Re: [PATCH 3/9] drm/ttm: use per BO cleanup workers

[Christian König <christian.koenig@xxxxxxx>]

Am 29.11.22 um 22:14 schrieb Felix Kuehling:
> On 2022-11-25 05:21, Christian König wrote:
> > Instead of a single worker going over the list of delete BOs in regular
> > intervals use a per BO worker which blocks for the resv object and
> > locking of the BO.
> >
> > This not only simplifies the handling massively, but also results in
> > much better response time when cleaning up buffers.
> >
> > Signed-off-by: Christian König <christian.koenig@xxxxxxx>
>
> Just thinking out loud: If I understand it correctly, this can cause a 
> lot of sleeping worker threads when 
> AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE is used and many BOs are freed 
> at the same time. This happens e.g. when a KFD process terminates or 
> crashes. I guess with a concurrency-managed workqueue this isn't going 
> to be excessive. And since it's on a per device workqueue, it doesn't 
> stall work items on the system work queue or from other devices.

Yes, exactly that. The last parameter to alloc_workqueue() limits how 
many work items can be sleeping.

> I'm trying to understand why you set WQ_MEM_RECLAIM. This work queue 
> is not about freeing ttm_resources but about freeing the BOs. But it 
> affects freeing of ghost_objs that are holding the ttm_resources being 
> freed.

Well if the BO is idle, but not immediately lockable we delegate freeing 
the backing pages in the TT object to those workers as well. It might 
even be a good idea to use a separate wq for this case.

> If those assumptions all make sense, patches 1-3 are
>
> Reviewed-by: Felix Kuehling <Felix.Kuehling@xxxxxxx>

Thanks,
Christian.

