[Intel-gfx] [PATCH 09/13 v4] drm/i915: Implementation of GuC client

[O'Rourke, Tom]

On Thu, Jul 09, 2015 at 07:29:10PM +0100, Dave Gordon wrote:
> A GuC client has its own doorbell and workqueue. It maintains the
> doorbell cache line, process description object and work queue item.
> 
> A default guc_client is created for the i915 driver to use for
> normal-priority in-order submission.
> 
> Note that the created client is not yet ready for use; doorbell
> allocation will fail as we haven't yet linked the GuC's context
> descriptor to the default contexts for each ring (see later patch).
> 
> v2:
>     Defer adding structure members until needed [Chris Wilson]
>     Rationalise type declarations [Chris Wilson]
> 
> v4:
>     Rebased
> 
> Issue: VIZ-4884
> Signed-off-by: Alex Dai <yu.dai at intel.com>
> Signed-off-by: Dave Gordon <david.s.gordon at intel.com>

[TOR:] I had some non-critical questions below.

Reviewed-by: Tom O'Rourke <Tom.O'Rourke at intel.com>
> ---
>  drivers/gpu/drm/i915/i915_guc_submission.c | 649 +++++++++++++++++++++++++++++
>  drivers/gpu/drm/i915/intel_guc.h           |  42 ++
>  drivers/gpu/drm/i915/intel_guc_loader.c    |  12 +
>  3 files changed, 703 insertions(+)
> 
> diff --git a/drivers/gpu/drm/i915/i915_guc_submission.c b/drivers/gpu/drm/i915/i915_guc_submission.c
> index e9d46d6..25d8807 100644
> --- a/drivers/gpu/drm/i915/i915_guc_submission.c
> +++ b/drivers/gpu/drm/i915/i915_guc_submission.c
> @@ -27,6 +27,512 @@
>  #include "intel_guc.h"
>  
>  /**
> + * DOC: GuC Client
> + *
> + * i915_guc_client:
> + * We use the term client to avoid confusion with contexts. A i915_guc_client is
> + * equivalent to GuC object guc_context_desc. This context descriptor is
> + * allocated from a pool of 1024 entries. Kernel driver will allocate doorbell
> + * and workqueue for it. Also the process descriptor (guc_process_desc), which
> + * is mapped to client space. So the client can write Work Item then ring the
> + * doorbell.
> + *
> + * To simplify the implementation, we allocate one gem object that contains all
> + * pages for doorbell, process descriptor and workqueue.
> + *
> + * The Scratch registers:
> + * There are 16 MMIO-based registers start from 0xC180. The kernel driver writes
> + * a value to the action register (SOFT_SCRATCH_0) along with any data. It then
> + * triggers an interrupt on the GuC via another register write (0xC4C8).
> + * Firmware writes a success/fail code back to the action register after
> + * processes the request. The kernel driver polls waiting for this update and
> + * then proceeds.
> + * See host2guc_action()
> + *
> + * Doorbells:
> + * Doorbells are interrupts to uKernel. A doorbell is a single cache line (QW)
> + * mapped into process space.
> + *
> + * Work Items:
> + * There are several types of work items that the host may place into a
> + * workqueue, each with its own requirements and limitations. Currently only
> + * WQ_TYPE_INORDER is needed to support legacy submission via GuC, which
> + * represents in-order queue. The kernel driver packs ring tail pointer and an
> + * ELSP context descriptor dword into Work Item.
> + * See guc_add_workqueue_item()
> + *
> + */
> +
> +/*
> + * Read GuC command/status register (SOFT_SCRATCH_0)
> + * Return true if it contains a response rather than a command
> + */
> +static inline bool host2guc_action_response(struct drm_i915_private *dev_priv,
> +					    u32 *status)
> +{
> +	u32 val = I915_READ(SOFT_SCRATCH(0));
> +	*status = val;
> +	return GUC2HOST_IS_RESPONSE(val);
> +}
> +
> +static int host2guc_action(struct intel_guc *guc, u32 *data, u32 len)
> +{
> +	struct drm_i915_private *dev_priv = guc_to_i915(guc);
> +	u32 status;
> +	int i;
> +	int ret;
> +
> +	if (WARN_ON(len < 1 || len > 15))
> +		return -EINVAL;
> +

[TOR:] Would it be good for host2guc_action to take a 
forcewake?  There are several writes and polling reads 
for completion.  Taking a forcewake could avoid surplus 
forcewakes for each register access.

