 2 lines

And checking if the LRU is populated is mandatory here

How to check it outside of TTM? because the code is in dm.

Well just use a static cast on the first entry of the LRU.

Quoted 198 lines

We can't upstream that solution anyway, so just a hack should do for now.

Quoted 3 lines
Quoted 2 lines

or otherwise you can run into an endless loop.

I already add a timeout for that.

Thinking more about it we most likely actually need to grab the lock on the first BO entry from the LRU.

Quoted 189 lines

-David

Quoted 186 lines

Christian.

Am 24.04.19 um 09:59 schrieb zhoucm1:

Quoted 181 lines

how about new attached?

-David

Quoted 167 lines · Christian König

That would change the semantics of ttm_bo_mem_space() and so could change the return code in an IOCTL as well. Not a good idea, cause that could have a lot of side effects.

Instead in the calling DC code you could check if you get an -ENOMEM and then call schedule().

If after the schedule() we see that we have now BOs on the LRU we can try again and see if pinning the frame buffer now succeeds.

Christian.

Am 24.04.19 um 09:17 schrieb zhoucm1:

Quoted 156 lines

I made a patch as attached.

I'm not sure how to make patch as your proposal, Could you make a patch for that if mine isn't enough?

-David

Quoted 142 lines · Christian König
Quoted 1 line

how about just adding a wrapper for pin function as below?

I considered this as well and don't think it will work reliable.

We could use it as a band aid for this specific problem, but in general we need to improve the handling in TTM to resolve those kind of resource conflicts.

Regards,
Christian.

Am 23.04.19 um 17:09 schrieb Zhou, David(ChunMing):

Quoted 132 lines
Quoted 1 line

3. If we have a ticket we grab a reference to the first BO on

the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

The BO on LRU is already locked by cs user, can it be dropped here by DC user? and then DC user grab its lock with ticket, how does CS grab it again?

If you think waiting in ttm has this risk, how about just adding a wrapper for pin function as below? amdgpu_get_pin_bo_timeout()
{
do {
amdgpo_bo_reserve();
r=amdgpu_bo_pin();

if(!r)
        break;
amdgpu_bo_unreserve();
timeout--;

} while(timeout>0);

}

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Zhou, David(ChunMing)" ,"Koenig, Christian" ,"Liang, Prike" ,[email protected] CC:

Well that's not so easy of hand.

The basic problem here is that when you busy wait at this place you can easily run into situations where application A busy waits for B while B busy waits for A -> deadlock.

So what we need here is the deadlock detection logic of the ww_mutex. To use this we at least need to do the following steps:

1. Reserve the BO in DC using a ww_mutex ticket (trivial).

2. If we then run into this EBUSY condition in TTM check if the BO we need memory for (or rather the ww_mutex of its reservation object) has a ticket assigned.

3. If we have a ticket we grab a reference to the first BO on the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

4. If getting the reservation lock with the ticket succeeded we check if the BO is still the first one on the LRU in question (the BO could have moved).

5. If the BO is still the first one on the LRU in question we try to evict it as we would evict any other BO.

6. If any of the "If's" above fail we just back off and return -EBUSY.

Steps 2-5 are certainly not trivial, but doable as far as I can see.

Have fun :)
Christian.

Am 23.04.19 um 15:19 schrieb Zhou, David(ChunMing):

Quoted 70 lines

How about adding more condition ctx->resv inline to address your concern? As well as don't wait from same user, shouldn't lead to deadlock.

Otherwise, any other idea?

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Liang, Prike" ,"Zhou, David(ChunMing)" ,[email protected] CC:

Well that is certainly a NAK because it can lead to deadlock in the memory management.

You can't just busy wait with all those locks held.

Regards,
Christian.

Am 23.04.19 um 03:45 schrieb Liang, Prike:

Quoted 9 lines

Thanks,
Prike
-----Original Message-----
From: Chunming Zhou <[email protected]>
Sent: Monday, April 22, 2019 6:39 PM
To: [email protected]
Cc: Liang, Prike <[email protected]>; Zhou,

David(ChunMing) <[email protected]>

Quoted 1 line

Subject: [PATCH] ttm: wait mem space if user allow while gpu

busy

Quoted 2 lines

heavy gpu job could occupy memory long time, which could

lead to other user fail to get memory.

