[Intel-gfx] [PATCH v2] drm/i915: Track the last-active inside the i915_vma

[Chris Wilson]

Quoting Tvrtko Ursulin (2018-07-04 12:34:04)
> 
> On 04/07/2018 10:39, Tvrtko Ursulin wrote:
> > 
> > On 04/07/2018 09:34, Chris Wilson wrote:
> >> Using a VMA on more than one timeline concurrently is the exception
> >> rather than the rule (using it concurrently on multiple engines). As we
> >> expect to only use one active tracker, store the most recently used
> >> tracker inside the i915_vma itself and only fallback to the rbtree if
> >> we need a second or more concurrent active trackers.
> >>
> >> v2: Comments on how we overwrite any existing last_active cache.
> >>
> >> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> >> Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> >> ---
> >>   drivers/gpu/drm/i915/i915_vma.c | 50 +++++++++++++++++++++++++++++++--
> >>   drivers/gpu/drm/i915/i915_vma.h |  1 +
> >>   2 files changed, 49 insertions(+), 2 deletions(-)
> >>
> >> diff --git a/drivers/gpu/drm/i915/i915_vma.c 
> >> b/drivers/gpu/drm/i915/i915_vma.c
> >> index cd94ffc7f079..33925e00f7e8 100644
> >> --- a/drivers/gpu/drm/i915/i915_vma.c
> >> +++ b/drivers/gpu/drm/i915/i915_vma.c
> >> @@ -119,6 +119,12 @@ i915_vma_retire(struct i915_gem_active *base, 
> >> struct i915_request *rq)
> >>       __i915_vma_retire(active->vma, rq);
> >>   }
> >> +static void
> >> +i915_vma_last_retire(struct i915_gem_active *base, struct 
> >> i915_request *rq)
> >> +{
> >> +    __i915_vma_retire(container_of(base, struct i915_vma, 
> >> last_active), rq);
> >> +}
> >> +
> >>   static struct i915_vma *
> >>   vma_create(struct drm_i915_gem_object *obj,
> >>          struct i915_address_space *vm,
> >> @@ -136,6 +142,7 @@ vma_create(struct drm_i915_gem_object *obj,
> >>       vma->active = RB_ROOT;
> >> +    init_request_active(&vma->last_active, i915_vma_last_retire);
> >>       init_request_active(&vma->last_fence, NULL);
> >>       vma->vm = vm;
> >>       vma->ops = &vm->vma_ops;
> >> @@ -895,6 +902,22 @@ static struct i915_gem_active 
> >> *lookup_active(struct i915_vma *vma, u64 idx)
> >>   {
> >>       struct i915_vma_active *active;
> >>       struct rb_node **p, *parent;
> >> +    struct i915_request *old;
> >> +
> >> +    /*
> >> +     * We track the most recently used timeline to skip a rbtree search
> >> +     * for the common case, under typical loads we never need the rbtree
> >> +     * at all. We can reuse the last_active slot if it is empty, that is
> >> +     * after the previous activity has been retired, or if the active
> >> +     * matches the current timeline.
> >> +     */
> >> +    old = i915_gem_active_raw(&vma->last_active,
> >> +                  &vma->vm->i915->drm.struct_mutex);
> >> +    if (!old || old->fence.context == idx)
> >> +        goto out;
> >> +
> >> +    /* Move the currently active fence into the rbtree */
> >> +    idx = old->fence.context;
> >>       parent = NULL;
> >>       p = &vma->active.rb_node;
> >> @@ -903,7 +926,7 @@ static struct i915_gem_active 
> >> *lookup_active(struct i915_vma *vma, u64 idx)
> >>           active = rb_entry(parent, struct i915_vma_active, node);
> >>           if (active->timeline == idx)
> >> -            return &active->base;
> >> +            goto replace;
> >>           if (active->timeline < idx)
> >>               p = &parent->rb_right;
> >> @@ -922,7 +945,25 @@ static struct i915_gem_active 
> >> *lookup_active(struct i915_vma *vma, u64 idx)
> >>       rb_link_node(&active->node, parent, p);
> >>       rb_insert_color(&active->node, &vma->active);
> >> -    return &active->base;
> >> +replace:
> >> +    /*
> >> +     * Overwrite the previous active slot in the rbtree with 
> >> last_active,
> >> +     * leaving last_active zeroed. If the previous slot is still active,
> >> +     * we must be careful as we now only expect to recieve one retire
> > 
> > typo in receive
> > 
> >> +     * callback not two, and so much undo the active counting for the
> >> +     * overwritten slot.
> >> +     */
> >> +    if (i915_gem_active_isset(&active->base)) {
> >> +        __list_del_entry(&active->base.link);
> >> +        vma->active_count--;
> >  > +        GEM_BUG_ON(!vma->active_count);
> > 
> > I still don't get this. The cache is exclusive, so when transferring a 
> > record from rbtree to last_active, why do we need to decrement the 
> > vma->active_count here? Don't get the part in the comment about two 
> > retires - do you really sometimes expect two - ie cache is not exclusive?
> > 
> > But the fact that lookup of a cached entry is a straight return, meaning 
> > vma->active_count is manipulated elsewhere, makes me think it is 
> > avoidable messing with it on this path as well.
> > 
> > Maybe the separation of duties between the callers and this function 
> > needs to be stronger.
> 
> Hmm or is your cache actually inclusive? Don't see no rbtree 
> manipulation on migration to and from last_active/rbtree..

Both. Inclusive in the sense that both last_active and its timeline slot
in the rbtree may be active tracking different requests and so receive
retirement callbacks independently. Exclusive in that we don't store
last_active in the cache slot and in the rbtree.
 
> And since rbtree lookup is always for the last_active context id, you 
> would otherwise never hit the the "goto replace" path.
> 
> How do you ever look up an id which is not cached in last_active then?

We don't. We only lookup on evicting a still active request from
last_active. The MRU recent request always goes into last_active.
-Chris