drm/i915: Eliminate lots of iterations over the execobjects array

The major scaling bottleneck in execbuffer is the processing of the
execobjects. Creating an auxiliary list is inefficient when compared to
using the execobject array we already have allocated.

Reservation is then split into phases. As we lookup up the VMA, we
try and bind it back into active location. Only if that fails, do we add
it to the unbound list for phase 2. In phase 2, we try and add all those
objects that could not fit into their previous location, with fallback
to retrying all objects and evicting the VM in case of severe
fragmentation. (This is the same as before, except that phase 1 is now
done inline with looking up the VMA to avoid an iteration over the
execobject array. In the ideal case, we eliminate the separate reservation
phase). During the reservation phase, we only evict from the VM between
passes (rather than currently as we try to fit every new VMA). In
testing with Unreal Engine's Atlantis demo which stresses the eviction
logic on gen7 class hardware, this speed up the framerate by a factor of
2.

The second loop amalgamation is between move_to_gpu and move_to_active.
As we always submit the request, even if incomplete, we can use the
current request to track active VMA as we perform the flushes and
synchronisation required.

The next big advancement is to avoid copying back to the user any
execobjects and relocations that are not changed.

v2: Add a Theory of Operation spiel.
v3: Fall back to slow relocations in preparation for flushing userptrs.
v4: Document struct members, factor out eb_validate_vma(), add a few
more comments to explain some magic and hide other magic behind macros.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>

7 files changed, 1239 insertions, 916 deletions

diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index af2a54672396..7e182dd7e356 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -3581,7 +3581,7 @@ int __must_check i915_gem_evict_something(struct i915_address_space *vm,
 int __must_check i915_gem_evict_for_node(struct i915_address_space *vm,
 					 struct drm_mm_node *node,
 					 unsigned int flags);
-int i915_gem_evict_vm(struct i915_address_space *vm, bool do_idle);
+int i915_gem_evict_vm(struct i915_address_space *vm);
 
 /* belongs in i915_gem_gtt.h */
 static inline void i915_gem_chipset_flush(struct drm_i915_private *dev_priv)
diff --git a/drivers/gpu/drm/i915/i915_gem_evict.c b/drivers/gpu/drm/i915/i915_gem_evict.c
index 204a2d9288ae..a193f1b36c67 100644
--- a/drivers/gpu/drm/i915/i915_gem_evict.c
+++ b/drivers/gpu/drm/i915/i915_gem_evict.c
@@ -50,6 +50,29 @@ static bool ggtt_is_idle(struct drm_i915_private *dev_priv)
 	return true;
 }
 
+static int ggtt_flush(struct drm_i915_private *i915)
+{
+	int err;
+
+	/* Not everything in the GGTT is tracked via vma (otherwise we
+	 * could evict as required with minimal stalling) so we are forced
+	 * to idle the GPU and explicitly retire outstanding requests in
+	 * the hopes that we can then remove contexts and the like only
+	 * bound by their active reference.
+	 */
+	err = i915_gem_switch_to_kernel_context(i915);
+	if (err)
+		return err;
+
+	err = i915_gem_wait_for_idle(i915,
+				     I915_WAIT_INTERRUPTIBLE |
+				     I915_WAIT_LOCKED);
+	if (err)
+		return err;
+
+	return 0;
+}
+
 static bool
 mark_free(struct drm_mm_scan *scan,
 	  struct i915_vma *vma,
@@ -175,19 +198,7 @@ search_again:
 		return intel_has_pending_fb_unpin(dev_priv) ? -EAGAIN : -ENOSPC;
 	}
 
-	/* Not everything in the GGTT is tracked via vma (otherwise we
-	 * could evict as required with minimal stalling) so we are forced
-	 * to idle the GPU and explicitly retire outstanding requests in
-	 * the hopes that we can then remove contexts and the like only
-	 * bound by their active reference.
-	 */
-	ret = i915_gem_switch_to_kernel_context(dev_priv);
-	if (ret)
-		return ret;
-
-	ret = i915_gem_wait_for_idle(dev_priv,
-				     I915_WAIT_INTERRUPTIBLE |
-				     I915_WAIT_LOCKED);
+	ret = ggtt_flush(dev_priv);
 	if (ret)
 		return ret;
 
@@ -337,10 +348,8 @@ int i915_gem_evict_for_node(struct i915_address_space *vm,
 /**
  * i915_gem_evict_vm - Evict all idle vmas from a vm
  * @vm: Address space to cleanse
- * @do_idle: Boolean directing whether to idle first.
  *
- * This function evicts all idles vmas from a vm. If all unpinned vmas should be
- * evicted the @do_idle needs to be set to true.
+ * This function evicts all vmas from a vm.
  *
  * This is used by the execbuf code as a last-ditch effort to defragment the
  * address space.
@@ -348,37 +357,50 @@ int i915_gem_evict_for_node(struct i915_address_space *vm,
  * To clarify: This is for freeing up virtual address space, not for freeing
  * memory in e.g. the shrinker.
  */
-int i915_gem_evict_vm(struct i915_address_space *vm, bool do_idle)
+int i915_gem_evict_vm(struct i915_address_space *vm)
 {
+	struct list_head *phases[] = {
+		&vm->inactive_list,
+		&vm->active_list,
+		NULL
+	}, **phase;
+	struct list_head eviction_list;
 	struct i915_vma *vma, *next;
 	int ret;
 
 	lockdep_assert_held(&vm->i915->drm.struct_mutex);
 	trace_i915_gem_evict_vm(vm);
 
