Quoting Tvrtko Ursulin (2020-12-18 15:52:05)
>
> On 18/12/2020 12:24, Chris Wilson wrote:
> > Since we allow removing the timeline map at runtime, there is a risk
> > that rq->hwsp points into a stale page. To control that risk, we hold
> > the RCU read lock while reading *rq->hwsp, but we missed a couple of
> > important barriers. First, the unpinning / removal of the timeline map
> > must be after all RCU readers into that map are complete, i.e. after an
> > rcu barrier (in this case courtesy of call_rcu()). Secondly, we must
> > make sure that the rq->hwsp we are about to dereference under the RCU
> > lock is valid. In this case, we make the rq->hwsp pointer safe during
> > i915_request_retire() and so we know that rq->hwsp may become invalid
> > only after the request has been signaled. Therefore is the request is
> > not yet signaled when we acquire rq->hwsp under the RCU, we know that
> > rq->hwsp will remain valid for the duration of the RCU read lock.
> >
> > This is a very small window that may lead to either considering the
> > request not completed (causing a delay until the request is checked
> > again, any wait for the request is not affected) or dereferencing an
> > invalid pointer.
> >
> > Fixes: 3adac4689f58 ("drm/i915: Introduce concept of per-timeline (context) HWSP")
> > Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> > Cc: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
> > Cc: <stable@xxxxxxxxxxxxxxx> # v5.1+
> > ---
> > drivers/gpu/drm/i915/gt/intel_breadcrumbs.c | 11 ++----
> > drivers/gpu/drm/i915/gt/intel_timeline.c | 10 +++---
> > drivers/gpu/drm/i915/i915_request.h | 37 ++++++++++++++++++---
> > 3 files changed, 39 insertions(+), 19 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/i915/gt/intel_breadcrumbs.c b/drivers/gpu/drm/i915/gt/intel_breadcrumbs.c
> > index 3c62fd6daa76..f96cd7d9b419 100644
> > --- a/drivers/gpu/drm/i915/gt/intel_breadcrumbs.c
> > +++ b/drivers/gpu/drm/i915/gt/intel_breadcrumbs.c
> > @@ -134,11 +134,6 @@ static bool remove_signaling_context(struct intel_breadcrumbs *b,
> > return true;
> > }
> >
> > -static inline bool __request_completed(const struct i915_request *rq)
> > -{
> > - return i915_seqno_passed(__hwsp_seqno(rq), rq->fence.seqno);
> > -}
> > -
> > __maybe_unused static bool
> > check_signal_order(struct intel_context *ce, struct i915_request *rq)
> > {
> > @@ -245,7 +240,7 @@ static void signal_irq_work(struct irq_work *work)
> > list_for_each_entry_rcu(rq, &ce->signals, signal_link) {
> > bool release;
> >
> > - if (!__request_completed(rq))
> > + if (!__i915_request_is_complete(rq))
> > break;
> >
> > if (!test_and_clear_bit(I915_FENCE_FLAG_SIGNAL,
> > @@ -380,7 +375,7 @@ static void insert_breadcrumb(struct i915_request *rq)
> > * straight onto a signaled list, and queue the irq worker for
> > * its signal completion.
> > */
> > - if (__request_completed(rq)) {
> > + if (__i915_request_is_complete(rq)) {
> > irq_signal_request(rq, b);
> > return;
> > }
> > @@ -468,7 +463,7 @@ void i915_request_cancel_breadcrumb(struct i915_request *rq)
> > if (release)
> > intel_context_put(ce);
> >
> > - if (__request_completed(rq))
> > + if (__i915_request_is_complete(rq))
> > irq_signal_request(rq, b);
> >
> > i915_request_put(rq);
> > diff --git a/drivers/gpu/drm/i915/gt/intel_timeline.c b/drivers/gpu/drm/i915/gt/intel_timeline.c
> > index 512afacd2bdc..a005d0165bf4 100644
> > --- a/drivers/gpu/drm/i915/gt/intel_timeline.c
> > +++ b/drivers/gpu/drm/i915/gt/intel_timeline.c
> > @@ -126,6 +126,10 @@ static void __rcu_cacheline_free(struct rcu_head *rcu)
> > struct intel_timeline_cacheline *cl =
> > container_of(rcu, typeof(*cl), rcu);
> >
> > + /* Must wait until after all *rq->hwsp are complete before removing */
> > + i915_gem_object_unpin_map(cl->hwsp->vma->obj);
> > + __idle_hwsp_free(cl->hwsp, ptr_unmask_bits(cl->vaddr, CACHELINE_BITS));
> > +
> > i915_active_fini(&cl->active);
> > kfree(cl);
> > }
> > @@ -133,11 +137,6 @@ static void __rcu_cacheline_free(struct rcu_head *rcu)
> > static void __idle_cacheline_free(struct intel_timeline_cacheline *cl)
> > {
> > GEM_BUG_ON(!i915_active_is_idle(&cl->active));
> > -
> > - i915_gem_object_unpin_map(cl->hwsp->vma->obj);
