Re: [Intel-gfx] [PATCH 10/20] drm/i915: Export a preallocate variant of i915_active_acquire()

[Maarten Lankhorst <maarten.lankhorst@linux.intel.com>]

Op 10-07-2020 om 14:24 schreef Tvrtko Ursulin:
>
> On 09/07/2020 15:36, Maarten Lankhorst wrote:
>> Op 06-07-2020 om 08:19 schreef Chris Wilson:
>>> Sometimes we have to be very careful not to allocate underneath a mutex
>>> (or spinlock) and yet still want to track activity. Enter
>>> i915_active_acquire_for_context(). This raises the activity counter on
>>> i915_active prior to use and ensures that the fence-tree contains a slot
>>> for the context.
>>>
>>> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
>>> ---
>>>   .../gpu/drm/i915/gem/i915_gem_execbuffer.c    |   2 +-
>>>   drivers/gpu/drm/i915/gt/intel_timeline.c      |   4 +-
>>>   drivers/gpu/drm/i915/i915_active.c            | 113 +++++++++++++++---
>>>   drivers/gpu/drm/i915/i915_active.h            |  14 ++-
>>>   4 files changed, 113 insertions(+), 20 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
>>> index 1015c4fd9f3e..6d20be29ff3c 100644
>>> --- a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
>>> +++ b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
>>> @@ -1789,7 +1789,7 @@ __parser_mark_active(struct i915_vma *vma,
>>>   {
>>>       struct intel_gt_buffer_pool_node *node = vma->private;
>>>   -    return i915_active_ref(&node->active, tl, fence);
>>> +    return i915_active_ref(&node->active, tl->fence_context, fence);
>>>   }
>>>     static int
>>> diff --git a/drivers/gpu/drm/i915/gt/intel_timeline.c b/drivers/gpu/drm/i915/gt/intel_timeline.c
>>> index 4546284fede1..e4a5326633b8 100644
>>> --- a/drivers/gpu/drm/i915/gt/intel_timeline.c
>>> +++ b/drivers/gpu/drm/i915/gt/intel_timeline.c
>>> @@ -479,7 +479,9 @@ __intel_timeline_get_seqno(struct intel_timeline *tl,
>>>        * free it after the current request is retired, which ensures that
>>>        * all writes into the cacheline from previous requests are complete.
>>>        */
>>> -    err = i915_active_ref(&tl->hwsp_cacheline->active, tl, &rq->fence);
>>> +    err = i915_active_ref(&tl->hwsp_cacheline->active,
>>> +                  tl->fence_context,
>>> +                  &rq->fence);
>>>       if (err)
>>>           goto err_cacheline;
>>>   diff --git a/drivers/gpu/drm/i915/i915_active.c b/drivers/gpu/drm/i915/i915_active.c
>>> index d960d0be5bd2..3f595446fd44 100644
>>> --- a/drivers/gpu/drm/i915/i915_active.c
>>> +++ b/drivers/gpu/drm/i915/i915_active.c
>>> @@ -217,11 +217,10 @@ excl_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
>>>   }
>>>     static struct i915_active_fence *
>>> -active_instance(struct i915_active *ref, struct intel_timeline *tl)
>>> +active_instance(struct i915_active *ref, u64 idx)
>>>   {
>>>       struct active_node *node, *prealloc;
>>>       struct rb_node **p, *parent;
>>> -    u64 idx = tl->fence_context;
>>>         /*
>>>        * We track the most recently used timeline to skip a rbtree search
>>> @@ -353,21 +352,17 @@ __active_del_barrier(struct i915_active *ref, struct active_node *node)
>>>       return ____active_del_barrier(ref, node, barrier_to_engine(node));
>>>   }
>>>   -int i915_active_ref(struct i915_active *ref,
>>> -            struct intel_timeline *tl,
>>> -            struct dma_fence *fence)
>>> +int i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence)
>>>   {
>>>       struct i915_active_fence *active;
>>>       int err;
>>>   -    lockdep_assert_held(&tl->mutex);
>>> -
>>>       /* Prevent reaping in case we malloc/wait while building the tree */
>>>       err = i915_active_acquire(ref);
