[v3,0/2] drm/panic: Add support to scanout buffer as array of pages

Message

Jocelyn Falempe April 7, 2025, 1:42 p.m. UTC

Some drivers like virtio-gpu, don't map the scanout buffer in the
kernel. Calling vmap() in a panic handler is not safe, and writing an
atomic_vmap() API is more complex than expected [1].
So instead, pass the array of pages of the scanout buffer to the
panic handler, and map only one page at a time to draw the pixels.
This is obviously slow, but acceptable for a panic handler.
As kmap_local_page() is not safe to call from a panic handler,
introduce a kmap_local_page_try_from_panic() that will avoid unsafe
operations.

[1] https://lore.kernel.org/dri-devel/[e-mail]/

v2:
 * Add kmap_local_page_try_from_panic() (Simona Vetter)
 * Correctly handle the case if kmap_local_page_try_from_panic()
   returns NULL
 * Check that the current page is not NULL before trying to map it.
 * Add a comment in struct drm_scanout_buffer, that the array of
   pages shouldn't be allocated in the get_scanout_buffer() callback.

v3:
 * Replace DRM_WARN_ONCE with pr_debug_once (Simona Vetter)
 * Add a comment in kmap_local_page_try_from_panic() (Thomas Gleixner)

Jocelyn Falempe (2):
  mm/kmap: Add kmap_local_page_try_from_panic()
  drm/panic: Add support to scanout buffer as array of pages

 drivers/gpu/drm/drm_panic.c      | 142 +++++++++++++++++++++++++++++--
 include/drm/drm_panic.h          |  12 ++-
 include/linux/highmem-internal.h |  13 +++
 3 files changed, 160 insertions(+), 7 deletions(-)


base-commit: fbe43810d563a293e3de301141d33caf1f5d5c5a

Comments

Jocelyn Falempe April 10, 2025, 9:06 a.m. UTC | #1

On 07/04/2025 15:42, Jocelyn Falempe wrote:
> Some drivers like virtio-gpu, don't map the scanout buffer in the
> kernel. Calling vmap() in a panic handler is not safe, and writing an
> atomic_vmap() API is more complex than expected [1].
> So instead, pass the array of pages of the scanout buffer to the
> panic handler, and map only one page at a time to draw the pixels.
> This is obviously slow, but acceptable for a panic handler.
> As kmap_local_page() is not safe to call from a panic handler,
> introduce a kmap_local_page_try_from_panic() that will avoid unsafe
> operations.

I applied both patches to drm-misc/drm-misc-next.

Thanks for the reviews.

> 
> [1] https://lore.kernel.org/dri-devel/[e-mail]/
> 
> v2:
>   * Add kmap_local_page_try_from_panic() (Simona Vetter)
>   * Correctly handle the case if kmap_local_page_try_from_panic()
>     returns NULL
>   * Check that the current page is not NULL before trying to map it.
>   * Add a comment in struct drm_scanout_buffer, that the array of
>     pages shouldn't be allocated in the get_scanout_buffer() callback.
> 
> v3:
>   * Replace DRM_WARN_ONCE with pr_debug_once (Simona Vetter)
>   * Add a comment in kmap_local_page_try_from_panic() (Thomas Gleixner)
> 
> Jocelyn Falempe (2):
>    mm/kmap: Add kmap_local_page_try_from_panic()
>    drm/panic: Add support to scanout buffer as array of pages
> 
>   drivers/gpu/drm/drm_panic.c      | 142 +++++++++++++++++++++++++++++--
>   include/drm/drm_panic.h          |  12 ++-
>   include/linux/highmem-internal.h |  13 +++
>   3 files changed, 160 insertions(+), 7 deletions(-)
> 
> 
> base-commit: fbe43810d563a293e3de301141d33caf1f5d5c5a