> > - i915_vma_put(cl->hwsp->vma);
> > - __idle_hwsp_free(cl->hwsp, ptr_unmask_bits(cl->vaddr, CACHELINE_BITS));
> > -
> > call_rcu(&cl->rcu, __rcu_cacheline_free);
> > }
> >
> > @@ -179,7 +178,6 @@ cacheline_alloc(struct intel_timeline_hwsp *hwsp, unsigned int cacheline)
> > return ERR_CAST(vaddr);
> > }
> >
> > - i915_vma_get(hwsp->vma);
> > cl->hwsp = hwsp;
> > cl->vaddr = page_pack_bits(vaddr, cacheline);
> >
> > diff --git a/drivers/gpu/drm/i915/i915_request.h b/drivers/gpu/drm/i915/i915_request.h
> > index 92e4320c50c4..7c4453e60323 100644
> > --- a/drivers/gpu/drm/i915/i915_request.h
> > +++ b/drivers/gpu/drm/i915/i915_request.h
> > @@ -440,7 +440,7 @@ static inline u32 hwsp_seqno(const struct i915_request *rq)
> >
> > static inline bool __i915_request_has_started(const struct i915_request *rq)
> > {
> > - return i915_seqno_passed(hwsp_seqno(rq), rq->fence.seqno - 1);
> > + return i915_seqno_passed(__hwsp_seqno(rq), rq->fence.seqno - 1);
> > }
> >
> > /**
> > @@ -471,11 +471,19 @@ static inline bool __i915_request_has_started(const struct i915_request *rq)
> > */
> > static inline bool i915_request_started(const struct i915_request *rq)
> > {
> > + bool result;
> > +
> > if (i915_request_signaled(rq))
> > return true;
> >
> > - /* Remember: started but may have since been preempted! */
> > - return __i915_request_has_started(rq);
> > + result = true;
> > + rcu_read_lock(); /* the HWSP may be freed at runtime */
> > + if (likely(!i915_request_signaled(rq)))
> > + /* Remember: started but may have since been preempted! */
> > + result = __i915_request_has_started(rq);
> > + rcu_read_unlock();
> > +
> > + return result;
> > }
> >
> > /**
> > @@ -488,10 +496,16 @@ static inline bool i915_request_started(const struct i915_request *rq)
> > */
> > static inline bool i915_request_is_running(const struct i915_request *rq)
> > {
> > + bool result;
> > +
> > if (!i915_request_is_active(rq))
> > return false;
> >
> > - return __i915_request_has_started(rq);
> > + rcu_read_lock();
> > + result = __i915_request_has_started(rq) && i915_request_is_active(rq);
> > + rcu_read_unlock();
> > +
> > + return result;
> > }
> >
> > /**
> > @@ -515,12 +529,25 @@ static inline bool i915_request_is_ready(const struct i915_request *rq)
> > return !list_empty(&rq->sched.link);
> > }
> >
> > +static inline bool __i915_request_is_complete(const struct i915_request *rq)
> > +{
> > + return i915_seqno_passed(__hwsp_seqno(rq), rq->fence.seqno);
> > +}
> > +
> > static inline bool i915_request_completed(const struct i915_request *rq)
> > {
> > + bool result;
> > +
> > if (i915_request_signaled(rq))
> > return true;
> >
> > - return i915_seqno_passed(hwsp_seqno(rq), rq->fence.seqno);
> > + result = true;
> > + rcu_read_lock(); /* the HWSP may be freed at runtime */
> > + if (likely(!i915_request_signaled(rq)))
> > + result = __i915_request_is_complete(rq);
> > + rcu_read_unlock();
> > +
> > + return result; > }
> >
> > static inline void i915_request_mark_complete(struct i915_request *rq)
> >
>
> Should rq->hwsp_seqno be marked as rcu pointer?
50:50; we're careful enough with the helper, but we don't always call it
from under a rcu read lock. Hmm, that used to be true but now the
interrupt handler is under rcu context. That might not be a huge task to
annotate now, but still a few places that know the request->hwsp is valid
since they are before request submit.
> We reset the fence signaled status before re-assigning the timeline. So
> if we were to query a request in the process of being allocated, do we
> need to do something extra to make sure !signaled status is not seen
> before the rq->hwsp_seqno is replaced? Like should the order of re-init
> be inverted?
You're thinking of request reuse. The question is where do we look at
i915_request_competed outside of an explicit ref. Off the top of my head
that is only gem_busy. (Similar thought exercise for plain
i915_request_is_signaled, just less pointer dancing.) gem_busy is
protected by the seqlock, that is if the dma-resv is changed and the
fence freed, we restart the inspection.
So... I don't think we have a use where the order matters across
recycling, and so nothing to define the rules.
-Chris
