Skip to contentSkip to conversation

drm/radeon: TTM must be init with cpu-visible VRAM

PATCH In review Lauri Kasanen 2 participants 7 messages Feb 27, 2014 – Feb 28, 2014

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/radeon: TTM must be init with cpu-visible VRAM

Without this, a bo may get created in the cpu-inaccessible vram. Before the CP engines get setup, all copies are done via cpu memcpy.

This means that the cpu tries to read from inaccessible memory, fails, and the radeon module proceeds to disable acceleration.

Doing this has no downsides, as the real VRAM size gets set as soon as the CP engines get init.

This is a candidate for 3.14 fixes.

---
 drivers/gpu/drm/radeon/radeon_ttm.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/gpu/drm/radeon/radeon_ttm.c +2 -0
index 3aa853c..35eae91 100644
@@ -715,6 +715,8 @@ int radeon_ttm_init(struct radeon_device *rdev)
 		DRM_ERROR("Failed initializing VRAM heap.\n");
 		return r;
 	}
+	radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
+
 	r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true,
 			     RADEON_GEM_DOMAIN_VRAM,
 			     NULL, &rdev->stollen_vga_memory);
Signature · 1 line
1.8.3.1
#1 lore raw

Am 27.02.2014 22:38, schrieb Lauri Kasanen:

Quoted 10 lines

Without this, a bo may get created in the cpu-inaccessible vram. Before the CP engines get setup, all copies are done via cpu memcpy.

This means that the cpu tries to read from inaccessible memory, fails, and the radeon module proceeds to disable acceleration.

Doing this has no downsides, as the real VRAM size gets set as soon as the CP engines get init.

This is a candidate for 3.14 fixes.

This should be unnecessary, since TTM gets initialized only seeing the visible VRAM and later on radeon_ttm_set_active_vram_size gets called to increase the limit.

If this isn't the case any more we should figure out why instead of working around it like this.

Christian.

Quoted 19 lines
---
  drivers/gpu/drm/radeon/radeon_ttm.c | 2 ++
  1 file changed, 2 insertions(+)
drivers/gpu/drm/radeon/radeon_ttm.c +2 -0
index 3aa853c..35eae91 100644
@@ -715,6 +715,8 @@ int radeon_ttm_init(struct radeon_device *rdev)
  		DRM_ERROR("Failed initializing VRAM heap.\n");
  		return r;
  	}
+	radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
+
  	r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true,
  			     RADEON_GEM_DOMAIN_VRAM,
  			     NULL, &rdev->stollen_vga_memory);
#2 lore raw
Quoted 10 lines · On Fri, 28 Feb 2014 10:36:59 +0100 Christian König <[email protected]>…

Am 27.02.2014 22:38, schrieb Lauri Kasanen:

Quoted 9 lines

Without this, a bo may get created in the cpu-inaccessible vram. Before the CP engines get setup, all copies are done via cpu memcpy.

This means that the cpu tries to read from inaccessible memory, fails, and the radeon module proceeds to disable acceleration.

Doing this has no downsides, as the real VRAM size gets set as soon as the CP engines get init.

Quoted 8 lines
Quoted 1 line

This is a candidate for 3.14 fixes.

This should be unnecessary, since TTM gets initialized only seeing the visible VRAM and later on radeon_ttm_set_active_vram_size gets called to increase the limit.

If this isn't the case any more we should figure out why instead of working around it like this.

Negative, TTM gets initialized with real_vram just a few lines above this patch, not visible_vram.

- Lauri

#3 lore raw
Quoted 19 lines · On Fri, 28 Feb 2014 17:30:39 +0200 Lauri Kasanen <[email protected]> wrote:
Quoted 16 lines · On Fri, 28 Feb 2014 10:36:59 +0100 Christian König <[email protected]>…

Am 27.02.2014 22:38, schrieb Lauri Kasanen:

Quoted 10 lines

Without this, a bo may get created in the cpu-inaccessible vram. Before the CP engines get setup, all copies are done via cpu memcpy.

This means that the cpu tries to read from inaccessible memory, fails, and the radeon module proceeds to disable acceleration.

Doing this has no downsides, as the real VRAM size gets set as soon as the CP engines get init.

This is a candidate for 3.14 fixes.

This should be unnecessary, since TTM gets initialized only seeing the visible VRAM and later on radeon_ttm_set_active_vram_size gets called to increase the limit.

Quoted 5 lines
Quoted 2 lines

If this isn't the case any more we should figure out why instead of working around it like this.

Negative, TTM gets initialized with real_vram just a few lines above this patch, not visible_vram.

git blame shows 7a50f01a from 2009, "drm/radeon/kms: vram sizing on certain r100 chips needs workaround." by Dave Airlie.

So the TTM VRAM init has been wrong for five years, and only worked by accident because until now all allocations were done bottom-up.

- Lauri

#4 lore raw

Am 28.02.2014 16:36, schrieb Lauri Kasanen:

Quoted 23 lines
Quoted 20 lines · On Fri, 28 Feb 2014 17:30:39 +0200 Lauri Kasanen <[email protected]> wrote:
Quoted 17 lines · On Fri, 28 Feb 2014 10:36:59 +0100 Christian König <[email protected]>…

Am 27.02.2014 22:38, schrieb Lauri Kasanen:

Quoted 10 lines

Without this, a bo may get created in the cpu-inaccessible vram. Before the CP engines get setup, all copies are done via cpu memcpy.