-	if (do_idle) {
-		struct drm_i915_private *dev_priv = vm->i915;
-
-		if (i915_is_ggtt(vm)) {
-			ret = i915_gem_switch_to_kernel_context(dev_priv);
-			if (ret)
-				return ret;
-		}
-
-		ret = i915_gem_wait_for_idle(dev_priv,
-					     I915_WAIT_INTERRUPTIBLE |
-					     I915_WAIT_LOCKED);
+	/* Switch back to the default context in order to unpin
+	 * the existing context objects. However, such objects only
+	 * pin themselves inside the global GTT and performing the
+	 * switch otherwise is ineffective.
+	 */
+	if (i915_is_ggtt(vm)) {
+		ret = ggtt_flush(vm->i915);
 		if (ret)
 			return ret;
-
-		WARN_ON(!list_empty(&vm->active_list));
 	}
 
-	list_for_each_entry_safe(vma, next, &vm->inactive_list, vm_link)
-		if (!i915_vma_is_pinned(vma))
-			WARN_ON(i915_vma_unbind(vma));
+	INIT_LIST_HEAD(&eviction_list);
+	phase = phases;
+	do {
+		list_for_each_entry(vma, *phase, vm_link) {
+			if (i915_vma_is_pinned(vma))
+				continue;
 
-	return 0;
+			__i915_vma_pin(vma);
+			list_add(&vma->evict_link, &eviction_list);
+		}
+	} while (*++phase);
+
+	ret = 0;
+	list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
+		__i915_vma_unpin(vma);
+		if (ret == 0)
+			ret = i915_vma_unbind(vma);
+	}
+	return ret;
 }
 
 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
diff --git a/drivers/gpu/drm/i915/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
index 9c3f6c40270f..a052072fe8b3 100644
--- a/drivers/gpu/drm/i915/i915_gem_execbuffer.c
+++ b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
@@ -42,41 +42,195 @@
 
 #define DBG_USE_CPU_RELOC 0 /* -1 force GTT relocs; 1 force CPU relocs */
 
-#define  __EXEC_OBJECT_HAS_PIN		(1<<31)
-#define  __EXEC_OBJECT_HAS_FENCE	(1<<30)
-#define  __EXEC_OBJECT_NEEDS_MAP	(1<<29)
-#define  __EXEC_OBJECT_NEEDS_BIAS	(1<<28)
-#define  __EXEC_OBJECT_INTERNAL_FLAGS (0xf<<28) /* all of the above */
+#define __EXEC_OBJECT_HAS_PIN		BIT(31)
+#define __EXEC_OBJECT_HAS_FENCE		BIT(30)
+#define __EXEC_OBJECT_NEEDS_MAP		BIT(29)
+#define __EXEC_OBJECT_NEEDS_BIAS	BIT(28)
+#define __EXEC_OBJECT_INTERNAL_FLAGS	(~0u << 28) /* all of the above */
+#define __EXEC_OBJECT_RESERVED (__EXEC_OBJECT_HAS_PIN | __EXEC_OBJECT_HAS_FENCE)
+
+#define __EXEC_HAS_RELOC	BIT(31)
+#define __EXEC_VALIDATED	BIT(30)
+#define UPDATE			PIN_OFFSET_FIXED
 
 #define BATCH_OFFSET_BIAS (256*1024)
 
 #define __I915_EXEC_ILLEGAL_FLAGS \
 	(__I915_EXEC_UNKNOWN_FLAGS | I915_EXEC_CONSTANTS_MASK)
 
