]> git.baikalelectronics.ru Git - kernel.git/commitdiff
drm/i915/guc: Defer context unpin until scheduling is disabled
authorMatthew Brost <matthew.brost@intel.com>
Wed, 21 Jul 2021 21:50:51 +0000 (14:50 -0700)
committerJohn Harrison <John.C.Harrison@Intel.com>
Thu, 22 Jul 2021 17:07:13 +0000 (10:07 -0700)
With GuC scheduling, it isn't safe to unpin a context while scheduling
is enabled for that context as the GuC may touch some of the pinned
state (e.g. LRC). To ensure scheduling isn't enabled when an unpin is
done, a call back is added to intel_context_unpin when pin count == 1
to disable scheduling for that context. When the response CTB is
received it is safe to do the final unpin.

Future patches may add a heuristic / delay to schedule the disable
call back to avoid thrashing on schedule enable / disable.

v2:
 (John H)
  - s/drm_dbg/drm_err
 (Daneiel)
  - Clean up sched state function

Cc: John Harrison <john.c.harrison@intel.com>
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: John Harrison <John.C.Harrison@Intel.com>
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210721215101.139794-9-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/uc/intel_guc.h
drivers/gpu/drm/i915/gt/uc/intel_guc_ct.c
drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c

index 7c9159517f8a618bbf7b0f2f16b7431652b57c6b..dca25b6569aa42bb35622bf8b523545500235d3f 100644 (file)
@@ -306,9 +306,9 @@ retry:
        return err;
 }
 
-void intel_context_unpin(struct intel_context *ce)
+void __intel_context_do_unpin(struct intel_context *ce, int sub)
 {
-       if (!atomic_dec_and_test(&ce->pin_count))
+       if (!atomic_sub_and_test(sub, &ce->pin_count))
                return;
 
        CE_TRACE(ce, "unpin\n");
index b10cbe8fee9923f8e2a62572d2a4515d6936bb02..974ef85320c28835463f323334b36f08b0311143 100644 (file)
@@ -113,7 +113,32 @@ static inline void __intel_context_pin(struct intel_context *ce)
        atomic_inc(&ce->pin_count);
 }
 
-void intel_context_unpin(struct intel_context *ce);
+void __intel_context_do_unpin(struct intel_context *ce, int sub);
+
+static inline void intel_context_sched_disable_unpin(struct intel_context *ce)
+{
+       __intel_context_do_unpin(ce, 2);
+}
+
+static inline void intel_context_unpin(struct intel_context *ce)
+{
+       if (!ce->ops->sched_disable) {
+               __intel_context_do_unpin(ce, 1);
+       } else {
+               /*
+                * Move ownership of this pin to the scheduling disable which is
+                * an async operation. When that operation completes the above
+                * intel_context_sched_disable_unpin is called potentially
+                * unpinning the context.
+                */
+               while (!atomic_add_unless(&ce->pin_count, -1, 1)) {
+                       if (atomic_cmpxchg(&ce->pin_count, 1, 2) == 1) {
+                               ce->ops->sched_disable(ce);
+                               break;
+                       }
+               }
+       }
+}
 
 void intel_context_enter_engine(struct intel_context *ce);
 void intel_context_exit_engine(struct intel_context *ce);
index e0e3a937f7093681e335e3995c2c73b2d133fb8f..4a5518d295c2c0ff8c1aa8d84fa68e29ffeacd9b 100644 (file)
@@ -43,6 +43,8 @@ struct intel_context_ops {
        void (*enter)(struct intel_context *ce);
        void (*exit)(struct intel_context *ce);
 