> +	spin_lock(&dev_priv->guc.host2guc_lock);
> +
> +	dev_priv->guc.action_count += 1;
> +	dev_priv->guc.action_cmd = data[0];
> +
> +	for (i = 0; i < len; i++)
> +		I915_WRITE(SOFT_SCRATCH(i), data[i]);
> +
> +	POSTING_READ(SOFT_SCRATCH(i - 1));
> +
> +	I915_WRITE(HOST2GUC_INTERRUPT, HOST2GUC_TRIGGER);
> +
> +	ret = wait_for_atomic(host2guc_action_response(dev_priv, &status), 10);
[TOR:] Why 10?

> +	if (status != GUC2HOST_STATUS_SUCCESS) {
> +		/* either GuC doesn't respond, which is a TIMEOUT,
> +		 * or a failure code is returned. */
> +		if (ret != -ETIMEDOUT)
> +			ret = -EIO;
> +
> +		DRM_ERROR("GUC: host2guc action 0x%X failed. ret=%d "
> +				"status=0x%08X response=0x%08X\n",
> +				data[0], ret, status,
> +				I915_READ(SOFT_SCRATCH(15)));
> +
> +		dev_priv->guc.action_fail += 1;
> +		dev_priv->guc.action_err = ret;
> +	}
> +	dev_priv->guc.action_status = status;
> +
> +	spin_unlock(&dev_priv->guc.host2guc_lock);
> +
> +	return ret;
> +}
> +
> +/*
> + * Tell the GuC to allocate or deallocate a specific doorbell
> + */
> +
> +static int host2guc_allocate_doorbell(struct intel_guc *guc,
> +				      struct i915_guc_client *client)
> +{
> +	u32 data[2];
> +
> +	data[0] = HOST2GUC_ACTION_ALLOCATE_DOORBELL;
> +	data[1] = client->ctx_index;
> +
> +	return host2guc_action(guc, data, 2);
> +}
> +
> +static int host2guc_release_doorbell(struct intel_guc *guc,
> +				     struct i915_guc_client *client)
> +{
> +	u32 data[2];
> +
> +	data[0] = HOST2GUC_ACTION_DEALLOCATE_DOORBELL;
> +	data[1] = client->ctx_index;
> +
> +	return host2guc_action(guc, data, 2);
> +}
> +
> +/*
> + * Initialise, update, or clear doorbell data shared with the GuC
> + *
> + * These functions modify shared data and so need access to the mapped
> + * client object which contains the page being used for the doorbell
> + */
> +
> +static void guc_init_doorbell(struct intel_guc *guc,
> +			      struct i915_guc_client *client)
> +{
> +	struct guc_doorbell_info *doorbell;
> +	void *base;
> +
> +	base = kmap_atomic(i915_gem_object_get_page(client->client_obj, 0));
> +	doorbell = base + client->doorbell_offset;
> +
> +	doorbell->db_status = 1;
> +	doorbell->cookie = 0;
> +
> +	kunmap_atomic(base);
> +}
> +
> +static int guc_ring_doorbell(struct i915_guc_client *gc)
> +{
> +	struct guc_process_desc *desc;
> +	union guc_doorbell_qw db_cmp, db_exc, db_ret;
> +	union guc_doorbell_qw *db;
> +	void *base;
> +	int attempt = 2, ret = -EAGAIN;
> +
> +	base = kmap_atomic(i915_gem_object_get_page(gc->client_obj, 0));
> +	desc = base + gc->proc_desc_offset;
> +
> +	/* Update the tail so it is visible to GuC */
> +	desc->tail = gc->wq_tail;
> +
> +	/* current cookie */
> +	db_cmp.db_status = GUC_DOORBELL_ENABLED;
> +	db_cmp.cookie = gc->cookie;
> +
> +	/* cookie to be updated */
> +	db_exc.db_status = GUC_DOORBELL_ENABLED;
> +	db_exc.cookie = gc->cookie + 1;
> +	if (db_exc.cookie == 0)
> +		db_exc.cookie = 1;
> +
> +	/* pointer of current doorbell cacheline */
> +	db = base + gc->doorbell_offset;
> +
> +	while (attempt--) {
> +		/* lets ring the doorbell */
> +		db_ret.value_qw = atomic64_cmpxchg((atomic64_t *)db,
