On 26/01/2024 09:39, Boris Brezillon wrote:
> On Thu, 25 Jan 2024 16:47:24 +0000
> Steven Price <steven.price@xxxxxxx> wrote:
>
>> On 05/01/2024 18:46, Dmitry Osipenko wrote:
>>> To simplify the drm-shmem refcnt handling, we're moving away from
>>> the implicit get_pages() that is used by get_pages_sgt(). From now on
>>> drivers will have to pin pages while they use sgt. Panfrost's shrinker
>>> doesn't support swapping out BOs, hence pages are pinned and sgt is valid
>>> as long as pages' use-count > 0.
>>>
>>> In Panfrost, panfrost_gem_mapping, which is the object representing a
>>> GPU mapping of a BO, owns a pages ref. This guarantees that any BO being
>>> mapped GPU side has its pages retained till the mapping is destroyed.
>>>
>>> Since pages are no longer guaranteed to stay pinned for the BO lifetime,
>>> and MADVISE(DONT_NEED) flagging remains after the GEM handle has been
>>> destroyed, we need to add an extra 'is_purgeable' check in
>>> panfrost_gem_purge(), to make sure we're not trying to purge a BO that
>>> already had its pages released.
>>>
>>> Signed-off-by: Dmitry Osipenko <dmitry.osipenko@xxxxxxxxxxxxx>
>>
>> Reviewed-by: Steven Price <steven.price@xxxxxxx>
>>
>> Although I don't like the condition in panfrost_gem_mapping_release()
>> for drm_gem_shmem_put_pages() and assigning NULL to bo->sgts - it feels
>> very fragile. See below.
>>
>>> ---
>>> drivers/gpu/drm/panfrost/panfrost_gem.c | 63 ++++++++++++++-----
>>> .../gpu/drm/panfrost/panfrost_gem_shrinker.c | 6 ++
>>> 2 files changed, 52 insertions(+), 17 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/panfrost/panfrost_gem.c b/drivers/gpu/drm/panfrost/panfrost_gem.c
>>> index f268bd5c2884..7edfc12f7c1f 100644
>>> --- a/drivers/gpu/drm/panfrost/panfrost_gem.c
>>> +++ b/drivers/gpu/drm/panfrost/panfrost_gem.c
>>> @@ -35,20 +35,6 @@ static void panfrost_gem_free_object(struct drm_gem_object *obj)
>>> */
>>> WARN_ON_ONCE(!list_empty(&bo->mappings.list));
>>>
>>> - if (bo->sgts) {
>>> - int i;
>>> - int n_sgt = bo->base.base.size / SZ_2M;
>>> -
>>> - for (i = 0; i < n_sgt; i++) {
>>> - if (bo->sgts[i].sgl) {
>>> - dma_unmap_sgtable(pfdev->dev, &bo->sgts[i],
>>> - DMA_BIDIRECTIONAL, 0);
>>> - sg_free_table(&bo->sgts[i]);
>>> - }
>>> - }
>>> - kvfree(bo->sgts);
>>> - }
>>> -
>>> drm_gem_shmem_free(&bo->base);
>>> }
>>>
>>> @@ -85,11 +71,40 @@ panfrost_gem_teardown_mapping(struct panfrost_gem_mapping *mapping)
>>>
>>> static void panfrost_gem_mapping_release(struct kref *kref)
>>> {
>>> - struct panfrost_gem_mapping *mapping;
>>> -
>>> - mapping = container_of(kref, struct panfrost_gem_mapping, refcount);
>>> + struct panfrost_gem_mapping *mapping =
>>> + container_of(kref, struct panfrost_gem_mapping, refcount);
>>> + struct panfrost_gem_object *bo = mapping->obj;
>>> + struct panfrost_device *pfdev = bo->base.base.dev->dev_private;
>>>
>>> panfrost_gem_teardown_mapping(mapping);
>>> +
>>> + /* On heap BOs, release the sgts created in the fault handler path. */
>>> + if (bo->sgts) {
>>> + int i, n_sgt = bo->base.base.size / SZ_2M;
>>> +
>>> + for (i = 0; i < n_sgt; i++) {
>>> + if (bo->sgts[i].sgl) {
>>> + dma_unmap_sgtable(pfdev->dev, &bo->sgts[i],
>>> + DMA_BIDIRECTIONAL, 0);
>>> + sg_free_table(&bo->sgts[i]);
>>> + }
>>> + }
>>> + kvfree(bo->sgts);
>>> + }
>>> +
>>> + /* Pages ref is owned by the panfrost_gem_mapping object. We must
>>> + * release our pages ref (if any), before releasing the object
>>> + * ref.
