drivers/gpu/drm/i915/i915_gem.c

/*
 * Copyright © 2008 Intel Corporation
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 * IN THE SOFTWARE.
 *
 * Authors:
 *    Eric Anholt <eric@anholt.net>
 *
 */

#include "drmP.h"
#include "drm.h"
#include "i915_drm.h"
#include "i915_drv.h"
#include <linux/swap.h>

#define I915_GEM_GPU_DOMAINS	(~(I915_GEM_DOMAIN_CPU | I915_GEM_DOMAIN_GTT))

static int
i915_gem_object_set_domain(struct drm_gem_object *obj,
			    uint32_t read_domains,
			    uint32_t write_domain);
static int
i915_gem_object_set_domain_range(struct drm_gem_object *obj,
				 uint64_t offset,
				 uint64_t size,
				 uint32_t read_domains,
				 uint32_t write_domain);
static int
i915_gem_set_domain(struct drm_gem_object *obj,
		    struct drm_file *file_priv,
		    uint32_t read_domains,
		    uint32_t write_domain);
static int i915_gem_object_set_to_gtt_domain(struct drm_gem_object *obj,
					     int write);
static int i915_gem_object_get_page_list(struct drm_gem_object *obj);
static void i915_gem_object_free_page_list(struct drm_gem_object *obj);
static int i915_gem_object_wait_rendering(struct drm_gem_object *obj);

static void
i915_gem_cleanup_ringbuffer(struct drm_device *dev);

int
i915_gem_init_ioctl(struct drm_device *dev, void *data,
		    struct drm_file *file_priv)
{
	drm_i915_private_t *dev_priv = dev->dev_private;
	struct drm_i915_gem_init *args = data;

	mutex_lock(&dev->struct_mutex);

	if (args->gtt_start >= args->gtt_end ||
	    (args->gtt_start & (PAGE_SIZE - 1)) != 0 ||
	    (args->gtt_end & (PAGE_SIZE - 1)) != 0) {
		mutex_unlock(&dev->struct_mutex);
		return -EINVAL;
	}

	drm_mm_init(&dev_priv->mm.gtt_space, args->gtt_start,
	    args->gtt_end - args->gtt_start);

	dev->gtt_total = (uint32_t) (args->gtt_end - args->gtt_start);

	mutex_unlock(&dev->struct_mutex);

	return 0;
}

int
i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
			    struct drm_file *file_priv)
{
	struct drm_i915_gem_get_aperture *args = data;

	if (!(dev->driver->driver_features & DRIVER_GEM))
		return -ENODEV;

	args->aper_size = dev->gtt_total;
	args->aper_available_size = (args->aper_size -
				     atomic_read(&dev->pin_memory));

	return 0;
}


/**
 * Creates a new mm object and returns a handle to it.
 */
int
i915_gem_create_ioctl(struct drm_device *dev, void *data,
		      struct drm_file *file_priv)
{
	struct drm_i915_gem_create *args = data;
	struct drm_gem_object *obj;
	int handle, ret;

	args->size = roundup(args->size, PAGE_SIZE);

	/* Allocate the new object */
	obj = drm_gem_object_alloc(dev, args->size);
	if (obj == NULL)
		return -ENOMEM;

	ret = drm_gem_handle_create(file_priv, obj, &handle);
	mutex_lock(&dev->struct_mutex);
	drm_gem_object_handle_unreference(obj);
	mutex_unlock(&dev->struct_mutex);

	if (ret)
		return ret;

	args->handle = handle;

	return 0;
}

/**
 * Reads data from the object referenced by handle.
 *
 * On error, the contents of *data are undefined.
 */
int
i915_gem_pread_ioctl(struct drm_device *dev, void *data,
		     struct drm_file *file_priv)
{
	struct drm_i915_gem_pread *args = data;
	struct drm_gem_object *obj;
	struct drm_i915_gem_object *obj_priv;
	ssize_t read;
	loff_t offset;
	int ret;

	obj = drm_gem_object_lookup(dev, file_priv, args->handle);
	if (obj == NULL)
		return -EBADF;
	obj_priv = obj->driver_private;

	/* Bounds check source.
	 *
	 * XXX: This could use review for overflow issues...
	 */
	if (args->offset > obj->size || args->size > obj->size ||
	    args->offset + args->size > obj->size) {
		drm_gem_object_unreference(obj);
		return -EINVAL;
	}

	mutex_lock(&dev->struct_mutex);

	ret = i915_gem_object_set_domain_range(obj, args->offset, args->size,
					       I915_GEM_DOMAIN_CPU, 0);
	if (ret != 0) {
		drm_gem_object_unreference(obj);
		mutex_unlock(&dev->struct_mutex);
		return ret;
	}

	offset = args->offset;

	read = vfs_read(obj->filp, (char __user *)(uintptr_t)args->data_ptr,
			args->size, &offset);
	if (read != args->size) {
		drm_gem_object_unreference(obj);
		mutex_unlock(&dev->struct_mutex);
		if (read < 0)
			return read;
		else
			return -EINVAL;
	}

	drm_gem_object_unreference(obj);
	mutex_unlock(&dev->struct_mutex);

	return 0;
}

/* This is the fast write path which cannot handle
 * page faults in the source data
 */

static inline int
fast_user_write(struct io_mapping *mapping,
		loff_t page_base, int page_offset,
		char __user *user_data,
		int length)
{
	char *vaddr_atomic;
	unsigned long unwritten;

	vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
	unwritten = __copy_from_user_inatomic_nocache(vaddr_atomic + page_offset,
						      user_data, length);
	io_mapping_unmap_atomic(vaddr_atomic);
	if (unwritten)
		return -EFAULT;
	return 0;
}

/* Here's the write path which can sleep for
 * page faults
 */

static inline int
slow_user_write(struct io_mapping *mapping,
		loff_t page_base, int page_offset,
		char __user *user_data,
		int length)
{
	char __iomem *vaddr;
	unsigned long unwritten;

	vaddr = io_mapping_map_wc(mapping, page_base);
	if (vaddr == NULL)
		return -EFAULT;
	unwritten = __copy_from_user(vaddr + page_offset,
				     user_data, length);
	io_mapping_unmap(vaddr);
	if (unwritten)
		return -EFAULT;
	return 0;
}

static int
i915_gem_gtt_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
		    struct drm_i915_gem_pwrite *args,
		    struct drm_file *file_priv)
{
	struct drm_i915_gem_object *obj_priv = obj->driver_private;
	drm_i915_private_t *dev_priv = dev->dev_private;
	ssize_t remain;
	loff_t offset, page_base;
	char __user *user_data;
	int page_offset, page_length;
	int ret;

	user_data = (char __user *) (uintptr_t) args->data_ptr;
	remain = args->size;
	if (!access_ok(VERIFY_READ, user_data, remain))
		return -EFAULT;


	mutex_lock(&dev->struct_mutex);
	ret = i915_gem_object_pin(obj, 0);
	if (ret) {
		mutex_unlock(&dev->struct_mutex);
		return ret;
	}
	ret = i915_gem_object_set_to_gtt_domain(obj, 1);
	if (ret)
		goto fail;

	obj_priv = obj->driver_private;
	offset = obj_priv->gtt_offset + args->offset;
	obj_priv->dirty = 1;

	while (remain > 0) {
		/* Operation in this page
		 *
		 * page_base = page offset within aperture
		 * page_offset = offset within page
		 * page_length = bytes to copy for this page
		 */
		page_base = (offset & ~(PAGE_SIZE-1));
		page_offset = offset & (PAGE_SIZE-1);
		page_length = remain;
		if ((page_offset + remain) > PAGE_SIZE)
			page_length = PAGE_SIZE - page_offset;

		ret = fast_user_write (dev_priv->mm.gtt_mapping, page_base,
				       page_offset, user_data, page_length);

		/* If we get a fault while copying data, then (presumably) our
		 * source page isn't available. In this case, use the
		 * non-atomic function
		 */
		if (ret) {
			ret = slow_user_write (dev_priv->mm.gtt_mapping,
					       page_base, page_offset,
					       user_data, page_length);
			if (ret)
				goto fail;
		}

		remain -= page_length;
		user_data += page_length;
		offset += page_length;
	}

fail:
	i915_gem_object_unpin(obj);
	mutex_unlock(&dev->struct_mutex);

	return ret;
}

static int
i915_gem_shmem_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
		      struct drm_i915_gem_pwrite *args,
		      struct drm_file *file_priv)
{
	int ret;
	loff_t offset;
	ssize_t written;

	mutex_lock(&dev->struct_mutex);

	ret = i915_gem_set_domain(obj, file_priv,
				  I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
	if (ret) {
		mutex_unlock(&dev->struct_mutex);
		return ret;
	}

	offset = args->offset;

	written = vfs_write(obj->filp,
			    (char __user *)(uintptr_t) args->data_ptr,
			    args->size, &offset);
	if (written != args->size) {
		mutex_unlock(&dev->struct_mutex);
		if (written < 0)
			return written;
		else
			return -EINVAL;
	}

	mutex_unlock(&dev->struct_mutex);

	return 0;
}

/**
 * Writes data to the object referenced by handle.
 *
 * On error, the contents of the buffer that were to be modified are undefined.
 */
int
i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
		      struct drm_file *file_priv)
{
	struct drm_i915_gem_pwrite *args = data;
	struct drm_gem_object *obj;
	struct drm_i915_gem_object *obj_priv;
	int ret = 0;

	obj = drm_gem_object_lookup(dev, file_priv, args->handle);
	if (obj == NULL)
		return -EBADF;
	obj_priv = obj->driver_private;

	/* Bounds check destination.
	 *
	 * XXX: This could use review for overflow issues...
	 */
	if (args->offset > obj->size || args->size > obj->size ||
	    args->offset + args->size > obj->size) {
		drm_gem_object_unreference(obj);
		return -EINVAL;
	}

	/* We can only do the GTT pwrite on untiled buffers, as otherwise
	 * it would end up going through the fenced access, and we'll get
	 * different detiling behavior between reading and writing.
	 * pread/pwrite currently are reading and writing from the CPU
	 * perspective, requiring manual detiling by the client.
	 */
	if (obj_priv->tiling_mode == I915_TILING_NONE &&
	    dev->gtt_total != 0)
		ret = i915_gem_gtt_pwrite(dev, obj, args, file_priv);
	else
		ret = i915_gem_shmem_pwrite(dev, obj, args, file_priv);

#if WATCH_PWRITE
	if (ret)
		DRM_INFO("pwrite failed %d\n", ret);
#endif

	drm_gem_object_unreference(obj);

	return ret;
}

/**
 * Called when user space prepares to use an object with the CPU, either
 * through the mmap ioctl's mapping or a GTT mapping.
 */
int
i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
			  struct drm_file *file_priv)
{
	struct drm_i915_gem_set_domain *args = data;
	struct drm_gem_object *obj;
	uint32_t read_domains = args->read_domains;
	uint32_t write_domain = args->write_domain;
	int ret;

	if (!(dev->driver->driver_features & DRIVER_GEM))
		return -ENODEV;

