[RFC 05/10] drm/i915/vm_bind: Handle persistent vmas

Hellstrom, Thomas thomas.hellstrom at intel.com
Thu Jul 7 11:27:23 UTC 2022


On Fri, 2022-07-01 at 15:50 -0700, Niranjana Vishwanathapura wrote:
> Treat VM_BIND vmas as persistent and handle them during the
> request submission in the execbuff path.
> 
> Support eviction by maintaining a list of evicted persistent vmas
> for rebinding during next submission.
> 
> Signed-off-by: Niranjana Vishwanathapura
> <niranjana.vishwanathapura at intel.com>
> ---
>  drivers/gpu/drm/i915/gem/i915_gem_object.c    |  1 +
>  drivers/gpu/drm/i915/gem/i915_gem_vm_bind.h   |  3 +
>  .../drm/i915/gem/i915_gem_vm_bind_object.c    | 12 ++-
>  drivers/gpu/drm/i915/gt/intel_gtt.c           |  2 +
>  drivers/gpu/drm/i915/gt/intel_gtt.h           |  2 +
>  drivers/gpu/drm/i915/i915_gem_gtt.h           | 22 ++++++
>  drivers/gpu/drm/i915/i915_vma.c               | 32 +++++++-
>  drivers/gpu/drm/i915/i915_vma.h               | 78
> +++++++++++++++++--
>  drivers/gpu/drm/i915/i915_vma_types.h         | 23 ++++++
>  9 files changed, 163 insertions(+), 12 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.c
> b/drivers/gpu/drm/i915/gem/i915_gem_object.c
> index ccec4055fde3..5121f02ba95c 100644
> --- a/drivers/gpu/drm/i915/gem/i915_gem_object.c
> +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.c
> @@ -38,6 +38,7 @@
>  #include "i915_gem_mman.h"
>  #include "i915_gem_object.h"
>  #include "i915_gem_ttm.h"
> +#include "i915_gem_vm_bind.h"
>  #include "i915_memcpy.h"
>  #include "i915_trace.h"
>  
> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_vm_bind.h
> b/drivers/gpu/drm/i915/gem/i915_gem_vm_bind.h
> index 849bf3c1061e..eaadf5a6ab09 100644
> --- a/drivers/gpu/drm/i915/gem/i915_gem_vm_bind.h
> +++ b/drivers/gpu/drm/i915/gem/i915_gem_vm_bind.h
> @@ -6,6 +6,7 @@
>  #ifndef __I915_GEM_VM_BIND_H
>  #define __I915_GEM_VM_BIND_H
>  
> +#include <linux/dma-resv.h>
>  #include "i915_drv.h"
>  
>  #define assert_vm_bind_held(vm)   lockdep_assert_held(&(vm)-
> >vm_bind_lock)
> @@ -26,6 +27,8 @@ static inline void i915_gem_vm_bind_unlock(struct
> i915_address_space *vm)
>         mutex_unlock(&vm->vm_bind_lock);
>  }
>  
> +#define assert_vm_priv_held(vm)   assert_object_held((vm)->root_obj)
> +
>  static inline int i915_gem_vm_priv_lock(struct i915_address_space
> *vm,
>                                         struct i915_gem_ww_ctx *ww)
>  {
> 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 96f139cc8060..1a8efa83547f 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 +85,13 @@ void i915_gem_vm_bind_remove(struct i915_vma *vma,
> bool release_obj)
>  {
>         assert_vm_bind_held(vma->vm);
>  
> +       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);
> +       i915_vma_set_freed(vma);

Following the vma destruction discussion, can we ditch the freed flag
now?

> +       spin_unlock(&vma->vm->vm_rebind_lock);
> +
>         if (!list_empty(&vma->vm_bind_link)) {
>                 list_del_init(&vma->vm_bind_link);
>                 list_del_init(&vma->non_priv_vm_bind_link);
> @@ -220,6 +227,7 @@ static struct i915_vma *vm_bind_get_vma(struct
> i915_address_space *vm,
>  
>         vma->start = va->start;
>         vma->last = va->start + va->length - 1;
> +       i915_vma_set_persistent(vma);
>  
>         return vma;
>  }
> @@ -304,8 +312,10 @@ int i915_gem_vm_bind_obj(struct
> i915_address_space *vm,
>  
>         i915_vm_bind_put_fence(vma);
>  put_vma:
> -       if (ret)
> +       if (ret) {
> +               i915_vma_set_freed(vma);
>                 i915_vma_destroy(vma);
> +       }
>  
>         i915_gem_ww_ctx_fini(&ww);
>  unlock_vm:
> diff --git a/drivers/gpu/drm/i915/gt/intel_gtt.c
> b/drivers/gpu/drm/i915/gt/intel_gtt.c
> index df0a8459c3c6..55d5389b2c6c 100644
> --- a/drivers/gpu/drm/i915/gt/intel_gtt.c
> +++ b/drivers/gpu/drm/i915/gt/intel_gtt.c
> @@ -293,6 +293,8 @@ void i915_address_space_init(struct
> i915_address_space *vm, int subclass)
>         INIT_LIST_HEAD(&vm->non_priv_vm_bind_list);
>         vm->root_obj = i915_gem_object_create_internal(vm->i915,
> PAGE_SIZE);
>         GEM_BUG_ON(IS_ERR(vm->root_obj));
> +       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 f538ce9115c9..fe5485c4a1cd 100644
> --- a/drivers/gpu/drm/i915/gt/intel_gtt.h
> +++ b/drivers/gpu/drm/i915/gt/intel_gtt.h
> @@ -265,6 +265,8 @@ struct i915_address_space {
>         struct mutex vm_bind_lock;  /* Protects vm_bind lists */
>         struct list_head vm_bind_list;
>         struct list_head vm_bound_list;
> +       struct list_head vm_rebind_list;
> +       spinlock_t vm_rebind_lock;   /* Protects vm_rebind_list */
>         /* 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_gem_gtt.h
> b/drivers/gpu/drm/i915/i915_gem_gtt.h
> index 8c2f57eb5dda..09b89d1913fc 100644
> --- a/drivers/gpu/drm/i915/i915_gem_gtt.h
> +++ b/drivers/gpu/drm/i915/i915_gem_gtt.h
> @@ -51,4 +51,26 @@ int i915_gem_gtt_insert(struct i915_address_space
> *vm,
>  
>  #define PIN_OFFSET_MASK                I915_GTT_PAGE_MASK
>  
> +static inline int i915_vm_sync(struct i915_address_space *vm)

No need to make this an inline function. Only for performance reasons.
Kerneldoc.

> +{
> +       int ret;
> +
> +       /* Wait for all requests under this vm to finish */
> +       ret = dma_resv_wait_timeout(vm->root_obj->base.resv,
> +                                   DMA_RESV_USAGE_BOOKKEEP, false,
> +                                   MAX_SCHEDULE_TIMEOUT);
> +       if (ret < 0)
> +               return ret;
> +       else if (ret > 0)
> +               return 0;
> +       else
> +               return -ETIMEDOUT;
> +}
> +
> +static inline bool i915_vm_is_active(const struct i915_address_space
> *vm)
> +{
> +       return !dma_resv_test_signaled(vm->root_obj->base.resv,
> +                                      DMA_RESV_USAGE_BOOKKEEP);
> +}
> +
>  #endif
> diff --git a/drivers/gpu/drm/i915/i915_vma.c
> b/drivers/gpu/drm/i915/i915_vma.c
> index 6737236b7884..6adb013579be 100644
> --- a/drivers/gpu/drm/i915/i915_vma.c
> +++ b/drivers/gpu/drm/i915/i915_vma.c
> @@ -237,6 +237,7 @@ 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:
> @@ -1622,7 +1623,8 @@ void i915_vma_close(struct i915_vma *vma)
>         if (atomic_dec_and_lock_irqsave(&vma->open_count,
>                                         &gt->closed_lock,
>                                         flags)) {
> -               __vma_close(vma, gt);
> +               if (!i915_vma_is_persistent(vma))
> +                       __vma_close(vma, gt);

This change is not needed since persistent vmas shouldn't take part in
the other vma open-close life-time management.

