[Intel-gfx] [PATCH v3 4/8] drm/i915/guc: Remove GuC wq reservation
Michał Winiarski
michal.winiarski at intel.com
Mon May 22 22:07:51 UTC 2017
Now that we have an upper bound on the number of work items being sent
to GuC, we can remove the reservation.
v2: Multiply by number of engines in compile time assert (Daniele)
v3: Comment on reasoning behind the compile time assert (Chris)
Cc: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio at intel.com>
Cc: Jeff McGee <jeff.mcgee at intel.com>
Cc: Michal Wajdeczko <michal.wajdeczko at intel.com>
Cc: Oscar Mateo <oscar.mateo at intel.com>
Signed-off-by: Michał Winiarski <michal.winiarski at intel.com>
---
drivers/gpu/drm/i915/i915_debugfs.c | 2 -
drivers/gpu/drm/i915/i915_guc_submission.c | 89 +++++-------------------------
drivers/gpu/drm/i915/intel_lrc.c | 25 +--------
drivers/gpu/drm/i915/intel_uc.h | 8 ---
4 files changed, 17 insertions(+), 107 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_debugfs.c b/drivers/gpu/drm/i915/i915_debugfs.c
index 0969b8d..83366a4 100644
--- a/drivers/gpu/drm/i915/i915_debugfs.c
+++ b/drivers/gpu/drm/i915/i915_debugfs.c
@@ -2481,8 +2481,6 @@ static void i915_guc_client_info(struct seq_file *m,
seq_printf(m, "\tWQ size %d, offset: 0x%x, tail %d\n",
client->wq_size, client->wq_offset, client->wq_tail);
- seq_printf(m, "\tWork queue full: %u\n", client->no_wq_space);
-
for_each_engine(engine, dev_priv, id) {
u64 submissions = client->submissions[id];
tot += submissions;
diff --git a/drivers/gpu/drm/i915/i915_guc_submission.c b/drivers/gpu/drm/i915/i915_guc_submission.c
index 2a0c3161..66d8eaf 100644
--- a/drivers/gpu/drm/i915/i915_guc_submission.c
+++ b/drivers/gpu/drm/i915/i915_guc_submission.c
@@ -406,63 +406,6 @@ static void guc_stage_desc_fini(struct intel_guc *guc,
memset(desc, 0, sizeof(*desc));
}
-/**
- * i915_guc_wq_reserve() - reserve space in the GuC's workqueue
- * @request: request associated with the commands
- *
- * Return: 0 if space is available
- * -EAGAIN if space is not currently available
- *
- * This function must be called (and must return 0) before a request
- * is submitted to the GuC via i915_guc_submit() below. Once a result
- * of 0 has been returned, it must be balanced by a corresponding
- * call to submit().
- *
- * Reservation allows the caller to determine in advance that space
- * will be available for the next submission before committing resources
- * to it, and helps avoid late failures with complicated recovery paths.
- */
-int i915_guc_wq_reserve(struct drm_i915_gem_request *request)
-{
- const size_t wqi_size = sizeof(struct guc_wq_item);
- struct i915_guc_client *client = request->i915->guc.execbuf_client;
- struct guc_process_desc *desc = __get_process_desc(client);
- u32 freespace;
- int ret;
-
- spin_lock_irq(&client->wq_lock);
- freespace = CIRC_SPACE(client->wq_tail, desc->head, client->wq_size);
- freespace -= client->wq_rsvd;
- if (likely(freespace >= wqi_size)) {
- client->wq_rsvd += wqi_size;
- ret = 0;
- } else {
- client->no_wq_space++;
- ret = -EAGAIN;
- }
- spin_unlock_irq(&client->wq_lock);
-
- return ret;
-}
-
-static void guc_client_update_wq_rsvd(struct i915_guc_client *client, int size)
-{
- unsigned long flags;
-
- spin_lock_irqsave(&client->wq_lock, flags);
- client->wq_rsvd += size;
- spin_unlock_irqrestore(&client->wq_lock, flags);
-}
-
-void i915_guc_wq_unreserve(struct drm_i915_gem_request *request)
-{
- const int wqi_size = sizeof(struct guc_wq_item);
- struct i915_guc_client *client = request->i915->guc.execbuf_client;
-
- GEM_BUG_ON(READ_ONCE(client->wq_rsvd) < wqi_size);
- guc_client_update_wq_rsvd(client, -wqi_size);
-}
-
/* Construct a Work Item and append it to the GuC's Work Queue */
static void guc_wq_item_append(struct i915_guc_client *client,
struct drm_i915_gem_request *rq)
@@ -475,7 +418,7 @@ static void guc_wq_item_append(struct i915_guc_client *client,
struct guc_wq_item *wqi;
u32 freespace, tail, wq_off;
- /* Free space is guaranteed, see i915_guc_wq_reserve() above */
+ /* Free space is guaranteed */
freespace = CIRC_SPACE(client->wq_tail, desc->head, client->wq_size);
GEM_BUG_ON(freespace < wqi_size);
@@ -581,14 +524,6 @@ static int guc_ring_doorbell(struct i915_guc_client *client)
* i915_guc_submit() - Submit commands through GuC
* @rq: request associated with the commands
*
- * The caller must have already called i915_guc_wq_reserve() above with
- * a result of 0 (success), guaranteeing that there is space in the work
- * queue for the new request, so enqueuing the item cannot fail.
- *
- * Bad Things Will Happen if the caller violates this protocol e.g. calls
- * submit() when _reserve() says there's no space, or calls _submit()
- * a different number of times from (successful) calls to _reserve().