> +			db_cmp.value_qw, db_exc.value_qw);
> +
> +		/* if the exchange was successfully executed */
> +		if (db_ret.value_qw == db_cmp.value_qw) {
> +			/* db was successfully rung */
> +			gc->cookie = db_exc.cookie;
> +			ret = 0;
> +			break;
> +		}
> +
> +		/* XXX: doorbell was lost and need to acquire it again */
> +		if (db_ret.db_status == GUC_DOORBELL_DISABLED)
> +			break;
> +
> +		DRM_ERROR("Cookie mismatch. Expected %d, returned %d\n",
> +			  db_cmp.cookie, db_ret.cookie);
> +
> +		/* update the cookie to newly read cookie from GuC */
> +		db_cmp.cookie = db_ret.cookie;
> +		db_exc.cookie = db_ret.cookie + 1;
> +		if (db_exc.cookie == 0)
> +			db_exc.cookie = 1;
> +	}
> +
> +	kunmap_atomic(base);
> +	return ret;
> +}
> +
> +static void guc_disable_doorbell(struct intel_guc *guc,
> +				 struct i915_guc_client *client)
> +{
> +	struct drm_i915_private *dev_priv = guc_to_i915(guc);
> +	struct guc_doorbell_info *doorbell;
> +	void *base;
> +	int drbreg = GEN8_DRBREGL(client->doorbell_id);
> +	int value;
> +
> +	base = kmap_atomic(i915_gem_object_get_page(client->client_obj, 0));
> +	doorbell = base + client->doorbell_offset;
> +
> +	doorbell->db_status = 0;
> +
> +	kunmap_atomic(base);
> +
> +	I915_WRITE(drbreg, I915_READ(drbreg) & ~GEN8_DRB_VALID);
> +
> +	value = I915_READ(drbreg);
> +	WARN_ON((value & GEN8_DRB_VALID) != 0);
> +
> +	I915_WRITE(GEN8_DRBREGU(client->doorbell_id), 0);
> +	I915_WRITE(drbreg, 0);
> +
> +	/* XXX: wait for any interrupts */
> +	/* XXX: wait for workqueue to drain */
> +}
> +
> +/*
> + * Select, assign and relase doorbell cachelines
> + *
> + * These functions track which doorbell cachelines are in use.
> + * The data they manipulate is protected by the host2guc lock.
> + */
> +
> +static uint32_t select_doorbell_cacheline(struct intel_guc *guc)
> +{
> +	const uint32_t cacheline_size = boot_cpu_data.x86_clflush_size;
> +	uint32_t offset;
> +
> +	spin_lock(&guc->host2guc_lock);
> +
> +	/* Doorbell uses a single cache line within a page */
> +	offset = guc->db_cacheline & PAGE_MASK;
> +
> +	/* Moving to next cache line to reduce contention */
> +	guc->db_cacheline += cacheline_size;
> +
> +	spin_unlock(&guc->host2guc_lock);
> +
> +	return offset;
> +}
> +
> +static uint16_t assign_doorbell(struct intel_guc *guc, uint32_t priority)
> +{
> +	/* The bitmap is split into two halves - high and normal priority. */
> +	const uint16_t half = GUC_MAX_DOORBELLS / 2;
> +	const uint16_t start = (priority <= GUC_CTX_PRIORITY_HIGH) ? half : 0;
> +	const uint16_t end = start + half;
> +	uint16_t id;
> +
> +	spin_lock(&guc->host2guc_lock);
> +	id = find_next_zero_bit(guc->doorbell_bitmap, end, start);
> +	if (id == end)
> +		id = GUC_INVALID_DOORBELL_ID;
> +	else
> +		bitmap_set(guc->doorbell_bitmap, id, 1);
> +	spin_unlock(&guc->host2guc_lock);
> +
> +	return id;
> +}
> +
> +static void release_doorbell(struct intel_guc *guc, uint16_t id)
> +{
> +	spin_lock(&guc->host2guc_lock);
> +	bitmap_clear(guc->doorbell_bitmap, id, 1);
> +	spin_unlock(&guc->host2guc_lock);
> +}
> +
> +/*
> + * Initialise the process descriptor shared with the GuC firmware.