+/**
+ * DOC: User command execution
+ *
+ * Userspace submits commands to be executed on the GPU as an instruction
+ * stream within a GEM object we call a batchbuffer. This instructions may
+ * refer to other GEM objects containing auxiliary state such as kernels,
+ * samplers, render targets and even secondary batchbuffers. Userspace does
+ * not know where in the GPU memory these objects reside and so before the
+ * batchbuffer is passed to the GPU for execution, those addresses in the
+ * batchbuffer and auxiliary objects are updated. This is known as relocation,
+ * or patching. To try and avoid having to relocate each object on the next
+ * execution, userspace is told the location of those objects in this pass,
+ * but this remains just a hint as the kernel may choose a new location for
+ * any object in the future.
+ *
+ * Processing an execbuf ioctl is conceptually split up into a few phases.
+ *
+ * 1. Validation - Ensure all the pointers, handles and flags are valid.
+ * 2. Reservation - Assign GPU address space for every object
+ * 3. Relocation - Update any addresses to point to the final locations
+ * 4. Serialisation - Order the request with respect to its dependencies
+ * 5. Construction - Construct a request to execute the batchbuffer
+ * 6. Submission (at some point in the future execution)
+ *
+ * Reserving resources for the execbuf is the most complicated phase. We
+ * neither want to have to migrate the object in the address space, nor do
+ * we want to have to update any relocations pointing to this object. Ideally,
+ * we want to leave the object where it is and for all the existing relocations
+ * to match. If the object is given a new address, or if userspace thinks the
+ * object is elsewhere, we have to parse all the relocation entries and update
+ * the addresses. Userspace can set the I915_EXEC_NORELOC flag to hint that
+ * all the target addresses in all of its objects match the value in the
+ * relocation entries and that they all match the presumed offsets given by the
+ * list of execbuffer objects. Using this knowledge, we know that if we haven't
+ * moved any buffers, all the relocation entries are valid and we can skip
+ * the update. (If userspace is wrong, the likely outcome is an impromptu GPU
+ * hang.) The requirement for using I915_EXEC_NO_RELOC are:
+ *
+ *      The addresses written in the objects must match the corresponding
+ *      reloc.presumed_offset which in turn must match the corresponding
+ *      execobject.offset.
+ *
+ *      Any render targets written to in the batch must be flagged with
+ *      EXEC_OBJECT_WRITE.
+ *
+ *      To avoid stalling, execobject.offset should match the current
+ *      address of that object within the active context.
+ *
+ * The reservation is done is multiple phases. First we try and keep any
+ * object already bound in its current location - so as long as meets the
+ * constraints imposed by the new execbuffer. Any object left unbound after the
+ * first pass is then fitted into any available idle space. If an object does
+ * not fit, all objects are removed from the reservation and the process rerun
+ * after sorting the objects into a priority order (more difficult to fit
+ * objects are tried first). Failing that, the entire VM is cleared and we try
+ * to fit the execbuf once last time before concluding that it simply will not
+ * fit.
+ *
+ * A small complication to all of this is that we allow userspace not only to
+ * specify an alignment and a size for the object in the address space, but
+ * we also allow userspace to specify the exact offset. This objects are
+ * simpler to place (the location is known a priori) all we have to do is make
+ * sure the space is available.
+ *
+ * Once all the objects are in place, patching up the buried pointers to point
+ * to the final locations is a fairly simple job of walking over the relocation
+ * entry arrays, looking up the right address and rewriting the value into
+ * the object. Simple! ... The relocation entries are stored in user memory
+ * and so to access them we have to copy them into a local buffer. That copy
+ * has to avoid taking any pagefaults as they may lead back to a GEM object
+ * requiring the struct_mutex (i.e. recursive deadlock). So once again we split
+ * the relocation into multiple passes. First we try to do everything within an
+ * atomic context (avoid the pagefaults) which requires that we never wait. If
+ * we detect that we may wait, or if we need to fault, then we have to fallback
+ * to a slower path. The slowpath has to drop the mutex. (Can you hear alarm
+ * bells yet?) Dropping the mutex means that we lose all the state we have
+ * built up so far for the execbuf and we must reset any global data. However,
+ * we do leave the objects pinned in their final locations - which is a
+ * potential issue for concurrent execbufs. Once we have left the mutex, we can
+ * allocate and copy all the relocation entries into a large array at our
+ * leisure, reacquire the mutex, reclaim all the objects and other state and
+ * then proceed to update any incorrect addresses with the objects.
+ *
+ * As we process the relocation entries, we maintain a record of whether the
+ * object is being written to. Using NORELOC, we expect userspace to provide
+ * this information instead. We also check whether we can skip the relocation
+ * by comparing the expected value inside the relocation entry with the target's
+ * final address. If they differ, we have to map the current object and rewrite
+ * the 4 or 8 byte pointer within.
+ *
+ * Serialising an execbuf is quite simple according to the rules of the GEM
+ * ABI. Execution within each context is ordered by the order of submission.
+ * Writes to any GEM object are in order of submission and are exclusive. Reads
+ * from a GEM object are unordered with respect to other reads, but ordered by
+ * writes. A write submitted after a read cannot occur before the read, and
+ * similarly any read submitted after a write cannot occur before the write.
+ * Writes are ordered between engines such that only one write occurs at any
+ * time (completing any reads beforehand) - using semaphores where available
+ * and CPU serialisation otherwise. Other GEM access obey the same rules, any
+ * write (either via mmaps using set-domain, or via pwrite) must flush all GPU
+ * reads before starting, and any read (either using set-domain or pread) must
+ * flush all GPU writes before starting. (Note we only employ a barrier before,
+ * we currently rely on userspace not concurrently starting a new execution
+ * whilst reading or writing to an object. This may be an advantage or not
+ * depending on how much you trust userspace not to shoot themselves in the
+ * foot.) Serialisation may just result in the request being inserted into
+ * a DAG awaiting its turn, but most simple is to wait on the CPU until
+ * all dependencies are resolved.
+ *
+ * After all of that, is just a matter of closing the request and handing it to
+ * the hardware (well, leaving it in a queue to be executed). However, we also
+ * offer the ability for batchbuffers to be run with elevated privileges so
+ * that they access otherwise hidden registers. (Used to adjust L3 cache etc.)
+ * Before any batch is given extra privileges we first must check that it
+ * contains no nefarious instructions, we check that each instruction is from
+ * our whitelist and all registers are also from an allowed list. We first
+ * copy the user's batchbuffer to a shadow (so that the user doesn't have
+ * access to it, either by the CPU or GPU as we scan it) and then parse each
+ * instruction. If everything is ok, we set a flag telling the hardware to run
+ * the batchbuffer in trusted mode, otherwise the ioctl is rejected.
+ */
+
 struct i915_execbuffer {
-	struct drm_i915_private *i915;
-	struct drm_file *file;
-	struct drm_i915_gem_execbuffer2 *args;
-	struct drm_i915_gem_exec_object2 *exec;
-	struct intel_engine_cs *engine;
-	struct i915_gem_context *ctx;
-	struct i915_address_space *vm;
-	struct i915_vma *batch;
-	struct drm_i915_gem_request *request;
-	u32 batch_start_offset;
-	u32 batch_len;
-	unsigned int dispatch_flags;
-	struct drm_i915_gem_exec_object2 shadow_exec_entry;
-	bool need_relocs;
-	struct list_head vmas;
+	struct drm_i915_private *i915; /** i915 backpointer */
+	struct drm_file *file; /** per-file lookup tables and limits */
+	struct drm_i915_gem_execbuffer2 *args; /** ioctl parameters */
+	struct drm_i915_gem_exec_object2 *exec; /** ioctl execobj[] */
+
+	struct intel_engine_cs *engine; /** engine to queue the request to */
+	struct i915_gem_context *ctx; /** context for building the request */
+	struct i915_address_space *vm; /** GTT and vma for the request */
+
+	struct drm_i915_gem_request *request; /** our request to build */
+	struct i915_vma *batch; /** identity of the batch obj/vma */
+
+	/** actual size of execobj[] as we may extend it for the cmdparser */
+	unsigned int buffer_count;
+
+	/** list of vma not yet bound during reservation phase */
+	struct list_head unbound;
+
+	/** list of vma that have execobj.relocation_count */
+	struct list_head relocs;
+
+	/**
+	 * Track the most recently used object for relocations, as we
+	 * frequently have to perform multiple relocations within the same
+	 * obj/page
+	 */
 	struct reloc_cache {
-		struct drm_mm_node node;
-		unsigned long vaddr;
-		unsigned int page;
+		struct drm_mm_node node; /** temporary GTT binding */
+		unsigned long vaddr; /** Current kmap address */
+		unsigned long page; /** Currently mapped page index */
 		bool use_64bit_reloc : 1;
+		bool has_llc : 1;
+		bool has_fence : 1;
+		bool needs_unfenced : 1;
 	} reloc_cache;
-	int lut_mask;
-	struct hlist_head *buckets;
+
+	u64 invalid_flags; /** Set of execobj.flags that are invalid */
+	u32 context_flags; /** Set of execobj.flags to insert from the ctx */
+
+	u32 batch_start_offset; /** Location within object of batch */
+	u32 batch_len; /** Length of batch within object */
+	u32 batch_flags; /** Flags composed for emit_bb_start() */
+
+	/**
+	 * Indicate either the size of the hastable used to resolve
+	 * relocation handles, or if negative that we are using a direct
+	 * index into the execobj[].
+	 */
+	int lut_size;
+	struct hlist_head *buckets; /** ht for relocation handles */
 };
 