Review

Quoted 8 lines
Change-Id I0b322d98cd76e5ac32b00462bbae8008d76c5e11

---
   drivers/gpu/drm/ttm/ttm_bo.c | 6 ++++--

   1 file changed, 4 insertions(+), 2 deletions(-)
(quoted diff) +0 -0

b/drivers/gpu/drm/ttm/ttm_bo.c index 7c484729f9b2..6c596cc24bec 100644

drivers/gpu/drm/ttm/ttm_bo.c static int

Quoted 3 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -0
@@ -830,8 +830,10 @@ static int 

ttm_bo_mem_force_space(struct ttm_buffer_object *bo,

drivers/gpu/drm/ttm/ttm_bo.c static int · line 832

Quoted 3 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -0
                 if (mem->mm_node)
                         break;
                 ret = ttm_mem_evict_first(bdev, mem_type, 

place, ctx);

drivers/gpu/drm/ttm/ttm_bo.c static int · line 834

Quoted 4 lines
drivers/gpu/drm/ttm/ttm_bo.c +2 -2
-             if (unlikely(ret != 0))
-                     return ret;
+             if (unlikely(ret != 0)) {
+                     if (!ctx || ctx->no_wait_gpu || ret != 

-EBUSY)

Quoted 8 lines
drivers/gpu/drm/ttm/ttm_bo.c +2 -1
+                             return ret;
+             }
         } while (1);
         mem->mem_type = mem_type;
         return ttm_bo_add_move_fence(bo, man, mem);
--

2.17.1

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#14 lore raw

No, I'm busy with P2P and recoverable page faults.

Christian.

Am 24.04.19 um 11:03 schrieb zhoucm1:

Quoted 221 lines

OK, Let's drop mine, then Could you draft a solution for that?

-David

Quoted 13 lines · Christian

Am 24.04.19 um 10:11 schrieb zhoucm1:

Quoted 9 lines · Christian König
Quoted 1 line

This is used in a work item, so you don't need to check for signals.

will remove.

Quoted 2 lines

And checking if the LRU is populated is mandatory here

How to check it outside of TTM? because the code is in dm.

Well just use a static cast on the first entry of the LRU.

Quoted 198 lines

We can't upstream that solution anyway, so just a hack should do for now.

Quoted 3 lines
Quoted 2 lines

or otherwise you can run into an endless loop.

I already add a timeout for that.

Thinking more about it we most likely actually need to grab the lock on the first BO entry from the LRU.

Quoted 189 lines

-David

Quoted 186 lines

Christian.

Am 24.04.19 um 09:59 schrieb zhoucm1:

Quoted 181 lines

how about new attached?

-David

Quoted 167 lines · Christian König

That would change the semantics of ttm_bo_mem_space() and so could change the return code in an IOCTL as well. Not a good idea, cause that could have a lot of side effects.

Instead in the calling DC code you could check if you get an -ENOMEM and then call schedule().

If after the schedule() we see that we have now BOs on the LRU we can try again and see if pinning the frame buffer now succeeds.

Christian.

Am 24.04.19 um 09:17 schrieb zhoucm1:

Quoted 156 lines

I made a patch as attached.

I'm not sure how to make patch as your proposal, Could you make a patch for that if mine isn't enough?

-David

Quoted 142 lines · Christian König
Quoted 1 line

how about just adding a wrapper for pin function as below?

I considered this as well and don't think it will work reliable.

We could use it as a band aid for this specific problem, but in general we need to improve the handling in TTM to resolve those kind of resource conflicts.

Regards,
Christian.

Am 23.04.19 um 17:09 schrieb Zhou, David(ChunMing):

Quoted 132 lines
Quoted 1 line

3. If we have a ticket we grab a reference to the first BO on

the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

The BO on LRU is already locked by cs user, can it be dropped here by DC user? and then DC user grab its lock with ticket, how does CS grab it again?

If you think waiting in ttm has this risk, how about just adding a wrapper for pin function as below? amdgpu_get_pin_bo_timeout()
{
do {
amdgpo_bo_reserve();
r=amdgpu_bo_pin();

if(!r)
        break;
amdgpu_bo_unreserve();
timeout--;

} while(timeout>0);

}

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Zhou, David(ChunMing)" ,"Koenig, Christian" ,"Liang, Prike" ,[email protected] CC:

Well that's not so easy of hand.

The basic problem here is that when you busy wait at this place you can easily run into situations where application A busy waits for B while B busy waits for A -> deadlock.

