OSDN Git Service

drm/i915/guc: Support request cancellation
authorMatthew Brost <matthew.brost@intel.com>
Tue, 27 Jul 2021 00:23:40 +0000 (17:23 -0700)
committerJohn Harrison <John.C.Harrison@Intel.com>
Wed, 28 Jul 2021 00:32:14 +0000 (17:32 -0700)
This adds GuC backend support for i915_request_cancel(), which in turn
makes CONFIG_DRM_I915_REQUEST_TIMEOUT work.

This implementation makes use of fence while there are likely simplier
options. A fence was chosen because of another feature coming soon
which requires a user to block on a context until scheduling is
disabled. In that case we return the fence to the user and the user can
wait on that fence.

v2:
 (Daniele)
  - A comment about locking the blocked incr / decr
  - A comments about the use of the fence
  - Update commit message explaining why fence
  - Delete redundant check blocked count in unblock function
  - Ring buffer implementation
  - Comment about blocked in submission path
  - Shorter rpm path
v3:
 (Checkpatch)
  - Fix typos in commit message
 (Daniel)
  - Rework to simplier locking structure in guc_context_block / unblock

Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210727002348.97202-26-matthew.brost@intel.com
drivers/gpu/drm/i915/gt/intel_context.c
drivers/gpu/drm/i915/gt/intel_context.h
drivers/gpu/drm/i915/gt/intel_context_types.h
drivers/gpu/drm/i915/gt/intel_execlists_submission.c
drivers/gpu/drm/i915/gt/intel_ring_submission.c
drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
drivers/gpu/drm/i915/i915_request.c

index 237b70e..477c42d 100644 (file)
@@ -366,6 +366,12 @@ static int __intel_context_active(struct i915_active *active)
        return 0;
 }
 
+static int sw_fence_dummy_notify(struct i915_sw_fence *sf,
+                                enum i915_sw_fence_notify state)
+{
+       return NOTIFY_DONE;
+}
+
 void
 intel_context_init(struct intel_context *ce, struct intel_engine_cs *engine)
 {
@@ -399,6 +405,13 @@ intel_context_init(struct intel_context *ce, struct intel_engine_cs *engine)
        ce->guc_id = GUC_INVALID_LRC_ID;
        INIT_LIST_HEAD(&ce->guc_id_link);
 
+       /*
+        * Initialize fence to be complete as this is expected to be complete
+        * unless there is a pending schedule disable outstanding.
+        */
+       i915_sw_fence_init(&ce->guc_blocked, sw_fence_dummy_notify);
+       i915_sw_fence_commit(&ce->guc_blocked);
+
        i915_active_init(&ce->active,
                         __intel_context_active, __intel_context_retire, 0);
 }
index 814d927..876bdb0 100644 (file)
@@ -70,6 +70,13 @@ intel_context_is_pinned(struct intel_context *ce)
        return atomic_read(&ce->pin_count);
 }
 
+static inline void intel_context_cancel_request(struct intel_context *ce,
+                                               struct i915_request *rq)
+{
+       GEM_BUG_ON(!ce->ops->cancel_request);
+       return ce->ops->cancel_request(ce, rq);
+}
+
 /**
  * intel_context_unlock_pinned - Releases the earlier locking of 'pinned' status
  * @ce - the context
index 57c19ee..a5bc876 100644 (file)
@@ -13,6 +13,7 @@
 #include <linux/types.h>
 
 #include "i915_active_types.h"
+#include "i915_sw_fence.h"
 #include "i915_utils.h"
 #include "intel_engine_types.h"
 #include "intel_sseu.h"
@@ -42,6 +43,9 @@ struct intel_context_ops {
        void (*unpin)(struct intel_context *ce);
        void (*post_unpin)(struct intel_context *ce);
 
+       void (*cancel_request)(struct intel_context *ce,
+                              struct i915_request *rq);
+
        void (*enter)(struct intel_context *ce);
        void (*exit)(struct intel_context *ce);
 
@@ -156,7 +160,7 @@ struct intel_context {
                 * sched_state: scheduling state of this context using GuC
                 * submission
                 */
-               u8 sched_state;
+               u16 sched_state;
                /*
                 * fences: maintains of list of requests that have a submit
                 * fence related to GuC submission
@@ -184,6 +188,9 @@ struct intel_context {
         * GuC ID link - in list when unpinned but guc_id still valid in GuC
         */
        struct list_head guc_id_link;
+
+       /* GuC context blocked fence */
+       struct i915_sw_fence guc_blocked;
 };
 