> > ---
> >   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c |   2 +-
> >   drivers/gpu/drm/i915/i915_gem.c            |   2 +-
> >   drivers/gpu/drm/i915/intel_region_ttm.c    |   2 +-
> >   drivers/gpu/drm/ttm/ttm_bo.c               | 112 ++++++++-------------
> >   drivers/gpu/drm/ttm/ttm_bo_util.c          |   1 -
> >   drivers/gpu/drm/ttm/ttm_device.c           |  24 ++---
> >   include/drm/ttm/ttm_bo_api.h               |  18 +---
> >   include/drm/ttm/ttm_device.h               |   7 +-
> >   8 files changed, 57 insertions(+), 111 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c 
> > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> > index 2b1db37e25c1..74ccbd566777 100644
> > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> > @@ -3984,7 +3984,7 @@ void amdgpu_device_fini_hw(struct amdgpu_device 
> > *adev)
> >       amdgpu_fence_driver_hw_fini(adev);
> >         if (adev->mman.initialized)
> > -        flush_delayed_work(&adev->mman.bdev.wq);
> > +        drain_workqueue(adev->mman.bdev.wq);
> >         if (adev->pm_sysfs_en)
> >           amdgpu_pm_sysfs_fini(adev);
> > diff --git a/drivers/gpu/drm/i915/i915_gem.c 
> > b/drivers/gpu/drm/i915/i915_gem.c
> > index 8468ca9885fd..c38306f156d6 100644
> > --- a/drivers/gpu/drm/i915/i915_gem.c
> > +++ b/drivers/gpu/drm/i915/i915_gem.c
> > @@ -1099,7 +1099,7 @@ void i915_gem_drain_freed_objects(struct 
> > drm_i915_private *i915)
> >   {
> >       while (atomic_read(&i915->mm.free_count)) {
> >           flush_work(&i915->mm.free_work);
> > -        flush_delayed_work(&i915->bdev.wq);
> > +        drain_workqueue(i915->bdev.wq);
> >           rcu_barrier();
> >       }
> >   }
> > diff --git a/drivers/gpu/drm/i915/intel_region_ttm.c 
> > b/drivers/gpu/drm/i915/intel_region_ttm.c
> > index cf89d0c2a2d9..657bbc16a48a 100644
> > --- a/drivers/gpu/drm/i915/intel_region_ttm.c
> > +++ b/drivers/gpu/drm/i915/intel_region_ttm.c
> > @@ -132,7 +132,7 @@ int intel_region_ttm_fini(struct 
> > intel_memory_region *mem)
> >               break;
> >             msleep(20);
> > -        flush_delayed_work(&mem->i915->bdev.wq);
> > +        drain_workqueue(mem->i915->bdev.wq);
> >       }
> >         /* If we leaked objects, Don't free the region causing use 
> > after free */
> > diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
> > index b77262a623e0..4749b65bedc4 100644
> > --- a/drivers/gpu/drm/ttm/ttm_bo.c
> > +++ b/drivers/gpu/drm/ttm/ttm_bo.c
> > @@ -280,14 +280,13 @@ static int ttm_bo_cleanup_refs(struct 
> > ttm_buffer_object *bo,
> >           ret = 0;
> >       }
> >   -    if (ret || unlikely(list_empty(&bo->ddestroy))) {
> > +    if (ret) {
> >           if (unlock_resv)
> >               dma_resv_unlock(bo->base.resv);
> >           spin_unlock(&bo->bdev->lru_lock);
> >           return ret;
> >       }
> >   -    list_del_init(&bo->ddestroy);
> >       spin_unlock(&bo->bdev->lru_lock);
> >       ttm_bo_cleanup_memtype_use(bo);
> >   @@ -300,47 +299,21 @@ static int ttm_bo_cleanup_refs(struct 
> > ttm_buffer_object *bo,
> >   }
> >     /*
> > - * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
> > - * encountered buffers.
> > + * Block for the dma_resv object to become idle, lock the buffer and 
> > clean up
> > + * the resource and tt object.
> >    */
> > -bool ttm_bo_delayed_delete(struct ttm_device *bdev, bool remove_all)
> > +static void ttm_bo_delayed_delete(struct work_struct *work)
> >   {
> > -    struct list_head removed;
> > -    bool empty;
> > -
> > -    INIT_LIST_HEAD(&removed);
> > -
> > -    spin_lock(&bdev->lru_lock);
> > -    while (!list_empty(&bdev->ddestroy)) {
> > -        struct ttm_buffer_object *bo;
> > -
> > -        bo = list_first_entry(&bdev->ddestroy, struct 
> > ttm_buffer_object,
> > -                      ddestroy);
> > -        list_move_tail(&bo->ddestroy, &removed);
> > -        if (!ttm_bo_get_unless_zero(bo))
> > -            continue;
> > -
> > -        if (remove_all || bo->base.resv != &bo->base._resv) {
> > -            spin_unlock(&bdev->lru_lock);
> > -            dma_resv_lock(bo->base.resv, NULL);
> > -
> > -            spin_lock(&bdev->lru_lock);
> > -            ttm_bo_cleanup_refs(bo, false, !remove_all, true);
> > -
> > -        } else if (dma_resv_trylock(bo->base.resv)) {
> > -            ttm_bo_cleanup_refs(bo, false, !remove_all, true);
> > -        } else {
> > -            spin_unlock(&bdev->lru_lock);
> > -        }
> > +    struct ttm_buffer_object *bo;
