Re: [Intel-gfx] [PATCH v3 07/17] drm/i915/vm_bind: Add support to handle object evictions

[Matthew Auld <matthew.auld@intel.com>]

On 10/10/2022 17:11, Niranjana Vishwanathapura wrote:
> On Mon, Oct 10, 2022 at 02:30:49PM +0100, Matthew Auld wrote:
>> On 10/10/2022 07:58, Niranjana Vishwanathapura wrote:
>>> Support eviction by maintaining a list of evicted persistent vmas
>>> for rebinding during next submission. Ensure the list do not
>>> include persistent vmas that are being purged.
>>>
>>> v2: Remove unused I915_VMA_PURGED definition.
>>> v3: Properly handle __i915_vma_unbind_async() case.
>>>
>>> Acked-by: Matthew Auld <matthew.auld@intel.com>
>>> Signed-off-by: Niranjana Vishwanathapura 
>>> <niranjana.vishwanathapura@intel.com>
>>> Signed-off-by: Andi Shyti <andi.shyti@linux.intel.com>
>>> ---
>>>  .../drm/i915/gem/i915_gem_vm_bind_object.c    |  6 ++++
>>>  drivers/gpu/drm/i915/gt/intel_gtt.c           |  2 ++
>>>  drivers/gpu/drm/i915/gt/intel_gtt.h           |  4 +++
>>>  drivers/gpu/drm/i915/i915_vma.c               | 31 +++++++++++++++++--
>>>  drivers/gpu/drm/i915/i915_vma.h               | 10 ++++++
>>>  drivers/gpu/drm/i915/i915_vma_types.h         |  8 +++++
>>>  6 files changed, 59 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_vm_bind_object.c 
>>> b/drivers/gpu/drm/i915/gem/i915_gem_vm_bind_object.c
>>> index 8e3e6ceb9442..c435d49af2c8 100644
>>> --- a/drivers/gpu/drm/i915/gem/i915_gem_vm_bind_object.c
>>> +++ b/drivers/gpu/drm/i915/gem/i915_gem_vm_bind_object.c
>>> @@ -85,6 +85,12 @@ static void i915_gem_vm_bind_remove(struct 
>>> i915_vma *vma, bool release_obj)
>>>  {
>>>      lockdep_assert_held(&vma->vm->vm_bind_lock);
>>> +    spin_lock(&vma->vm->vm_rebind_lock);
>>> +    if (!list_empty(&vma->vm_rebind_link))
>>> +        list_del_init(&vma->vm_rebind_link);
>>> +    i915_vma_set_purged(vma);
>>> +    spin_unlock(&vma->vm->vm_rebind_lock);
>>> +
>>>      list_del_init(&vma->vm_bind_link);
>>>      list_del_init(&vma->non_priv_vm_bind_link);
>>>      i915_vm_bind_it_remove(vma, &vma->vm->va);
>>> diff --git a/drivers/gpu/drm/i915/gt/intel_gtt.c 
>>> b/drivers/gpu/drm/i915/gt/intel_gtt.c
>>> index 422394f8fb40..2fa37f46750b 100644
>>> --- a/drivers/gpu/drm/i915/gt/intel_gtt.c
>>> +++ b/drivers/gpu/drm/i915/gt/intel_gtt.c
>>> @@ -295,6 +295,8 @@ void i915_address_space_init(struct 
>>> i915_address_space *vm, int subclass)
>>>      INIT_LIST_HEAD(&vm->vm_bound_list);
>>>      mutex_init(&vm->vm_bind_lock);
>>>      INIT_LIST_HEAD(&vm->non_priv_vm_bind_list);
>>> +    INIT_LIST_HEAD(&vm->vm_rebind_list);
>>> +    spin_lock_init(&vm->vm_rebind_lock);
>>>  }
>>>  void *__px_vaddr(struct drm_i915_gem_object *p)
>>> diff --git a/drivers/gpu/drm/i915/gt/intel_gtt.h 
>>> b/drivers/gpu/drm/i915/gt/intel_gtt.h
>>> index 4ae5734f7d6b..443d1918ad4e 100644
>>> --- a/drivers/gpu/drm/i915/gt/intel_gtt.h
>>> +++ b/drivers/gpu/drm/i915/gt/intel_gtt.h
>>> @@ -265,6 +265,10 @@ struct i915_address_space {
>>>      struct list_head vm_bind_list;
>>>      /** @vm_bound_list: List of vm_binding completed */
>>>      struct list_head vm_bound_list;
>>> +    /* @vm_rebind_list: list of vmas to be rebinded */
>>> +    struct list_head vm_rebind_list;
>>> +    /* @vm_rebind_lock: protects vm_rebound_list */
>>> +    spinlock_t vm_rebind_lock;
>>>      /* @va: tree of persistent vmas */
>>>      struct rb_root_cached va;