	/* Only handle setting domains to types used by the CPU. */
	if (write_domain & ~(I915_GEM_DOMAIN_CPU | I915_GEM_DOMAIN_GTT))
		return -EINVAL;

	if (read_domains & ~(I915_GEM_DOMAIN_CPU | I915_GEM_DOMAIN_GTT))
		return -EINVAL;

	/* Having something in the write domain implies it's in the read
	 * domain, and only that read domain.  Enforce that in the request.
	 */
	if (write_domain != 0 && read_domains != write_domain)
		return -EINVAL;

	obj = drm_gem_object_lookup(dev, file_priv, args->handle);
	if (obj == NULL)
		return -EBADF;

	mutex_lock(&dev->struct_mutex);
#if WATCH_BUF
	DRM_INFO("set_domain_ioctl %p(%d), %08x %08x\n",
		 obj, obj->size, read_domains, write_domain);
#endif
	if (read_domains & I915_GEM_DOMAIN_GTT) {
		ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
	} else {
		ret = i915_gem_set_domain(obj, file_priv,
					  read_domains, write_domain);
	}

	drm_gem_object_unreference(obj);
	mutex_unlock(&dev->struct_mutex);
	return ret;
}

/**
 * Called when user space has done writes to this buffer
 */
int
i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
		      struct drm_file *file_priv)
{
	struct drm_i915_gem_sw_finish *args = data;
	struct drm_gem_object *obj;
	struct drm_i915_gem_object *obj_priv;
	int ret = 0;

	if (!(dev->driver->driver_features & DRIVER_GEM))
		return -ENODEV;

	mutex_lock(&dev->struct_mutex);
	obj = drm_gem_object_lookup(dev, file_priv, args->handle);
	if (obj == NULL) {
		mutex_unlock(&dev->struct_mutex);
		return -EBADF;
	}

#if WATCH_BUF
	DRM_INFO("%s: sw_finish %d (%p %d)\n",
		 __func__, args->handle, obj, obj->size);
#endif
	obj_priv = obj->driver_private;

	/* Pinned buffers may be scanout, so flush the cache */
	if ((obj->write_domain & I915_GEM_DOMAIN_CPU) && obj_priv->pin_count) {
		i915_gem_clflush_object(obj);
		drm_agp_chipset_flush(dev);
	}
	drm_gem_object_unreference(obj);
	mutex_unlock(&dev->struct_mutex);
	return ret;
}

/**
 * Maps the contents of an object, returning the address it is mapped
 * into.
 *
 * While the mapping holds a reference on the contents of the object, it doesn't
 * imply a ref on the object itself.
 */
int
i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
		   struct drm_file *file_priv)
{
	struct drm_i915_gem_mmap *args = data;
	struct drm_gem_object *obj;
	loff_t offset;
	unsigned long addr;

	if (!(dev->driver->driver_features & DRIVER_GEM))
		return -ENODEV;

	obj = drm_gem_object_lookup(dev, file_priv, args->handle);
	if (obj == NULL)
		return -EBADF;

	offset = args->offset;

	down_write(&current->mm->mmap_sem);
	addr = do_mmap(obj->filp, 0, args->size,
		       PROT_READ | PROT_WRITE, MAP_SHARED,
		       args->offset);
	up_write(&current->mm->mmap_sem);
	mutex_lock(&dev->struct_mutex);
	drm_gem_object_unreference(obj);
	mutex_unlock(&dev->struct_mutex);
	if (IS_ERR((void *)addr))
		return addr;

	args->addr_ptr = (uint64_t) addr;

	return 0;
}

static void
i915_gem_object_free_page_list(struct drm_gem_object *obj)
{
	struct drm_i915_gem_object *obj_priv = obj->driver_private;
	int page_count = obj->size / PAGE_SIZE;
	int i;

	if (obj_priv->page_list == NULL)
		return;


	for (i = 0; i < page_count; i++)
		if (obj_priv->page_list[i] != NULL) {
			if (obj_priv->dirty)
				set_page_dirty(obj_priv->page_list[i]);
			mark_page_accessed(obj_priv->page_list[i]);
			page_cache_release(obj_priv->page_list[i]);
		}
	obj_priv->dirty = 0;

	drm_free(obj_priv->page_list,
		 page_count * sizeof(struct page *),
		 DRM_MEM_DRIVER);
	obj_priv->page_list = NULL;
}

static void
i915_gem_object_move_to_active(struct drm_gem_object *obj, uint32_t seqno)
{
	struct drm_device *dev = obj->dev;
	drm_i915_private_t *dev_priv = dev->dev_private;
	struct drm_i915_gem_object *obj_priv = obj->driver_private;

	/* Add a reference if we're newly entering the active list. */
	if (!obj_priv->active) {
		drm_gem_object_reference(obj);
		obj_priv->active = 1;
	}
	/* Move from whatever list we were on to the tail of execution. */
	list_move_tail(&obj_priv->list,
		       &dev_priv->mm.active_list);
	obj_priv->last_rendering_seqno = seqno;
}

static void
i915_gem_object_move_to_flushing(struct drm_gem_object *obj)
{
	struct drm_device *dev = obj->dev;
	drm_i915_private_t *dev_priv = dev->dev_private;
	struct drm_i915_gem_object *obj_priv = obj->driver_private;

	BUG_ON(!obj_priv->active);
	list_move_tail(&obj_priv->list, &dev_priv->mm.flushing_list);
	obj_priv->last_rendering_seqno = 0;
}

static void
i915_gem_object_move_to_inactive(struct drm_gem_object *obj)
{
	struct drm_device *dev = obj->dev;
	drm_i915_private_t *dev_priv = dev->dev_private;
	struct drm_i915_gem_object *obj_priv = obj->driver_private;

	i915_verify_inactive(dev, __FILE__, __LINE__);
	if (obj_priv->pin_count != 0)
		list_del_init(&obj_priv->list);
	else
		list_move_tail(&obj_priv->list, &dev_priv->mm.inactive_list);

	obj_priv->last_rendering_seqno = 0;
	if (obj_priv->active) {
		obj_priv->active = 0;
		drm_gem_object_unreference(obj);
	}
	i915_verify_inactive(dev, __FILE__, __LINE__);
}

/**
 * Creates a new sequence number, emitting a write of it to the status page
 * plus an interrupt, which will trigger i915_user_interrupt_handler.
 *
 * Must be called with struct_lock held.
 *
 * Returned sequence numbers are nonzero on success.
 */
static uint32_t
i915_add_request(struct drm_device *dev, uint32_t flush_domains)
{
	drm_i915_private_t *dev_priv = dev->dev_private;
	struct drm_i915_gem_request *request;
	uint32_t seqno;
	int was_empty;
	RING_LOCALS;

	request = drm_calloc(1, sizeof(*request), DRM_MEM_DRIVER);
	if (request == NULL)
		return 0;

	/* Grab the seqno we're going to make this request be, and bump the
	 * next (skipping 0 so it can be the reserved no-seqno value).
	 */
	seqno = dev_priv->mm.next_gem_seqno;
	dev_priv->mm.next_gem_seqno++;
	if (dev_priv->mm.next_gem_seqno == 0)
		dev_priv->mm.next_gem_seqno++;

	BEGIN_LP_RING(4);
	OUT_RING(MI_STORE_DWORD_INDEX);
	OUT_RING(I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
	OUT_RING(seqno);

	OUT_RING(MI_USER_INTERRUPT);
	ADVANCE_LP_RING();

	DRM_DEBUG("%d\n", seqno);

	request->seqno = seqno;
	request->emitted_jiffies = jiffies;
	was_empty = list_empty(&dev_priv->mm.request_list);
	list_add_tail(&request->list, &dev_priv->mm.request_list);

	/* Associate any objects on the flushing list matching the write
	 * domain we're flushing with our flush.
	 */
	if (flush_domains != 0) {
		struct drm_i915_gem_object *obj_priv, *next;

		list_for_each_entry_safe(obj_priv, next,
					 &dev_priv->mm.flushing_list, list) {
			struct drm_gem_object *obj = obj_priv->obj;

			if ((obj->write_domain & flush_domains) ==
			    obj->write_domain) {
				obj->write_domain = 0;
				i915_gem_object_move_to_active(obj, seqno);
			}
		}

	}

	if (was_empty && !dev_priv->mm.suspended)
		schedule_delayed_work(&dev_priv->mm.retire_work, HZ);
	return seqno;
}

/**
 * Command execution barrier
 *
 * Ensures that all commands in the ring are finished
 * before signalling the CPU
 */
static uint32_t
i915_retire_commands(struct drm_device *dev)
{
	drm_i915_private_t *dev_priv = dev->dev_private;
	uint32_t cmd = MI_FLUSH | MI_NO_WRITE_FLUSH;
	uint32_t flush_domains = 0;
	RING_LOCALS;

	/* The sampler always gets flushed on i965 (sigh) */
	if (IS_I965G(dev))
		flush_domains |= I915_GEM_DOMAIN_SAMPLER;
	BEGIN_LP_RING(2);
	OUT_RING(cmd);
	OUT_RING(0); /* noop */
	ADVANCE_LP_RING();
	return flush_domains;
}

/**
 * Moves buffers associated only with the given active seqno from the active
 * to inactive list, potentially freeing them.
 */
static void
i915_gem_retire_request(struct drm_device *dev,
			struct drm_i915_gem_request *request)
{
	drm_i915_private_t *dev_priv = dev->dev_private;

	/* Move any buffers on the active list that are no longer referenced
	 * by the ringbuffer to the flushing/inactive lists as appropriate.
	 */
	while (!list_empty(&dev_priv->mm.active_list)) {
		struct drm_gem_object *obj;
		struct drm_i915_gem_object *obj_priv;

		obj_priv = list_first_entry(&dev_priv->mm.active_list,
					    struct drm_i915_gem_object,
					    list);
		obj = obj_priv->obj;

		/* If the seqno being retired doesn't match the oldest in the
		 * list, then the oldest in the list must still be newer than
		 * this seqno.
		 */
		if (obj_priv->last_rendering_seqno != request->seqno)
			return;
#if WATCH_LRU
		DRM_INFO("%s: retire %d moves to inactive list %p\n",
			 __func__, request->seqno, obj);
#endif

		if (obj->write_domain != 0)
			i915_gem_object_move_to_flushing(obj);
		else
			i915_gem_object_move_to_inactive(obj);
	}
}