 /*
@@ -87,11 +241,41 @@ struct i915_execbuffer {
 #define __exec_to_vma(ee) (ee)->rsvd2
 #define exec_to_vma(ee) u64_to_ptr(struct i915_vma, __exec_to_vma(ee))
 
+/*
+ * Used to convert any address to canonical form.
+ * Starting from gen8, some commands (e.g. STATE_BASE_ADDRESS,
+ * MI_LOAD_REGISTER_MEM and others, see Broadwell PRM Vol2a) require the
+ * addresses to be in a canonical form:
+ * "GraphicsAddress[63:48] are ignored by the HW and assumed to be in correct
+ * canonical form [63:48] == [47]."
+ */
+#define GEN8_HIGH_ADDRESS_BIT 47
+static inline u64 gen8_canonical_addr(u64 address)
+{
+	return sign_extend64(address, GEN8_HIGH_ADDRESS_BIT);
+}
+
+static inline u64 gen8_noncanonical_addr(u64 address)
+{
+	return address & GENMASK_ULL(GEN8_HIGH_ADDRESS_BIT, 0);
+}
+
 static int eb_create(struct i915_execbuffer *eb)
 {
-	if ((eb->args->flags & I915_EXEC_HANDLE_LUT) == 0) {
-		unsigned int size = 1 + ilog2(eb->args->buffer_count);
+	if (!(eb->args->flags & I915_EXEC_HANDLE_LUT)) {
+		unsigned int size = 1 + ilog2(eb->buffer_count);
 
+		/*
+		 * Without a 1:1 association between relocation handles and
+		 * the execobject[] index, we instead create a hashtable.
+		 * We size it dynamically based on available memory, starting
+		 * first with 1:1 assocative hash and scaling back until
+		 * the allocation succeeds.
+		 *
+		 * Later on we use a positive lut_size to indicate we are
+		 * using this hashtable, and a negative value to indicate a
+		 * direct lookup.
+		 */
 		do {
 			eb->buckets = kzalloc(sizeof(struct hlist_head) << size,
 					      GFP_TEMPORARY |
@@ -108,112 +292,411 @@ static int eb_create(struct i915_execbuffer *eb)
 				return -ENOMEM;
 		}
 