This means that the cpu tries to read from inaccessible memory, fails, and the radeon module proceeds to disable acceleration.

Doing this has no downsides, as the real VRAM size gets set as soon as the CP engines get init.

This is a candidate for 3.14 fixes.

This should be unnecessary, since TTM gets initialized only seeing the visible VRAM and later on radeon_ttm_set_active_vram_size gets called to increase the limit.

If this isn't the case any more we should figure out why instead of working around it like this.

Quoted 7 lines
Quoted 2 lines

Negative, TTM gets initialized with real_vram just a few lines above this patch, not visible_vram.

git blame shows 7a50f01a from 2009, "drm/radeon/kms: vram sizing on certain r100 chips needs workaround." by Dave Airlie.

So the TTM VRAM init has been wrong for five years, and only worked by accident because until now all allocations were done bottom-up.

Yeah, just came to the same conclusion. We probably never hit the case in the last five years because we don't really access the memory before we start the copy ring.

Please fix ttm_bo_init_mm to use rdev->mc.visible_vram_size instead.

That allocations are made bottom-up is relied upon in a couple of other cases as well, the stolen VGA memory and the UVD firmware handling for example.

Christian.

Quoted 1 line

- Lauri

#5 lore raw
Quoted 29 lines · On Fri, 28 Feb 2014 16:43:54 +0100 Christian König <[email protected]>…
Quoted 17 lines
Quoted 17 lines
Quoted 17 lines

Am 27.02.2014 22:38, schrieb Lauri Kasanen:

Quoted 10 lines

Without this, a bo may get created in the cpu-inaccessible vram. Before the CP engines get setup, all copies are done via cpu memcpy.

This means that the cpu tries to read from inaccessible memory, fails, and the radeon module proceeds to disable acceleration.

Doing this has no downsides, as the real VRAM size gets set as soon as the CP engines get init.

This is a candidate for 3.14 fixes.

This should be unnecessary, since TTM gets initialized only seeing the visible VRAM and later on radeon_ttm_set_active_vram_size gets called to increase the limit.

If this isn't the case any more we should figure out why instead of working around it like this.

Quoted 7 lines
Quoted 2 lines

Negative, TTM gets initialized with real_vram just a few lines above this patch, not visible_vram.

git blame shows 7a50f01a from 2009, "drm/radeon/kms: vram sizing on certain r100 chips needs workaround." by Dave Airlie.

So the TTM VRAM init has been wrong for five years, and only worked by accident because until now all allocations were done bottom-up.

Yeah, just came to the same conclusion. We probably never hit the case in the last five years because we don't really access the memory before we start the copy ring.

Quoted 5 lines

Please fix ttm_bo_init_mm to use rdev->mc.visible_vram_size instead.

That allocations are made bottom-up is relied upon in a couple of other cases as well, the stolen VGA memory and the UVD firmware handling for example.

I initially did that, but it caused some issues. I didn't investigage it further, but I guess it has to be init to the maximum size, and then only the size lowered, so that the change only affects lpfn.

- Lauri

#6 lore raw

Am 28.02.2014 17:52, schrieb Lauri Kasanen:

Quoted 31 lines
Quoted 28 lines · On Fri, 28 Feb 2014 16:43:54 +0100 Christian König <[email protected]>…
Quoted 17 lines
Quoted 17 lines
Quoted 17 lines

Am 27.02.2014 22:38, schrieb Lauri Kasanen:

Quoted 10 lines

Without this, a bo may get created in the cpu-inaccessible vram. Before the CP engines get setup, all copies are done via cpu memcpy.

This means that the cpu tries to read from inaccessible memory, fails, and the radeon module proceeds to disable acceleration.

Doing this has no downsides, as the real VRAM size gets set as soon as the CP engines get init.

This is a candidate for 3.14 fixes.

This should be unnecessary, since TTM gets initialized only seeing the visible VRAM and later on radeon_ttm_set_active_vram_size gets called to increase the limit.

If this isn't the case any more we should figure out why instead of working around it like this.

Quoted 7 lines
Quoted 2 lines

Negative, TTM gets initialized with real_vram just a few lines above this patch, not visible_vram.

git blame shows 7a50f01a from 2009, "drm/radeon/kms: vram sizing on certain r100 chips needs workaround." by Dave Airlie.

So the TTM VRAM init has been wrong for five years, and only worked by accident because until now all allocations were done bottom-up.

Yeah, just came to the same conclusion. We probably never hit the case in the last five years because we don't really access the memory before we start the copy ring.

Quoted 8 lines
Quoted 5 lines

Please fix ttm_bo_init_mm to use rdev->mc.visible_vram_size instead.

That allocations are made bottom-up is relied upon in a couple of other cases as well, the stolen VGA memory and the UVD firmware handling for example.

I initially did that, but it caused some issues. I didn't investigage it further, but I guess it has to be init to the maximum size, and then only the size lowered, so that the change only affects lpfn.

Fair enough, probably something gets allocated on init or something like that.

Please document that with a comment and resend the patch.

Christian.

Quoted 2 lines

- Lauri