/**
 * Returns true if seq1 is later than seq2.
 */
static int
i915_seqno_passed(uint32_t seq1, uint32_t seq2)
{
	return (int32_t)(seq1 - seq2) >= 0;
}

uint32_t
i915_get_gem_seqno(struct drm_device *dev)
{
	drm_i915_private_t *dev_priv = dev->dev_private;

	return READ_HWSP(dev_priv, I915_GEM_HWS_INDEX);
}

/**
 * This function clears the request list as sequence numbers are passed.
 */
void
i915_gem_retire_requests(struct drm_device *dev)
{
	drm_i915_private_t *dev_priv = dev->dev_private;
	uint32_t seqno;

	seqno = i915_get_gem_seqno(dev);

	while (!list_empty(&dev_priv->mm.request_list)) {
		struct drm_i915_gem_request *request;
		uint32_t retiring_seqno;

		request = list_first_entry(&dev_priv->mm.request_list,
					   struct drm_i915_gem_request,
					   list);
		retiring_seqno = request->seqno;

		if (i915_seqno_passed(seqno, retiring_seqno) ||
		    dev_priv->mm.wedged) {
			i915_gem_retire_request(dev, request);

			list_del(&request->list);
			drm_free(request, sizeof(*request), DRM_MEM_DRIVER);
		} else
			break;
	}
}

void
i915_gem_retire_work_handler(struct work_struct *work)
{
	drm_i915_private_t *dev_priv;
	struct drm_device *dev;

	dev_priv = container_of(work, drm_i915_private_t,
				mm.retire_work.work);
	dev = dev_priv->dev;

	mutex_lock(&dev->struct_mutex);
	i915_gem_retire_requests(dev);
	if (!dev_priv->mm.suspended &&
	    !list_empty(&dev_priv->mm.request_list))
		schedule_delayed_work(&dev_priv->mm.retire_work, HZ);
	mutex_unlock(&dev->struct_mutex);
}

/**
 * Waits for a sequence number to be signaled, and cleans up the
 * request and object lists appropriately for that event.
 */
static int
i915_wait_request(struct drm_device *dev, uint32_t seqno)
{
	drm_i915_private_t *dev_priv = dev->dev_private;
	int ret = 0;

	BUG_ON(seqno == 0);

	if (!i915_seqno_passed(i915_get_gem_seqno(dev), seqno)) {
		dev_priv->mm.waiting_gem_seqno = seqno;
		i915_user_irq_get(dev);
		ret = wait_event_interruptible(dev_priv->irq_queue,
					       i915_seqno_passed(i915_get_gem_seqno(dev),
								 seqno) ||
					       dev_priv->mm.wedged);
		i915_user_irq_put(dev);
		dev_priv->mm.waiting_gem_seqno = 0;
	}
	if (dev_priv->mm.wedged)
		ret = -EIO;

	if (ret && ret != -ERESTARTSYS)
		DRM_ERROR("%s returns %d (awaiting %d at %d)\n",
			  __func__, ret, seqno, i915_get_gem_seqno(dev));

	/* Directly dispatch request retiring.  While we have the work queue
	 * to handle this, the waiter on a request often wants an associated
	 * buffer to have made it to the inactive list, and we would need
	 * a separate wait queue to handle that.
	 */
	if (ret == 0)
		i915_gem_retire_requests(dev);

	return ret;
}

static void
i915_gem_flush(struct drm_device *dev,
	       uint32_t invalidate_domains,
	       uint32_t flush_domains)
{
	drm_i915_private_t *dev_priv = dev->dev_private;
	uint32_t cmd;
	RING_LOCALS;

#if WATCH_EXEC
	DRM_INFO("%s: invalidate %08x flush %08x\n", __func__,
		  invalidate_domains, flush_domains);
#endif

	if (flush_domains & I915_GEM_DOMAIN_CPU)
		drm_agp_chipset_flush(dev);

	if ((invalidate_domains | flush_domains) & ~(I915_GEM_DOMAIN_CPU |
						     I915_GEM_DOMAIN_GTT)) {
		/*
		 * read/write caches:
		 *
		 * I915_GEM_DOMAIN_RENDER is always invalidated, but is
		 * only flushed if MI_NO_WRITE_FLUSH is unset.  On 965, it is
		 * also flushed at 2d versus 3d pipeline switches.
		 *
		 * read-only caches:
		 *
		 * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
		 * MI_READ_FLUSH is set, and is always flushed on 965.
		 *
		 * I915_GEM_DOMAIN_COMMAND may not exist?
		 *
		 * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
		 * invalidated when MI_EXE_FLUSH is set.
		 *
		 * I915_GEM_DOMAIN_VERTEX, which exists on 965, is
		 * invalidated with every MI_FLUSH.
		 *
		 * TLBs:
		 *
		 * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
		 * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
		 * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
		 * are flushed at any MI_FLUSH.
		 */

		cmd = MI_FLUSH | MI_NO_WRITE_FLUSH;
		if ((invalidate_domains|flush_domains) &
		    I915_GEM_DOMAIN_RENDER)
			cmd &= ~MI_NO_WRITE_FLUSH;
		if (!IS_I965G(dev)) {
			/*
			 * On the 965, the sampler cache always gets flushed
			 * and this bit is reserved.
			 */
			if (invalidate_domains & I915_GEM_DOMAIN_SAMPLER)
				cmd |= MI_READ_FLUSH;
		}
		if (invalidate_domains & I915_GEM_DOMAIN_INSTRUCTION)
			cmd |= MI_EXE_FLUSH;

#if WATCH_EXEC
		DRM_INFO("%s: queue flush %08x to ring\n", __func__, cmd);
#endif
		BEGIN_LP_RING(2);
		OUT_RING(cmd);
		OUT_RING(0); /* noop */
		ADVANCE_LP_RING();
	}
}

/**
 * Ensures that all rendering to the object has completed and the object is
 * safe to unbind from the GTT or access from the CPU.
 */
static int
i915_gem_object_wait_rendering(struct drm_gem_object *obj)
{
	struct drm_device *dev = obj->dev;
	struct drm_i915_gem_object *obj_priv = obj->driver_private;
	int ret;

	/* If there are writes queued to the buffer, flush and
	 * create a new seqno to wait for.
	 */
	if (obj->write_domain & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT)) {
		uint32_t seqno, write_domain = obj->write_domain;
#if WATCH_BUF
		DRM_INFO("%s: flushing object %p from write domain %08x\n",
			  __func__, obj, write_domain);
#endif
		i915_gem_flush(dev, 0, write_domain);

		seqno = i915_add_request(dev, write_domain);
		i915_gem_object_move_to_active(obj, seqno);
#if WATCH_LRU
		DRM_INFO("%s: flush moves to exec list %p\n", __func__, obj);
#endif
	}

	/* If there is rendering queued on the buffer being evicted, wait for
	 * it.
	 */
	if (obj_priv->active) {
#if WATCH_BUF
		DRM_INFO("%s: object %p wait for seqno %08x\n",
			  __func__, obj, obj_priv->last_rendering_seqno);
#endif
		ret = i915_wait_request(dev, obj_priv->last_rendering_seqno);
		if (ret != 0)
			return ret;
	}

	return 0;
}

/**
 * Unbinds an object from the GTT aperture.
 */
static int
i915_gem_object_unbind(struct drm_gem_object *obj)
{
	struct drm_device *dev = obj->dev;
	struct drm_i915_gem_object *obj_priv = obj->driver_private;
	int ret = 0;

#if WATCH_BUF
	DRM_INFO("%s:%d %p\n", __func__, __LINE__, obj);
	DRM_INFO("gtt_space %p\n", obj_priv->gtt_space);
#endif
	if (obj_priv->gtt_space == NULL)
		return 0;

	if (obj_priv->pin_count != 0) {
		DRM_ERROR("Attempting to unbind pinned buffer\n");
		return -EINVAL;
	}

	/* Wait for any rendering to complete
	 */
	ret = i915_gem_object_wait_rendering(obj);
	if (ret) {
		DRM_ERROR("wait_rendering failed: %d\n", ret);
		return ret;
	}

	/* Move the object to the CPU domain to ensure that
	 * any possible CPU writes while it's not in the GTT
	 * are flushed when we go to remap it. This will
	 * also ensure that all pending GPU writes are finished
	 * before we unbind.
	 */
	ret = i915_gem_object_set_domain(obj, I915_GEM_DOMAIN_CPU,
					 I915_GEM_DOMAIN_CPU);
	if (ret) {
		DRM_ERROR("set_domain failed: %d\n", ret);
		return ret;
	}

	if (obj_priv->agp_mem != NULL) {
		drm_unbind_agp(obj_priv->agp_mem);
		drm_free_agp(obj_priv->agp_mem, obj->size / PAGE_SIZE);
		obj_priv->agp_mem = NULL;
	}

	BUG_ON(obj_priv->active);

	i915_gem_object_free_page_list(obj);

	if (obj_priv->gtt_space) {
		atomic_dec(&dev->gtt_count);
		atomic_sub(obj->size, &dev->gtt_memory);

		drm_mm_put_block(obj_priv->gtt_space);
		obj_priv->gtt_space = NULL;
	}

	/* Remove ourselves from the LRU list if present. */
	if (!list_empty(&obj_priv->list))
		list_del_init(&obj_priv->list);

	return 0;
}

static int
i915_gem_evict_something(struct drm_device *dev)
{
	drm_i915_private_t *dev_priv = dev->dev_private;
	struct drm_gem_object *obj;
	struct drm_i915_gem_object *obj_priv;
	int ret = 0;

	for (;;) {
		/* If there's an inactive buffer available now, grab it
		 * and be done.
		 */
		if (!list_empty(&dev_priv->mm.inactive_list)) {
			obj_priv = list_first_entry(&dev_priv->mm.inactive_list,
						    struct drm_i915_gem_object,
						    list);
			obj = obj_priv->obj;
			BUG_ON(obj_priv->pin_count != 0);
#if WATCH_LRU
			DRM_INFO("%s: evicting %p\n", __func__, obj);
#endif
			BUG_ON(obj_priv->active);

			/* Wait on the rendering and unbind the buffer. */
			ret = i915_gem_object_unbind(obj);
			break;
		}

		/* If we didn't get anything, but the ring is still processing
		 * things, wait for one of those things to finish and hopefully
		 * leave us a buffer to evict.
		 */
		if (!list_empty(&dev_priv->mm.request_list)) {
			struct drm_i915_gem_request *request;

			request = list_first_entry(&dev_priv->mm.request_list,
						   struct drm_i915_gem_request,
						   list);

			ret = i915_wait_request(dev, request->seqno);
			if (ret)
				break;

			/* if waiting caused an object to become inactive,
			 * then loop around and wait for it. Otherwise, we
			 * assume that waiting freed and unbound something,
			 * so there should now be some space in the GTT
			 */
			if (!list_empty(&dev_priv->mm.inactive_list))
				continue;
			break;
		}

