Re: [PATCH v6 4/8] drm/i915/gem: Add i915_gem_object_panic_map()

["Ville Syrjälä" <ville.syrjala@linux.intel.com>]

On Tue, Apr 01, 2025 at 08:47:50PM +0300, Ville Syrjälä wrote:
> On Tue, Apr 01, 2025 at 02:51:10PM +0200, Jocelyn Falempe wrote:
> > Prepare the work for drm_panic support. This is used to map the
> > current framebuffer, so the CPU can overwrite it with the panic
> > message.
> > 
> > Signed-off-by: Jocelyn Falempe <jfalempe@redhat.com>
> > ---
> > 
> > v5:
> >  * Use iosys_map for intel_bo_panic_map().
> > 
> >  drivers/gpu/drm/i915/display/intel_bo.c    |  5 ++++
> >  drivers/gpu/drm/i915/display/intel_bo.h    |  1 +
> >  drivers/gpu/drm/i915/gem/i915_gem_object.h |  2 ++
> >  drivers/gpu/drm/i915/gem/i915_gem_pages.c  | 29 ++++++++++++++++++++++
> >  drivers/gpu/drm/xe/display/intel_bo.c      | 10 ++++++++
> >  5 files changed, 47 insertions(+)
> > 
> > diff --git a/drivers/gpu/drm/i915/display/intel_bo.c b/drivers/gpu/drm/i915/display/intel_bo.c
> > index fbd16d7b58d9..ac904e9ec7d5 100644
> > --- a/drivers/gpu/drm/i915/display/intel_bo.c
> > +++ b/drivers/gpu/drm/i915/display/intel_bo.c
> > @@ -57,3 +57,8 @@ void intel_bo_describe(struct seq_file *m, struct drm_gem_object *obj)
> >  {
> >  	i915_debugfs_describe_obj(m, to_intel_bo(obj));
> >  }
> > +
> > +void intel_bo_panic_map(struct drm_gem_object *obj, struct iosys_map *map)
> > +{
> > +	i915_gem_object_panic_map(to_intel_bo(obj), map);
> > +}
> > diff --git a/drivers/gpu/drm/i915/display/intel_bo.h b/drivers/gpu/drm/i915/display/intel_bo.h
> > index ea7a2253aaa5..5b6c63d99786 100644
> > --- a/drivers/gpu/drm/i915/display/intel_bo.h
> > +++ b/drivers/gpu/drm/i915/display/intel_bo.h
> > @@ -23,5 +23,6 @@ struct intel_frontbuffer *intel_bo_set_frontbuffer(struct drm_gem_object *obj,
> >  						   struct intel_frontbuffer *front);
> >  
> >  void intel_bo_describe(struct seq_file *m, struct drm_gem_object *obj);
> > +void intel_bo_panic_map(struct drm_gem_object *obj, struct iosys_map *map);
> >  
> >  #endif /* __INTEL_BO__ */
> > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.h b/drivers/gpu/drm/i915/gem/i915_gem_object.h
> > index a5f34542135c..b16092707ea5 100644
> > --- a/drivers/gpu/drm/i915/gem/i915_gem_object.h
> > +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.h
> > @@ -692,6 +692,8 @@ i915_gem_object_unpin_pages(struct drm_i915_gem_object *obj)
> >  int __i915_gem_object_put_pages(struct drm_i915_gem_object *obj);
> >  int i915_gem_object_truncate(struct drm_i915_gem_object *obj);
> >  
> > +void i915_gem_object_panic_map(struct drm_i915_gem_object *obj, struct iosys_map *map);
> > +
> >  /**
> >   * i915_gem_object_pin_map - return a contiguous mapping of the entire object
> >   * @obj: the object to map into kernel address space
> > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_pages.c b/drivers/gpu/drm/i915/gem/i915_gem_pages.c
> > index 8780aa243105..718bea6474d7 100644
> > --- a/drivers/gpu/drm/i915/gem/i915_gem_pages.c
> > +++ b/drivers/gpu/drm/i915/gem/i915_gem_pages.c
> > @@ -355,6 +355,35 @@ static void *i915_gem_object_map_pfn(struct drm_i915_gem_object *obj,
> >  	return vaddr ?: ERR_PTR(-ENOMEM);
> >  }
> >  
> > +/* Map the current framebuffer for CPU access. Called from panic handler, so no
> > + * need to pin or cleanup.
> > + */
> > +void i915_gem_object_panic_map(struct drm_i915_gem_object *obj, struct iosys_map *map)
> > +{
> > +	enum i915_map_type has_type;
> > +	void *ptr;
> > +
> > +	ptr = page_unpack_bits(obj->mm.mapping, &has_type);
> > +
> > +
> > +	if (!ptr) {
> > +		if (i915_gem_object_has_struct_page(obj))
> > +			ptr = i915_gem_object_map_page(obj, I915_MAP_WB);
> > +		else
> > +			ptr = i915_gem_object_map_pfn(obj, I915_MAP_WB);
> 
> WB mapping would require clflushing to make it to the display.
> Is that being done somewhere?

This also seems to have a bunch of race conditions:
- what happens if the oops happens before the pages have
  even been swapped in?
- what happens if the oops happens before we've committed
  the fb to the hardware?

> 
> > +
> > +		if (IS_ERR(ptr))
> > +			return;
> 
> What happens when the mapping fails?
> 
> > +
> > +		obj->mm.mapping = page_pack_bits(ptr, I915_MAP_WB);
> > +	}
> > +
> > +	if (i915_gem_object_has_iomem(obj))
> > +		iosys_map_set_vaddr_iomem(map, (void __iomem *) ptr);
> > +	else
> > +		iosys_map_set_vaddr(map, ptr);
> > +}
> > +
> >  /* get, pin, and map the pages of the object into kernel space */
> >  void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj,
> >  			      enum i915_map_type type)
> > diff --git a/drivers/gpu/drm/xe/display/intel_bo.c b/drivers/gpu/drm/xe/display/intel_bo.c
> > index 27437c22bd70..c68166a64336 100644
> > --- a/drivers/gpu/drm/xe/display/intel_bo.c
> > +++ b/drivers/gpu/drm/xe/display/intel_bo.c
> > @@ -59,3 +59,13 @@ void intel_bo_describe(struct seq_file *m, struct drm_gem_object *obj)
> >  {
> >  	/* FIXME */
> >  }
> > +
> > +void intel_bo_panic_map(struct drm_gem_object *obj, struct iosys_map *map)
> > +{
> > +	struct xe_bo *bo = gem_to_xe_bo(obj);
> > +	int ret;
> > +
> > +	ret = ttm_bo_vmap(&bo->ttm, map);
> > +	if (ret)
> > +		iosys_map_clear(map);
> > +}
> > -- 
> > 2.49.0
> 
> -- 
> Ville Syrjälä
> Intel

-- 
Ville Syrjälä
Intel