[Intel-gfx] [PATCH] drm/i915/gem: Lazily acquire the device wakeref for freeing objects
Janusz Krzysztofik
janusz.krzysztofik at linux.intel.com
Tue Apr 28 15:30:59 UTC 2020
On Tue, 2020-04-28 at 15:48 +0100, Chris Wilson wrote:
> Quoting Janusz Krzysztofik (2020-04-28 15:35:03)
> > On Tue, 2020-04-28 at 13:59 +0100, Chris Wilson wrote:
> > > 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.
> > > */
> >
> > For me, that still doesn't explain why we are free to request PPGTT
> > lock subclass, not the one dedicated to GGTT as the comment suggests I
> > believe, when locking GGTT vm only because that GGTT vm is closed.
>
> The choice of subclass is not affected by i915_vm_is_closed. It is based
> solely on whether to not the vma was only used as PIN_USER.
That was true before the change, now the subclass argument also depends
on vma->vm->open:
- 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);
err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref);
>
> > > > 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.
> >
> > As I haven't found this explanation sufficiently clear for me, I've had
> > a closer look as my homework. As a result, I think the answer why this
> > may be safe can be found in commits b6422694c585 ("drm/i915/gt: Only
> > wait for register chipset flush if active") and 0d86ee35097a
> > ("drm/i915/gt: Make fence revocation unequivocal").
>
> No. We were always free to ignore discarding the fence if not awake.
> This wakeref is only about the GGTT unbind as we pulled the wakeref from
> inside the ggtt_unbind to here to avoid a lock inversion via the
> vm->mutex + shrinker.
OK, but anyway, looking at those changes, especially the first one, I
can see we find hardware operations we are going to perform as safe
even if the device is not powered on.
Thanks,
Janusz
> -Chris
More information about the Intel-gfx
mailing list