		/* If we didn't have anything on the request list but there
		 * are buffers awaiting a flush, emit one and try again.
		 * When we wait on it, those buffers waiting for that flush
		 * will get moved to inactive.
		 */
		if (!list_empty(&dev_priv->mm.flushing_list)) {
			obj_priv = list_first_entry(&dev_priv->mm.flushing_list,
						    struct drm_i915_gem_object,
						    list);
			obj = obj_priv->obj;

			i915_gem_flush(dev,
				       obj->write_domain,
				       obj->write_domain);
			i915_add_request(dev, obj->write_domain);

			obj = NULL;
			continue;
		}

		DRM_ERROR("inactive empty %d request empty %d "
			  "flushing empty %d\n",
			  list_empty(&dev_priv->mm.inactive_list),
			  list_empty(&dev_priv->mm.request_list),
			  list_empty(&dev_priv->mm.flushing_list));
		/* If we didn't do any of the above, there's nothing to be done
		 * and we just can't fit it in.
		 */
		return -ENOMEM;
	}
	return ret;
}

static int
i915_gem_object_get_page_list(struct drm_gem_object *obj)
{
	struct drm_i915_gem_object *obj_priv = obj->driver_private;
	int page_count, i;
	struct address_space *mapping;
	struct inode *inode;
	struct page *page;
	int ret;

	if (obj_priv->page_list)
		return 0;

	/* Get the list of pages out of our struct file.  They'll be pinned
	 * at this point until we release them.
	 */
	page_count = obj->size / PAGE_SIZE;
	BUG_ON(obj_priv->page_list != NULL);
	obj_priv->page_list = drm_calloc(page_count, sizeof(struct page *),
					 DRM_MEM_DRIVER);
	if (obj_priv->page_list == NULL) {
		DRM_ERROR("Faled to allocate page list\n");
		return -ENOMEM;
	}

	inode = obj->filp->f_path.dentry->d_inode;
	mapping = inode->i_mapping;
	for (i = 0; i < page_count; i++) {
		page = read_mapping_page(mapping, i, NULL);
		if (IS_ERR(page)) {
			ret = PTR_ERR(page);
			DRM_ERROR("read_mapping_page failed: %d\n", ret);
			i915_gem_object_free_page_list(obj);
			return ret;
		}
		obj_priv->page_list[i] = page;
	}
	return 0;
}

/**
 * Finds free space in the GTT aperture and binds the object there.
 */
static int
i915_gem_object_bind_to_gtt(struct drm_gem_object *obj, unsigned alignment)
{
	struct drm_device *dev = obj->dev;
	drm_i915_private_t *dev_priv = dev->dev_private;
	struct drm_i915_gem_object *obj_priv = obj->driver_private;
	struct drm_mm_node *free_space;
	int page_count, ret;

	if (alignment == 0)
		alignment = PAGE_SIZE;
	if (alignment & (PAGE_SIZE - 1)) {
		DRM_ERROR("Invalid object alignment requested %u\n", alignment);
		return -EINVAL;
	}

 search_free:
	free_space = drm_mm_search_free(&dev_priv->mm.gtt_space,
					obj->size, alignment, 0);
	if (free_space != NULL) {
		obj_priv->gtt_space = drm_mm_get_block(free_space, obj->size,
						       alignment);
		if (obj_priv->gtt_space != NULL) {
			obj_priv->gtt_space->private = obj;
			obj_priv->gtt_offset = obj_priv->gtt_space->start;
		}
	}
	if (obj_priv->gtt_space == NULL) {
		/* If the gtt is empty and we're still having trouble
		 * fitting our object in, we're out of memory.
		 */
#if WATCH_LRU
		DRM_INFO("%s: GTT full, evicting something\n", __func__);
#endif
		if (list_empty(&dev_priv->mm.inactive_list) &&
		    list_empty(&dev_priv->mm.flushing_list) &&
		    list_empty(&dev_priv->mm.active_list)) {
			DRM_ERROR("GTT full, but LRU list empty\n");
			return -ENOMEM;
		}

		ret = i915_gem_evict_something(dev);
		if (ret != 0) {
			DRM_ERROR("Failed to evict a buffer %d\n", ret);
			return ret;
		}
		goto search_free;
	}

#if WATCH_BUF
	DRM_INFO("Binding object of size %d at 0x%08x\n",
		 obj->size, obj_priv->gtt_offset);
#endif
	ret = i915_gem_object_get_page_list(obj);
	if (ret) {
		drm_mm_put_block(obj_priv->gtt_space);
		obj_priv->gtt_space = NULL;
		return ret;
	}

	page_count = obj->size / PAGE_SIZE;
	/* Create an AGP memory structure pointing at our pages, and bind it
	 * into the GTT.
	 */
	obj_priv->agp_mem = drm_agp_bind_pages(dev,
					       obj_priv->page_list,
					       page_count,
					       obj_priv->gtt_offset,
					       obj_priv->agp_type);
	if (obj_priv->agp_mem == NULL) {
		i915_gem_object_free_page_list(obj);
		drm_mm_put_block(obj_priv->gtt_space);
		obj_priv->gtt_space = NULL;
		return -ENOMEM;
	}
	atomic_inc(&dev->gtt_count);
	atomic_add(obj->size, &dev->gtt_memory);

	/* Assert that the object is not currently in any GPU domain. As it
	 * wasn't in the GTT, there shouldn't be any way it could have been in
	 * a GPU cache
	 */
	BUG_ON(obj->read_domains & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT));
	BUG_ON(obj->write_domain & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT));

	return 0;
}

void
i915_gem_clflush_object(struct drm_gem_object *obj)
{
	struct drm_i915_gem_object	*obj_priv = obj->driver_private;

	/* If we don't have a page list set up, then we're not pinned
	 * to GPU, and we can ignore the cache flush because it'll happen
	 * again at bind time.
	 */
	if (obj_priv->page_list == NULL)
		return;

	drm_clflush_pages(obj_priv->page_list, obj->size / PAGE_SIZE);
}

/**
 * Moves a single object to the GTT read, and possibly write domain.
 *
 * This function returns when the move is complete, including waiting on
 * flushes to occur.
 */
static int
i915_gem_object_set_to_gtt_domain(struct drm_gem_object *obj, int write)
{
	struct drm_device *dev = obj->dev;
	struct drm_i915_gem_object *obj_priv = obj->driver_private;
	uint32_t flush_domains;

	/* Figure out what GPU domains we need to flush or invalidate for
	 * moving to GTT.
	 */
	flush_domains = obj->write_domain & I915_GEM_GPU_DOMAINS;

	/* Queue the GPU write cache flushing we need. */
	if (flush_domains != 0) {
		uint32_t seqno;

		obj->write_domain &= ~I915_GEM_GPU_DOMAINS;
		i915_gem_flush(dev, 0, flush_domains);
		seqno = i915_add_request(dev, flush_domains);
		i915_gem_object_move_to_active(obj, seqno);
	}

	/* Wait on any GPU rendering and flushing to occur. */
	if (obj_priv->active) {
		int ret;

		ret = i915_wait_request(dev, obj_priv->last_rendering_seqno);
		if (ret != 0)
			return ret;
	}

	/* If we're writing through the GTT domain, then CPU and GPU caches
	 * will need to be invalidated at next use.
	 */
	if (write)
		obj->read_domains &= ~(I915_GEM_GPU_DOMAINS |
					    I915_GEM_DOMAIN_CPU);

	/* Flush the CPU domain if it's dirty. */
	if (obj->write_domain & I915_GEM_DOMAIN_CPU) {
		i915_gem_clflush_object(obj);
		drm_agp_chipset_flush(dev);

		obj->write_domain &= ~I915_GEM_DOMAIN_CPU;
	}

	/* It should now be out of any other write domains, and we can update
	 * the domain values for our changes.
	 */
	BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
	obj->read_domains |= I915_GEM_DOMAIN_GTT;
	if (write)
		obj->write_domain = I915_GEM_DOMAIN_GTT;

	return 0;
}

/*
 * Set the next domain for the specified object. This
 * may not actually perform the necessary flushing/invaliding though,
 * as that may want to be batched with other set_domain operations
 *
 * This is (we hope) the only really tricky part of gem. The goal
 * is fairly simple -- track which caches hold bits of the object
 * and make sure they remain coherent. A few concrete examples may
 * help to explain how it works. For shorthand, we use the notation
 * (read_domains, write_domain), e.g. (CPU, CPU) to indicate the
 * a pair of read and write domain masks.
 *
 * Case 1: the batch buffer
 *
 *	1. Allocated
 *	2. Written by CPU
 *	3. Mapped to GTT
 *	4. Read by GPU
 *	5. Unmapped from GTT
 *	6. Freed
 *
 *	Let's take these a step at a time
 *
 *	1. Allocated
 *		Pages allocated from the kernel may still have
 *		cache contents, so we set them to (CPU, CPU) always.
 *	2. Written by CPU (using pwrite)
 *		The pwrite function calls set_domain (CPU, CPU) and
 *		this function does nothing (as nothing changes)
 *	3. Mapped by GTT
 *		This function asserts that the object is not
 *		currently in any GPU-based read or write domains
 *	4. Read by GPU
 *		i915_gem_execbuffer calls set_domain (COMMAND, 0).
 *		As write_domain is zero, this function adds in the
 *		current read domains (CPU+COMMAND, 0).
 *		flush_domains is set to CPU.
 *		invalidate_domains is set to COMMAND
 *		clflush is run to get data out of the CPU caches
 *		then i915_dev_set_domain calls i915_gem_flush to
 *		emit an MI_FLUSH and drm_agp_chipset_flush
 *	5. Unmapped from GTT
 *		i915_gem_object_unbind calls set_domain (CPU, CPU)
 *		flush_domains and invalidate_domains end up both zero
 *		so no flushing/invalidating happens
 *	6. Freed
 *		yay, done
 *
 * Case 2: The shared render buffer
 *
 *	1. Allocated
 *	2. Mapped to GTT
 *	3. Read/written by GPU
 *	4. set_domain to (CPU,CPU)
 *	5. Read/written by CPU
 *	6. Read/written by GPU
 *
 *	1. Allocated
 *		Same as last example, (CPU, CPU)
 *	2. Mapped to GTT
 *		Nothing changes (assertions find that it is not in the GPU)
 *	3. Read/written by GPU
 *		execbuffer calls set_domain (RENDER, RENDER)
 *		flush_domains gets CPU
 *		invalidate_domains gets GPU
 *		clflush (obj)
 *		MI_FLUSH and drm_agp_chipset_flush
 *	4. set_domain (CPU, CPU)
 *		flush_domains gets GPU
 *		invalidate_domains gets CPU
 *		wait_rendering (obj) to make sure all drawing is complete.
 *		This will include an MI_FLUSH to get the data from GPU
 *		to memory
 *		clflush (obj) to invalidate the CPU cache
 *		Another MI_FLUSH in i915_gem_flush (eliminate this somehow?)
 *	5. Read/written by CPU
 *		cache lines are loaded and dirtied
 *	6. Read written by GPU
 *		Same as last GPU access
 *
 * Case 3: The constant buffer
 *
 *	1. Allocated
 *	2. Written by CPU
 *	3. Read by GPU
 *	4. Updated (written) by CPU again
 *	5. Read by GPU
 *
 *	1. Allocated
 *		(CPU, CPU)
 *	2. Written by CPU
 *		(CPU, CPU)
 *	3. Read by GPU
 *		(CPU+RENDER, 0)
 *		flush_domains = CPU
 *		invalidate_domains = RENDER
 *		clflush (obj)
 *		MI_FLUSH
 *		drm_agp_chipset_flush
 *	4. Updated (written) by CPU again
 *		(CPU, CPU)
 *		flush_domains = 0 (no previous write domain)
 *		invalidate_domains = 0 (no new read domains)
 *	5. Read by GPU
 *		(CPU+RENDER, 0)
 *		flush_domains = CPU
 *		invalidate_domains = RENDER
 *		clflush (obj)
 *		MI_FLUSH
 *		drm_agp_chipset_flush
 */
static int
i915_gem_object_set_domain(struct drm_gem_object *obj,
			    uint32_t read_domains,
			    uint32_t write_domain)
{
	struct drm_device		*dev = obj->dev;
	struct drm_i915_gem_object	*obj_priv = obj->driver_private;
	uint32_t			invalidate_domains = 0;
	uint32_t			flush_domains = 0;
	int				ret;