+       void (*sched_disable)(struct intel_context *ce);
+
        void (*reset)(struct intel_context *ce);
        void (*destroy)(struct kref *kref);
 };
index 7fd6c3e343e454e5e2e892a9c7771c9fef365a26..4d470ebeda95e9174234be7c5420bc9ebbd7249e 100644 (file)
@@ -248,6 +248,8 @@ int intel_guc_reset_engine(struct intel_guc *guc,
 
 int intel_guc_deregister_done_process_msg(struct intel_guc *guc,
                                          const u32 *msg, u32 len);
+int intel_guc_sched_done_process_msg(struct intel_guc *guc,
+                                    const u32 *msg, u32 len);
 
 void intel_guc_load_status(struct intel_guc *guc, struct drm_printer *p);
 
index 28ff82c5be45e0332800bd14d840a5e5a7539169..019b25ff18888aaf965a5faa34795b45561ef654 100644 (file)
@@ -932,6 +932,9 @@ static int ct_process_request(struct intel_guc_ct *ct, struct ct_incoming_msg *r
                ret = intel_guc_deregister_done_process_msg(guc, payload,
                                                            len);
                break;
+       case INTEL_GUC_ACTION_SCHED_CONTEXT_MODE_DONE:
+               ret = intel_guc_sched_done_process_msg(guc, payload, len);
+               break;
        default:
                ret = -EOPNOTSUPP;
                break;
index a0871b800153dde8e883f9987eed5846897ba20a..2f393d9dba0d0d55773c9b50d1553c27468f71c6 100644 (file)
@@ -70,6 +70,7 @@
  * possible for some of the bits to changing at the same time though.
  */
 #define SCHED_STATE_NO_LOCK_ENABLED                    BIT(0)
+#define SCHED_STATE_NO_LOCK_PENDING_ENABLE             BIT(1)
 static inline bool context_enabled(struct intel_context *ce)
 {
        return (atomic_read(&ce->guc_sched_state_no_lock) &
@@ -87,6 +88,24 @@ static inline void clr_context_enabled(struct intel_context *ce)
                   &ce->guc_sched_state_no_lock);
 }
 
+static inline bool context_pending_enable(struct intel_context *ce)
+{
+       return (atomic_read(&ce->guc_sched_state_no_lock) &
+               SCHED_STATE_NO_LOCK_PENDING_ENABLE);
+}
+
+static inline void set_context_pending_enable(struct intel_context *ce)
+{
+       atomic_or(SCHED_STATE_NO_LOCK_PENDING_ENABLE,
+                 &ce->guc_sched_state_no_lock);
+}
+
+static inline void clr_context_pending_enable(struct intel_context *ce)
+{
+       atomic_and((u32)~SCHED_STATE_NO_LOCK_PENDING_ENABLE,
+                  &ce->guc_sched_state_no_lock);
+}
+
 /*
  * Below is a set of functions which control the GuC scheduling state which
  * require a lock, aside from the special case where the functions are called
@@ -95,6 +114,7 @@ static inline void clr_context_enabled(struct intel_context *ce)
  */
 #define SCHED_STATE_WAIT_FOR_DEREGISTER_TO_REGISTER    BIT(0)
 #define SCHED_STATE_DESTROYED                          BIT(1)
+#define SCHED_STATE_PENDING_DISABLE                    BIT(2)
 static inline void init_sched_state(struct intel_context *ce)
 {
        /* Only should be called from guc_lrc_desc_pin() */
@@ -138,6 +158,23 @@ set_context_destroyed(struct intel_context *ce)
        ce->guc_state.sched_state |= SCHED_STATE_DESTROYED;
 }
 