>>> + * Non-heap BOs acquired the pages at panfrost_gem_mapping creation
>>> + * time, and heap BOs may have acquired pages if the fault handler
>>> + * was called, in which case bo->sgts should be non-NULL.
>>> + */
>>> + if (!bo->base.base.import_attach && (!bo->is_heap || bo->sgts) &&
>>> + bo->base.madv >= 0) {
>>> + drm_gem_shmem_put_pages(&bo->base);
>>> + bo->sgts = NULL;
>>
>> The assignment of NULL here really ought to be unconditional - it isn't
>> a valid pointer because of the kvfree() above.
>
> Fair enough. How about we drop the '|| bo->sgts' and add an
> drm_gem_shmem_put_pages() to the above if (bo->sgts) block, where we'll
> also assign bo->sgts to NULL?
Yes that would be good.
>>
>> I also feel that the big condition above suggests there's a need for a
>> better state machine to keep track of what's going on.
>
> I'm planning to extend drm_gem_shmem to support the alloc-on-fault use
> case that all Mali GPUs seem to rely on (lima, panfrost and soon
> panthor would use those helpers). The idea is to:
>
> - make the allocation non-blocking, so we can kill the blocking
> allocation in the dma signalling path (basically what intel does)
> - allow dynamic extension of the pages array using an xarray instead of
> a plain array
>
> Hopefully this makes the state tracking a lot easier, and we can also
> get rid of the hack we have in panfrost/lima where we manipulate
> drm_gem_shmem_object refcounts directly.
That sounds great - it would definitely be good to get rid of the
refcount hack, it confuses me everytime ;)
Thanks,
Steve
>>
>> But having said that I do think this series as a whole is an
>> improvement, it's nice to get the shrinker code generic. And sadly I
>> don't have an immediate idea for cleaning this up, hence my R-b.
>>
>> Steve
>>
>>> + }
>>> +
>>> drm_gem_object_put(&mapping->obj->base.base);
>>> panfrost_mmu_ctx_put(mapping->mmu);
>>> kfree(mapping);
>>> @@ -125,6 +140,20 @@ int panfrost_gem_open(struct drm_gem_object *obj, struct drm_file *file_priv)
>>> if (!mapping)
>>> return -ENOMEM;
>>>
>>> + if (!bo->is_heap && !bo->base.base.import_attach) {
>>> + /* Pages ref is owned by the panfrost_gem_mapping object.
>>> + * For non-heap BOs, we request pages at mapping creation
>>> + * time, such that the panfrost_mmu_map() call, further down in
>>> + * this function, is guaranteed to have pages_use_count > 0
>>> + * when drm_gem_shmem_get_pages_sgt() is called.
>>> + */
>>> + ret = drm_gem_shmem_get_pages(&bo->base);
>>> + if (ret) {
>>> + kfree(mapping);
>>> + return ret;
>>> + }
>>> + }
>>> +
>>> INIT_LIST_HEAD(&mapping->node);
>>> kref_init(&mapping->refcount);
>>> drm_gem_object_get(obj);
>>> diff --git a/drivers/gpu/drm/panfrost/panfrost_gem_shrinker.c b/drivers/gpu/drm/panfrost/panfrost_gem_shrinker.c
>>> index 02b60ea1433a..d4fb0854cf2f 100644
>>> --- a/drivers/gpu/drm/panfrost/panfrost_gem_shrinker.c
>>> +++ b/drivers/gpu/drm/panfrost/panfrost_gem_shrinker.c
>>> @@ -50,6 +50,12 @@ static bool panfrost_gem_purge(struct drm_gem_object *obj)
>>> if (!dma_resv_trylock(shmem->base.resv))
>>> goto unlock_mappings;
>>>
>>> + /* BO might have become unpurgeable if the last pages_use_count ref
>>> + * was dropped, but the BO hasn't been destroyed yet.
>>> + */
>>> + if (!drm_gem_shmem_is_purgeable(shmem))
>>> + goto unlock_mappings;
>>> +
>>> panfrost_gem_teardown_mappings_locked(bo);
>>> drm_gem_shmem_purge_locked(&bo->base);
>>> ret = true;
>>
>