 #endif /* __INTEL_CONTEXT_TYPES__ */
index b4a8767..de5f9c8 100644 (file)
 #include "gen8_engine_cs.h"
 #include "intel_breadcrumbs.h"
 #include "intel_context.h"
+#include "intel_engine_heartbeat.h"
 #include "intel_engine_pm.h"
 #include "intel_engine_stats.h"
 #include "intel_execlists_submission.h"
@@ -2587,11 +2588,26 @@ static int execlists_context_alloc(struct intel_context *ce)
        return lrc_alloc(ce, ce->engine);
 }
 
+static void execlists_context_cancel_request(struct intel_context *ce,
+                                            struct i915_request *rq)
+{
+       struct intel_engine_cs *engine = NULL;
+
+       i915_request_active_engine(rq, &engine);
+
+       if (engine && intel_engine_pulse(engine))
+               intel_gt_handle_error(engine->gt, engine->mask, 0,
+                                     "request cancellation by %s",
+                                     current->comm);
+}
+
 static const struct intel_context_ops execlists_context_ops = {
        .flags = COPS_HAS_INFLIGHT,
 
        .alloc = execlists_context_alloc,
 
+       .cancel_request = execlists_context_cancel_request,
+
        .pre_pin = execlists_context_pre_pin,
        .pin = execlists_context_pin,
        .unpin = lrc_unpin,
@@ -3608,6 +3624,8 @@ static const struct intel_context_ops virtual_context_ops = {
 
        .alloc = virtual_context_alloc,
 
+       .cancel_request = execlists_context_cancel_request,
+
        .pre_pin = virtual_context_pre_pin,
        .pin = virtual_context_pin,
        .unpin = lrc_unpin,
index 05bb9f4..2958e2f 100644 (file)
@@ -16,6 +16,7 @@
 #include "intel_reset.h"
 #include "intel_ring.h"
 #include "shmem_utils.h"
+#include "intel_engine_heartbeat.h"
 
 /* Rough estimate of the typical request size, performing a flush,
  * set-context and then emitting the batch.
@@ -604,9 +605,24 @@ static void ring_context_ban(struct intel_context *ce,
                }
 }
 
+static void ring_context_cancel_request(struct intel_context *ce,
+                                       struct i915_request *rq)
+{
+       struct intel_engine_cs *engine = NULL;
+
+       i915_request_active_engine(rq, &engine);
+
+       if (engine && intel_engine_pulse(engine))
+               intel_gt_handle_error(engine->gt, engine->mask, 0,
+                                     "request cancellation by %s",
+                                     current->comm);
+}
+
 static const struct intel_context_ops ring_context_ops = {
        .alloc = ring_context_alloc,
 
+       .cancel_request = ring_context_cancel_request,
+
        .ban = ring_context_ban,
 
        .pre_pin = ring_context_pre_pin,
index ad9a38a..ee4f1f9 100644 (file)
@@ -126,6 +126,9 @@ static inline void clr_context_pending_enable(struct intel_context *ce)
 #define SCHED_STATE_DESTROYED                          BIT(1)
 #define SCHED_STATE_PENDING_DISABLE                    BIT(2)
 #define SCHED_STATE_BANNED                             BIT(3)
+#define SCHED_STATE_BLOCKED_SHIFT                      4
+#define SCHED_STATE_BLOCKED            BIT(SCHED_STATE_BLOCKED_SHIFT)
+#define SCHED_STATE_BLOCKED_MASK       (0xfff << SCHED_STATE_BLOCKED_SHIFT)
 static inline void init_sched_state(struct intel_context *ce)
 {
        /* Only should be called from guc_lrc_desc_pin() */
@@ -203,6 +206,32 @@ static inline void clr_context_banned(struct intel_context *ce)
        ce->guc_state.sched_state &= ~SCHED_STATE_BANNED;
 }
 
+static inline u32 context_blocked(struct intel_context *ce)
+{
+       return (ce->guc_state.sched_state & SCHED_STATE_BLOCKED_MASK) >>
+               SCHED_STATE_BLOCKED_SHIFT;
+}
+
+static inline void incr_context_blocked(struct intel_context *ce)
+{
+       lockdep_assert_held(&ce->engine->sched_engine->lock);
+       lockdep_assert_held(&ce->guc_state.lock);
+
+       ce->guc_state.sched_state += SCHED_STATE_BLOCKED;
+
+       GEM_BUG_ON(!context_blocked(ce));       /* Overflow check */
+}
+
+static inline void decr_context_blocked(struct intel_context *ce)
+{
+       lockdep_assert_held(&ce->engine->sched_engine->lock);
+       lockdep_assert_held(&ce->guc_state.lock);
+
+       GEM_BUG_ON(!context_blocked(ce));       /* Underflow check */
+
+       ce->guc_state.sched_state -= SCHED_STATE_BLOCKED;
+}
+
 static inline bool context_guc_id_invalid(struct intel_context *ce)
 {
        return ce->guc_id == GUC_INVALID_LRC_ID;
@@ -404,6 +433,14 @@ static int guc_add_request(struct intel_guc *guc, struct i915_request *rq)
                if (unlikely(err))
                        goto out;
        }
+
+       /*
+        * The request / context will be run on the hardware when scheduling
+        * gets enabled in the unblock.
+        */
+       if (unlikely(context_blocked(ce)))
+               goto out;
+
        enabled = context_enabled(ce);
 