> + */
> +static void guc_init_proc_desc(struct intel_guc *guc,
> +			       struct i915_guc_client *client)
> +{
> +	struct guc_process_desc *desc;
> +	void *base;
> +
> +	base = kmap_atomic(i915_gem_object_get_page(client->client_obj, 0));
> +	desc = base + client->proc_desc_offset;
> +
> +	memset(desc, 0, sizeof(*desc));
> +
> +	/*
> +	 * XXX: pDoorbell and WQVBaseAddress are pointers in process address
> +	 * space for ring3 clients (set them as in mmap_ioctl) or kernel
> +	 * space for kernel clients (map on demand instead? May make debug
> +	 * easier to have it mapped).
> +	 */
> +	desc->wq_base_addr = 0;
> +	desc->db_base_addr = 0;
> +
> +	desc->context_id = client->ctx_index;
> +	desc->wq_size_bytes = client->wq_size;
> +	desc->wq_status = WQ_STATUS_ACTIVE;
> +	desc->priority = client->priority;
> +
> +	kunmap_atomic(base);
> +}
> +
> +/*
> + * Initialise/clear the context descriptor shared with the GuC firmware.
> + *
> + * This descriptor tells the GuC where (in GGTT space) to find the important
> + * data structures relating to this client (doorbell, process descriptor,
> + * write queue, etc).
> + */
> +
> +static void guc_init_ctx_desc(struct intel_guc *guc,
> +			      struct i915_guc_client *client)
> +{
> +	struct guc_context_desc desc;
> +	struct sg_table *sg;
> +
> +	memset(&desc, 0, sizeof(desc));
> +
> +	desc.attribute = GUC_CTX_DESC_ATTR_ACTIVE | GUC_CTX_DESC_ATTR_KERNEL;
> +	desc.context_id = client->ctx_index;
> +	desc.priority = client->priority;
> +	desc.engines_used = (1 << RCS) | (1 << VCS) | (1 << BCS) |
> +			    (1 << VECS) | (1 << VCS2); /* all engines */
> +	desc.db_id = client->doorbell_id;
> +
> +	/*
> +	 * The CPU address is only needed at certain points, so kmap_atomic on
> +	 * demand instead of storing it in the ctx descriptor.
> +	 * XXX: May make debug easier to have it mapped
> +	 */
> +	desc.db_trigger_cpu = 0;
> +	desc.db_trigger_uk = client->doorbell_offset +
> +		i915_gem_obj_ggtt_offset(client->client_obj);
> +	desc.db_trigger_phy = client->doorbell_offset +
> +		sg_dma_address(client->client_obj->pages->sgl);
> +
> +	desc.process_desc = client->proc_desc_offset +
> +		i915_gem_obj_ggtt_offset(client->client_obj);
> +
> +	desc.wq_addr = client->wq_offset +
> +		i915_gem_obj_ggtt_offset(client->client_obj);
> +
> +	desc.wq_size = client->wq_size;
> +
> +	/*
> +	 * XXX: Take LRCs from an existing intel_context if this is not an
> +	 * IsKMDCreatedContext client
> +	 */
> +	desc.desc_private = (uintptr_t)client;
> +
> +	/* Pool context is pinned already */
> +	sg = guc->ctx_pool_obj->pages;
> +	sg_pcopy_from_buffer(sg->sgl, sg->nents, &desc, sizeof(desc),
> +			     sizeof(desc) * client->ctx_index);
> +}
> +
> +static void guc_fini_ctx_desc(struct intel_guc *guc,
> +			      struct i915_guc_client *client)
> +{
> +	struct guc_context_desc desc;
> +	struct sg_table *sg;
> +
> +	memset(&desc, 0, sizeof(desc));
> +
> +	sg = guc->ctx_pool_obj->pages;
> +	sg_pcopy_from_buffer(sg->sgl, sg->nents, &desc, sizeof(desc),
> +			     sizeof(desc) * client->ctx_index);
> +}
> +
> +/* Get valid workqueue item and return it back to offset */
> +static int guc_get_workqueue_space(struct i915_guc_client *gc, u32 *offset)
> +{
> +	struct guc_process_desc *desc;
> +	void *base;
> +	u32 size = sizeof(struct guc_wq_item);
> +	int ret = 0, timeout_counter = 200;
> +	unsigned long flags;
> +
> +	base = kmap_atomic(i915_gem_object_get_page(gc->client_obj, 0));
> +	desc = base + gc->proc_desc_offset;