>>>       if (err)
>>>           return err;
>>>   -    active = active_instance(ref, tl);
>>> +    active = active_instance(ref, idx);
>>>       if (!active) {
>>>           err = -ENOMEM;
>>>           goto out;
>>> @@ -384,32 +379,104 @@ int i915_active_ref(struct i915_active *ref,
>>>           atomic_dec(&ref->count);
>>>       }
>>>       if (!__i915_active_fence_set(active, fence))
>>> -        atomic_inc(&ref->count);
>>> +        __i915_active_acquire(ref);
>>>     out:
>>>       i915_active_release(ref);
>>>       return err;
>>>   }
>>>   -struct dma_fence *
>>> -i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f)
>>> +static struct dma_fence *
>>> +__i915_active_set_fence(struct i915_active *ref,
>>> +            struct i915_active_fence *active,
>>> +            struct dma_fence *fence)
>>>   {
>>>       struct dma_fence *prev;
>>>         /* We expect the caller to manage the exclusive timeline ordering */
>>>       GEM_BUG_ON(i915_active_is_idle(ref));
>>>   +    if (is_barrier(active)) { /* proto-node used by our idle barrier */
>>> +        /*
>>> +         * This request is on the kernel_context timeline, and so
>>> +         * we can use it to substitute for the pending idle-barrer
>>> +         * request that we want to emit on the kernel_context.
>>> +         */
>>> +        __active_del_barrier(ref, node_from_active(active));
>>> +        RCU_INIT_POINTER(active->fence, NULL);
>>> +        atomic_dec(&ref->count);
>>> +    }
>>> +
>>>       rcu_read_lock();
>>> -    prev = __i915_active_fence_set(&ref->excl, f);
>>> +    prev = __i915_active_fence_set(active, fence);
>>>       if (prev)
>>>           prev = dma_fence_get_rcu(prev);
>>>       else
>>> -        atomic_inc(&ref->count);
>>> +        __i915_active_acquire(ref);
>>>       rcu_read_unlock();
>>>         return prev;
>>>   }
>>>   +static struct i915_active_fence *
>>> +__active_lookup(struct i915_active *ref, u64 idx)
>>> +{
>>> +    struct active_node *node;
>>> +    struct rb_node *p;
>>> +
>>> +    /* Like active_instance() but with no malloc */
>>> +
>>> +    node = READ_ONCE(ref->cache);
>>> +    if (node && node->timeline == idx)
>>> +        return &node->base;
>>> +
>>> +    spin_lock_irq(&ref->tree_lock);
>>> +    GEM_BUG_ON(i915_active_is_idle(ref));
>>> +
>>> +    p = ref->tree.rb_node;
>>> +    while (p) {
>>> +        node = rb_entry(p, struct active_node, node);
>>> +        if (node->timeline == idx) {
>>> +            ref->cache = node;
>>> +            spin_unlock_irq(&ref->tree_lock);
>>> +            return &node->base;
>>> +        }
>>> +
>>> +        if (node->timeline < idx)
>>> +            p = p->rb_right;
>>> +        else
>>> +            p = p->rb_left;
>>> +    }
>>> +
>>> +    spin_unlock_irq(&ref->tree_lock);
>>> +
>>> +    return NULL;
>>> +}
>>> +
>>> +struct dma_fence *
>>> +__i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence)
>>> +{
>>> +    struct dma_fence *prev = ERR_PTR(-ENOENT);
>>> +    struct i915_active_fence *active;
>>> +
>>> +    if (!i915_active_acquire_if_busy(ref))
>>> +        return ERR_PTR(-EINVAL);
>>> +
>>> +    active = __active_lookup(ref, idx);
>>> +    if (active)
>>> +        prev = __i915_active_set_fence(ref, active, fence);
>>> +
>>> +    i915_active_release(ref);
>>> +    return prev;
>>> +}
>>> +
>>> +struct dma_fence *
>>> +i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f)
>>> +{
>>> +    /* We expect the caller to manage the exclusive timeline ordering */
>>> +    return __i915_active_set_fence(ref, &ref->excl, f);
>>> +}
>>> +
>>>   bool i915_active_acquire_if_busy(struct i915_active *ref)