- *
* The only error here arises if the doorbell hardware isn't functioning
* as expected, which really shouln't happen.
*/
@@ -691,7 +626,6 @@ static void i915_guc_dequeue(struct intel_engine_cs *engine)
trace_i915_gem_request_in(rq, port_index(port, engine));
last = rq;
submit = true;
- i915_guc_wq_unreserve(rq);
}
rb = rb_next(rb);
@@ -1216,6 +1150,19 @@ int i915_guc_submission_enable(struct drm_i915_private *dev_priv)
enum intel_engine_id id;
int err;
+ /*
+ * We're using GuC work items for submitting work through GuC. Since
+ * we're coalescing multiple requests from a single context into a
+ * single work item prior to assigning it to execlist_port, we can
+ * never have more work items than the total number of ports (for all
+ * engines). The GuC firmware is controlling the HEAD of work queue,
+ * and it is guaranteed that it will remove the work item from the
+ * queue before our request is completed.
+ */
+ BUILD_BUG_ON(ARRAY_SIZE(engine->execlist_port) *
+ sizeof(struct guc_wq_item) *
+ I915_NUM_ENGINES > GUC_WQ_SIZE);
+
if (!client) {
client = guc_client_alloc(dev_priv,
INTEL_INFO(dev_priv)->ring_mask,
@@ -1243,8 +1190,6 @@ int i915_guc_submission_enable(struct drm_i915_private *dev_priv)
guc_interrupts_capture(dev_priv);
for_each_engine(engine, dev_priv, id) {
- const int wqi_size = sizeof(struct guc_wq_item);
- struct drm_i915_gem_request *rq;
struct execlist_port *port = engine->execlist_port;
int n;
@@ -1256,12 +1201,6 @@ int i915_guc_submission_enable(struct drm_i915_private *dev_priv)
engine->irq_tasklet.func = i915_guc_irq_handler;
clear_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted);
- /* Replay the current set of previously submitted requests */
- spin_lock_irq(&engine->timeline->lock);
- list_for_each_entry(rq, &engine->timeline->requests, link)
- guc_client_update_wq_rsvd(client, wqi_size);
- spin_unlock_irq(&engine->timeline->lock);
-
for (n = 0; n < ARRAY_SIZE(engine->execlist_port); n++) {
if (!port_isset(&port[n]))
break;
diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
index 014b30a..5d4f23c 100644
--- a/drivers/gpu/drm/i915/intel_lrc.c
+++ b/drivers/gpu/drm/i915/intel_lrc.c
@@ -913,27 +913,14 @@ static int execlists_request_alloc(struct drm_i915_gem_request *request)
*/
request->reserved_space += EXECLISTS_REQUEST_SIZE;
- if (i915.enable_guc_submission) {
- /*
- * Check that the GuC has space for the request before
- * going any further, as the i915_add_request() call
- * later on mustn't fail ...
- */
- ret = i915_guc_wq_reserve(request);
- if (ret)
- goto err;
- }
-
cs = intel_ring_begin(request, 0);
- if (IS_ERR(cs)) {
- ret = PTR_ERR(cs);
- goto err_unreserve;
- }
+ if (IS_ERR(cs))
+ return PTR_ERR(cs);
if (!ce->initialised) {
ret = engine->init_context(request);
if (ret)
- goto err_unreserve;
+ return ret;
ce->initialised = true;
}
@@ -947,12 +934,6 @@ static int execlists_request_alloc(struct drm_i915_gem_request *request)
request->reserved_space -= EXECLISTS_REQUEST_SIZE;
return 0;
-
-err_unreserve:
- if (i915.enable_guc_submission)
- i915_guc_wq_unreserve(request);
-err:
- return ret;
}
/*
diff --git a/drivers/gpu/drm/i915/intel_uc.h b/drivers/gpu/drm/i915/intel_uc.h
index 930f2e1..b86f866 100644
--- a/drivers/gpu/drm/i915/intel_uc.h
+++ b/drivers/gpu/drm/i915/intel_uc.h
@@ -55,10 +55,6 @@ struct drm_i915_gem_request;
*
* We also keep a few statistics on failures. Ideally, these should all
* be zero!
- * no_wq_space: times that the submission pre-check found no space was
- * available in the work queue (note, the queue is shared,
- * not per-engine). It is OK for this to be nonzero, but
- * it should not be huge!
*/
struct i915_guc_client {
struct i915_vma *vma;
@@ -79,8 +75,6 @@ struct i915_guc_client {
uint32_t wq_offset;
uint32_t wq_size;
uint32_t wq_tail;
- uint32_t wq_rsvd;
- uint32_t no_wq_space;
/* Per-engine counts of GuC submissions */
uint64_t submissions[I915_NUM_ENGINES];
@@ -242,8 +236,6 @@ u32 intel_guc_wopcm_size(struct drm_i915_private *dev_priv);
/* i915_guc_submission.c */
int i915_guc_submission_init(struct drm_i915_private *dev_priv);
int i915_guc_submission_enable(struct drm_i915_private *dev_priv);
-int i915_guc_wq_reserve(struct drm_i915_gem_request *rq);
-void i915_guc_wq_unreserve(struct drm_i915_gem_request *request);
void i915_guc_submission_disable(struct drm_i915_private *dev_priv);
void i915_guc_submission_fini(struct drm_i915_private *dev_priv);
struct i915_vma *intel_guc_allocate_vma(struct intel_guc *guc, u32 size);
--
2.9.4
More information about the Intel-gfx
mailing list