Re: [PATCH v19 22/30] drm/shmem-helper: Add common memory shrinker

[Boris Brezillon <boris.brezillon@collabora.com>]

On Fri,  5 Jan 2024 21:46:16 +0300
Dmitry Osipenko <dmitry.osipenko@collabora.com> wrote:

>   *
>   * This function Increases the use count and allocates the backing pages if
>   * use-count equals to zero.
> + *
> + * Note that this function doesn't pin pages in memory. If your driver
> + * uses drm-shmem shrinker, then it's free to relocate pages to swap.
> + * Getting pages only guarantees that pages are allocated, and not that
> + * pages reside in memory. In order to pin pages use drm_gem_shmem_pin().

I still find this explanation confusing, if pages are allocated, they
reside in memory. The only difference between drm_gem_shmem_get_pages()
and drm_gem_shmem_pin_pages() is that the former lets the system
reclaim the memory if the buffer is idle (no unsignalled fence attached
to the dma_resv).

We also need to describe the workflow for GEM validation (that's the
TTM term for the swapin process happening when a GPU job is submitted).

1. Prepare the GPU job and initialize its fence
2. Lock the GEM resv
3. Add the GPU job fence to the resv object
4. If the GEM is evicted
   a. call drm_gem_shmem_swapin_locked()
   b. get the new sgt with drm_gem_shmem_get_pages_sgt_locked()
   c. repopulate the MMU table (driver internals)
5. Unlock the GEM dma_resv
6. Submit the GPU job

With this sequence, the GEM pages are guaranteed to stay around until
the GPU job is finished.

>   */
>  int drm_gem_shmem_get_pages(struct drm_gem_shmem_object *shmem)