>                 spin_unlock_irqrestore(&gt->closed_lock, flags);
>         }
>  }
> @@ -1647,6 +1649,13 @@ static void force_unbind(struct i915_vma *vma)
>         if (!drm_mm_node_allocated(&vma->node))
>                 return;
>  
> +       /*
> +        * Mark persistent vma as purged to avoid it waiting
> +        * for VM to be released.
> +        */
> +       if (i915_vma_is_persistent(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));
> @@ -1666,9 +1675,12 @@ static void release_references(struct i915_vma
> *vma, bool vm_ddestroy)
>  
>         spin_unlock(&obj->vma.lock);
>  
> -       i915_gem_vm_bind_lock(vma->vm);
> -       i915_gem_vm_bind_remove(vma, true);
> -       i915_gem_vm_bind_unlock(vma->vm);
> +       if (i915_vma_is_persistent(vma) &&
> +           !i915_vma_is_freed(vma)) {
> +               i915_gem_vm_bind_lock(vma->vm);
> +               i915_gem_vm_bind_remove(vma, true);
> +               i915_gem_vm_bind_unlock(vma->vm);
> +       }
>  
>         spin_lock_irq(&gt->closed_lock);
>         __i915_vma_remove_closed(vma);
> @@ -1839,6 +1851,8 @@ int _i915_vma_move_to_active(struct i915_vma
> *vma,
>         int err;
>  
>         assert_object_held(obj);
> +       if (i915_vma_is_persistent(vma))
> +               return -EINVAL;
>  
>         GEM_BUG_ON(!vma->pages);
>  
> @@ -1999,6 +2013,16 @@ 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;
>  }
>  
> diff --git a/drivers/gpu/drm/i915/i915_vma.h
> b/drivers/gpu/drm/i915/i915_vma.h
> index dcb49f79ff7e..6c1369a40e03 100644
> --- a/drivers/gpu/drm/i915/i915_vma.h
> +++ b/drivers/gpu/drm/i915/i915_vma.h
> @@ -47,12 +47,6 @@ i915_vma_instance(struct drm_i915_gem_object *obj,
>  
>  void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned
> int flags);
>  #define I915_VMA_RELEASE_MAP BIT(0)
> -
> -static inline bool i915_vma_is_active(const struct i915_vma *vma)
> -{
> -       return !i915_active_is_idle(&vma->active);
> -}
> -
>  /* do not reserve memory to prevent deadlocks */
>  #define __EXEC_OBJECT_NO_RESERVE BIT(31)
>  
> @@ -138,6 +132,48 @@ static inline u32 i915_ggtt_pin_bias(struct
> i915_vma *vma)
>         return i915_vm_to_ggtt(vma->vm)->pin_bias;
>  }
>  
> +static inline bool i915_vma_is_persistent(const struct i915_vma
> *vma)
> +{
> +       return test_bit(I915_VMA_PERSISTENT_BIT,
> __i915_vma_flags(vma));
> +}
> +
> +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 bool i915_vma_is_freed(const struct i915_vma *vma)
> +{
> +       return test_bit(I915_VMA_FREED_BIT, __i915_vma_flags(vma));
> +}
> +
> +static inline void i915_vma_set_freed(struct i915_vma *vma)
> +{
> +       set_bit(I915_VMA_FREED_BIT, __i915_vma_flags(vma));
> +}
> +
> +static inline bool i915_vma_is_active(const struct i915_vma *vma)
> +{
> +       if (i915_vma_is_persistent(vma)) {
> +               if (i915_vma_is_purged(vma))
> +                       return false;
> +
> +               return i915_vm_is_active(vma->vm);
> +       }
> +
> +       return !i915_active_is_idle(&vma->active);
> +}
> +
>  static inline struct i915_vma *i915_vma_get(struct i915_vma *vma)
>  {
>         i915_gem_object_get(vma->obj);
> @@ -408,8 +444,36 @@ int i915_vma_wait_for_bind(struct i915_vma
> *vma);
>  
>  static inline int i915_vma_sync(struct i915_vma *vma)
>  {
> +       int ret;
> +
>         /* Wait for the asynchronous bindings and pending GPU reads
> */
> -       return i915_active_wait(&vma->active);
> +       ret = i915_active_wait(&vma->active);
> +       if (ret || !i915_vma_is_persistent(vma) ||
> i915_vma_is_purged(vma))
> +               return ret;
> +
> +       return i915_vm_sync(vma->vm);
> +}
> +
> +static inline bool i915_vma_is_bind_complete(struct i915_vma *vma)
> +{
> +       /* Ensure vma bind is initiated */
> +       if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK))
> +               return false;
> +
> +       /* Ensure any binding started is complete */
> +       if (rcu_access_pointer(vma->active.excl.fence)) {
> +               struct dma_fence *fence;
> +
> +               rcu_read_lock();
> +               fence = dma_fence_get_rcu_safe(&vma-
> >active.excl.fence);
> +               rcu_read_unlock();
> +               if (fence) {
> +                       dma_fence_put(fence);
> +                       return false;
> +               }
> +       }