> >   -        ttm_bo_put(bo);
> > -        spin_lock(&bdev->lru_lock);
> > -    }
> > -    list_splice_tail(&removed, &bdev->ddestroy);
> > -    empty = list_empty(&bdev->ddestroy);
> > -    spin_unlock(&bdev->lru_lock);
> > +    bo = container_of(work, typeof(*bo), delayed_delete);
> >   -    return empty;
> > +    dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP, 
> > false,
> > +                  MAX_SCHEDULE_TIMEOUT);
> > +    dma_resv_lock(bo->base.resv, NULL);
> > +    ttm_bo_cleanup_memtype_use(bo);
> > +    dma_resv_unlock(bo->base.resv);
> > +    ttm_bo_put(bo);
> >   }
> >     static void ttm_bo_release(struct kref *kref)
> > @@ -369,44 +342,40 @@ static void ttm_bo_release(struct kref *kref)
> >             drm_vma_offset_remove(bdev->vma_manager, 
> > &bo->base.vma_node);
> >           ttm_mem_io_free(bdev, bo->resource);
> > -    }
> > -
> > -    if (!dma_resv_test_signaled(bo->base.resv, 
> > DMA_RESV_USAGE_BOOKKEEP) ||
> > -        !dma_resv_trylock(bo->base.resv)) {
> > -        /* The BO is not idle, resurrect it for delayed destroy */
> > -        ttm_bo_flush_all_fences(bo);
> > -        bo->deleted = true;
> >   -        spin_lock(&bo->bdev->lru_lock);
> > +        if (!dma_resv_test_signaled(bo->base.resv,
> > +                        DMA_RESV_USAGE_BOOKKEEP) ||
> > +            !dma_resv_trylock(bo->base.resv)) {
> > +            /* The BO is not idle, resurrect it for delayed destroy */
> > +            ttm_bo_flush_all_fences(bo);
> > +            bo->deleted = true;
> >   -        /*
> > -         * Make pinned bos immediately available to
> > -         * shrinkers, now that they are queued for
> > -         * destruction.
> > -         *
> > -         * FIXME: QXL is triggering this. Can be removed when the
> > -         * driver is fixed.
> > -         */
> > -        if (bo->pin_count) {
> > -            bo->pin_count = 0;
> > -            ttm_resource_move_to_lru_tail(bo->resource);
> > -        }
> > +            spin_lock(&bo->bdev->lru_lock);
> >   -        kref_init(&bo->kref);
> > -        list_add_tail(&bo->ddestroy, &bdev->ddestroy);
> > -        spin_unlock(&bo->bdev->lru_lock);
> > +            /*
> > +             * Make pinned bos immediately available to
> > +             * shrinkers, now that they are queued for
> > +             * destruction.
> > +             *
> > +             * FIXME: QXL is triggering this. Can be removed when the
> > +             * driver is fixed.
> > +             */
> > +            if (bo->pin_count) {
> > +                bo->pin_count = 0;
> > +                ttm_resource_move_to_lru_tail(bo->resource);
> > +            }
> >   -        schedule_delayed_work(&bdev->wq,
> > -                      ((HZ / 100) < 1) ? 1 : HZ / 100);
> > -        return;
> > -    }
> > +            kref_init(&bo->kref);
> > +            spin_unlock(&bo->bdev->lru_lock);
> >   -    spin_lock(&bo->bdev->lru_lock);
> > -    list_del(&bo->ddestroy);
> > -    spin_unlock(&bo->bdev->lru_lock);
> > +            INIT_WORK(&bo->delayed_delete, ttm_bo_delayed_delete);
> > +            queue_work(bdev->wq, &bo->delayed_delete);
> > +            return;
> > +        }
> >   -    ttm_bo_cleanup_memtype_use(bo);
> > -    dma_resv_unlock(bo->base.resv);
> > +        ttm_bo_cleanup_memtype_use(bo);
> > +        dma_resv_unlock(bo->base.resv);
> > +    }
> >         atomic_dec(&ttm_glob.bo_count);
> >       bo->destroy(bo);
> > @@ -946,7 +915,6 @@ int ttm_bo_init_reserved(struct ttm_device *bdev, 
> > struct ttm_buffer_object *bo,
> >       int ret;
> >         kref_init(&bo->kref);
> > -    INIT_LIST_HEAD(&bo->ddestroy);
> >       bo->bdev = bdev;
> >       bo->type = type;
> >       bo->page_alignment = alignment;
> > diff --git a/drivers/gpu/drm/ttm/ttm_bo_util.c 
> > b/drivers/gpu/drm/ttm/ttm_bo_util.c
> > index ba3aa0a0fc43..ae4b7922ee1a 100644
> > --- a/drivers/gpu/drm/ttm/ttm_bo_util.c
> > +++ b/drivers/gpu/drm/ttm/ttm_bo_util.c
> > @@ -230,7 +230,6 @@ static int ttm_buffer_object_transfer(struct 
> > ttm_buffer_object *bo,
> >        */
> >         atomic_inc(&ttm_glob.bo_count);
> > -    INIT_LIST_HEAD(&fbo->base.ddestroy);
> >       drm_vma_node_reset(&fbo->base.base.vma_node);
> >         kref_init(&fbo->base.kref);
> > diff --git a/drivers/gpu/drm/ttm/ttm_device.c 
> > b/drivers/gpu/drm/ttm/ttm_device.c
> > index e7147e304637..e9bedca4dfdc 100644
> > --- a/drivers/gpu/drm/ttm/ttm_device.c
> > +++ b/drivers/gpu/drm/ttm/ttm_device.c