        if (!enabled) {
@@ -532,6 +569,7 @@ static void __guc_context_destroy(struct intel_context *ce);
 static void release_guc_id(struct intel_guc *guc, struct intel_context *ce);
 static void guc_signal_context_fence(struct intel_context *ce);
 static void guc_cancel_context_requests(struct intel_context *ce);
+static void guc_blocked_fence_complete(struct intel_context *ce);
 
 static void scrub_guc_desc_for_outstanding_g2h(struct intel_guc *guc)
 {
@@ -579,6 +617,10 @@ static void scrub_guc_desc_for_outstanding_g2h(struct intel_guc *guc)
                        }
                        intel_context_sched_disable_unpin(ce);
                        atomic_dec(&guc->outstanding_submission_g2h);
+                       spin_lock_irqsave(&ce->guc_state.lock, flags);
+                       guc_blocked_fence_complete(ce);
+                       spin_unlock_irqrestore(&ce->guc_state.lock, flags);
+
                        intel_context_put(ce);
                }
        }
@@ -1354,6 +1396,21 @@ static void guc_context_post_unpin(struct intel_context *ce)
        lrc_post_unpin(ce);
 }
 
+static void __guc_context_sched_enable(struct intel_guc *guc,
+                                      struct intel_context *ce)
+{
+       u32 action[] = {
+               INTEL_GUC_ACTION_SCHED_CONTEXT_MODE_SET,
+               ce->guc_id,
+               GUC_CONTEXT_ENABLE
+       };
+
+       trace_intel_context_sched_enable(ce);
+
+       guc_submission_send_busy_loop(guc, action, ARRAY_SIZE(action),
+                                     G2H_LEN_DW_SCHED_CONTEXT_MODE_SET, true);
+}
+
 static void __guc_context_sched_disable(struct intel_guc *guc,
                                        struct intel_context *ce,
                                        u16 guc_id)
@@ -1372,17 +1429,143 @@ static void __guc_context_sched_disable(struct intel_guc *guc,
                                      G2H_LEN_DW_SCHED_CONTEXT_MODE_SET, true);
 }
 
+static void guc_blocked_fence_complete(struct intel_context *ce)
+{
+       lockdep_assert_held(&ce->guc_state.lock);
+
+       if (!i915_sw_fence_done(&ce->guc_blocked))
+               i915_sw_fence_complete(&ce->guc_blocked);
+}
+
+static void guc_blocked_fence_reinit(struct intel_context *ce)
+{
+       lockdep_assert_held(&ce->guc_state.lock);
+       GEM_BUG_ON(!i915_sw_fence_done(&ce->guc_blocked));
+
+       /*
+        * This fence is always complete unless a pending schedule disable is
+        * outstanding. We arm the fence here and complete it when we receive
+        * the pending schedule disable complete message.
+        */
+       i915_sw_fence_fini(&ce->guc_blocked);
+       i915_sw_fence_reinit(&ce->guc_blocked);
+       i915_sw_fence_await(&ce->guc_blocked);
+       i915_sw_fence_commit(&ce->guc_blocked);
+}
+
 static u16 prep_context_pending_disable(struct intel_context *ce)
 {
        lockdep_assert_held(&ce->guc_state.lock);
 
        set_context_pending_disable(ce);
        clr_context_enabled(ce);
+       guc_blocked_fence_reinit(ce);
        intel_context_get(ce);
 