> +
> +	while (timeout_counter-- > 0) {
> +		spin_lock_irqsave(&gc->wq_lock, flags);
> +
> +		ret = wait_for_atomic(CIRC_SPACE(gc->wq_tail, desc->head,
> +				gc->wq_size) >= size, 1);
> +
> +		if (!ret) {
> +			*offset = gc->wq_tail;
> +
> +			/* advance the tail for next workqueue item */
> +			gc->wq_tail += size;
> +			gc->wq_tail &= gc->wq_size - 1;
> +
> +			/* this will break the loop */
> +			timeout_counter = 0;
> +		}
> +
> +		spin_unlock_irqrestore(&gc->wq_lock, flags);
> +	};
> +
> +	kunmap_atomic(base);
> +
> +	return ret;
> +}
> +
> +static int guc_add_workqueue_item(struct i915_guc_client *gc,
> +				  struct drm_i915_gem_request *rq)
> +{
> +	enum intel_ring_id ring_id = rq->ring->id;
> +	struct guc_wq_item *wqi;
> +	void *base;
> +	u32 tail, wq_len, wq_off = 0;
> +	int ret;
> +
> +	ret = guc_get_workqueue_space(gc, &wq_off);
> +	if (ret)
> +		return ret;
> +
> +	/* For now workqueue item is 4 DWs; workqueue buffer is 2 pages. So we
> +	 * should not have the case where structure wqi is across page, neither
> +	 * wrapped to the beginning. This simplifies the implementation below.
> +	 *
> +	 * XXX: if not the case, we need save data to a temp wqi and copy it to
> +	 * workqueue buffer dw by dw.
> +	 */
> +	WARN_ON(sizeof(struct guc_wq_item) != 16);
> +	WARN_ON(wq_off & 3);
> +
> +	/* wq starts from the page after doorbell / process_desc */
> +	base = kmap_atomic(i915_gem_object_get_page(gc->client_obj,
> +			(wq_off + GUC_DB_SIZE) >> PAGE_SHIFT));
> +	wq_off &= PAGE_SIZE - 1;
> +	wqi = (struct guc_wq_item *)((char *)base + wq_off);
> +
> +	/* len does not include the header */
> +	wq_len = sizeof(struct guc_wq_item) / sizeof(u32) - 1;
> +	wqi->header = WQ_TYPE_INORDER |
> +			(wq_len << WQ_LEN_SHIFT) |
> +			(ring_id << WQ_TARGET_SHIFT) |
> +			WQ_NO_WCFLUSH_WAIT;
> +
> +	/* The GuC wants only the low-order word of the context descriptor */
> +	wqi->context_desc = (u32)intel_lr_context_descriptor(rq->ctx, rq->ring);
> +
> +	/* The GuC firmware wants the tail index in QWords, not bytes */
> +	tail = rq->ringbuf->tail >> 3;
> +	wqi->ring_tail = tail << WQ_RING_TAIL_SHIFT;
> +	wqi->fence_id = 0; /*XXX: what fence to be here */
> +
> +	kunmap_atomic(base);
> +
> +	return 0;
> +}
> +
> +/**
> + * i915_guc_submit() - Submit commands through GuC
> + * @client:	the guc client where commands will go through
> + * @ctx:	LRC where commands come from
> + * @ring:	HW engine that will excute the commands
> + *
> + * Return:	0 if succeed
> + */
> +int i915_guc_submit(struct i915_guc_client *client,
> +		    struct drm_i915_gem_request *rq)
> +{
> +	unsigned long flags;
> +	int q_ret, b_ret;
> +
> +	/* Need this because of the deferred pin ctx and ring */
> +	/* Shall we move this right after ring is pinned? */
> +	intel_lr_context_update(rq);
> +
> +	q_ret = guc_add_workqueue_item(client, rq);
> +	if (q_ret == 0)
> +		b_ret = guc_ring_doorbell(client);
> +
> +	spin_lock_irqsave(&client->wq_lock, flags);
> +	client->submissions += 1;
> +	if (q_ret) {
> +		client->q_fail += 1;
> +		client->retcode = q_ret;
> +	} else if (b_ret) {
> +		client->b_fail += 1;
> +		client->retcode = q_ret = b_ret;
> +	} else {
> +		client->retcode = 0;
> +	}
> +	spin_unlock_irqrestore(&client->wq_lock, flags);
> +
> +	return q_ret;
> +}
> +
> +/*
> + * Everything below here is concerned with setup & teardown, and is
> + * therefore not part of the somewhat time-critical batch-submission
> + * path of i915_guc_submit() above.