>>>   {
>>>       debug_active_assert(ref);
>>> @@ -443,6 +510,24 @@ int i915_active_acquire(struct i915_active *ref)
>>>       return err;
>>>   }
>>>   +int i915_active_acquire_for_context(struct i915_active *ref, u64 idx)
>>> +{
>>> +    struct i915_active_fence *active;
>>> +    int err;
>>> +
>>> +    err = i915_active_acquire(ref);
>>> +    if (err)
>>> +        return err;
>>> +
>>> +    active = active_instance(ref, idx);
>>> +    if (!active) {
>>> +        i915_active_release(ref);
>>> +        return -ENOMEM;
>>> +    }
>>> +
>>> +    return 0; /* return with active ref */
>>> +}
>>> +
>>>   void i915_active_release(struct i915_active *ref)
>>>   {
>>>       debug_active_assert(ref);
>>> @@ -804,7 +889,7 @@ int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
>>>                */
>>>               RCU_INIT_POINTER(node->base.fence, ERR_PTR(-EAGAIN));
>>>               node->base.cb.node.prev = (void *)engine;
>>> -            atomic_inc(&ref->count);
>>> +            __i915_active_acquire(ref);
>>>           }
>>>           GEM_BUG_ON(rcu_access_pointer(node->base.fence) != ERR_PTR(-EAGAIN));
>>>   diff --git a/drivers/gpu/drm/i915/i915_active.h b/drivers/gpu/drm/i915/i915_active.h
>>> index cf4058150966..2e0bcb3289ec 100644
>>> --- a/drivers/gpu/drm/i915/i915_active.h
>>> +++ b/drivers/gpu/drm/i915/i915_active.h
>>> @@ -163,14 +163,18 @@ void __i915_active_init(struct i915_active *ref,
>>>       __i915_active_init(ref, active, retire, &__mkey, &__wkey);    \
>>>   } while (0)
>>>   -int i915_active_ref(struct i915_active *ref,
>>> -            struct intel_timeline *tl,
>>> -            struct dma_fence *fence);
>>> +struct dma_fence *
>>> +__i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence);
>>> +int i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence);
>>>     static inline int
>>>   i915_active_add_request(struct i915_active *ref, struct i915_request *rq)
>>>   {
>>> -    return i915_active_ref(ref, i915_request_timeline(rq), &rq->fence);
>>> +    struct intel_timeline *tl = i915_request_timeline(rq);
>>> +
>>> +    lockdep_assert_held(&tl->mutex);
>>> +
>>> +    return i915_active_ref(ref, tl->fence_context, &rq->fence);
>>>   }
>>>     struct dma_fence *
>>> @@ -198,7 +202,9 @@ int i915_request_await_active(struct i915_request *rq,
>>>   #define I915_ACTIVE_AWAIT_BARRIER BIT(2)
>>>     int i915_active_acquire(struct i915_active *ref);
>>> +int i915_active_acquire_for_context(struct i915_active *ref, u64 idx);
>>>   bool i915_active_acquire_if_busy(struct i915_active *ref);
>>> +
>>>   void i915_active_release(struct i915_active *ref);
>>>     static inline void __i915_active_acquire(struct i915_active *ref)
>>
>> Somewhere this is hiding a locking inversion in the later part of this series.
>>
>> tl->mutex needs to be an inner lock of ww objects, see my patch series.
>>
>> Since this is not the case, you're somehow hiding a locking inversion.
>
> Which locks you think and from where? I want to make sure I am not missing something while reading the series..
>
> There used to be (still is in upstream) a possible deadlock with async get pages if batch buffer was an userptr object, but this series solves that by pre-allocating page directories in the reservation phase. That's not a plain locking inversion, but maybe that's what you had in mind?
>
> Does it take tl->mutex undeer dma_fence_wait? I can't see that it does.
>
> Can't think of any other options at the moment.
>
> Regards,
>
> Tvrtko

Ah never mind, context pinning/throttling is not reworked yet here.

_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx