drm/i915: Add soft-pinning API for execbuffer
authorChris Wilson <chris@chris-wilson.co.uk>
Tue, 8 Dec 2015 11:55:07 +0000 (11:55 +0000)
committerTvrtko Ursulin <tvrtko.ursulin@intel.com>
Wed, 9 Dec 2015 10:20:17 +0000 (10:20 +0000)
Userspace can pass in an offset that it presumes the object is located
at. The kernel will then do its utmost to fit the object into that
location. The assumption is that userspace is handling its own object
locations (for example along with full-ppgtt) and that the kernel will
rarely have to make space for the user's requests.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
v2: Fixed incorrect eviction found by Michal Winiarski - fix suggested by Chris
Wilson.  Fixed incorrect error paths causing crash found by Michal Winiarski.
(Not published externally)

v3: Rebased because of trivial conflict in object_bind_to_vm.  Fixed eviction
to allow eviction of soft-pinned objects when another soft-pinned object used
by a subsequent execbuffer overlaps reported by Michal Winiarski.
(Not published externally)

v4: Moved soft-pinned objects to the front of ordered_vmas so that they are
pinned first after an address conflict happens to avoid repeated conflicts in
rare cases (Suggested by Chris Wilson).  Expanded comment on
drm_i915_gem_exec_object2.offset to cover this new API.

v5: Added I915_PARAM_HAS_EXEC_SOFTPIN parameter for detecting this capability
(Kristian). Added check for multiple pinnings on eviction (Akash). Made sure
buffers are not considered misplaced without the user specifying
EXEC_OBJECT_SUPPORTS_48B_ADDRESS.  User must assume responsibility for any
addressing workarounds.  Updated object2.offset field comment again to clarify
NO_RELOC case (Chris).  checkpatch cleanup.

v6: Trivial rebase on latest drm-intel-nightly

v7: Catch attempts to pin above the max virtual address size and return
EINVAL (Tvrtko). Decouple EXEC_OBJECT_SUPPORTS_48B_ADDRESS and
EXEC_OBJECT_PINNED flags, user must pass both flags in any attempt to pin
something at an offset above 4GB (Chris, Daniel Vetter).

Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Akash Goel <akash.goel@intel.com>
Cc: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
Cc: Michal Winiarski <michal.winiarski@intel.com>
Cc: Zou Nanhai <nanhai.zou@intel.com>
Cc: Kristian Høgsberg <hoegsberg@gmail.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Reviewed-by: Michel Thierry <michel.thierry@intel.com>
Acked-by: PDT
Signed-off-by: Thomas Daniel <thomas.daniel@intel.com>
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1449575707-20933-1-git-send-email-thomas.daniel@intel.com
drivers/gpu/drm/i915/i915_dma.c
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_gem.c
drivers/gpu/drm/i915/i915_gem_evict.c
drivers/gpu/drm/i915/i915_gem_execbuffer.c
include/uapi/drm/i915_drm.h

index a81c766035448853b91b481cadf180a271ce7530..52b82893ba42620d06ac51bc45fddd652a3b8baa 100644 (file)
@@ -169,6 +169,9 @@ static int i915_getparam(struct drm_device *dev, void *data,
        case I915_PARAM_HAS_RESOURCE_STREAMER:
                value = HAS_RESOURCE_STREAMER(dev);
                break;
+       case I915_PARAM_HAS_EXEC_SOFTPIN:
+               value = 1;
+               break;
        default:
                DRM_DEBUG("Unknown parameter %d\n", param->param);
                return -EINVAL;
index 6dee97c0d5d03cefa8a6f20e3ac8821a27cef139..547c142692928b890d55995eff56d98426cb39de 100644 (file)
@@ -2868,6 +2868,7 @@ void i915_gem_vma_destroy(struct i915_vma *vma);
 #define PIN_UPDATE     (1<<5)
 #define PIN_ZONE_4G    (1<<6)
 #define PIN_HIGH       (1<<7)
+#define PIN_OFFSET_FIXED       (1<<8)
 #define PIN_OFFSET_MASK (~4095)
 int __must_check
 i915_gem_object_pin(struct drm_i915_gem_object *obj,
@@ -3213,6 +3214,7 @@ int __must_check i915_gem_evict_something(struct drm_device *dev,
                                          unsigned long start,
                                          unsigned long end,
                                          unsigned flags);
+int __must_check i915_gem_evict_for_vma(struct i915_vma *target);
 int i915_gem_evict_vm(struct i915_address_space *vm, bool do_idle);
 
 /* belongs in i915_gem_gtt.h */
index a531cb83295c13fd5e7852fc25a268774ec2c61b..d7b8d16ff6e63c6b4040846edef031e1df186818 100644 (file)
@@ -3468,30 +3468,50 @@ i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj,
        if (IS_ERR(vma))
                goto err_unpin;
 