> + */
> +
> +/**
>   * gem_allocate_guc_obj() - Allocate gem object for GuC usage
>   * @dev:	drm device
>   * @size:	size of object
> @@ -75,6 +581,121 @@ static void gem_release_guc_obj(struct drm_i915_gem_object *obj)
>  	drm_gem_object_unreference(&obj->base);
>  }
>  
> +static void guc_client_free(struct drm_device *dev,
> +			    struct i915_guc_client *client)
> +{
> +	struct drm_i915_private *dev_priv = dev->dev_private;
> +	struct intel_guc *guc = &dev_priv->guc;
> +
> +	if (!client)
> +		return;
> +
> +	if (client->doorbell_id != GUC_INVALID_DOORBELL_ID) {
> +		/*
> +		 * First disable the doorbell, then tell the GuC we've
> +		 * finished with it, finally deallocate it in our bitmap
> +		 */
> +		guc_disable_doorbell(guc, client);
> +		host2guc_release_doorbell(guc, client);
> +		release_doorbell(guc, client->doorbell_id);
> +	}
> +
> +	/*
> +	 * XXX: wait for any outstanding submissions before freeing memory.
> +	 * Be sure to drop any locks
> +	 */
> +
> +	gem_release_guc_obj(client->client_obj);
> +
> +	if (client->ctx_index != GUC_INVALID_CTX_ID) {
> +		guc_fini_ctx_desc(guc, client);
> +		ida_simple_remove(&guc->ctx_ids, client->ctx_index);
> +	}
> +
> +	kfree(client);
> +}
> +
> +/**
> + * guc_client_alloc() - Allocate an i915_guc_client
> + * @dev:	drm device
> + * @priority:	four levels priority _CRITICAL, _HIGH, _NORMAL and _LOW
> + * 		The kernel client to replace ExecList submission is created with
> + * 		NORMAL priority. Priority of a client for scheduler can be HIGH,
> + * 		while a preemption context can use CRITICAL.
> + *
> + * Return:	An i915_guc_client object if success.
> + */
> +static struct i915_guc_client *guc_client_alloc(struct drm_device *dev,
> +						uint32_t priority)
> +{
> +	struct i915_guc_client *client;
> +	struct drm_i915_private *dev_priv = dev->dev_private;
> +	struct intel_guc *guc = &dev_priv->guc;
> +	struct drm_i915_gem_object *obj;
> +
> +	client = kzalloc(sizeof(*client), GFP_KERNEL);
> +	if (!client)
> +		return NULL;
> +
> +	client->doorbell_id = GUC_INVALID_DOORBELL_ID;
> +	client->priority = priority;
> +
> +	client->ctx_index = (uint32_t)ida_simple_get(&guc->ctx_ids, 0,
> +			GUC_MAX_GPU_CONTEXTS, GFP_KERNEL);
> +	if (client->ctx_index >= GUC_MAX_GPU_CONTEXTS) {
> +		client->ctx_index = GUC_INVALID_CTX_ID;
> +		goto err;
> +	}
> +
> +	/* The first page is doorbell/proc_desc. Two followed pages are wq. */
> +	obj = gem_allocate_guc_obj(dev, GUC_DB_SIZE + GUC_WQ_SIZE);
> +	if (!obj)
> +		goto err;
> +
> +	client->client_obj = obj;
> +	client->wq_offset = GUC_DB_SIZE;
> +	client->wq_size = GUC_WQ_SIZE;
> +	spin_lock_init(&client->wq_lock);
> +
> +	client->doorbell_offset = select_doorbell_cacheline(guc);
> +
> +	/*
> +	 * Since the doorbell only requires a single cacheline, we can save
> +	 * space by putting the application process descriptor in the same
> +	 * page. Use the half of the page that doesn't include the doorbell.
