[PATCH hwc v2 17/18] drm_hwcomposer: Flatten scene synchronously

Alexandru-Cosmin Gheorghe Alexandru-Cosmin.Gheorghe at arm.com
Wed Apr 18 11:14:03 UTC 2018


On Tue, Apr 17, 2018 at 01:47:46PM -0400, Sean Paul wrote:
> On Wed, Apr 11, 2018 at 04:22:28PM +0100, Alexandru Gheorghe wrote:
> > Flatten scene on the same CRTC as the one driving the display.
> > The active composition is played back to the display with a buffer
> > attached to the writeback connector.
> > Then we build a composition that has only one plane enabled and that
> > uses the result of the writeback as the input.
> > 
> > Signed-off-by: Alexandru Gheorghe <alexandru-cosmin.gheorghe at arm.com>
> > ---
> >  drmdisplaycompositor.cpp | 203 +++++++++++++++++++++++++++++++++++++++++++++--
> >  drmdisplaycompositor.h   |   7 +-
> >  2 files changed, 204 insertions(+), 6 deletions(-)
> > 
> > diff --git a/drmdisplaycompositor.cpp b/drmdisplaycompositor.cpp
> > index e535e8a..cb670e6 100644
> > --- a/drmdisplaycompositor.cpp
> > +++ b/drmdisplaycompositor.cpp
> > @@ -36,6 +36,7 @@
> >  #include "drmplane.h"
> >  #include "drmresources.h"
> >  #include "glworker.h"
> > +static const uint32_t kWaitWritebackFence = 100;  // ms
> >  
> >  namespace android {
> >  
> > @@ -523,7 +524,9 @@ int DrmDisplayCompositor::PrepareFrame(DrmDisplayComposition *display_comp) {
> >  }
> >  
> >  int DrmDisplayCompositor::CommitFrame(DrmDisplayComposition *display_comp,
> > -                                      bool test_only) {
> > +                                      bool test_only,
> > +                                      DrmDisplayComposition *writeback_comp,
> > +                                      DrmConnector *writeback_conn) {
> >    ATRACE_CALL();
> >  
> >    int ret = 0;
> > @@ -532,6 +535,7 @@ int DrmDisplayCompositor::CommitFrame(DrmDisplayComposition *display_comp,
> >    std::vector<DrmCompositionPlane> &comp_planes =
> >        display_comp->composition_planes();
> >    uint64_t out_fences[drm_->crtcs().size()];
> > +  int writeback_fence = -1;
> >  
> >    DrmConnector *connector = drm_->GetConnectorForDisplay(display_);
> >    if (!connector) {
> > @@ -550,9 +554,37 @@ int DrmDisplayCompositor::CommitFrame(DrmDisplayComposition *display_comp,
> >      return -ENOMEM;
> >    }
> >  
> > +  if (writeback_comp != NULL) {
> > +    if (writeback_conn == NULL)
> > +      return -EINVAL;
> > +    if (writeback_conn->writeback_fb_id().id() == 0 ||
> > +        writeback_conn->writeback_out_fence().id() == 0) {
> > +      ALOGE("Writeback properties don't exit");
> > +      return -EINVAL;
> > +    }
> > +    if (writeback_comp->layers().size() != 1) {
> > +      ALOGE("Invalid number of layers for writeback composition");
> > +      return -EINVAL;
> > +    }
> > +    ret = drmModeAtomicAddProperty(
> > +        pset, writeback_conn->id(), writeback_conn->writeback_fb_id().id(),
> > +        writeback_comp->layers().back().buffer->fb_id);
> > +    if (ret < 0) {
> > +      ALOGE("Failed to add writeback_fb_id");
> > +      return ret;
> > +    }
> > +    ret = drmModeAtomicAddProperty(pset, writeback_conn->id(),
> > +                                   writeback_conn->writeback_out_fence().id(),
> > +                                   (uint64_t)&writeback_fence);
> 
> Upcasting int to u64 isn't a great idea, please go the other way (as we do with
> out_fences).

Genuinely curious about this, why does it make a difference I'm
upcasting the address not the int itself.
More, the kernel does this s32 __user *fence_ptr = u64_to_user_ptr(val);
To be completly fair it should have been int32_t.