-       if (flags & PIN_HIGH) {
-               search_flag = DRM_MM_SEARCH_BELOW;
-               alloc_flag = DRM_MM_CREATE_TOP;
+       if (flags & PIN_OFFSET_FIXED) {
+               uint64_t offset = flags & PIN_OFFSET_MASK;
+
+               if (offset & (alignment - 1) || offset + size > end) {
+                       ret = -EINVAL;
+                       goto err_free_vma;
+               }
+               vma->node.start = offset;
+               vma->node.size = size;
+               vma->node.color = obj->cache_level;
+               ret = drm_mm_reserve_node(&vm->mm, &vma->node);
+               if (ret) {
+                       ret = i915_gem_evict_for_vma(vma);
+                       if (ret == 0)
+                               ret = drm_mm_reserve_node(&vm->mm, &vma->node);
+               }
+               if (ret)
+                       goto err_free_vma;
        } else {
-               search_flag = DRM_MM_SEARCH_DEFAULT;
-               alloc_flag = DRM_MM_CREATE_DEFAULT;
-       }
+               if (flags & PIN_HIGH) {
+                       search_flag = DRM_MM_SEARCH_BELOW;
+                       alloc_flag = DRM_MM_CREATE_TOP;
+               } else {
+                       search_flag = DRM_MM_SEARCH_DEFAULT;
+                       alloc_flag = DRM_MM_CREATE_DEFAULT;
+               }
 
 search_free:
-       ret = drm_mm_insert_node_in_range_generic(&vm->mm, &vma->node,
-                                                 size, alignment,
-                                                 obj->cache_level,
-                                                 start, end,
-                                                 search_flag,
-                                                 alloc_flag);
-       if (ret) {
-               ret = i915_gem_evict_something(dev, vm, size, alignment,
-                                              obj->cache_level,
-                                              start, end,
-                                              flags);
-               if (ret == 0)
-                       goto search_free;
+               ret = drm_mm_insert_node_in_range_generic(&vm->mm, &vma->node,
+                                                         size, alignment,
+                                                         obj->cache_level,
+                                                         start, end,
+                                                         search_flag,
+                                                         alloc_flag);
+               if (ret) {
+                       ret = i915_gem_evict_something(dev, vm, size, alignment,
+                                                      obj->cache_level,
+                                                      start, end,
+                                                      flags);
+                       if (ret == 0)
+                               goto search_free;
 
-               goto err_free_vma;
+                       goto err_free_vma;
+               }
        }
        if (WARN_ON(!i915_gem_valid_gtt_space(vma, obj->cache_level))) {
                ret = -EINVAL;
@@ -4082,6 +4102,10 @@ i915_vma_misplaced(struct i915_vma *vma, uint32_t alignment, uint64_t flags)
            vma->node.start < (flags & PIN_OFFSET_MASK))
                return true;
 
+       if (flags & PIN_OFFSET_FIXED &&
+           vma->node.start != (flags & PIN_OFFSET_MASK))
+               return true;
+
        return false;
 }
 
index d71a133ceff5a25e87aa2888b45609577ce34939..07c6e4d320c91b954a4e8056bb780fee12d90b88 100644 (file)
@@ -199,6 +199,45 @@ found:
        return ret;
 }
 
+int
+i915_gem_evict_for_vma(struct i915_vma *target)
+{
+       struct drm_mm_node *node, *next;
+
+       list_for_each_entry_safe(node, next,
+                       &target->vm->mm.head_node.node_list,
+                       node_list) {
+               struct i915_vma *vma;
+               int ret;
+
+               if (node->start + node->size <= target->node.start)
+                       continue;
+               if (node->start >= target->node.start + target->node.size)
+                       break;
+
+               vma = container_of(node, typeof(*vma), node);
+
+               if (vma->pin_count) {
+                       if (!vma->exec_entry || (vma->pin_count > 1))
+                               /* Object is pinned for some other use */
+                               return -EBUSY;
+
+                       /* We need to evict a buffer in the same batch */
+                       if (vma->exec_entry->flags & EXEC_OBJECT_PINNED)
+                               /* Overlapping fixed objects in the same batch */
+                               return -EINVAL;
+
+                       return -ENOSPC;
+               }
+
+               ret = i915_vma_unbind(vma);
+               if (ret)
+                       return ret;
+       }
+
+       return 0;
+}
+
 /**
  * i915_gem_evict_vm - Evict all idle vmas from a vm
  * @vm: Address space to cleanse
index a4c243cec4aa018e88dbe9146f1318c3766fc80f..48ec4846e6f29b064a8b47bae8f3e1997699d3cd 100644 (file)
@@ -599,6 +599,8 @@ i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
                        flags |= PIN_GLOBAL | PIN_MAPPABLE;
                if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS)
                        flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
+               if (entry->flags & EXEC_OBJECT_PINNED)
+                       flags |= entry->offset | PIN_OFFSET_FIXED;
                if ((flags & PIN_MAPPABLE) == 0)
                        flags |= PIN_HIGH;
        }
@@ -670,6 +672,10 @@ eb_vma_misplaced(struct i915_vma *vma)
            vma->node.start & (entry->alignment - 1))
                return true;
 