So what we need here is the deadlock detection logic of the ww_mutex. To use this we at least need to do the following steps:

1. Reserve the BO in DC using a ww_mutex ticket (trivial).

2. If we then run into this EBUSY condition in TTM check if the BO we need memory for (or rather the ww_mutex of its reservation object) has a ticket assigned.

3. If we have a ticket we grab a reference to the first BO on the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

4. If getting the reservation lock with the ticket succeeded we check if the BO is still the first one on the LRU in question (the BO could have moved).

5. If the BO is still the first one on the LRU in question we try to evict it as we would evict any other BO.

6. If any of the "If's" above fail we just back off and return -EBUSY.

Steps 2-5 are certainly not trivial, but doable as far as I can see.

Have fun :)
Christian.

Am 23.04.19 um 15:19 schrieb Zhou, David(ChunMing):

Quoted 70 lines

How about adding more condition ctx->resv inline to address your concern? As well as don't wait from same user, shouldn't lead to deadlock.

Otherwise, any other idea?

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Liang, Prike" ,"Zhou, David(ChunMing)" ,[email protected] CC:

Well that is certainly a NAK because it can lead to deadlock in the memory management.

You can't just busy wait with all those locks held.

Regards,
Christian.

Am 23.04.19 um 03:45 schrieb Liang, Prike:

Quoted 9 lines

Thanks,
Prike
-----Original Message-----
From: Chunming Zhou <[email protected]>
Sent: Monday, April 22, 2019 6:39 PM
To: [email protected]
Cc: Liang, Prike <[email protected]>; Zhou,

David(ChunMing) <[email protected]>

Quoted 1 line

Subject: [PATCH] ttm: wait mem space if user allow while

gpu busy

Quoted 2 lines

heavy gpu job could occupy memory long time, which could

lead to other user fail to get memory.

Review

Quoted 8 lines
Change-Id I0b322d98cd76e5ac32b00462bbae8008d76c5e11

---
   drivers/gpu/drm/ttm/ttm_bo.c | 6 ++++--

   1 file changed, 4 insertions(+), 2 deletions(-)
(quoted diff) +0 -0

b/drivers/gpu/drm/ttm/ttm_bo.c index 7c484729f9b2..6c596cc24bec 100644

drivers/gpu/drm/ttm/ttm_bo.c static int

Quoted 3 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -0
@@ -830,8 +830,10 @@ static int 

ttm_bo_mem_force_space(struct ttm_buffer_object *bo,

drivers/gpu/drm/ttm/ttm_bo.c static int · line 832

Quoted 3 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -0
                 if (mem->mm_node)
                         break;
                 ret = ttm_mem_evict_first(bdev, mem_type, 

place, ctx);

drivers/gpu/drm/ttm/ttm_bo.c static int · line 834

Quoted 4 lines
drivers/gpu/drm/ttm/ttm_bo.c +2 -2
-             if (unlikely(ret != 0))
-                     return ret;
+             if (unlikely(ret != 0)) {
+                     if (!ctx || ctx->no_wait_gpu || ret 

!= -EBUSY)

Quoted 8 lines
drivers/gpu/drm/ttm/ttm_bo.c +2 -1
+ return ret;
+             }
         } while (1);
         mem->mem_type = mem_type;
         return ttm_bo_add_move_fence(bo, man, mem);
--

2.17.1

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#15 lore raw
In reply to Christian König
Quoted 7 lines · Christian König

Well that's not so easy of hand.

The basic problem here is that when you busy wait at this place you can easily run into situations where application A busy waits for B while B busy waits for A -> deadlock.

So what we need here is the deadlock detection logic of the ww_mutex. To use this we at least need to do the following steps:

Quoted 7 lines

1. Reserve the BO in DC using a ww_mutex ticket (trivial).

2. If we then run into this EBUSY condition in TTM check if the BO we need memory for (or rather the ww_mutex of its reservation object) has a ticket assigned.

3. If we have a ticket we grab a reference to the first BO on the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

4. If getting the reservation lock with the ticket succeeded we check if the BO is still the first one on the LRU in question (the BO could have moved).