        return ce->guc_id;
 }
 
+static struct i915_sw_fence *guc_context_block(struct intel_context *ce)
+{
+       struct intel_guc *guc = ce_to_guc(ce);
+       struct i915_sched_engine *sched_engine = ce->engine->sched_engine;
+       unsigned long flags;
+       struct intel_runtime_pm *runtime_pm = ce->engine->uncore->rpm;
+       intel_wakeref_t wakeref;
+       u16 guc_id;
+       bool enabled;
+
+       spin_lock_irqsave(&ce->guc_state.lock, flags);
+
+       /*
+        * Sync with submission path, increment before below changes to context
+        * state.
+        */
+       spin_lock(&sched_engine->lock);
+       incr_context_blocked(ce);
+       spin_unlock(&sched_engine->lock);
+
+       enabled = context_enabled(ce);
+       if (unlikely(!enabled || submission_disabled(guc))) {
+               if (enabled)
+                       clr_context_enabled(ce);
+               spin_unlock_irqrestore(&ce->guc_state.lock, flags);
+               return &ce->guc_blocked;
+       }
+
+       /*
+        * We add +2 here as the schedule disable complete CTB handler calls
+        * intel_context_sched_disable_unpin (-2 to pin_count).
+        */
+       atomic_add(2, &ce->pin_count);
+
+       guc_id = prep_context_pending_disable(ce);
+
+       spin_unlock_irqrestore(&ce->guc_state.lock, flags);
+
+       with_intel_runtime_pm(runtime_pm, wakeref)
+               __guc_context_sched_disable(guc, ce, guc_id);
+
+       return &ce->guc_blocked;
+}
+
+static void guc_context_unblock(struct intel_context *ce)
+{
+       struct intel_guc *guc = ce_to_guc(ce);
+       struct i915_sched_engine *sched_engine = ce->engine->sched_engine;
+       unsigned long flags;
+       struct intel_runtime_pm *runtime_pm = ce->engine->uncore->rpm;
+       intel_wakeref_t wakeref;
+       bool enable;
+
+       GEM_BUG_ON(context_enabled(ce));
+
+       spin_lock_irqsave(&ce->guc_state.lock, flags);
+
+       if (unlikely(submission_disabled(guc) ||
+                    !intel_context_is_pinned(ce) ||
+                    context_pending_disable(ce) ||
+                    context_blocked(ce) > 1)) {
+               enable = false;
+       } else {
+               enable = true;
+               set_context_pending_enable(ce);
+               set_context_enabled(ce);
+               intel_context_get(ce);
+       }
+
+       /*
+        * Sync with submission path, decrement after above changes to context
+        * state.
+        */
+       spin_lock(&sched_engine->lock);
+       decr_context_blocked(ce);
+       spin_unlock(&sched_engine->lock);
+
+       spin_unlock_irqrestore(&ce->guc_state.lock, flags);
+
+       if (enable) {
+               with_intel_runtime_pm(runtime_pm, wakeref)
+                       __guc_context_sched_enable(guc, ce);
+       }
+}
+
+static void guc_context_cancel_request(struct intel_context *ce,
+                                      struct i915_request *rq)
+{
+       if (i915_sw_fence_signaled(&rq->submit)) {
+               struct i915_sw_fence *fence = guc_context_block(ce);
+
+               i915_sw_fence_wait(fence);
+               if (!i915_request_completed(rq)) {
+                       __i915_request_skip(rq);
+                       guc_reset_state(ce, intel_ring_wrap(ce->ring, rq->head),
+                                       true);
+               }
+               guc_context_unblock(ce);
+       }
+}
+
 static void __guc_context_set_preemption_timeout(struct intel_guc *guc,
                                                 u16 guc_id,
                                                 u32 preemption_timeout)
@@ -1642,6 +1825,8 @@ static const struct intel_context_ops guc_context_ops = {
 
        .ban = guc_context_ban,
 
+       .cancel_request = guc_context_cancel_request,
+
        .enter = intel_context_enter_engine,
        .exit = intel_context_exit_engine,
 
@@ -1837,6 +2022,8 @@ static const struct intel_context_ops virtual_guc_context_ops = {
 
        .ban = guc_context_ban,
 
+       .cancel_request = guc_context_cancel_request,
+
        .enter = guc_virtual_context_enter,
        .exit = guc_virtual_context_exit,
 
@@ -2295,6 +2482,7 @@ int intel_guc_sched_done_process_msg(struct intel_guc *guc,
                clr_context_banned(ce);
                clr_context_pending_disable(ce);
                __guc_signal_context_fence(ce);
+               guc_blocked_fence_complete(ce);
                spin_unlock_irqrestore(&ce->guc_state.lock, flags);
 
                if (banned) {
index 28f38b0..541a203 100644 (file)
@@ -710,18 +710,6 @@ void i915_request_unsubmit(struct i915_request *request)
        spin_unlock_irqrestore(&engine->sched_engine->lock, flags);
 }
 
-static void __cancel_request(struct i915_request *rq)
-{
-       struct intel_engine_cs *engine = NULL;
-
-       i915_request_active_engine(rq, &engine);
-
-       if (engine && intel_engine_pulse(engine))
-               intel_gt_handle_error(engine->gt, engine->mask, 0,
-                                     "request cancellation by %s",
-                                     current->comm);
-}
-
 void i915_request_cancel(struct i915_request *rq, int error)
 {
        if (!i915_request_set_error_once(rq, error))
@@ -729,7 +717,7 @@ void i915_request_cancel(struct i915_request *rq, int error)
 
        set_bit(I915_FENCE_FLAG_SENTINEL, &rq->fence.flags);
 
-       __cancel_request(rq);
+       intel_context_cancel_request(rq->context, rq);
 }
 
 static int __i915_sw_fence_call