+       if (entry->flags & EXEC_OBJECT_PINNED &&
+           vma->node.start != entry->offset)
+               return true;
+
        if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS &&
            vma->node.start < BATCH_OFFSET_BIAS)
                return true;
@@ -695,6 +701,7 @@ i915_gem_execbuffer_reserve(struct intel_engine_cs *ring,
        struct i915_vma *vma;
        struct i915_address_space *vm;
        struct list_head ordered_vmas;
+       struct list_head pinned_vmas;
        bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
        int retry;
 
@@ -703,6 +710,7 @@ i915_gem_execbuffer_reserve(struct intel_engine_cs *ring,
        vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
 
        INIT_LIST_HEAD(&ordered_vmas);
+       INIT_LIST_HEAD(&pinned_vmas);
        while (!list_empty(vmas)) {
                struct drm_i915_gem_exec_object2 *entry;
                bool need_fence, need_mappable;
@@ -721,7 +729,9 @@ i915_gem_execbuffer_reserve(struct intel_engine_cs *ring,
                        obj->tiling_mode != I915_TILING_NONE;
                need_mappable = need_fence || need_reloc_mappable(vma);
 
-               if (need_mappable) {
+               if (entry->flags & EXEC_OBJECT_PINNED)
+                       list_move_tail(&vma->exec_list, &pinned_vmas);
+               else if (need_mappable) {
                        entry->flags |= __EXEC_OBJECT_NEEDS_MAP;
                        list_move(&vma->exec_list, &ordered_vmas);
                } else
@@ -731,6 +741,7 @@ i915_gem_execbuffer_reserve(struct intel_engine_cs *ring,
                obj->base.pending_write_domain = 0;
        }
        list_splice(&ordered_vmas, vmas);
+       list_splice(&pinned_vmas, vmas);
 
        /* Attempt to pin all of the buffers into the GTT.
         * This is done in 3 phases:
@@ -1317,7 +1328,8 @@ eb_get_batch(struct eb_vmas *eb)
         * Note that actual hangs have only been observed on gen7, but for
         * paranoia do it everywhere.
         */
-       vma->exec_entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
+       if ((vma->exec_entry->flags & EXEC_OBJECT_PINNED) == 0)
+               vma->exec_entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
 
        return vma->obj;
 }
index 67ef73a5d6ebdffacb32b10dc453afbbabcb5b8d..d727b49f07aca19031483b291c3dee2e9b68d1c8 100644 (file)
@@ -356,6 +356,7 @@ typedef struct drm_i915_irq_wait {
 #define I915_PARAM_EU_TOTAL             34
 #define I915_PARAM_HAS_GPU_RESET        35
 #define I915_PARAM_HAS_RESOURCE_STREAMER 36
+#define I915_PARAM_HAS_EXEC_SOFTPIN     37
 
 typedef struct drm_i915_getparam {
        __s32 param;
@@ -682,8 +683,12 @@ struct drm_i915_gem_exec_object2 {
        __u64 alignment;
 
        /**
-        * Returned value of the updated offset of the object, for future
-        * presumed_offset writes.
+        * When the EXEC_OBJECT_PINNED flag is specified this is populated by
+        * the user with the GTT offset at which this object will be pinned.
+        * When the I915_EXEC_NO_RELOC flag is specified this must contain the
+        * presumed_offset of the object.
+        * During execbuffer2 the kernel populates it with the value of the
+        * current GTT offset of the object, for future presumed_offset writes.
         */
        __u64 offset;
 
@@ -691,7 +696,8 @@ struct drm_i915_gem_exec_object2 {
 #define EXEC_OBJECT_NEEDS_GTT  (1<<1)
 #define EXEC_OBJECT_WRITE      (1<<2)
 #define EXEC_OBJECT_SUPPORTS_48B_ADDRESS (1<<3)
-#define __EXEC_OBJECT_UNKNOWN_FLAGS -(EXEC_OBJECT_SUPPORTS_48B_ADDRESS<<1)
+#define EXEC_OBJECT_PINNED     (1<<4)
+#define __EXEC_OBJECT_UNKNOWN_FLAGS -(EXEC_OBJECT_PINNED<<1)
        __u64 flags;
 
        __u64 rsvd1;