> > @@ -175,16 +175,6 @@ int ttm_device_swapout(struct ttm_device *bdev, 
> > struct ttm_operation_ctx *ctx,
> >   }
> >   EXPORT_SYMBOL(ttm_device_swapout);
> >   -static void ttm_device_delayed_workqueue(struct work_struct *work)
> > -{
> > -    struct ttm_device *bdev =
> > -        container_of(work, struct ttm_device, wq.work);
> > -
> > -    if (!ttm_bo_delayed_delete(bdev, false))
> > -        schedule_delayed_work(&bdev->wq,
> > -                      ((HZ / 100) < 1) ? 1 : HZ / 100);
> > -}
> > -
> >   /**
> >    * ttm_device_init
> >    *
> > @@ -215,15 +205,19 @@ int ttm_device_init(struct ttm_device *bdev, 
> > struct ttm_device_funcs *funcs,
> >       if (ret)
> >           return ret;
> >   +    bdev->wq = alloc_workqueue("ttm", WQ_MEM_RECLAIM | WQ_HIGHPRI, 
> > 16);
> > +    if (!bdev->wq) {
> > +        ttm_global_release();
> > +        return -ENOMEM;
> > +    }
> > +
> >       bdev->funcs = funcs;
> >         ttm_sys_man_init(bdev);
> >       ttm_pool_init(&bdev->pool, dev, use_dma_alloc, use_dma32);
> >         bdev->vma_manager = vma_manager;
> > -    INIT_DELAYED_WORK(&bdev->wq, ttm_device_delayed_workqueue);
> >       spin_lock_init(&bdev->lru_lock);
> > -    INIT_LIST_HEAD(&bdev->ddestroy);
> >       INIT_LIST_HEAD(&bdev->pinned);
> >       bdev->dev_mapping = mapping;
> >       mutex_lock(&ttm_global_mutex);
> > @@ -247,10 +241,8 @@ void ttm_device_fini(struct ttm_device *bdev)
> >       list_del(&bdev->device_list);
> >       mutex_unlock(&ttm_global_mutex);
> >   -    cancel_delayed_work_sync(&bdev->wq);
> > -
> > -    if (ttm_bo_delayed_delete(bdev, true))
> > -        pr_debug("Delayed destroy list was clean\n");
> > +    drain_workqueue(bdev->wq);
> > +    destroy_workqueue(bdev->wq);
> >         spin_lock(&bdev->lru_lock);
> >       for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
> > diff --git a/include/drm/ttm/ttm_bo_api.h b/include/drm/ttm/ttm_bo_api.h
> > index 7758347c461c..69e62bbb01e3 100644
> > --- a/include/drm/ttm/ttm_bo_api.h
> > +++ b/include/drm/ttm/ttm_bo_api.h
> > @@ -92,7 +92,6 @@ struct ttm_tt;
> >    * @ttm: TTM structure holding system pages.
> >    * @evicted: Whether the object was evicted without user-space 
> > knowing.
> >    * @deleted: True if the object is only a zombie and already deleted.
> > - * @ddestroy: List head for the delayed destroy list.
> >    * @swap: List head for swap LRU list.
> >    * @offset: The current GPU offset, which can have different meanings
> >    * depending on the memory type. For SYSTEM type memory, it should 
> > be 0.
> > @@ -135,19 +134,14 @@ struct ttm_buffer_object {
> >       struct ttm_tt *ttm;
> >       bool deleted;
> >       struct ttm_lru_bulk_move *bulk_move;
> > +    unsigned priority;
> > +    unsigned pin_count;
> >         /**
> > -     * Members protected by the bdev::lru_lock.
> > -     */
> > -
> > -    struct list_head ddestroy;
> > -
> > -    /**
> > -     * Members protected by a bo reservation.
> > +     * @delayed_delete: Work item used when we can't delete the BO
> > +     * immediately
> >        */
> > -
> > -    unsigned priority;
> > -    unsigned pin_count;
> > +    struct work_struct delayed_delete;
> >         /**
> >        * Special members that are protected by the reserve lock
> > @@ -448,8 +442,6 @@ void ttm_bo_vm_close(struct vm_area_struct *vma);
> >     int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
> >                void *buf, int len, int write);
> > -bool ttm_bo_delayed_delete(struct ttm_device *bdev, bool remove_all);
> > -
> >   vm_fault_t ttm_bo_vm_dummy_page(struct vm_fault *vmf, pgprot_t prot);
> >     #endif
> > diff --git a/include/drm/ttm/ttm_device.h b/include/drm/ttm/ttm_device.h
> > index 95b3c04b1ab9..4f3e81eac6f3 100644
> > --- a/include/drm/ttm/ttm_device.h
> > +++ b/include/drm/ttm/ttm_device.h
> > @@ -251,11 +251,6 @@ struct ttm_device {
> >        */
> >       spinlock_t lru_lock;
> >   -    /**
> > -     * @ddestroy: Destroyed but not yet cleaned up buffer objects.
> > -     */
> > -    struct list_head ddestroy;
> > -
> >       /**
> >        * @pinned: Buffer objects which are pinned and so not on any 
> > LRU list.
> >        */
> > @@ -270,7 +265,7 @@ struct ttm_device {
> >       /**
> >        * @wq: Work queue structure for the delayed delete workqueue.
> >        */
> > -    struct delayed_work wq;
> > +    struct workqueue_struct *wq;
> >   };
> >     int ttm_global_swapout(struct ttm_operation_ctx *ctx, gfp_t 
> > gfp_flags);