Could we use i915_active_fence_get() instead of first testing the
pointer and then open-coding i915_active_fence_get()? Also no need to
inline this.

> +
> +       return true;
>  }
>  
>  /**
> diff --git a/drivers/gpu/drm/i915/i915_vma_types.h
> b/drivers/gpu/drm/i915/i915_vma_types.h
> index 7d830a6a0b51..405c82e1bc30 100644
> --- a/drivers/gpu/drm/i915/i915_vma_types.h
> +++ b/drivers/gpu/drm/i915/i915_vma_types.h
> @@ -264,6 +264,28 @@ struct i915_vma {
>  #define I915_VMA_SCANOUT_BIT   17
>  #define I915_VMA_SCANOUT       ((int)BIT(I915_VMA_SCANOUT_BIT))
>  
> +  /**
> +   * 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.
> +   *
> +   * I915_VMA_FREED_BIT:
> +   * The persistent vma is being released by UMD via VM_UNBIND call.
> +   * While releasing the vma, do not take VM_BIND lock as VM_UNBIND
> call
> +   * already holds the lock.
> +   */
> +#define I915_VMA_PERSISTENT_BIT        19
> +#define I915_VMA_PURGED_BIT    20
> +#define I915_VMA_FREED_BIT     21
> +
> +#define I915_VMA_PERSISTENT    ((int)BIT(I915_VMA_PERSISTENT_BIT))
> +#define
> I915_VMA_PURGED                ((int)BIT(I915_VMA_PURGED_BIT))
> +#define I915_VMA_FREED         ((int)BIT(I915_VMA_FREED_BIT))
> +
>         struct i915_active active;
>  
>  #define I915_VMA_PAGES_BIAS 24
> @@ -292,6 +314,7 @@ struct i915_vma {
>         struct list_head vm_bind_link; /* Link in persistent VMA list
> */
Protected by..

>         /* Link in non-private persistent VMA list */
>         struct list_head non_priv_vm_bind_link;
> +       struct list_head vm_rebind_link; /* Link in vm_rebind_list */
>  
>         /** Timeline fence for vm_bind completion notification */
>         struct {



More information about the dri-devel mailing list