> +	 */
> +	if (client->doorbell_offset >= (GUC_DB_SIZE / 2))
> +		client->proc_desc_offset = 0;
> +	else
> +		client->proc_desc_offset = (GUC_DB_SIZE / 2);
> +
> +	client->doorbell_id = assign_doorbell(guc, client->priority);
> +	if (client->doorbell_id == GUC_INVALID_DOORBELL_ID)
> +		/* XXX: evict a doorbell instead */
> +		goto err;
> +
> +	guc_init_proc_desc(guc, client);
> +	guc_init_ctx_desc(guc, client);
> +	guc_init_doorbell(guc, client);
> +
> +	/* Invalidate GuC TLB to let GuC take the latest updates to GTT. */
> +	I915_WRITE(GEN8_GTCR, GEN8_GTCR_INVALIDATE);
> +
> +	/* XXX: Any cache flushes needed? General domain mgmt calls? */
> +
> +	if (host2guc_allocate_doorbell(guc, client))
> +		goto err;
> +
> +	DRM_DEBUG_DRIVER("new priority %u client %p: ctx_index %u db_id %u\n",
> +		priority, client, client->ctx_index, client->doorbell_id);
> +
> +	return client;
> +
> +err:
> +	guc_client_free(dev, client);
> +	return NULL;
> +}
> +
>  static void guc_create_log(struct intel_guc *guc)
>  {
>  	struct drm_i915_private *dev_priv = guc_to_i915(guc);
> @@ -138,6 +759,8 @@ int i915_guc_submission_init(struct drm_device *dev)
>  	if (!guc->ctx_pool_obj)
>  		return -ENOMEM;
>  
> +	spin_lock_init(&dev_priv->guc.host2guc_lock);
> +
>  	ida_init(&guc->ctx_ids);
>  
>  	guc_create_log(guc);
> @@ -145,6 +768,32 @@ int i915_guc_submission_init(struct drm_device *dev)
>  	return 0;
>  }
>  
> +int i915_guc_submission_enable(struct drm_device *dev)
> +{
> +	struct drm_i915_private *dev_priv = dev->dev_private;
> +	struct intel_guc *guc = &dev_priv->guc;
> +	struct i915_guc_client *client;
> +
> +	/* client for execbuf submission */
> +	client = guc_client_alloc(dev, GUC_CTX_PRIORITY_NORMAL);
> +	if (!client) {
> +		DRM_ERROR("Failed to create execbuf guc_client\n");
> +		return -ENOMEM;
> +	}
> +
> +	guc->execbuf_client = client;
> +	return 0;
> +}
> +
> +void i915_guc_submission_disable(struct drm_device *dev)
> +{
> +	struct drm_i915_private *dev_priv = dev->dev_private;
> +	struct intel_guc *guc = &dev_priv->guc;
> +
> +	guc_client_free(dev, guc->execbuf_client);
> +	guc->execbuf_client = NULL;
> +}
> +
>  void i915_guc_submission_fini(struct drm_device *dev)
>  {
>  	struct drm_i915_private *dev_priv = dev->dev_private;
> diff --git a/drivers/gpu/drm/i915/intel_guc.h b/drivers/gpu/drm/i915/intel_guc.h
> index 5b51b05..d249326 100644
> --- a/drivers/gpu/drm/i915/intel_guc.h
> +++ b/drivers/gpu/drm/i915/intel_guc.h
> @@ -27,6 +27,30 @@
>  #include "intel_guc_fwif.h"
>  #include "i915_guc_reg.h"
>  
> +struct i915_guc_client {
> +	struct drm_i915_gem_object *client_obj;
> +	uint32_t priority;
> +	uint32_t ctx_index;
> +
> +	uint32_t proc_desc_offset;
> +	uint32_t doorbell_offset;
> +	uint32_t cookie;
> +	uint16_t doorbell_id;
> +	uint16_t padding;		/* Maintain alignment		*/
> +
> +	uint32_t wq_offset;
> +	uint32_t wq_size;
> +
> +	spinlock_t wq_lock;		/* Protects all data below	*/
