[v2,1/2] drm/i915: Call dma_set_max_seg_size() in i915_ggtt_probe_hw()
diff mbox series

Message ID 20190822203127.24648-1-lyude@redhat.com
State New
Headers show
Series
  • [v2,1/2] drm/i915: Call dma_set_max_seg_size() in i915_ggtt_probe_hw()
Related show

Commit Message

Lyude Paul Aug. 22, 2019, 8:31 p.m. UTC
Currently, we don't call dma_set_max_seg_size() for i915 because we
intentionally do not limit the segment length that the device supports.
However, this results in a warning being emitted if we try to map
anything larger than SZ_64K on a kernel with CONFIG_DMA_API_DEBUG_SG
enabled:

[    7.751926] DMA-API: i915 0000:00:02.0: mapping sg segment longer
than device claims to support [len=98304] [max=65536]
[    7.751934] WARNING: CPU: 5 PID: 474 at kernel/dma/debug.c:1220
debug_dma_map_sg+0x20f/0x340

This was originally brought up on
https://bugs.freedesktop.org/show_bug.cgi?id=108517 , and the consensus
there was it wasn't really useful to set a limit (and that dma-debug
isn't really all that useful for i915 in the first place). Unfortunately
though, CONFIG_DMA_API_DEBUG_SG is enabled in the debug configs for
various distro kernels. Since a WARN_ON() will disable automatic problem
reporting (and cause any CI with said option enabled to start
complaining), we really should just fix the problem.

Note that as me and Chris Wilson discussed, the other solution for this
would be to make DMA-API not make such assumptions when a driver hasn't
explicitly set a maximum segment size. But, taking a look at the commit
which originally introduced this behavior, commit 78c47830a5cb
("dma-debug: check scatterlist segments"), there is an explicit mention
of this assumption and how it applies to devices with no segment size:

	Conversely, devices which are less limited than the rather
	conservative defaults, or indeed have no limitations at all
	(e.g. GPUs with their own internal MMU), should be encouraged to
	set appropriate dma_parms, as they may get more efficient DMA
	mapping performance out of it.

So unless there's any concerns (I'm open to discussion!), let's just
follow suite and call dma_set_max_seg_size() with UINT_MAX as our limit
to silence any warnings.

Signed-off-by: Lyude Paul <lyude@redhat.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: <stable@vger.kernel.org> # v4.18+
---
 drivers/gpu/drm/i915/i915_gem_gtt.c | 5 +++++
 1 file changed, 5 insertions(+)

Comments

Chris Wilson Aug. 23, 2019, 7:53 p.m. UTC | #1
Quoting Lyude Paul (2019-08-22 21:31:26)
> Currently, we don't call dma_set_max_seg_size() for i915 because we
> intentionally do not limit the segment length that the device supports.
> However, this results in a warning being emitted if we try to map
> anything larger than SZ_64K on a kernel with CONFIG_DMA_API_DEBUG_SG
> enabled:
> 
> [    7.751926] DMA-API: i915 0000:00:02.0: mapping sg segment longer
> than device claims to support [len=98304] [max=65536]
> [    7.751934] WARNING: CPU: 5 PID: 474 at kernel/dma/debug.c:1220
> debug_dma_map_sg+0x20f/0x340
> 
> This was originally brought up on
> https://bugs.freedesktop.org/show_bug.cgi?id=108517 , and the consensus
> there was it wasn't really useful to set a limit (and that dma-debug
> isn't really all that useful for i915 in the first place). Unfortunately
> though, CONFIG_DMA_API_DEBUG_SG is enabled in the debug configs for
> various distro kernels. Since a WARN_ON() will disable automatic problem
> reporting (and cause any CI with said option enabled to start
> complaining), we really should just fix the problem.
> 
> Note that as me and Chris Wilson discussed, the other solution for this
> would be to make DMA-API not make such assumptions when a driver hasn't
> explicitly set a maximum segment size. But, taking a look at the commit
> which originally introduced this behavior, commit 78c47830a5cb
> ("dma-debug: check scatterlist segments"), there is an explicit mention
> of this assumption and how it applies to devices with no segment size:
> 
>         Conversely, devices which are less limited than the rather
>         conservative defaults, or indeed have no limitations at all
>         (e.g. GPUs with their own internal MMU), should be encouraged to
>         set appropriate dma_parms, as they may get more efficient DMA
>         mapping performance out of it.
> 
> So unless there's any concerns (I'm open to discussion!), let's just
> follow suite and call dma_set_max_seg_size() with UINT_MAX as our limit
> to silence any warnings.
> 
> Signed-off-by: Lyude Paul <lyude@redhat.com>
> Cc: Chris Wilson <chris@chris-wilson.co.uk>
> Cc: <stable@vger.kernel.org> # v4.18+
> ---
>  drivers/gpu/drm/i915/i915_gem_gtt.c | 5 +++++
>  1 file changed, 5 insertions(+)
> 
> diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c b/drivers/gpu/drm/i915/i915_gem_gtt.c
> index 0b81e0b64393..a1475039d182 100644
> --- a/drivers/gpu/drm/i915/i915_gem_gtt.c
> +++ b/drivers/gpu/drm/i915/i915_gem_gtt.c
> @@ -3152,6 +3152,11 @@ static int ggtt_probe_hw(struct i915_ggtt *ggtt, struct intel_gt *gt)
>         if (ret)
>                 return ret;
>  
> +       /* We don't have a max segment size, so set it to the max so sg's
> +        * debugging layer doesn't complain
> +        */
> +       dma_set_max_seg_size(ggtt->vm.dma, UINT_MAX);

The rest of the dma setup is in i915_driver_hw_probe, so I would put it
there just after dma_set_coherent_mask() (maybe one day even being brave
enough to pull out those to their own function).

I think I've made my point about the futility of dma-debug and you've
made yours about the simplicity of the patch to shut it up, so move it
across and
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
-Chris

Patch
diff mbox series

diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c b/drivers/gpu/drm/i915/i915_gem_gtt.c
index 0b81e0b64393..a1475039d182 100644
--- a/drivers/gpu/drm/i915/i915_gem_gtt.c
+++ b/drivers/gpu/drm/i915/i915_gem_gtt.c
@@ -3152,6 +3152,11 @@  static int ggtt_probe_hw(struct i915_ggtt *ggtt, struct intel_gt *gt)
 	if (ret)
 		return ret;
 
+	/* We don't have a max segment size, so set it to the max so sg's
+	 * debugging layer doesn't complain
+	 */
+	dma_set_max_seg_size(ggtt->vm.dma, UINT_MAX);
+
 	if ((ggtt->vm.total - 1) >> 32) {
 		DRM_ERROR("We never expected a Global GTT with more than 32bits"
 			  " of address space! Found %lldM!\n",