>>>      struct list_head non_priv_vm_bind_list;
>>> diff --git a/drivers/gpu/drm/i915/i915_vma.c 
>>> b/drivers/gpu/drm/i915/i915_vma.c
>>> index 5d3d67a4bf47..b4be2cbe8382 100644
>>> --- a/drivers/gpu/drm/i915/i915_vma.c
>>> +++ b/drivers/gpu/drm/i915/i915_vma.c
>>> @@ -241,6 +241,7 @@ vma_create(struct drm_i915_gem_object *obj,
>>>      INIT_LIST_HEAD(&vma->vm_bind_link);
>>>      INIT_LIST_HEAD(&vma->non_priv_vm_bind_link);
>>> +    INIT_LIST_HEAD(&vma->vm_rebind_link);
>>>      return vma;
>>>  err_unlock:
>>> @@ -1686,6 +1687,14 @@ static void force_unbind(struct i915_vma *vma)
>>>      if (!drm_mm_node_allocated(&vma->node))
>>>          return;
>>> +    /*
>>> +     * Persistent vma should have been purged by now.
>>> +     * If not, issue a warning and purge it.
>>> +     */
>>> +    if (GEM_WARN_ON(i915_vma_is_persistent(vma) &&
>>> +            !i915_vma_is_purged(vma)))
>>> +        i915_vma_set_purged(vma);
>>> +
>>>      atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
>>>      WARN_ON(__i915_vma_unbind(vma));
>>>      GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
>>> @@ -2047,6 +2056,16 @@ int __i915_vma_unbind(struct i915_vma *vma)
>>>      __i915_vma_evict(vma, false);
>>>      drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
>>> +
>>> +    if (i915_vma_is_persistent(vma)) {
>>> +        spin_lock(&vma->vm->vm_rebind_lock);
>>> +        if (list_empty(&vma->vm_rebind_link) &&
>>> +            !i915_vma_is_purged(vma))
>>> +            list_add_tail(&vma->vm_rebind_link,
>>> +                      &vma->vm->vm_rebind_list);
>>> +        spin_unlock(&vma->vm->vm_rebind_lock);
>>> +    }
>>> +
>>>      return 0;
>>>  }
>>> @@ -2059,8 +2078,7 @@ static struct dma_fence 
>>> *__i915_vma_unbind_async(struct i915_vma *vma)
>>>      if (!drm_mm_node_allocated(&vma->node))
>>>          return NULL;
>>> -    if (i915_vma_is_pinned(vma) ||
>>> -        &vma->obj->mm.rsgt->table != vma->resource->bi.pages)
>>
>> Hmm that's looks interesting. IIRC we only keep a ref on the rsgt for 
>> the object pages, and not the vma->bi.pages, where the vma pages can 
>> be destroyed before the async unbind completes, which I guess was the 
>> idea behind this check.
>>
>> But in practice it looks the vma->bi.pages are always just some subset 
>> or rearrangement of the objects rsgt pages, if not the same table, so 
>> the device mapping pointed at by the PTEs should still be valid here 
>> (assuming rsgt in not NULL), even if bi.pages gets nuked? I guess this 
>> change should rather be a patch by itself, with proper explanation in 
>> commit message, since this looks mostly orthogonal?
>>
> 
> Yah, I am not sure about the intent of this check. It is expecting the
> vma->resource->bi.pages to just point to the sg table of the object
> (vma->obj->mm.rsgt->table) which is reference counted, instead of
> decoupling it as you mentioned above. Also, the return code -EAGAIN
> is bit confusing to me as I am not sure how trying again will fix it.
> 
> This check was preventing eviction of objects with persistent (vm_bind)
> vmas (Hence the update is in this patch).
> Persistent vmas have I915_GTT_VIEW_PARTIAL, so they will get their sg table
> by calling intel_partial_pages() which creates a new sg table instead of
> pointing to object's sg table (as done in the I915_GTT_VIEW_NORMAL case).]