> 
> > +    if (ret < 0) {
> > +      ALOGE("Failed to add writeback_out_fence");
> > +      return ret;
> > +    }
> > +  }
> 
> This would be more readable if it was split off into a function.
> 
> >    if (crtc->out_fence_ptr_property().id() != 0) {
> > -    ret = drmModeAtomicAddProperty(pset, crtc->id(), crtc->out_fence_ptr_property().id(),
> > -                                   (uint64_t) &out_fences[crtc->pipe()]);
> > +    ret = drmModeAtomicAddProperty(pset, crtc->id(),
> > +                                   crtc->out_fence_ptr_property().id(),
> > +                                   (uint64_t)&out_fences[crtc->pipe()]);
> >      if (ret < 0) {
> >        ALOGE("Failed to add OUT_FENCE_PTR property to pset: %d", ret);
> >        drmModeAtomicFree(pset);
> > @@ -580,6 +612,15 @@ int DrmDisplayCompositor::CommitFrame(DrmDisplayComposition *display_comp,
> >      }
> >    }
> >  
> > +  if (writeback_conn != NULL) {
> > +    ret = drmModeAtomicAddProperty(pset, writeback_conn->id(),
> > +                                   writeback_conn->crtc_id_property().id(),
> > +                                   crtc->id());
> > +    if (ret < 0) {
> > +      ALOGE("Failed to  attach writeback");
> > +    }
> > +  }
> 
> Can you do this above with the rest of the writeback properties?
> 
> > +
> >    for (DrmCompositionPlane &comp_plane : comp_planes) {
> >      DrmPlane *plane = comp_plane.plane();
> >      DrmCrtc *crtc = comp_plane.crtc();
> > @@ -729,8 +770,18 @@ int DrmDisplayCompositor::CommitFrame(DrmDisplayComposition *display_comp,
> >  
> >    if (!ret) {
> >      uint32_t flags = DRM_MODE_ATOMIC_ALLOW_MODESET;
> > -    if (test_only)
> > +    if (test_only) {
> >        flags |= DRM_MODE_ATOMIC_TEST_ONLY;
> > +    } else {
> > +      if (writeback_comp != NULL) {
> > +        if (!CountdownExpired() && active_composition_) {
> 
> Given that we're holding the lock throughout this function, can't you just
> abort at the start?

Yes, we could/should.

> 
> > +          ALOGE("Writeback composition not needed, abort commit");
> > +          drmModeAtomicFree(pset);
> > +          return -EINVAL;
> > +        };
> > +        flags |= DRM_MODE_ATOMIC_NONBLOCK;
> 
> I'm guessing this is the cause of the active_composition race you're fixing
> earlier in the series? Could you please pull this out and squash it into that
> patch as a "Use non-blocking commits" standalone? It might be useful for
> testing, and this is something that's substantial enough to warrant the
> additional visibility of its own patch.

Not really, that race could happen  even without non blocking. The
current implementation is:
1. Commit
2. lock.
3. Set active_composition
4. Unlock.

With two threads there is no guarantee active_composition will contain
what it's been comitted.


> 
> > +      }
> > +    }
> >  
> >      ret = drmModeAtomicCommit(drm_->fd(), pset, flags, drm_);
> >      if (ret) {
> > @@ -769,6 +820,13 @@ int DrmDisplayCompositor::CommitFrame(DrmDisplayComposition *display_comp,
> >    if (crtc->out_fence_ptr_property().id()) {
> >      display_comp->set_out_fence((int) out_fences[crtc->pipe()]);
> >    }
> > +  if (writeback_fence >= 0) {
> > +    if (writeback_comp->layers().size() != 1) {
> > +      ALOGE("Invalid numbers of layer for writeback_comp");
> > +      return -EINVAL;
> > +    }
> 
> You already test this above?
> 
> > +    writeback_comp->layers()[0].acquire_fence.Set(writeback_fence);
> > +  }
> >  
> >    return ret;
> >  }
> > @@ -837,6 +895,8 @@ void DrmDisplayCompositor::ApplyFrame(
> >        if (active_composition_)
> >          active_composition_->SignalCompositionDone();
> >        active_composition_.swap(composition);
> > +      flatten_countdown_ = FLATTEN_COUNTDOWN_INIT;
> > +      vsync_worker_.VSyncControl(!writeback);
> >      }
> >    }
> >  
> > @@ -913,8 +973,141 @@ int DrmDisplayCompositor::ApplyComposition(
> >    return ret;
> >  }
> >  
> > +int DrmDisplayCompositor::WritebackComposite(DrmDisplayComposition *src,
> > +                                             DrmDisplayComposition *dst,
> > +                                             DrmConnector *writeback_conn) {
> > +  int ret = 0;
> > +  if (src == NULL || dst == NULL)
> > +    return -EINVAL;
> > +  std::vector<DrmCompositionPlane> &src_planes = src->composition_planes();
> > +  DrmCompositionPlane squashed_comp(DrmCompositionPlane::Type::kPrecomp, NULL,
> > +                                    src->crtc());
> > +  for (DrmCompositionPlane &comp_plane : src_planes) {
> > +    if (comp_plane.plane() == NULL) {
> > +      ALOGE("Skipping squash all because of NULL plane");
> > +      ret = -EINVAL;
> > +    }
> > +    if (!squashed_comp.plane() &&
> > +        comp_plane.plane()->type() == DRM_PLANE_TYPE_PRIMARY)
> > +      squashed_comp.set_plane(comp_plane.plane());
> > +    else
> > +      dst->AddPlaneDisable(comp_plane.plane());
> > +  }
> > +
> > +  DrmFramebuffer *writeback_fb = NULL;
> > +  AutoLock lock(&lock_, __FUNCTION__);
> > +  if ((ret = lock.Lock()))
> 
> Same comments regarding assignment in a conditional.
> 
> > +    return ret;
> > +  writeback_fb = &framebuffers_[framebuffer_index_];
> > +  framebuffer_index_ = (framebuffer_index_ + 1) % DRM_DISPLAY_BUFFERS;
> > +  ret = PrepareFramebuffer(*writeback_fb, dst, mode_.mode.h_display(),
> > +                           mode_.mode.v_display());
> > +  if (ret) {
> > +    ALOGE("Failed to prepare destination buffer");
> > +    return ret;
> > +  }
> > +  lock.Unlock();
> > +  ret = CommitFrame(src, true, dst, writeback_conn);
> > +  if (ret) {
> > +    ALOGE("Atomic check failed");
> > +    return ret;
> > +  }
> > +  if ((ret = lock.Lock()))
> 
> All of these locks and unlocks are going to cause races, as mentioned below,
> please re-evaluate.
> 
> > +    return ret;
> > +  if (!CountdownExpired() && active_composition_) {
> > +    ALOGE("Writeback composition not needed abort");
> > +    return -EINVAL;
> > +  }
> > +  ret = CommitFrame(src, false, dst, writeback_conn);
> > +  lock.Unlock();
> > +  if (ret || dst->layers().size() != 1) {
> > +    ALOGE("Failed to flatten scene using writeback");
> > +    return -EINVAL;
> > +  }
> > +  squashed_comp.source_layers().push_back(0);
> > +  ret =
> > +      sync_wait(dst->layers()[0].acquire_fence.Release(), kWaitWritebackFence);
> 
> line break
> 
> > +  if (ret) {
> > +    ALOGE("Failed to wait on writeback fence");
> > +    return ret;
> > +  }
> > +  ret = dst->AddPlaneComposition(std::move(squashed_comp));
> > +  if (ret) {
> > +    ALOGE("Failed to add flatten scene");
> > +    return ret;
> > +  }
> > +  ret = dst->FinalizeComposition();
> > +  if (ret) {
> > +    ALOGE("Failed to finalize composition");
> > +    return ret;
> > +  }
> > +  return 0;
> > +}
> > +
> > +int DrmDisplayCompositor::FlattenSynchronously(DrmConnector *writeback_conn) {
> > +  if (writeback_conn->display() != display_) {
> > +    ALOGE("Cannot flatten synchronously on different display");
> > +    return -EINVAL;
> > +  }
> 
> You check this right before calling, so this isn't needed.
> 
> > +  ALOGI("FlattenSynchronously using the same display");
> 
> I think you should downgrade this log level.
> 
> > +  int ret = 0;
> > +  /* Flattened composition with only one layer that is built
> > +   * using the writeback connector
> > +   */
> > +  std::unique_ptr<DrmDisplayComposition> writeback_comp =
> > +      CreateInitializedComposition();
> > +  /* Copy of the active_composition_, we need a copy because
> > +   * if we use the active composition we have to hold the lock
> > +   * for the entire sequence of flattening.
> > +   */
> > +  std::unique_ptr<DrmDisplayComposition> copy_comp =
> > +      CreateInitializedComposition();
> > +
> > +  if (!copy_comp || !writeback_comp)
> > +    return -EINVAL;
> > +  AutoLock lock(&lock_, __FUNCTION__);
> > +  if ((ret = lock.Lock()))
> 
> Assignments in if statements make me nervous. Since you don't need the
> declaration above, just do:
> 
>   int ret = lock.Lock();
>   if (ret)
> 
> > +    return ret;
> > +  if (CountdownExpired()) {
> > +    ret = copy_comp->CopyLayers(active_composition_.get());
> > +    if (ret)
> > +      return ret;
> > +    copy_comp->CopyCompPlanes(active_composition_.get());
> > +  } else {
> > +    return -EINVAL;
> > +  }
> 
>   if (!CountdownExpired())
>     return -EINVAL;
> 
>   ret = copy_comp->CopyLayers(active_composition_.get());
>   if (ret)
>     return ret;
>   copy_comp->CopyCompPlanes(active_composition_.get());
> 
> 
> > +  lock.Unlock();
> 
> Just hold the lock through WritebackComposite(), and consider creating a
> ApplyFrameLocked() to hold it through ApplyFrame(). You should be able to reduce
> the number of times you have to check CountdownExpired() (or remove it)
> significantly by just consistently locking things.
> 
> > +  ret =
> > +      WritebackComposite(copy_comp.get(), writeback_comp.get(), writeback_conn);
> > +  if (ret) {
> > +    ALOGE("Failed to prepare writebackScene");
> > +    return ret;
> > +  }
> > +
> > +  ApplyFrame(std::move(writeback_comp), 0, true);
> > +  return 0;
> > +}
> > +
> >  int DrmDisplayCompositor::FlattenScene() {
> > -  return -EINVAL;
> > +  DrmConnector *writeback_conn =
> > +      drm_->resource_manager()->AvailableWritebackConnector(display_);
> > +  if (!active_composition_ || !writeback_conn)
> > +    return -EINVAL;
> > +  std::vector<DrmCompositionPlane> &src_planes =
> > +      active_composition_->composition_planes();
> > +  size_t src_planes_with_layer = std::count_if(
> > +      src_planes.begin(), src_planes.end(), [](DrmCompositionPlane &p) {
> > +        return p.type() != DrmCompositionPlane::Type::kDisable;
> > +      });
> > +
> > +  if (src_planes_with_layer <= 1)
> > +    return -EALREADY;
> > +
> > +  if (writeback_conn->display() == display_) {
> > +    return FlattenSynchronously(writeback_conn);
> > +  }
> > +
> > +  return 0;
> >  }
> >  
> >  int DrmDisplayCompositor::SquashAll() {
> > diff --git a/drmdisplaycompositor.h b/drmdisplaycompositor.h
> > index 26201b9..4cc4a5e 100644
> > --- a/drmdisplaycompositor.h
> > +++ b/drmdisplaycompositor.h
> > @@ -125,7 +125,9 @@ class DrmDisplayCompositor {
> >    int ApplySquash(DrmDisplayComposition *display_comp);
> >    int ApplyPreComposite(DrmDisplayComposition *display_comp);
> >    int PrepareFrame(DrmDisplayComposition *display_comp);
> > -  int CommitFrame(DrmDisplayComposition *display_comp, bool test_only);
> > +  int CommitFrame(DrmDisplayComposition *display_comp, bool test_only,
> > +                  DrmDisplayComposition *writeback_comp = NULL,
> > +                  DrmConnector *writeback_conn = NULL);
> >    int SquashFrame(DrmDisplayComposition *src, DrmDisplayComposition *dst);
> >    int ApplyDpms(DrmDisplayComposition *display_comp);
> >    int DisablePlanes(DrmDisplayComposition *display_comp);
> > @@ -134,7 +136,10 @@ class DrmDisplayCompositor {
> >    void ApplyFrame(std::unique_ptr<DrmDisplayComposition> composition,
> >                    int status, bool writeback = false);
> >    int FlattenScene();
> > +  int FlattenSynchronously(DrmConnector *writeback_conn);
> 
> Can we come up with a better name than Synchonrous/Asynchronous? Or at the very
> least comment on what that means?
> 
> >  
> > +  int WritebackComposite(DrmDisplayComposition *src, DrmDisplayComposition *dst,
> > +                         DrmConnector *writeback_conn);
> >    bool CountdownExpired() const;
> >  
> >    std::tuple<int, uint32_t> CreateModeBlob(const DrmMode &mode);
> > -- 
> > 2.7.4
> > 
> 
> -- 
> Sean Paul, Software Engineer, Google / Chromium OS

-- 
Cheers,
Alex G


More information about the dri-devel mailing list