#if WATCH_BUF
	DRM_INFO("%s: object %p read %08x -> %08x write %08x -> %08x\n",
		 __func__, obj,
		 obj->read_domains, read_domains,
		 obj->write_domain, write_domain);
#endif
	/*
	 * If the object isn't moving to a new write domain,
	 * let the object stay in multiple read domains
	 */
	if (write_domain == 0)
		read_domains |= obj->read_domains;
	else
		obj_priv->dirty = 1;

	/*
	 * Flush the current write domain if
	 * the new read domains don't match. Invalidate
	 * any read domains which differ from the old
	 * write domain
	 */
	if (obj->write_domain && obj->write_domain != read_domains) {
		flush_domains |= obj->write_domain;
		invalidate_domains |= read_domains & ~obj->write_domain;
	}
	/*
	 * Invalidate any read caches which may have
	 * stale data. That is, any new read domains.
	 */
	invalidate_domains |= read_domains & ~obj->read_domains;
	if ((flush_domains | invalidate_domains) & I915_GEM_DOMAIN_CPU) {
#if WATCH_BUF
		DRM_INFO("%s: CPU domain flush %08x invalidate %08x\n",
			 __func__, flush_domains, invalidate_domains);
#endif
		/*
		 * If we're invaliding the CPU cache and flushing a GPU cache,
		 * then pause for rendering so that the GPU caches will be
		 * flushed before the cpu cache is invalidated
		 */
		if ((invalidate_domains & I915_GEM_DOMAIN_CPU) &&
		    (flush_domains & ~(I915_GEM_DOMAIN_CPU |
				       I915_GEM_DOMAIN_GTT))) {
			ret = i915_gem_object_wait_rendering(obj);
			if (ret)
				return ret;
		}
		i915_gem_clflush_object(obj);
	}

	if ((write_domain | flush_domains) != 0)
		obj->write_domain = write_domain;

	/* If we're invalidating the CPU domain, clear the per-page CPU
	 * domain list as well.
	 */
	if (obj_priv->page_cpu_valid != NULL &&
	    (write_domain != 0 ||
	     read_domains & I915_GEM_DOMAIN_CPU)) {
		drm_free(obj_priv->page_cpu_valid, obj->size / PAGE_SIZE,
			 DRM_MEM_DRIVER);
		obj_priv->page_cpu_valid = NULL;
	}
	obj->read_domains = read_domains;

	dev->invalidate_domains |= invalidate_domains;
	dev->flush_domains |= flush_domains;
#if WATCH_BUF
	DRM_INFO("%s: read %08x write %08x invalidate %08x flush %08x\n",
		 __func__,
		 obj->read_domains, obj->write_domain,
		 dev->invalidate_domains, dev->flush_domains);
#endif
	return 0;
}

/**
 * Set the read/write domain on a range of the object.
 *
 * Currently only implemented for CPU reads, otherwise drops to normal
 * i915_gem_object_set_domain().
 */
static int
i915_gem_object_set_domain_range(struct drm_gem_object *obj,
				 uint64_t offset,
				 uint64_t size,
				 uint32_t read_domains,
				 uint32_t write_domain)
{
	struct drm_i915_gem_object *obj_priv = obj->driver_private;
	int ret, i;

	if (obj->read_domains & I915_GEM_DOMAIN_CPU)
		return 0;

	if (read_domains != I915_GEM_DOMAIN_CPU ||
	    write_domain != 0)
		return i915_gem_object_set_domain(obj,
						  read_domains, write_domain);

	/* Wait on any GPU rendering to the object to be flushed. */
	ret = i915_gem_object_wait_rendering(obj);
	if (ret)
		return ret;

	if (obj_priv->page_cpu_valid == NULL) {
		obj_priv->page_cpu_valid = drm_calloc(1, obj->size / PAGE_SIZE,
						      DRM_MEM_DRIVER);
	}

	/* Flush the cache on any pages that are still invalid from the CPU's
	 * perspective.
	 */
	for (i = offset / PAGE_SIZE; i <= (offset + size - 1) / PAGE_SIZE; i++) {
		if (obj_priv->page_cpu_valid[i])
			continue;

		drm_clflush_pages(obj_priv->page_list + i, 1);

		obj_priv->page_cpu_valid[i] = 1;
	}

	return 0;
}

/**
 * Once all of the objects have been set in the proper domain,
 * perform the necessary flush and invalidate operations.
 *
 * Returns the write domains flushed, for use in flush tracking.
 */
static uint32_t
i915_gem_dev_set_domain(struct drm_device *dev)
{
	uint32_t flush_domains = dev->flush_domains;

	/*
	 * Now that all the buffers are synced to the proper domains,
	 * flush and invalidate the collected domains
	 */
	if (dev->invalidate_domains | dev->flush_domains) {
#if WATCH_EXEC
		DRM_INFO("%s: invalidate_domains %08x flush_domains %08x\n",
			  __func__,
			 dev->invalidate_domains,
			 dev->flush_domains);
#endif
		i915_gem_flush(dev,
			       dev->invalidate_domains,
			       dev->flush_domains);
		dev->invalidate_domains = 0;
		dev->flush_domains = 0;
	}

	return flush_domains;
}

/**
 * Pin an object to the GTT and evaluate the relocations landing in it.
 */
static int
i915_gem_object_pin_and_relocate(struct drm_gem_object *obj,
				 struct drm_file *file_priv,
				 struct drm_i915_gem_exec_object *entry)
{
	struct drm_device *dev = obj->dev;
	drm_i915_private_t *dev_priv = dev->dev_private;
	struct drm_i915_gem_relocation_entry reloc;
	struct drm_i915_gem_relocation_entry __user *relocs;
	struct drm_i915_gem_object *obj_priv = obj->driver_private;
	int i, ret;
	void __iomem *reloc_page;

	/* Choose the GTT offset for our buffer and put it there. */
	ret = i915_gem_object_pin(obj, (uint32_t) entry->alignment);
	if (ret)
		return ret;

	entry->offset = obj_priv->gtt_offset;

	relocs = (struct drm_i915_gem_relocation_entry __user *)
		 (uintptr_t) entry->relocs_ptr;
	/* Apply the relocations, using the GTT aperture to avoid cache
	 * flushing requirements.
	 */
	for (i = 0; i < entry->relocation_count; i++) {
		struct drm_gem_object *target_obj;
		struct drm_i915_gem_object *target_obj_priv;
		uint32_t reloc_val, reloc_offset;
		uint32_t __iomem *reloc_entry;

		ret = copy_from_user(&reloc, relocs + i, sizeof(reloc));
		if (ret != 0) {
			i915_gem_object_unpin(obj);
			return ret;
		}

		target_obj = drm_gem_object_lookup(obj->dev, file_priv,
						   reloc.target_handle);
		if (target_obj == NULL) {
			i915_gem_object_unpin(obj);
			return -EBADF;
		}
		target_obj_priv = target_obj->driver_private;

		/* The target buffer should have appeared before us in the
		 * exec_object list, so it should have a GTT space bound by now.
		 */
		if (target_obj_priv->gtt_space == NULL) {
			DRM_ERROR("No GTT space found for object %d\n",
				  reloc.target_handle);
			drm_gem_object_unreference(target_obj);
			i915_gem_object_unpin(obj);
			return -EINVAL;
		}

		if (reloc.offset > obj->size - 4) {
			DRM_ERROR("Relocation beyond object bounds: "
				  "obj %p target %d offset %d size %d.\n",
				  obj, reloc.target_handle,
				  (int) reloc.offset, (int) obj->size);
			drm_gem_object_unreference(target_obj);
			i915_gem_object_unpin(obj);
			return -EINVAL;
		}
		if (reloc.offset & 3) {
			DRM_ERROR("Relocation not 4-byte aligned: "
				  "obj %p target %d offset %d.\n",
				  obj, reloc.target_handle,
				  (int) reloc.offset);
			drm_gem_object_unreference(target_obj);
			i915_gem_object_unpin(obj);
			return -EINVAL;
		}

		if (reloc.write_domain && target_obj->pending_write_domain &&
		    reloc.write_domain != target_obj->pending_write_domain) {
			DRM_ERROR("Write domain conflict: "
				  "obj %p target %d offset %d "
				  "new %08x old %08x\n",
				  obj, reloc.target_handle,
				  (int) reloc.offset,
				  reloc.write_domain,
				  target_obj->pending_write_domain);
			drm_gem_object_unreference(target_obj);
			i915_gem_object_unpin(obj);
			return -EINVAL;
		}

#if WATCH_RELOC
		DRM_INFO("%s: obj %p offset %08x target %d "
			 "read %08x write %08x gtt %08x "
			 "presumed %08x delta %08x\n",
			 __func__,
			 obj,
			 (int) reloc.offset,
			 (int) reloc.target_handle,
			 (int) reloc.read_domains,
			 (int) reloc.write_domain,
			 (int) target_obj_priv->gtt_offset,
			 (int) reloc.presumed_offset,
			 reloc.delta);
#endif

		target_obj->pending_read_domains |= reloc.read_domains;
		target_obj->pending_write_domain |= reloc.write_domain;