> +	uint32_t wq_tail;
> +
> +	/* GuC submission statistics & status */
> +	uint64_t submissions;
> +	uint32_t q_fail;
> +	uint32_t b_fail;
> +	int retcode;
> +};
> +
>  enum intel_guc_fw_status {
>  	GUC_FIRMWARE_FAIL = -1,
>  	GUC_FIRMWARE_NONE = 0,
> @@ -60,6 +84,20 @@ struct intel_guc {
>  
>  	struct drm_i915_gem_object *ctx_pool_obj;
>  	struct ida ctx_ids;
> +
> +	struct i915_guc_client *execbuf_client;
> +
> +	spinlock_t host2guc_lock;	/* Protects all data below	*/
> +
> +	DECLARE_BITMAP(doorbell_bitmap, GUC_MAX_DOORBELLS);
> +	int db_cacheline;
> +
> +	/* Action status & statistics */
> +	uint64_t action_count;		/* Total commands issued	*/
> +	uint32_t action_cmd;		/* Last command word		*/
> +	uint32_t action_status;		/* Last return status		*/
> +	uint32_t action_fail;		/* Total number of failures	*/
> +	int32_t action_err;		/* Last error code		*/
>  };
>  
>  /* intel_guc_loader.c */
> @@ -70,6 +108,10 @@ extern const char *intel_guc_fw_status_repr(enum intel_guc_fw_status status);
>  
>  /* i915_guc_submission.c */
>  int i915_guc_submission_init(struct drm_device *dev);
> +int i915_guc_submission_enable(struct drm_device *dev);
> +int i915_guc_submit(struct i915_guc_client *client,
> +		    struct drm_i915_gem_request *rq);
> +void i915_guc_submission_disable(struct drm_device *dev);
>  void i915_guc_submission_fini(struct drm_device *dev);
>  
>  #endif
> diff --git a/drivers/gpu/drm/i915/intel_guc_loader.c b/drivers/gpu/drm/i915/intel_guc_loader.c
> index e5d7136..25ba29f 100644
> --- a/drivers/gpu/drm/i915/intel_guc_loader.c
> +++ b/drivers/gpu/drm/i915/intel_guc_loader.c
> @@ -427,6 +427,8 @@ int intel_guc_ucode_load(struct drm_device *dev)
>  		intel_guc_fw_status_repr(guc_fw->guc_fw_fetch_status),
>  		intel_guc_fw_status_repr(guc_fw->guc_fw_load_status));
>  
> +	i915_guc_submission_disable(dev);
> +
>  	if (guc_fw->guc_fw_fetch_status == GUC_FIRMWARE_NONE)
>  		return 0;
>  
> @@ -479,12 +481,20 @@ int intel_guc_ucode_load(struct drm_device *dev)
>  		intel_guc_fw_status_repr(guc_fw->guc_fw_fetch_status),
>  		intel_guc_fw_status_repr(guc_fw->guc_fw_load_status));
>  
> +	if (i915.enable_guc_submission) {
> +		err = i915_guc_submission_enable(dev);
> +		if (err)
> +			goto fail;
> +	}
> +
>  	return 0;
>  
>  fail:
>  	if (guc_fw->guc_fw_load_status == GUC_FIRMWARE_PENDING)
>  		guc_fw->guc_fw_load_status = GUC_FIRMWARE_FAIL;
>  
> +	i915_guc_submission_disable(dev);
> +
>  	DRM_ERROR("Failed to initialize GuC, error %d\n", err);
>  
>  	return err;
> @@ -547,6 +557,8 @@ void intel_guc_ucode_fini(struct drm_device *dev)
>  	struct drm_i915_private *dev_priv = dev->dev_private;
>  	struct intel_guc_fw *guc_fw = &dev_priv->guc.guc_fw;
>  
> +	i915_guc_submission_fini(dev);
> +
>  	if (guc_fw->guc_fw_obj)
>  		drm_gem_object_unreference(&guc_fw->guc_fw_obj->base);
>  	guc_fw->guc_fw_obj = NULL;
> -- 
> 1.9.1
>