I don't think you actually need this check. Once you're in this slow reclaim mode all hope for performance is pretty much lost (you're thrashin vram terribly), forward progress matters. Also, less code :-)

Quoted 3 lines

5. If the BO is still the first one on the LRU in question we try to evict it as we would evict any other BO.

6. If any of the "If's" above fail we just back off and return -EBUSY.

Another idea I pondered (but never implemented) is a slow reclaim lru lock. Essentially there'd be two ways to walk the lru and evict bo:

- fast path: spinlock + trylock, like today

- slow path: ww_mutex lru lock, plus every bo is reserved (nested within that ww_mutex lru lock) with a full ww_mutex_lock. Guaranteed forward progress.

Transition would be that as soon as someone hits an EBUSY, they set the slow reclaim flag (while holding the quick reclaim spinlock quickly, which will drain anyone still stuck in fast reclaim path). Everytime fast reclaim acquires the spinlock it needs to check for the slow reclaim flag, and if that's set, fall back to slow reclaim.

Transitioning out of slow reclaim would only happen once the thread (with it's ww ticket) that hit the EBUSY has completed whatever it was trying to do (either successfully, or failed because even evicting everyone else didn't give you enough vram). Tricky part here is making sure threads still in slow reclaim don't blow up if we switch back. Since only ever one thread can be actually doing slow reclaim (everyone is serialized on the single ww mutex lru lock) should be doable by checking for the slow reclaim conditiona once you have the lru ww_mutex and if the slow reclaim condition is lifted, switch back to fast reclaim.

The slow reclaim conditiona might also need to be a full reference count, to handle multiple threads hitting EBUSY/slow reclaim without the book-keeping getting all confused.

Upshot of this is that it's guranteeing forward progress, but the perf cliff might be too steep if this happens too often. You might need to round it off with 1-2 retries when you hit EBUSY, before forcing slow reclaim.
-Daniel

Quoted 80 lines

Steps 2-5 are certainly not trivial, but doable as far as I can see.

Have fun :)
Christian.

Am 23.04.19 um 15:19 schrieb Zhou, David(ChunMing):

How about adding more condition ctx->resv inline to address your concern? As well as don't wait from same user, shouldn't lead to deadlock.

Otherwise, any other idea?

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Liang, Prike" ,"Zhou, David(ChunMing)" ,[email protected] CC:

Well that is certainly a NAK because it can lead to deadlock in the memory management.

You can't just busy wait with all those locks held.

Regards,
Christian.

Am 23.04.19 um 03:45 schrieb Liang, Prike:

Quoted 38 lines

Thanks,
Prike
-----Original Message-----
From: Chunming Zhou <[email protected]>
Sent: Monday, April 22, 2019 6:39 PM
To: [email protected]
Cc: Liang, Prike <[email protected]>; Zhou, David(ChunMing) <[email protected]>
Subject: [PATCH] ttm: wait mem space if user allow while gpu busy

heavy gpu job could occupy memory long time, which could lead to other user fail to get memory.

Change-Id I0b322d98cd76e5ac32b00462bbae8008d76c5e11
---
  drivers/gpu/drm/ttm/ttm_bo.c | 6 ++++--
  1 file changed, 4 insertions(+), 2 deletions(-)
drivers/gpu/drm/ttm/ttm_bo.c +4 -3
@@ -830,8 +830,10 @@ static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
               if (mem->mm_node)
                       break;
               ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
-             if (unlikely(ret != 0))
-                     return ret;
+             if (unlikely(ret != 0)) {
+                     if (!ctx || ctx->no_wait_gpu || ret != -EBUSY)
+                             return ret;
+             }
       } while (1);
       mem->mem_type = mem_type;
       return ttm_bo_add_move_fence(bo, man, mem);
--

2.17.1

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Signature · 7 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
#16 lore raw

Am 24.04.19 um 10:01 schrieb Daniel Vetter:

Quoted 10 lines
Quoted 8 lines · Christian König

Well that's not so easy of hand.

The basic problem here is that when you busy wait at this place you can easily run into situations where application A busy waits for B while B busy waits for A -> deadlock.

So what we need here is the deadlock detection logic of the ww_mutex. To use this we at least need to do the following steps:

1. Reserve the BO in DC using a ww_mutex ticket (trivial).

Quoted 8 lines
Quoted 5 lines

2. If we then run into this EBUSY condition in TTM check if the BO we need memory for (or rather the ww_mutex of its reservation object) has a ticket assigned.

3. If we have a ticket we grab a reference to the first BO on the LRU, drop the LRU lock and try to grab the reservation lock with the ticket.