		/* If the relocation already has the right value in it, no
		 * more work needs to be done.
		 */
		if (target_obj_priv->gtt_offset == reloc.presumed_offset) {
			drm_gem_object_unreference(target_obj);
			continue;
		}

		ret = i915_gem_object_set_to_gtt_domain(obj, 1);
		if (ret != 0) {
			drm_gem_object_unreference(target_obj);
			i915_gem_object_unpin(obj);
			return -EINVAL;
		}

		/* Map the page containing the relocation we're going to
		 * perform.
		 */
		reloc_offset = obj_priv->gtt_offset + reloc.offset;
		reloc_page = io_mapping_map_atomic_wc(dev_priv->mm.gtt_mapping,
						      (reloc_offset &
						       ~(PAGE_SIZE - 1)));
		reloc_entry = (uint32_t __iomem *)(reloc_page +
						   (reloc_offset & (PAGE_SIZE - 1)));
		reloc_val = target_obj_priv->gtt_offset + reloc.delta;

#if WATCH_BUF
		DRM_INFO("Applied relocation: %p@0x%08x %08x -> %08x\n",
			  obj, (unsigned int) reloc.offset,
			  readl(reloc_entry), reloc_val);
#endif
		writel(reloc_val, reloc_entry);
		io_mapping_unmap_atomic(reloc_page);

		/* Write the updated presumed offset for this entry back out
		 * to the user.
		 */
		reloc.presumed_offset = target_obj_priv->gtt_offset;
		ret = copy_to_user(relocs + i, &reloc, sizeof(reloc));
		if (ret != 0) {
			drm_gem_object_unreference(target_obj);
			i915_gem_object_unpin(obj);
			return ret;
		}

		drm_gem_object_unreference(target_obj);
	}

#if WATCH_BUF
	if (0)
		i915_gem_dump_object(obj, 128, __func__, ~0);
#endif
	return 0;
}

/** Dispatch a batchbuffer to the ring
 */
static int
i915_dispatch_gem_execbuffer(struct drm_device *dev,
			      struct drm_i915_gem_execbuffer *exec,
			      uint64_t exec_offset)
{
	drm_i915_private_t *dev_priv = dev->dev_private;
	struct drm_clip_rect __user *boxes = (struct drm_clip_rect __user *)
					     (uintptr_t) exec->cliprects_ptr;
	int nbox = exec->num_cliprects;
	int i = 0, count;
	uint32_t	exec_start, exec_len;
	RING_LOCALS;

	exec_start = (uint32_t) exec_offset + exec->batch_start_offset;
	exec_len = (uint32_t) exec->batch_len;

	if ((exec_start | exec_len) & 0x7) {
		DRM_ERROR("alignment\n");
		return -EINVAL;
	}

	if (!exec_start)
		return -EINVAL;

	count = nbox ? nbox : 1;

	for (i = 0; i < count; i++) {
		if (i < nbox) {
			int ret = i915_emit_box(dev, boxes, i,
						exec->DR1, exec->DR4);
			if (ret)
				return ret;
		}

		if (IS_I830(dev) || IS_845G(dev)) {
			BEGIN_LP_RING(4);
			OUT_RING(MI_BATCH_BUFFER);
			OUT_RING(exec_start | MI_BATCH_NON_SECURE);
			OUT_RING(exec_start + exec_len - 4);
			OUT_RING(0);
			ADVANCE_LP_RING();
		} else {
			BEGIN_LP_RING(2);
			if (IS_I965G(dev)) {
				OUT_RING(MI_BATCH_BUFFER_START |
					 (2 << 6) |
					 MI_BATCH_NON_SECURE_I965);
				OUT_RING(exec_start);
			} else {
				OUT_RING(MI_BATCH_BUFFER_START |
					 (2 << 6));
				OUT_RING(exec_start | MI_BATCH_NON_SECURE);
			}
			ADVANCE_LP_RING();
		}
	}

	/* XXX breadcrumb */
	return 0;
}

/* Throttle our rendering by waiting until the ring has completed our requests
 * emitted over 20 msec ago.
 *
 * This should get us reasonable parallelism between CPU and GPU but also
 * relatively low latency when blocking on a particular request to finish.
 */
static int
i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file_priv)
{
	struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv;
	int ret = 0;
	uint32_t seqno;

	mutex_lock(&dev->struct_mutex);
	seqno = i915_file_priv->mm.last_gem_throttle_seqno;
	i915_file_priv->mm.last_gem_throttle_seqno =
		i915_file_priv->mm.last_gem_seqno;
	if (seqno)
		ret = i915_wait_request(dev, seqno);
	mutex_unlock(&dev->struct_mutex);
	return ret;
}

int
i915_gem_execbuffer(struct drm_device *dev, void *data,
		    struct drm_file *file_priv)
{
	drm_i915_private_t *dev_priv = dev->dev_private;
	struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv;
	struct drm_i915_gem_execbuffer *args = data;
	struct drm_i915_gem_exec_object *exec_list = NULL;
	struct drm_gem_object **object_list = NULL;
	struct drm_gem_object *batch_obj;
	int ret, i, pinned = 0;
	uint64_t exec_offset;
	uint32_t seqno, flush_domains;

#if WATCH_EXEC
	DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
		  (int) args->buffers_ptr, args->buffer_count, args->batch_len);
#endif

	if (args->buffer_count < 1) {
		DRM_ERROR("execbuf with %d buffers\n", args->buffer_count);
		return -EINVAL;
	}
	/* Copy in the exec list from userland */
	exec_list = drm_calloc(sizeof(*exec_list), args->buffer_count,
			       DRM_MEM_DRIVER);
	object_list = drm_calloc(sizeof(*object_list), args->buffer_count,
				 DRM_MEM_DRIVER);
	if (exec_list == NULL || object_list == NULL) {
		DRM_ERROR("Failed to allocate exec or object list "
			  "for %d buffers\n",
			  args->buffer_count);
		ret = -ENOMEM;
		goto pre_mutex_err;
	}
	ret = copy_from_user(exec_list,
			     (struct drm_i915_relocation_entry __user *)
			     (uintptr_t) args->buffers_ptr,
			     sizeof(*exec_list) * args->buffer_count);
	if (ret != 0) {
		DRM_ERROR("copy %d exec entries failed %d\n",
			  args->buffer_count, ret);
		goto pre_mutex_err;
	}

	mutex_lock(&dev->struct_mutex);

	i915_verify_inactive(dev, __FILE__, __LINE__);

	if (dev_priv->mm.wedged) {
		DRM_ERROR("Execbuf while wedged\n");
		mutex_unlock(&dev->struct_mutex);
		return -EIO;
	}

	if (dev_priv->mm.suspended) {
		DRM_ERROR("Execbuf while VT-switched.\n");
		mutex_unlock(&dev->struct_mutex);
		return -EBUSY;
	}

	/* Zero the gloabl flush/invalidate flags. These
	 * will be modified as each object is bound to the
	 * gtt
	 */
	dev->invalidate_domains = 0;
	dev->flush_domains = 0;

	/* Look up object handles and perform the relocations */
	for (i = 0; i < args->buffer_count; i++) {
		object_list[i] = drm_gem_object_lookup(dev, file_priv,
						       exec_list[i].handle);
		if (object_list[i] == NULL) {
			DRM_ERROR("Invalid object handle %d at index %d\n",
				   exec_list[i].handle, i);
			ret = -EBADF;
			goto err;
		}

		object_list[i]->pending_read_domains = 0;
		object_list[i]->pending_write_domain = 0;
		ret = i915_gem_object_pin_and_relocate(object_list[i],
						       file_priv,
						       &exec_list[i]);
		if (ret) {
			DRM_ERROR("object bind and relocate failed %d\n", ret);
			goto err;
		}
		pinned = i + 1;
	}

	/* Set the pending read domains for the batch buffer to COMMAND */
	batch_obj = object_list[args->buffer_count-1];
	batch_obj->pending_read_domains = I915_GEM_DOMAIN_COMMAND;
	batch_obj->pending_write_domain = 0;

	i915_verify_inactive(dev, __FILE__, __LINE__);

	for (i = 0; i < args->buffer_count; i++) {
		struct drm_gem_object *obj = object_list[i];

		/* make sure all previous memory operations have passed */
		ret = i915_gem_object_set_domain(obj,
						 obj->pending_read_domains,
						 obj->pending_write_domain);
		if (ret) {
			/* As we've partially updated domains on our buffers,
			 * we have to emit the flush we've accumulated
			 * before exiting, or we'll have broken the
			 * active/flushing/inactive invariants.
			 *
			 * We'll potentially have some things marked as
			 * being in write domains that they actually aren't,
			 * but that should be merely a minor performance loss.
			 */
			flush_domains = i915_gem_dev_set_domain(dev);
			(void)i915_add_request(dev, flush_domains);
			goto err;
		}
	}

	i915_verify_inactive(dev, __FILE__, __LINE__);

	/* Flush/invalidate caches and chipset buffer */
	flush_domains = i915_gem_dev_set_domain(dev);

	i915_verify_inactive(dev, __FILE__, __LINE__);

#if WATCH_COHERENCY
	for (i = 0; i < args->buffer_count; i++) {
		i915_gem_object_check_coherency(object_list[i],
						exec_list[i].handle);
	}
#endif

	exec_offset = exec_list[args->buffer_count - 1].offset;

#if WATCH_EXEC
	i915_gem_dump_object(object_list[args->buffer_count - 1],
			      args->batch_len,
			      __func__,
			      ~0);
#endif

	(void)i915_add_request(dev, flush_domains);

	/* Exec the batchbuffer */
	ret = i915_dispatch_gem_execbuffer(dev, args, exec_offset);
	if (ret) {
		DRM_ERROR("dispatch failed %d\n", ret);
		goto err;
	}

	/*
	 * Ensure that the commands in the batch buffer are
	 * finished before the interrupt fires
	 */
	flush_domains = i915_retire_commands(dev);

	i915_verify_inactive(dev, __FILE__, __LINE__);

	/*
	 * Get a seqno representing the execution of the current buffer,
	 * which we can wait on.  We would like to mitigate these interrupts,
	 * likely by only creating seqnos occasionally (so that we have
	 * *some* interrupts representing completion of buffers that we can
	 * wait on when trying to clear up gtt space).
	 */
	seqno = i915_add_request(dev, flush_domains);
	BUG_ON(seqno == 0);
	i915_file_priv->mm.last_gem_seqno = seqno;
	for (i = 0; i < args->buffer_count; i++) {
		struct drm_gem_object *obj = object_list[i];

		i915_gem_object_move_to_active(obj, seqno);
#if WATCH_LRU
		DRM_INFO("%s: move to exec list %p\n", __func__, obj);
#endif
	}
#if WATCH_LRU
	i915_dump_lru(dev, __func__);
#endif

	i915_verify_inactive(dev, __FILE__, __LINE__);