-		eb->lut_mask = size;
+		eb->lut_size = size;
 	} else {
-		eb->lut_mask = -eb->args->buffer_count;
+		eb->lut_size = -eb->buffer_count;
 	}
 
 	return 0;
 }
 
+static bool
+eb_vma_misplaced(const struct drm_i915_gem_exec_object2 *entry,
+		 const struct i915_vma *vma)
+{
+	if (!(entry->flags & __EXEC_OBJECT_HAS_PIN))
+		return true;
+
+	if (vma->node.size < entry->pad_to_size)
+		return true;
+
+	if (entry->alignment && !IS_ALIGNED(vma->node.start, entry->alignment))
+		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;
+
+	if (!(entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) &&
+	    (vma->node.start + vma->node.size - 1) >> 32)
+		return true;
+
+	return false;
+}
+
+static inline void
+eb_pin_vma(struct i915_execbuffer *eb,
+	   struct drm_i915_gem_exec_object2 *entry,
+	   struct i915_vma *vma)
+{
+	u64 flags;
+
+	flags = vma->node.start;
+	flags |= PIN_USER | PIN_NONBLOCK | PIN_OFFSET_FIXED;
+	if (unlikely(entry->flags & EXEC_OBJECT_NEEDS_GTT))
+		flags |= PIN_GLOBAL;
+	if (unlikely(i915_vma_pin(vma, 0, 0, flags)))
+		return;
+
+	if (unlikely(entry->flags & EXEC_OBJECT_NEEDS_FENCE)) {
+		if (unlikely(i915_vma_get_fence(vma))) {
+			i915_vma_unpin(vma);
+			return;
+		}
+
+		if (i915_vma_pin_fence(vma))
+			entry->flags |= __EXEC_OBJECT_HAS_FENCE;
+	}
+
+	entry->flags |= __EXEC_OBJECT_HAS_PIN;
+}
+
 static inline void
 __eb_unreserve_vma(struct i915_vma *vma,
 		   const struct drm_i915_gem_exec_object2 *entry)
 {
+	GEM_BUG_ON(!(entry->flags & __EXEC_OBJECT_HAS_PIN));
+
 	if (unlikely(entry->flags & __EXEC_OBJECT_HAS_FENCE))
 		i915_vma_unpin_fence(vma);
 
-	if (entry->flags & __EXEC_OBJECT_HAS_PIN)
-		__i915_vma_unpin(vma);
+	__i915_vma_unpin(vma);
 }
 
-static void
-eb_unreserve_vma(struct i915_vma *vma)
+static inline void
+eb_unreserve_vma(struct i915_vma *vma,
+		 struct drm_i915_gem_exec_object2 *entry)
 {
-	struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
+	if (!(entry->flags & __EXEC_OBJECT_HAS_PIN))
+		return;
 
 	__eb_unreserve_vma(vma, entry);
-	entry->flags &= ~(__EXEC_OBJECT_HAS_FENCE | __EXEC_OBJECT_HAS_PIN);
+	entry->flags &= ~__EXEC_OBJECT_RESERVED;
 }
 
-static void
-eb_reset(struct i915_execbuffer *eb)
+static int
+eb_validate_vma(struct i915_execbuffer *eb,
+		struct drm_i915_gem_exec_object2 *entry,
+		struct i915_vma *vma)
 {
-	struct i915_vma *vma;
+	if (unlikely(entry->flags & eb->invalid_flags))
+		return -EINVAL;
 
-	list_for_each_entry(vma, &eb->vmas, exec_link) {
-		eb_unreserve_vma(vma);
-		i915_vma_put(vma);
-		vma->exec_entry = NULL;
+	if (unlikely(entry->alignment && !is_power_of_2(entry->alignment)))
+		return -EINVAL;
+
+	/*
+	 * Offset can be used as input (EXEC_OBJECT_PINNED), reject
+	 * any non-page-aligned or non-canonical addresses.
+	 */
+	if (unlikely(entry->flags & EXEC_OBJECT_PINNED &&
+		     entry->offset != gen8_canonical_addr(entry->offset & PAGE_MASK)))
+		return -EINVAL;
+
+	/* pad_to_size was once a reserved field, so sanitize it */
+	if (entry->flags & EXEC_OBJECT_PAD_TO_SIZE) {
+		if (unlikely(offset_in_page(entry->pad_to_size)))
+			return -EINVAL;
+	} else {
+		entry->pad_to_size = 0;
 	}
 
-	if (eb->lut_mask >= 0)
-		memset(eb->buckets, 0,
-		       sizeof(struct hlist_head) << eb->lut_mask);
+	if (unlikely(vma->exec_entry)) {
+		DRM_DEBUG("Object [handle %d, index %d] appears more than once in object list\n",
+			  entry->handle, (int)(entry - eb->exec));
+		return -EINVAL;
+	}
+
+	/*
+	 * From drm_mm perspective address space is continuous,
+	 * so from this point we're always using non-canonical
+	 * form internally.
+	 */
+	entry->offset = gen8_noncanonical_addr(entry->offset);
+
+	return 0;
 }
 
-static bool
-eb_add_vma(struct i915_execbuffer *eb, struct i915_vma *vma, int i)
+static int
+eb_add_vma(struct i915_execbuffer *eb,
+	   struct drm_i915_gem_exec_object2 *entry,
+	   struct i915_vma *vma)
 {
-	if (unlikely(vma->exec_entry)) {
-		DRM_DEBUG("Object [handle %d, index %d] appears more than once in object list\n",
-			  eb->exec[i].handle, i);
-		return false;
+	int err;
+
+	GEM_BUG_ON(i915_vma_is_closed(vma));
+
+	if (!(eb->args->flags & __EXEC_VALIDATED)) {
+		err = eb_validate_vma(eb, entry, vma);
+		if (unlikely(err))
+			return err;
 	}
-	list_add_tail(&vma->exec_link, &eb->vmas);
 
-	vma->exec_entry = &eb->exec[i];
-	if (eb->lut_mask >= 0) {
-		vma->exec_handle = eb->exec[i].handle;
+	if (eb->lut_size >= 0) {
+		vma->exec_handle = entry->handle;
 		hlist_add_head(&vma->exec_node,
-			       &eb->buckets[hash_32(vma->exec_handle,
-						    eb->lut_mask)]);
+			       &eb->buckets[hash_32(entry->handle,
+						    eb->lut_size)]);
 	}
 
-	i915_vma_get(vma);
-	__exec_to_vma(&eb->exec[i]) = (uintptr_t)vma;
-	return true;
+	if (entry->relocation_count)
+		list_add_tail(&vma->reloc_link, &eb->relocs);
+
+	if (!eb->reloc_cache.has_fence) {
+		entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
+	} else {
+		if ((entry->flags & EXEC_OBJECT_NEEDS_FENCE ||
+		     eb->reloc_cache.needs_unfenced) &&
+		    i915_gem_object_is_tiled(vma->obj))
+			entry->flags |= EXEC_OBJECT_NEEDS_GTT | __EXEC_OBJECT_NEEDS_MAP;
+	}
+
+	if (!(entry->flags & EXEC_OBJECT_PINNED))
+		entry->flags |= eb->context_flags;
+
+	/*
+	 * Stash a pointer from the vma to execobj, so we can query its flags,
+	 * size, alignment etc as provided by the user. Also we stash a pointer
+	 * to the vma inside the execobj so that we can use a direct lookup
+	 * to find the right target VMA when doing relocations.
+	 */
+	vma->exec_entry = entry;
+	__exec_to_vma(entry) = (uintptr_t)i915_vma_get(vma);
+
+	err = 0;
+	if (vma->node.size)
+		eb_pin_vma(eb, entry, vma);
+	if (eb_vma_misplaced(entry, vma)) {
+		eb_unreserve_vma(vma, entry);
+
+		list_add_tail(&vma->exec_link, &eb->unbound);
+		if (drm_mm_node_allocated(&vma->node))
+			err = i915_vma_unbind(vma);
+	} else {
+		if (entry->offset != vma->node.start) {
+			entry->offset = vma->node.start | UPDATE;
+			eb->args->flags |= __EXEC_HAS_RELOC;
+		}
+	}
+	return err;
+}
+
+static inline int use_cpu_reloc(const struct reloc_cache *cache,
+				const struct drm_i915_gem_object *obj)
+{
+	if (!i915_gem_object_has_struct_page(obj))
+		return false;
+
+	if (DBG_USE_CPU_RELOC)
+		return DBG_USE_CPU_RELOC > 0;
+
+	return (cache->has_llc ||
+		obj->cache_dirty ||
+		obj->cache_level != I915_CACHE_NONE);
+}
+
+static int eb_reserve_vma(const struct i915_execbuffer *eb,
+			  struct i915_vma *vma)
+{
+	struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
+	u64 flags;
+	int err;
+
+	flags = PIN_USER | PIN_NONBLOCK;
+	if (entry->flags & EXEC_OBJECT_NEEDS_GTT)
+		flags |= PIN_GLOBAL;
+
+	/*
+	 * Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset,
+	 * limit address to the first 4GBs for unflagged objects.
+	 */
+	if (!(entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS))
+		flags |= PIN_ZONE_4G;
+
+	if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
+		flags |= PIN_MAPPABLE;
+
+	if (entry->flags & EXEC_OBJECT_PINNED) {
+		flags |= entry->offset | PIN_OFFSET_FIXED;
+		flags &= ~PIN_NONBLOCK; /* force overlapping PINNED checks */
+	} else if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS) {
+		flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
+	}
+
+	err = i915_vma_pin(vma, entry->pad_to_size, entry->alignment, flags);
+	if (err)
+		return err;
+
+	if (entry->offset != vma->node.start) {
+		entry->offset = vma->node.start | UPDATE;
+		eb->args->flags |= __EXEC_HAS_RELOC;
+	}
+
+	entry->flags |= __EXEC_OBJECT_HAS_PIN;
+	GEM_BUG_ON(eb_vma_misplaced(entry, vma));
+
+	if (unlikely(entry->flags & EXEC_OBJECT_NEEDS_FENCE)) {
+		err = i915_vma_get_fence(vma);
+		if (unlikely(err)) {
+			i915_vma_unpin(vma);
+			return err;
+		}
+
+		if (i915_vma_pin_fence(vma))
+			entry->flags |= __EXEC_OBJECT_HAS_FENCE;
+	}
+
+	return 0;
+}
+
+static int eb_reserve(struct i915_execbuffer *eb)
+{
+	const unsigned int count = eb->buffer_count;
+	struct list_head last;
+	struct i915_vma *vma;
+	unsigned int i, pass;
+	int err;
+
+	/*
+	 * Attempt to pin all of the buffers into the GTT.
+	 * This is done in 3 phases:
+	 *
+	 * 1a. Unbind all objects that do not match the GTT constraints for
+	 *     the execbuffer (fenceable, mappable, alignment etc).
+	 * 1b. Increment pin count for already bound objects.
+	 * 2.  Bind new objects.
+	 * 3.  Decrement pin count.
+	 *
+	 * This avoid unnecessary unbinding of later objects in order to make
+	 * room for the earlier objects *unless* we need to defragment.
+	 */
+
+	pass = 0;
+	err = 0;
+	do {
+		list_for_each_entry(vma, &eb->unbound, exec_link) {
+			err = eb_reserve_vma(eb, vma);
+			if (err)
+				break;
+		}
+		if (err != -ENOSPC)
+			return err;
+
+		/* Resort *all* the objects into priority order */
+		INIT_LIST_HEAD(&eb->unbound);
+		INIT_LIST_HEAD(&last);
+		for (i = 0; i < count; i++) {
+			struct drm_i915_gem_exec_object2 *entry = &eb->exec[i];
+
+			if (entry->flags & EXEC_OBJECT_PINNED &&
+			    entry->flags & __EXEC_OBJECT_HAS_PIN)
+				continue;
+
+			vma = exec_to_vma(entry);
+			eb_unreserve_vma(vma, entry);
+
+			if (entry->flags & EXEC_OBJECT_PINNED)
+				list_add(&vma->exec_link, &eb->unbound);
+			else if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
+				list_add_tail(&vma->exec_link, &eb->unbound);
+			else
+				list_add_tail(&vma->exec_link, &last);
+		}
+		list_splice_tail(&last, &eb->unbound);
+
+		switch (pass++) {
+		case 0:
+			break;
+
+		case 1:
+			/* Too fragmented, unbind everything and retry */
+			err = i915_gem_evict_vm(eb->vm);
+			if (err)
+				return err;
+			break;
+
+		default:
+			return -ENOSPC;
+		}
+	} while (1);
 }
 
 static inline struct hlist_head *
-ht_head(const struct i915_gem_context *ctx, u32 handle)
+ht_head(const  struct i915_gem_context_vma_lut *lut, u32 handle)
 {
-	return &ctx->vma_lut.ht[hash_32(handle, ctx->vma_lut.ht_bits)];
+	return &lut->ht[hash_32(handle, lut->ht_bits)];
 }
 
 static inline bool
-ht_needs_resize(const struct i915_gem_context *ctx)
+ht_needs_resize(const struct i915_gem_context_vma_lut *lut)
 {
-	return (4*ctx->vma_lut.ht_count > 3*ctx->vma_lut.ht_size ||
-		4*ctx->vma_lut.ht_count + 1 < ctx->vma_lut.ht_size);
+	return (4*lut->ht_count > 3*lut->ht_size ||
+		4*lut->ht_count + 1 < lut->ht_size);
 }
 
-static int
-eb_lookup_vmas(struct i915_execbuffer *eb)
+static unsigned int eb_batch_index(const struct i915_execbuffer *eb)
+{
+	return eb->buffer_count - 1;
+}
+
+static int eb_select_context(struct i915_execbuffer *eb)
+{
+	struct i915_gem_context *ctx;
+
+	ctx = i915_gem_context_lookup(eb->file->driver_priv, eb->args->rsvd1);
+	if (unlikely(IS_ERR(ctx)))
+		return PTR_ERR(ctx);
+
+	if (unlikely(i915_gem_context_is_banned(ctx))) {
+		DRM_DEBUG("Context %u tried to submit while banned\n",
+			  ctx->user_handle);
+		return -EIO;
+	}
+
+	eb->ctx = i915_gem_context_get(ctx);
+	eb->vm = ctx->ppgtt ? &ctx->ppgtt->base : &eb->i915->ggtt.base;
+
+	eb->context_flags = 0;
+	if (ctx->flags & CONTEXT_NO_ZEROMAP)
+		eb->context_flags |= __EXEC_OBJECT_NEEDS_BIAS;
+
+	return 0;
+}
+
+static int eb_lookup_vmas(struct i915_execbuffer *eb)
 {
 #define INTERMEDIATE BIT(0)
-	const int count = eb->args->buffer_count;
+	const unsigned int count = eb->buffer_count;
+	struct i915_gem_context_vma_lut *lut = &eb->ctx->vma_lut;
 	struct i915_vma *vma;
+	struct idr *idr;
+	unsigned int i;
 	int slow_pass = -1;
-	int i;
+	int err;
 
-	INIT_LIST_HEAD(&eb->vmas);
+	INIT_LIST_HEAD(&eb->relocs);
+	INIT_LIST_HEAD(&eb->unbound);
 
-	if (unlikely(eb->ctx->vma_lut.ht_size & I915_CTX_RESIZE_IN_PROGRESS))
-		flush_work(&eb->ctx->vma_lut.resize);
-	GEM_BUG_ON(eb->ctx->vma_lut.ht_size & I915_CTX_RESIZE_IN_PROGRESS);
+	if (unlikely(lut->ht_size & I915_CTX_RESIZE_IN_PROGRESS))
+		flush_work(&lut->resize);
+	GEM_BUG_ON(lut->ht_size & I915_CTX_RESIZE_IN_PROGRESS);
 
 	for (i = 0; i < count; i++) {
 		__exec_to_vma(&eb->exec[i]) = 0;
 
 		hlist_for_each_entry(vma,
-				     ht_head(eb->ctx, eb->exec[i].handle),
+				     ht_head(lut, eb->exec[i].handle),
 				     ctx_node) {
 			if (vma->ctx_handle != eb->exec[i].handle)
 				continue;
 
-			if (!eb_add_vma(eb, vma, i))
-				return -EINVAL;
+			err = eb_add_vma(eb, &eb->exec[i], vma);
+			if (unlikely(err))
+				return err;
 
 			goto next_vma;
 		}
@@ -224,24 +707,27 @@ next_vma: ;
 	}
 
 	if (slow_pass < 0)
-		return 0;
+		goto out;
 
 	spin_lock(&eb->file->table_lock);
-	/* Grab a reference to the object and release the lock so we can lookup
-	 * or create the VMA without using GFP_ATOMIC */
+	/*
+	 * Grab a reference to the object and release the lock so we can lookup
+	 * or create the VMA without using GFP_ATOMIC
+	 */
+	idr = &eb->file->object_idr;
 	for (i = slow_pass; i < count; i++) {
 		struct drm_i915_gem_object *obj;
 
 		if (__exec_to_vma(&eb->exec[i]))
 			continue;
 
-		obj = to_intel_bo(idr_find(&eb->file->object_idr,
-					   eb->exec[i].handle));
+		obj = to_intel_bo(idr_find(idr, eb->exec[i].handle));
 		if (unlikely(!obj)) {
 			spin_unlock(&eb->file->table_lock);
 			DRM_DEBUG("Invalid object handle %d at index %d\n",
 				  eb->exec[i].handle, i);
-			return -ENOENT;
+			err = -ENOENT;
+			goto err;
 		}
 
 		__exec_to_vma(&eb->exec[i]) = INTERMEDIATE | (uintptr_t)obj;
@@ -251,7 +737,7 @@ next_vma: ;
 	for (i = slow_pass; i < count; i++) {
 		struct drm_i915_gem_object *obj;
 
-		if ((__exec_to_vma(&eb->exec[i]) & INTERMEDIATE) == 0)
+		if (!(__exec_to_vma(&eb->exec[i]) & INTERMEDIATE))
 			continue;
 
 		/*
@@ -262,12 +748,13 @@ next_vma: ;
 		 * from the (obj, vm) we don't run the risk of creating
 		 * duplicated vmas for the same vm.
 		 */
-		obj = u64_to_ptr(struct drm_i915_gem_object,
+		obj = u64_to_ptr(typeof(*obj),
 				 __exec_to_vma(&eb->exec[i]) & ~INTERMEDIATE);
 		vma = i915_vma_instance(obj, eb->vm, NULL);
 		if (unlikely(IS_ERR(vma))) {
 			DRM_DEBUG("Failed to lookup VMA\n");
-			return PTR_ERR(vma);
+			err = PTR_ERR(vma);
+			goto err;
 		}
 
 		/* First come, first served */
@@ -275,32 +762,31 @@ next_vma: ;
 			vma->ctx = eb->ctx;
 			vma->ctx_handle = eb->exec[i].handle;
 			hlist_add_head(&vma->ctx_node,
-				       ht_head(eb->ctx, eb->exec[i].handle));
-			eb->ctx->vma_lut.ht_count++;
+				       ht_head(lut, eb->exec[i].handle));
+			lut->ht_count++;
+			lut->ht_size |= I915_CTX_RESIZE_IN_PROGRESS;
 			if (i915_vma_is_ggtt(vma)) {
 				GEM_BUG_ON(obj->vma_hashed);
 				obj->vma_hashed = vma;
 			}
 		}
 
-		if (!eb_add_vma(eb, vma, i))
-			return -EINVAL;
+		err = eb_add_vma(eb, &eb->exec[i], vma);
+		if (unlikely(err))
+			goto err;
 	}
 
-	if (ht_needs_resize(eb->ctx)) {
-		eb->ctx->vma_lut.ht_size |= I915_CTX_RESIZE_IN_PROGRESS;