> 
> If the vma is removed before async unbind completes, we probably
> should wait for async unbind to complete before releaseing the vma
> pages? Other option is to have vma point to object's sg table even for
> partial gtt_view (instead of creating a new sg table) and handle
> partial binding during the page table update.

Yeah, it's a new sg_table, but nothing is actually remmapped in there it 
seems, so all the device addresses must map to something already in 
obj->mm.rsgt->table (which is ref counted), so all the PTEs will still 
be valid, even if the vma is nuked before the unbind actually completes, 
AFAICT.

> 
> I can also keep the above removed check and only don't check it for
> persistent vmas.
> 
> Any thoughts?
> 
> Niranjana
> 
>>> +    if (i915_vma_is_pinned(vma))
>>>          return ERR_PTR(-EAGAIN);
>>>      /*
>>> @@ -2082,6 +2100,15 @@ static struct dma_fence 
>>> *__i915_vma_unbind_async(struct i915_vma *vma)
>>>      drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
>>> +    if (i915_vma_is_persistent(vma)) {
>>> +        spin_lock(&vma->vm->vm_rebind_lock);
>>> +        if (list_empty(&vma->vm_rebind_link) &&
>>> +            !i915_vma_is_purged(vma))
>>> +            list_add_tail(&vma->vm_rebind_link,
>>> +                      &vma->vm->vm_rebind_list);
>>> +        spin_unlock(&vma->vm->vm_rebind_lock);
>>> +    }
>>> +
>>>      return fence;
>>>  }
>>> diff --git a/drivers/gpu/drm/i915/i915_vma.h 
>>> b/drivers/gpu/drm/i915/i915_vma.h
>>> index c5378ec2f70a..9a4a7a8dfe5b 100644
>>> --- a/drivers/gpu/drm/i915/i915_vma.h
>>> +++ b/drivers/gpu/drm/i915/i915_vma.h
>>> @@ -152,6 +152,16 @@ static inline void 
>>> i915_vma_set_persistent(struct i915_vma *vma)
>>>      set_bit(I915_VMA_PERSISTENT_BIT, __i915_vma_flags(vma));
>>>  }
>>> +static inline bool i915_vma_is_purged(const struct i915_vma *vma)
>>> +{
>>> +    return test_bit(I915_VMA_PURGED_BIT, __i915_vma_flags(vma));
>>> +}
>>> +
>>> +static inline void i915_vma_set_purged(struct i915_vma *vma)
>>> +{
>>> +    set_bit(I915_VMA_PURGED_BIT, __i915_vma_flags(vma));
>>> +}
>>> +
>>>  static inline struct i915_vma *i915_vma_get(struct i915_vma *vma)
>>>  {
>>>      i915_gem_object_get(vma->obj);
>>> diff --git a/drivers/gpu/drm/i915/i915_vma_types.h 
>>> b/drivers/gpu/drm/i915/i915_vma_types.h
>>> index b8176cca58c0..d32c72e8d242 100644
>>> --- a/drivers/gpu/drm/i915/i915_vma_types.h
>>> +++ b/drivers/gpu/drm/i915/i915_vma_types.h
>>> @@ -267,8 +267,14 @@ struct i915_vma {
>>>  /**
>>>   * I915_VMA_PERSISTENT_BIT:
>>>   * The vma is persistent (created with VM_BIND call).
>>> + *
>>> + * I915_VMA_PURGED_BIT:
>>> + * The persistent vma is force unbound either due to VM_UNBIND call
>>> + * from UMD or VM is released. Do not check/wait for VM activeness
>>> + * in i915_vma_is_active() and i915_vma_sync() calls.
>>>   */
>>>  #define I915_VMA_PERSISTENT_BIT    19
>>> +#define I915_VMA_PURGED_BIT    20
>>>      struct i915_active active;
>>> @@ -299,6 +305,8 @@ struct i915_vma {
>>>      struct list_head vm_bind_link;
>>>      /* @non_priv_vm_bind_link: Link in non-private persistent VMA 
>>> list */
>>>      struct list_head non_priv_vm_bind_link;
>>> +    /* @vm_rebind_link: link to vm_rebind_list and protected by 
>>> vm_rebind_lock */
>>> +    struct list_head vm_rebind_link; /* Link in vm_rebind_list */
>>>      /** Interval tree structures for persistent vma */