	/* Copy the new buffer offsets back to the user's exec list. */
	ret = copy_to_user((struct drm_i915_relocation_entry __user *)
			   (uintptr_t) args->buffers_ptr,
			   exec_list,
			   sizeof(*exec_list) * args->buffer_count);
	if (ret)
		DRM_ERROR("failed to copy %d exec entries "
			  "back to user (%d)\n",
			   args->buffer_count, ret);
err:
	if (object_list != NULL) {
		for (i = 0; i < pinned; i++)
			i915_gem_object_unpin(object_list[i]);

		for (i = 0; i < args->buffer_count; i++)
			drm_gem_object_unreference(object_list[i]);
	}
	mutex_unlock(&dev->struct_mutex);

pre_mutex_err:
	drm_free(object_list, sizeof(*object_list) * args->buffer_count,
		 DRM_MEM_DRIVER);
	drm_free(exec_list, sizeof(*exec_list) * args->buffer_count,
		 DRM_MEM_DRIVER);

	return ret;
}

int
i915_gem_object_pin(struct drm_gem_object *obj, uint32_t alignment)
{
	struct drm_device *dev = obj->dev;
	struct drm_i915_gem_object *obj_priv = obj->driver_private;
	int ret;

	i915_verify_inactive(dev, __FILE__, __LINE__);
	if (obj_priv->gtt_space == NULL) {
		ret = i915_gem_object_bind_to_gtt(obj, alignment);
		if (ret != 0) {
			DRM_ERROR("Failure to bind: %d", ret);
			return ret;
		}
	}
	obj_priv->pin_count++;

	/* If the object is not active and not pending a flush,
	 * remove it from the inactive list
	 */
	if (obj_priv->pin_count == 1) {
		atomic_inc(&dev->pin_count);
		atomic_add(obj->size, &dev->pin_memory);
		if (!obj_priv->active &&
		    (obj->write_domain & ~(I915_GEM_DOMAIN_CPU |
					   I915_GEM_DOMAIN_GTT)) == 0 &&
		    !list_empty(&obj_priv->list))
			list_del_init(&obj_priv->list);
	}
	i915_verify_inactive(dev, __FILE__, __LINE__);

	return 0;
}

void
i915_gem_object_unpin(struct drm_gem_object *obj)
{
	struct drm_device *dev = obj->dev;
	drm_i915_private_t *dev_priv = dev->dev_private;
	struct drm_i915_gem_object *obj_priv = obj->driver_private;

	i915_verify_inactive(dev, __FILE__, __LINE__);
	obj_priv->pin_count--;
	BUG_ON(obj_priv->pin_count < 0);
	BUG_ON(obj_priv->gtt_space == NULL);

	/* If the object is no longer pinned, and is
	 * neither active nor being flushed, then stick it on
	 * the inactive list
	 */
	if (obj_priv->pin_count == 0) {
		if (!obj_priv->active &&
		    (obj->write_domain & ~(I915_GEM_DOMAIN_CPU |
					   I915_GEM_DOMAIN_GTT)) == 0)
			list_move_tail(&obj_priv->list,
				       &dev_priv->mm.inactive_list);
		atomic_dec(&dev->pin_count);
		atomic_sub(obj->size, &dev->pin_memory);
	}
	i915_verify_inactive(dev, __FILE__, __LINE__);
}

int
i915_gem_pin_ioctl(struct drm_device *dev, void *data,
		   struct drm_file *file_priv)
{
	struct drm_i915_gem_pin *args = data;
	struct drm_gem_object *obj;
	struct drm_i915_gem_object *obj_priv;
	int ret;

	mutex_lock(&dev->struct_mutex);

	obj = drm_gem_object_lookup(dev, file_priv, args->handle);
	if (obj == NULL) {
		DRM_ERROR("Bad handle in i915_gem_pin_ioctl(): %d\n",
			  args->handle);
		mutex_unlock(&dev->struct_mutex);
		return -EBADF;
	}
	obj_priv = obj->driver_private;

	ret = i915_gem_object_pin(obj, args->alignment);
	if (ret != 0) {
		drm_gem_object_unreference(obj);
		mutex_unlock(&dev->struct_mutex);
		return ret;
	}

	/* XXX - flush the CPU caches for pinned objects
	 * as the X server doesn't manage domains yet
	 */
	if (obj->write_domain & I915_GEM_DOMAIN_CPU) {
		i915_gem_clflush_object(obj);
		drm_agp_chipset_flush(dev);
		obj->write_domain = 0;
	}
	args->offset = obj_priv->gtt_offset;
	drm_gem_object_unreference(obj);
	mutex_unlock(&dev->struct_mutex);

	return 0;
}

int
i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
		     struct drm_file *file_priv)
{
	struct drm_i915_gem_pin *args = data;
	struct drm_gem_object *obj;

	mutex_lock(&dev->struct_mutex);

	obj = drm_gem_object_lookup(dev, file_priv, args->handle);
	if (obj == NULL) {
		DRM_ERROR("Bad handle in i915_gem_unpin_ioctl(): %d\n",
			  args->handle);
		mutex_unlock(&dev->struct_mutex);
		return -EBADF;
	}

	i915_gem_object_unpin(obj);

	drm_gem_object_unreference(obj);
	mutex_unlock(&dev->struct_mutex);
	return 0;
}

int
i915_gem_busy_ioctl(struct drm_device *dev, void *data,
		    struct drm_file *file_priv)
{
	struct drm_i915_gem_busy *args = data;
	struct drm_gem_object *obj;
	struct drm_i915_gem_object *obj_priv;

	mutex_lock(&dev->struct_mutex);
	obj = drm_gem_object_lookup(dev, file_priv, args->handle);
	if (obj == NULL) {
		DRM_ERROR("Bad handle in i915_gem_busy_ioctl(): %d\n",
			  args->handle);
		mutex_unlock(&dev->struct_mutex);
		return -EBADF;
	}

	obj_priv = obj->driver_private;
	args->busy = obj_priv->active;

	drm_gem_object_unreference(obj);
	mutex_unlock(&dev->struct_mutex);
	return 0;
}

int
i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
			struct drm_file *file_priv)
{
    return i915_gem_ring_throttle(dev, file_priv);
}

int i915_gem_init_object(struct drm_gem_object *obj)
{
	struct drm_i915_gem_object *obj_priv;

	obj_priv = drm_calloc(1, sizeof(*obj_priv), DRM_MEM_DRIVER);
	if (obj_priv == NULL)
		return -ENOMEM;

	/*
	 * We've just allocated pages from the kernel,
	 * so they've just been written by the CPU with
	 * zeros. They'll need to be clflushed before we
	 * use them with the GPU.
	 */
	obj->write_domain = I915_GEM_DOMAIN_CPU;
	obj->read_domains = I915_GEM_DOMAIN_CPU;

	obj_priv->agp_type = AGP_USER_MEMORY;

	obj->driver_private = obj_priv;
	obj_priv->obj = obj;
	INIT_LIST_HEAD(&obj_priv->list);
	return 0;
}

void i915_gem_free_object(struct drm_gem_object *obj)
{
	struct drm_i915_gem_object *obj_priv = obj->driver_private;

	while (obj_priv->pin_count > 0)
		i915_gem_object_unpin(obj);

	i915_gem_object_unbind(obj);

	drm_free(obj_priv->page_cpu_valid, 1, DRM_MEM_DRIVER);
	drm_free(obj->driver_private, 1, DRM_MEM_DRIVER);
}

static int
i915_gem_set_domain(struct drm_gem_object *obj,
		    struct drm_file *file_priv,
		    uint32_t read_domains,
		    uint32_t write_domain)
{
	struct drm_device *dev = obj->dev;
	int ret;
	uint32_t flush_domains;

	BUG_ON(!mutex_is_locked(&dev->struct_mutex));

	ret = i915_gem_object_set_domain(obj, read_domains, write_domain);
	if (ret)
		return ret;
	flush_domains = i915_gem_dev_set_domain(obj->dev);

	if (flush_domains & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT))
		(void) i915_add_request(dev, flush_domains);

	return 0;
}

/** Unbinds all objects that are on the given buffer list. */
static int
i915_gem_evict_from_list(struct drm_device *dev, struct list_head *head)
{
	struct drm_gem_object *obj;
	struct drm_i915_gem_object *obj_priv;
	int ret;

	while (!list_empty(head)) {
		obj_priv = list_first_entry(head,
					    struct drm_i915_gem_object,
					    list);
		obj = obj_priv->obj;

		if (obj_priv->pin_count != 0) {
			DRM_ERROR("Pinned object in unbind list\n");
			mutex_unlock(&dev->struct_mutex);
			return -EINVAL;
		}

		ret = i915_gem_object_unbind(obj);
		if (ret != 0) {
			DRM_ERROR("Error unbinding object in LeaveVT: %d\n",
				  ret);
			mutex_unlock(&dev->struct_mutex);
			return ret;
		}
	}


	return 0;
}

static int
i915_gem_idle(struct drm_device *dev)
{
	drm_i915_private_t *dev_priv = dev->dev_private;
	uint32_t seqno, cur_seqno, last_seqno;
	int stuck, ret;

	mutex_lock(&dev->struct_mutex);

	if (dev_priv->mm.suspended || dev_priv->ring.ring_obj == NULL) {
		mutex_unlock(&dev->struct_mutex);
		return 0;
	}

	/* Hack!  Don't let anybody do execbuf while we don't control the chip.
	 * We need to replace this with a semaphore, or something.
	 */
	dev_priv->mm.suspended = 1;

	/* Cancel the retire work handler, wait for it to finish if running
	 */
	mutex_unlock(&dev->struct_mutex);
	cancel_delayed_work_sync(&dev_priv->mm.retire_work);
	mutex_lock(&dev->struct_mutex);

	i915_kernel_lost_context(dev);

	/* Flush the GPU along with all non-CPU write domains
	 */
	i915_gem_flush(dev, ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT),
		       ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT));
	seqno = i915_add_request(dev, ~(I915_GEM_DOMAIN_CPU |
					I915_GEM_DOMAIN_GTT));

	if (seqno == 0) {
		mutex_unlock(&dev->struct_mutex);
		return -ENOMEM;
	}

	dev_priv->mm.waiting_gem_seqno = seqno;
	last_seqno = 0;
	stuck = 0;
	for (;;) {
		cur_seqno = i915_get_gem_seqno(dev);
		if (i915_seqno_passed(cur_seqno, seqno))
			break;
		if (last_seqno == cur_seqno) {
			if (stuck++ > 100) {
				DRM_ERROR("hardware wedged\n");
				dev_priv->mm.wedged = 1;
				DRM_WAKEUP(&dev_priv->irq_queue);
				break;
			}
		}
		msleep(10);
		last_seqno = cur_seqno;
	}
	dev_priv->mm.waiting_gem_seqno = 0;

	i915_gem_retire_requests(dev);

	if (!dev_priv->mm.wedged) {
		/* Active and flushing should now be empty as we've
		 * waited for a sequence higher than any pending execbuffer
		 */
		WARN_ON(!list_empty(&dev_priv->mm.active_list));
		WARN_ON(!list_empty(&dev_priv->mm.flushing_list));
		/* Request should now be empty as we've also waited
		 * for the last request in the list
		 */
		WARN_ON(!list_empty(&dev_priv->mm.request_list));
	}

	/* Empty the active and flushing lists to inactive.  If there's
	 * anything left at this point, it means that we're wedged and
	 * nothing good's going to happen by leaving them there.  So strip
	 * the GPU domains and just stuff them onto inactive.
	 */
	while (!list_empty(&dev_priv->mm.active_list)) {
		struct drm_i915_gem_object *obj_priv;

		obj_priv = list_first_entry(&dev_priv->mm.active_list,
					    struct drm_i915_gem_object,
					    list);
		obj_priv->obj->write_domain &= ~I915_GEM_GPU_DOMAINS;
		i915_gem_object_move_to_inactive(obj_priv->obj);
	}

	while (!list_empty(&dev_priv->mm.flushing_list)) {
		struct drm_i915_gem_object *obj_priv;

		obj_priv = list_first_entry(&dev_priv->mm.flushing_list,
					    struct drm_i915_gem_object,
					    list);
		obj_priv->obj->write_domain &= ~I915_GEM_GPU_DOMAINS;
		i915_gem_object_move_to_inactive(obj_priv->obj);
	}


	/* Move all inactive buffers out of the GTT. */
	ret = i915_gem_evict_from_list(dev, &dev_priv->mm.inactive_list);
	WARN_ON(!list_empty(&dev_priv->mm.inactive_list));
	if (ret) {
		mutex_unlock(&dev->struct_mutex);
		return ret;
	}

	i915_gem_cleanup_ringbuffer(dev);
	mutex_unlock(&dev->struct_mutex);

	return 0;
}

static int
i915_gem_init_hws(struct drm_device *dev)
{
	drm_i915_private_t *dev_priv = dev->dev_private;
	struct drm_gem_object *obj;
	struct drm_i915_gem_object *obj_priv;
	int ret;

	/* If we need a physical address for the status page, it's already
	 * initialized at driver load time.
	 */
	if (!I915_NEED_GFX_HWS(dev))
		return 0;

	obj = drm_gem_object_alloc(dev, 4096);
	if (obj == NULL) {
		DRM_ERROR("Failed to allocate status page\n");
		return -ENOMEM;
	}
	obj_priv = obj->driver_private;
	obj_priv->agp_type = AGP_USER_CACHED_MEMORY;

	ret = i915_gem_object_pin(obj, 4096);
	if (ret != 0) {
		drm_gem_object_unreference(obj);
		return ret;
	}

	dev_priv->status_gfx_addr = obj_priv->gtt_offset;

	dev_priv->hw_status_page = kmap(obj_priv->page_list[0]);
	if (dev_priv->hw_status_page == NULL) {
		DRM_ERROR("Failed to map status page.\n");
		memset(&dev_priv->hws_map, 0, sizeof(dev_priv->hws_map));
		drm_gem_object_unreference(obj);
		return -EINVAL;
	}
	dev_priv->hws_obj = obj;
	memset(dev_priv->hw_status_page, 0, PAGE_SIZE);
	I915_WRITE(HWS_PGA, dev_priv->status_gfx_addr);
	I915_READ(HWS_PGA); /* posting read */
	DRM_DEBUG("hws offset: 0x%08x\n", dev_priv->status_gfx_addr);

	return 0;
}

static int
i915_gem_init_ringbuffer(struct drm_device *dev)
{
	drm_i915_private_t *dev_priv = dev->dev_private;
	struct drm_gem_object *obj;
	struct drm_i915_gem_object *obj_priv;
	int ret;
	u32 head;

	ret = i915_gem_init_hws(dev);
	if (ret != 0)
		return ret;

	obj = drm_gem_object_alloc(dev, 128 * 1024);
	if (obj == NULL) {
		DRM_ERROR("Failed to allocate ringbuffer\n");
		return -ENOMEM;
	}
	obj_priv = obj->driver_private;

	ret = i915_gem_object_pin(obj, 4096);
	if (ret != 0) {
		drm_gem_object_unreference(obj);
		return ret;
	}

	/* Set up the kernel mapping for the ring. */
	dev_priv->ring.Size = obj->size;
	dev_priv->ring.tail_mask = obj->size - 1;

	dev_priv->ring.map.offset = dev->agp->base + obj_priv->gtt_offset;
	dev_priv->ring.map.size = obj->size;
	dev_priv->ring.map.type = 0;
	dev_priv->ring.map.flags = 0;
	dev_priv->ring.map.mtrr = 0;

	drm_core_ioremap_wc(&dev_priv->ring.map, dev);
	if (dev_priv->ring.map.handle == NULL) {
		DRM_ERROR("Failed to map ringbuffer.\n");
		memset(&dev_priv->ring, 0, sizeof(dev_priv->ring));
		drm_gem_object_unreference(obj);
		return -EINVAL;
	}
	dev_priv->ring.ring_obj = obj;
	dev_priv->ring.virtual_start = dev_priv->ring.map.handle;

	/* Stop the ring if it's running. */
	I915_WRITE(PRB0_CTL, 0);
	I915_WRITE(PRB0_TAIL, 0);
	I915_WRITE(PRB0_HEAD, 0);

	/* Initialize the ring. */
	I915_WRITE(PRB0_START, obj_priv->gtt_offset);
	head = I915_READ(PRB0_HEAD) & HEAD_ADDR;

	/* G45 ring initialization fails to reset head to zero */
	if (head != 0) {
		DRM_ERROR("Ring head not reset to zero "
			  "ctl %08x head %08x tail %08x start %08x\n",
			  I915_READ(PRB0_CTL),
			  I915_READ(PRB0_HEAD),
			  I915_READ(PRB0_TAIL),
			  I915_READ(PRB0_START));
		I915_WRITE(PRB0_HEAD, 0);

		DRM_ERROR("Ring head forced to zero "
			  "ctl %08x head %08x tail %08x start %08x\n",
			  I915_READ(PRB0_CTL),
			  I915_READ(PRB0_HEAD),
			  I915_READ(PRB0_TAIL),
			  I915_READ(PRB0_START));
	}

	I915_WRITE(PRB0_CTL,
		   ((obj->size - 4096) & RING_NR_PAGES) |
		   RING_NO_REPORT |
		   RING_VALID);

	head = I915_READ(PRB0_HEAD) & HEAD_ADDR;

	/* If the head is still not zero, the ring is dead */
	if (head != 0) {
		DRM_ERROR("Ring initialization failed "
			  "ctl %08x head %08x tail %08x start %08x\n",
			  I915_READ(PRB0_CTL),
			  I915_READ(PRB0_HEAD),
			  I915_READ(PRB0_TAIL),
			  I915_READ(PRB0_START));
		return -EIO;
	}

	/* Update our cache of the ring state */
	i915_kernel_lost_context(dev);

	return 0;
}

static void
i915_gem_cleanup_ringbuffer(struct drm_device *dev)
{
	drm_i915_private_t *dev_priv = dev->dev_private;

	if (dev_priv->ring.ring_obj == NULL)
		return;

	drm_core_ioremapfree(&dev_priv->ring.map, dev);

	i915_gem_object_unpin(dev_priv->ring.ring_obj);
	drm_gem_object_unreference(dev_priv->ring.ring_obj);
	dev_priv->ring.ring_obj = NULL;
	memset(&dev_priv->ring, 0, sizeof(dev_priv->ring));

	if (dev_priv->hws_obj != NULL) {
		struct drm_gem_object *obj = dev_priv->hws_obj;
		struct drm_i915_gem_object *obj_priv = obj->driver_private;

		kunmap(obj_priv->page_list[0]);
		i915_gem_object_unpin(obj);
		drm_gem_object_unreference(obj);
		dev_priv->hws_obj = NULL;
		memset(&dev_priv->hws_map, 0, sizeof(dev_priv->hws_map));
		dev_priv->hw_status_page = NULL;

		/* Write high address into HWS_PGA when disabling. */
		I915_WRITE(HWS_PGA, 0x1ffff000);
	}
}

int
i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
		       struct drm_file *file_priv)
{
	drm_i915_private_t *dev_priv = dev->dev_private;
	int ret;

	if (dev_priv->mm.wedged) {
		DRM_ERROR("Reenabling wedged hardware, good luck\n");
		dev_priv->mm.wedged = 0;
	}

	ret = i915_gem_init_ringbuffer(dev);
	if (ret != 0)
		return ret;

	dev_priv->mm.gtt_mapping = io_mapping_create_wc(dev->agp->base,
							dev->agp->agp_info.aper_size
							* 1024 * 1024);

	mutex_lock(&dev->struct_mutex);
	BUG_ON(!list_empty(&dev_priv->mm.active_list));
	BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
	BUG_ON(!list_empty(&dev_priv->mm.inactive_list));
	BUG_ON(!list_empty(&dev_priv->mm.request_list));
	dev_priv->mm.suspended = 0;
	mutex_unlock(&dev->struct_mutex);

	drm_irq_install(dev);

	return 0;
}

int
i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
		       struct drm_file *file_priv)
{
	drm_i915_private_t *dev_priv = dev->dev_private;
	int ret;

	ret = i915_gem_idle(dev);
	drm_irq_uninstall(dev);

	io_mapping_free(dev_priv->mm.gtt_mapping);
	return ret;
}

void
i915_gem_lastclose(struct drm_device *dev)
{
	int ret;

	ret = i915_gem_idle(dev);
	if (ret)
		DRM_ERROR("failed to idle hardware: %d\n", ret);
}

void
i915_gem_load(struct drm_device *dev)
{
	drm_i915_private_t *dev_priv = dev->dev_private;

	INIT_LIST_HEAD(&dev_priv->mm.active_list);
	INIT_LIST_HEAD(&dev_priv->mm.flushing_list);
	INIT_LIST_HEAD(&dev_priv->mm.inactive_list);
	INIT_LIST_HEAD(&dev_priv->mm.request_list);
	INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
			  i915_gem_retire_work_handler);
	dev_priv->mm.next_gem_seqno = 1;

	i915_gem_detect_bit_6_swizzle(dev);
}