4. If getting the reservation lock with the ticket succeeded we check if the BO is still the first one on the LRU in question (the BO could have moved).

I don't think you actually need this check. Once you're in this slow reclaim mode all hope for performance is pretty much lost (you're thrashin vram terribly), forward progress matters. Also, less code :-)

This step is not for performance, but for correctness.

When you drop the LRU lock and grab the ww_mutex in the slow path you need to double check that this BO wasn't evicted by somebody else in the meantime.

Quoted 11 lines
Quoted 3 lines

5. If the BO is still the first one on the LRU in question we try to evict it as we would evict any other BO.

6. If any of the "If's" above fail we just back off and return -EBUSY.

Another idea I pondered (but never implemented) is a slow reclaim lru lock. Essentially there'd be two ways to walk the lru and evict bo:

- fast path: spinlock + trylock, like today

- slow path: ww_mutex lru lock, plus every bo is reserved (nested within that ww_mutex lru lock) with a full ww_mutex_lock. Guaranteed forward progress.

Of hand I don't see any advantage to this. So what is the benefit?

Regards,
Christian.

Quoted 22 lines

Transition would be that as soon as someone hits an EBUSY, they set the slow reclaim flag (while holding the quick reclaim spinlock quickly, which will drain anyone still stuck in fast reclaim path). Everytime fast reclaim acquires the spinlock it needs to check for the slow reclaim flag, and if that's set, fall back to slow reclaim.

Transitioning out of slow reclaim would only happen once the thread (with it's ww ticket) that hit the EBUSY has completed whatever it was trying to do (either successfully, or failed because even evicting everyone else didn't give you enough vram). Tricky part here is making sure threads still in slow reclaim don't blow up if we switch back. Since only ever one thread can be actually doing slow reclaim (everyone is serialized on the single ww mutex lru lock) should be doable by checking for the slow reclaim conditiona once you have the lru ww_mutex and if the slow reclaim condition is lifted, switch back to fast reclaim.

The slow reclaim conditiona might also need to be a full reference count, to handle multiple threads hitting EBUSY/slow reclaim without the book-keeping getting all confused.

Quoted 87 lines

Upshot of this is that it's guranteeing forward progress, but the perf cliff might be too steep if this happens too often. You might need to round it off with 1-2 retries when you hit EBUSY, before forcing slow reclaim.
-Daniel

Quoted 79 lines

Steps 2-5 are certainly not trivial, but doable as far as I can see.

Have fun :)
Christian.

Am 23.04.19 um 15:19 schrieb Zhou, David(ChunMing):

How about adding more condition ctx->resv inline to address your concern? As well as don't wait from same user, shouldn't lead to deadlock.

Otherwise, any other idea?

-------- Original Message --------
Subject: Re: [PATCH] ttm: wait mem space if user allow while gpu busy
From: Christian König
To: "Liang, Prike" ,"Zhou, David(ChunMing)" ,[email protected] CC:

Well that is certainly a NAK because it can lead to deadlock in the memory management.

You can't just busy wait with all those locks held.

Regards,
Christian.

Am 23.04.19 um 03:45 schrieb Liang, Prike:

Quoted 38 lines

Thanks,
Prike
-----Original Message-----
From: Chunming Zhou <[email protected]>
Sent: Monday, April 22, 2019 6:39 PM
To: [email protected]
Cc: Liang, Prike <[email protected]>; Zhou, David(ChunMing) <[email protected]>
Subject: [PATCH] ttm: wait mem space if user allow while gpu busy

heavy gpu job could occupy memory long time, which could lead to other user fail to get memory.

Change-Id I0b322d98cd76e5ac32b00462bbae8008d76c5e11
---
   drivers/gpu/drm/ttm/ttm_bo.c | 6 ++++--
   1 file changed, 4 insertions(+), 2 deletions(-)
drivers/gpu/drm/ttm/ttm_bo.c +4 -3
@@ -830,8 +830,10 @@ static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
                if (mem->mm_node)
                        break;
                ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
-             if (unlikely(ret != 0))
-                     return ret;
+             if (unlikely(ret != 0)) {
+                     if (!ctx || ctx->no_wait_gpu || ret != -EBUSY)
+                             return ret;
+             }
        } while (1);
        mem->mem_type = mem_type;
        return ttm_bo_add_move_fence(bo, man, mem);
--

2.17.1

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel