[Intel-gfx] [PATCH] drm/i915/gem: Lazily acquire the device wakeref for freeing objects

[Chris Wilson]

Quoting Janusz Krzysztofik (2020-04-28 13:45:13)
> Hi Chris,
> 
> On Fri, 2020-04-24 at 16:24 +0100, Chris Wilson wrote:
> > We only need the device wakeref on freeing the objects if we have to
> > unbind the object from the global GTT, or otherwise update device
> > information. If the objects are clean, we never need the wakeref, so
> > avoid taking until required.
> > 
> > For this to be effective in preventing us from waking the device after
> > it is unbind, we also need to mark the GGTT as closed on device removal.
> > The GGTT will be rebuilt from scratch the next time we need to open it
> > (on binding a new device).
> > 
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > Cc: Janusz Krzysztofik <janusz.krzysztofik at linux.intel.com>
> > ---
> >  drivers/gpu/drm/i915/gem/i915_gem_mman.c   | 3 +++
> >  drivers/gpu/drm/i915/gem/i915_gem_object.c | 3 ---
> >  drivers/gpu/drm/i915/gt/intel_ggtt.c       | 9 +++++++--
> >  drivers/gpu/drm/i915/gt/intel_gtt.h        | 1 +
> >  drivers/gpu/drm/i915/i915_drv.c            | 1 +
> >  drivers/gpu/drm/i915/i915_vma.c            | 3 ++-
> >  6 files changed, 14 insertions(+), 6 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_mman.c b/drivers/gpu/drm/i915/gem/i915_gem_mman.c
> > index b39c24dae64e..c6cead6f2b3e 100644
> > --- a/drivers/gpu/drm/i915/gem/i915_gem_mman.c
> > +++ b/drivers/gpu/drm/i915/gem/i915_gem_mman.c
> > @@ -421,6 +421,9 @@ static void i915_gem_object_release_mmap_gtt(struct drm_i915_gem_object *obj)
> >       struct drm_i915_private *i915 = to_i915(obj->base.dev);
> >       intel_wakeref_t wakeref;
> >  
> > +     if (!atomic_read(&i915->ggtt.vm.open))
> > +             return;
> > +
> >       /*
> >        * Serialisation between user GTT access and our code depends upon
> >        * revoking the CPU's PTE whilst the mutex is held. The next user
> > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.c b/drivers/gpu/drm/i915/gem/i915_gem_object.c
> > index 9d1d0131f7c2..99356c00c19e 100644
> > --- a/drivers/gpu/drm/i915/gem/i915_gem_object.c
> > +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.c
> > @@ -162,9 +162,7 @@ static void __i915_gem_free_objects(struct drm_i915_private *i915,
> >                                   struct llist_node *freed)
> >  {
> >       struct drm_i915_gem_object *obj, *on;
> > -     intel_wakeref_t wakeref;
> >  
> > -     wakeref = intel_runtime_pm_get(&i915->runtime_pm);
> >       llist_for_each_entry_safe(obj, on, freed, freed) {
> >               struct i915_mmap_offset *mmo, *mn;
> >  
> > @@ -224,7 +222,6 @@ static void __i915_gem_free_objects(struct drm_i915_private *i915,
> >               call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
> >               cond_resched();
> >       }
> > -     intel_runtime_pm_put(&i915->runtime_pm, wakeref);
> >  }
> >  
> >  void i915_gem_flush_free_objects(struct drm_i915_private *i915)
> > diff --git a/drivers/gpu/drm/i915/gt/intel_ggtt.c b/drivers/gpu/drm/i915/gt/intel_ggtt.c
> > index 66165b10256e..b65545182ef5 100644
> > --- a/drivers/gpu/drm/i915/gt/intel_ggtt.c
> > +++ b/drivers/gpu/drm/i915/gt/intel_ggtt.c
> > @@ -681,8 +681,6 @@ static void ggtt_cleanup_hw(struct i915_ggtt *ggtt)
> >  {
> >       struct i915_vma *vma, *vn;
> >  
> > -     atomic_set(&ggtt->vm.open, 0);
> > -
> >       rcu_barrier(); /* flush the RCU'ed__i915_vm_release */
> >       flush_workqueue(ggtt->vm.i915->wq);
> >  
> > @@ -709,6 +707,13 @@ static void ggtt_cleanup_hw(struct i915_ggtt *ggtt)
> >               io_mapping_fini(&ggtt->iomap);
> >  }
> >  
> > +void i915_ggtt_driver_remove(struct drm_i915_private *i915)
> > +{
> > +     struct i915_ggtt *ggtt = &i915->ggtt;
> > +
> > +     atomic_set(&ggtt->vm.open, 0);
> > +}
> > +
> >  /**
> >   * i915_ggtt_driver_release - Clean up GGTT hardware initialization
> >   * @i915: i915 device
> > diff --git a/drivers/gpu/drm/i915/gt/intel_gtt.h b/drivers/gpu/drm/i915/gt/intel_gtt.h
> > index d93ebdf3fa0e..f140ce5c171a 100644
> > --- a/drivers/gpu/drm/i915/gt/intel_gtt.h
> > +++ b/drivers/gpu/drm/i915/gt/intel_gtt.h
> > @@ -501,6 +501,7 @@ int i915_ggtt_enable_hw(struct drm_i915_private *i915);
> >  void i915_ggtt_enable_guc(struct i915_ggtt *ggtt);
> >  void i915_ggtt_disable_guc(struct i915_ggtt *ggtt);
> >  int i915_init_ggtt(struct drm_i915_private *i915);
> > +void i915_ggtt_driver_remove(struct drm_i915_private *i915);
> >  void i915_ggtt_driver_release(struct drm_i915_private *i915);
> >  
> >  static inline bool i915_ggtt_has_aperture(const struct i915_ggtt *ggtt)
> > diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
> > index 2d62efd9316f..bdf97a1cb7cc 100644
> > --- a/drivers/gpu/drm/i915/i915_drv.c
> > +++ b/drivers/gpu/drm/i915/i915_drv.c
> > @@ -768,6 +768,7 @@ static void i915_driver_hw_remove(struct drm_i915_private *dev_priv)
> >  
> >       i915_perf_fini(dev_priv);
> >  
> > +     i915_ggtt_driver_remove(dev_priv);
> >       if (pdev->msi_enabled)
> >               pci_disable_msi(pdev);
> >  
> > diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c
> > index fc14ebf9a0b7..6fe56ad2a542 100644
> > --- a/drivers/gpu/drm/i915/i915_vma.c
> > +++ b/drivers/gpu/drm/i915/i915_vma.c
> > @@ -1319,7 +1319,8 @@ int i915_vma_unbind(struct i915_vma *vma)
> >               return -EAGAIN;
> >       }
> >  
> > -     if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
> > +     if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND) &&
> > +         atomic_read(&vma->vm->open))
> >               /* XXX not always required: nop_clear_range */
> >               wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
> >  
> 
> Can you please explain why you think it is OK to call
> mutex_lock_interruptible_nested(&closed_GGTT->mutex, PPGTT_subclass)?

That should be explained in the comments

       /*
        * Differentiate between user/kernel vma inside the aliasing-ppgtt.
        *
        * We conflate the Global GTT with the user's vma when using the
        * aliasing-ppgtt, but it is still vitally important to try and
        * keep the use cases distinct. For example, userptr objects are
        * not allowed inside the Global GTT as that will cause lock
        * inversions when we have to evict them the mmu_notifier callbacks -
        * but they are allowed to be part of the user ppGTT which can never
        * be mapped. As such we try to give the distinct users of the same
        * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt
        * and i915_ppgtt separate].
        *
        * NB this may cause us to mask real lock inversions -- while the
        * code is safe today, lockdep may not be able to spot future
        * transgressions.
        */

> Can you please also explain why you think it is safe to call
> __i915_vma_unbind(vma_bound_to_GGTT) even if wakeref is not taken?

Because it has been closed, and only inside strictly serial code.
-Chris