drm/i915: Implementation of GuC submission client
authorDave Gordon <david.s.gordon@intel.com>
Wed, 12 Aug 2015 14:43:41 +0000 (15:43 +0100)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Fri, 14 Aug 2015 16:16:42 +0000 (18:16 +0200)
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]

v5:
    Add GuC per-engine submission & seqno statistics.
    Move wq locking to encompass both get_space() and add_item().
    Take forcewake lock in host2guc_action() [Tom O'Rourke]

v6:
    Fix GuC doorbell cacheline selection code (the
        cacheline-within-page calculation was wrong).
    Rename GuC priorities to make them closer to the names used in
        the GuC firmware source, matching what the autogenerated
        versions will (probably) be.
    Add per-ring statistics to client.

Issue: VIZ-4884
Signed-off-by: Alex Dai <yu.dai@intel.com>
Signed-off-by: Dave Gordon <david.s.gordon@intel.com>
Reviewed-by: Tom O'Rourke <Tom.O'Rourke@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_guc_submission.c
drivers/gpu/drm/i915/intel_guc.h
drivers/gpu/drm/i915/intel_guc_fwif.h
drivers/gpu/drm/i915/intel_guc_loader.c

index 669c889..3352b85 100644 (file)
 #include "i915_drv.h"
 #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;
+
+       intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
+       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);
+
+       /* No HOST2GUC command should take longer than 10ms */
+       ret = wait_for_atomic(host2guc_action_response(dev_priv, &status), 10);
+       if (status != GUC2HOST_STATUS_SUCCESS) {
+               /*
+                * Either the GuC explicitly returned an error (which
+                * we convert to -EIO here) or no response at all was
+                * received within the timeout limit (-ETIMEDOUT)
+                */
+               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);
+       intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
+
+       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 = cache_line_size();
+       uint32_t offset;
+
+       spin_lock(&guc->host2guc_lock);
+
+       /* Doorbell uses a single cache line within a page */
+       offset = offset_in_page(guc->db_cacheline);
+
+       /* Moving to next cache line to reduce contention */
+       guc->db_cacheline += cacheline_size;
+
+       spin_unlock(&guc->host2guc_lock);
+
+       DRM_DEBUG_DRIVER("selected doorbell cacheline 0x%x, next 0x%x, linesize %u\n",
+                       offset, guc->db_cacheline, cacheline_size);
+
+       return offset;
+}
+
+static uint16_t assign_doorbell(struct intel_guc *guc, uint32_t priority)
+{
+       /*
+        * The bitmap is split into two halves; the first half is used for
+        * normal priority contexts, the second half for high-priority ones.
+        * Note that logically higher priorities are numerically less than
+        * normal ones, so the test below means "is it high-priority?"
+        */
+       const bool hi_pri = (priority <= GUC_CTX_PRIORITY_HIGH);
+       const uint16_t half = GUC_MAX_DOORBELLS / 2;
+       const uint16_t start = hi_pri ? 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);
+
+       DRM_DEBUG_DRIVER("assigned %s priority doorbell id 0x%x\n",
+                       hi_pri ? "high" : "normal", id);
+
+       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;
+
+       base = kmap_atomic(i915_gem_object_get_page(gc->client_obj, 0));
+       desc = base + gc->proc_desc_offset;
+
+       while (timeout_counter-- > 0) {
+               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;
+               }
+       };
+
+       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)
+{
+       struct intel_guc *guc = client->guc;
+       enum intel_ring_id ring_id = rq->ring->id;
+       unsigned long flags;
+       int q_ret, b_ret;
+
+       spin_lock_irqsave(&client->wq_lock, flags);
+
+       q_ret = guc_add_workqueue_item(client, rq);
+       if (q_ret == 0)
+               b_ret = guc_ring_doorbell(client);
+
+       client->submissions[ring_id] += 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);
+
+       spin_lock(&guc->host2guc_lock);
+       guc->submissions[ring_id] += 1;
+       guc->last_seqno[ring_id] = rq->seqno;
+       spin_unlock(&guc->host2guc_lock);
+
+       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
@@ -79,6 +602,121 @@ 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->guc = guc;
+
+       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);
+
+       /* 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:
+       DRM_ERROR("FAILED to create priority %u GuC client!\n", priority);
+
+       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);
@@ -142,6 +780,8 @@ 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);
@@ -149,6 +789,32 @@ 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_KMD_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;
index 5b51b05..92658ce 100644 (file)
 #include "intel_guc_fwif.h"
 #include "i915_guc_reg.h"
 
+struct i915_guc_client {
+       struct drm_i915_gem_object *client_obj;
+       struct intel_guc *guc;
+       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[I915_NUM_RINGS];
+       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 +85,23 @@ 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);
+       uint32_t db_cacheline;          /* Cyclic counter mod pagesize  */
+
+       /* 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              */
+
+       uint64_t submissions[I915_NUM_RINGS];
+       uint32_t last_seqno[I915_NUM_RINGS];
 };
 
 /* intel_guc_loader.c */
@@ -70,6 +112,10 @@ 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
index 06aad6f..950c7e7 100644 (file)
 #define GFXCORE_FAMILY_GEN9            12
 #define GFXCORE_FAMILY_UNKNOWN         0x7fffffff
 
-#define GUC_CTX_PRIORITY_CRITICAL      0
+#define GUC_CTX_PRIORITY_KMD_HIGH      0
 #define GUC_CTX_PRIORITY_HIGH          1
-#define GUC_CTX_PRIORITY_NORMAL                2
-#define GUC_CTX_PRIORITY_LOW           3
+#define GUC_CTX_PRIORITY_KMD_NORMAL    2
+#define GUC_CTX_PRIORITY_NORMAL                3
 
 #define GUC_MAX_GPU_CONTEXTS           1024
 #define        GUC_INVALID_CTX_ID              (GUC_MAX_GPU_CONTEXTS + 1)
index 6ff7fea..8b4b057 100644 (file)
@@ -342,6 +342,8 @@ 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;
 
@@ -389,12 +391,20 @@ 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);
+
        return err;
 }