+static inline bool context_pending_disable(struct intel_context *ce)
+{
+       return ce->guc_state.sched_state & SCHED_STATE_PENDING_DISABLE;
+}
+
+static inline void set_context_pending_disable(struct intel_context *ce)
+{
+       lockdep_assert_held(&ce->guc_state.lock);
+       ce->guc_state.sched_state |= SCHED_STATE_PENDING_DISABLE;
+}
+
+static inline void clr_context_pending_disable(struct intel_context *ce)
+{
+       lockdep_assert_held(&ce->guc_state.lock);
+       ce->guc_state.sched_state &= ~SCHED_STATE_PENDING_DISABLE;
+}
+
 static inline bool context_guc_id_invalid(struct intel_context *ce)
 {
        return ce->guc_id == GUC_INVALID_LRC_ID;
@@ -230,6 +267,8 @@ static int guc_add_request(struct intel_guc *guc, struct i915_request *rq)
                action[len++] = INTEL_GUC_ACTION_SCHED_CONTEXT_MODE_SET;
                action[len++] = ce->guc_id;
                action[len++] = GUC_CONTEXT_ENABLE;
+               set_context_pending_enable(ce);
+               intel_context_get(ce);
        } else {
                action[len++] = INTEL_GUC_ACTION_SCHED_CONTEXT;
                action[len++] = ce->guc_id;
@@ -237,8 +276,12 @@ static int guc_add_request(struct intel_guc *guc, struct i915_request *rq)
 
        err = intel_guc_send_nb(guc, action, len);
 
-       if (!enabled && !err)
+       if (!enabled && !err) {
                set_context_enabled(ce);
+       } else if (!enabled) {
+               clr_context_pending_enable(ce);
+               intel_context_put(ce);
+       }
 
        return err;
 }
@@ -842,6 +885,62 @@ static void guc_context_post_unpin(struct intel_context *ce)
        lrc_post_unpin(ce);
 }
 
+static void __guc_context_sched_disable(struct intel_guc *guc,
+                                       struct intel_context *ce,
+                                       u16 guc_id)
+{
+       u32 action[] = {
+               INTEL_GUC_ACTION_SCHED_CONTEXT_MODE_SET,
+               guc_id, /* ce->guc_id not stable */
+               GUC_CONTEXT_DISABLE
+       };
+
+       GEM_BUG_ON(guc_id == GUC_INVALID_LRC_ID);
+
+       intel_context_get(ce);
+
+       intel_guc_send_busy_loop(guc, action, ARRAY_SIZE(action), true);
+}
+
+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);
+
+       return ce->guc_id;
+}
+
+static void guc_context_sched_disable(struct intel_context *ce)
+{
+       struct intel_guc *guc = ce_to_guc(ce);
+       struct intel_runtime_pm *runtime_pm = &ce->engine->gt->i915->runtime_pm;
+       unsigned long flags;
+       u16 guc_id;
+       intel_wakeref_t wakeref;
+
+       if (context_guc_id_invalid(ce) ||
+           !lrc_desc_registered(guc, ce->guc_id)) {
+               clr_context_enabled(ce);
+               goto unpin;
+       }
+
+       if (!context_enabled(ce))
+               goto unpin;
+
+       spin_lock_irqsave(&ce->guc_state.lock, flags);
+       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;
+unpin:
+       intel_context_sched_disable_unpin(ce);
+}
+
 static inline void guc_lrc_desc_unpin(struct intel_context *ce)
 {
        struct intel_guc *guc = ce_to_guc(ce);
@@ -849,6 +948,7 @@ static inline void guc_lrc_desc_unpin(struct intel_context *ce)
 
        GEM_BUG_ON(!lrc_desc_registered(guc, ce->guc_id));
        GEM_BUG_ON(ce != __get_context(guc, ce->guc_id));
+       GEM_BUG_ON(context_enabled(ce));
 
        spin_lock_irqsave(&ce->guc_state.lock, flags);
        set_context_destroyed(ce);
@@ -931,6 +1031,8 @@ static const struct intel_context_ops guc_context_ops = {
        .enter = intel_context_enter_engine,
        .exit = intel_context_exit_engine,
 
+       .sched_disable = guc_context_sched_disable,
+
        .reset = lrc_reset,
        .destroy = guc_context_destroy,
 };
@@ -1358,3 +1460,45 @@ int intel_guc_deregister_done_process_msg(struct intel_guc *guc,
 
        return 0;
 }
+
+int intel_guc_sched_done_process_msg(struct intel_guc *guc,
+                                    const u32 *msg,
+                                    u32 len)
+{
+       struct intel_context *ce;
+       unsigned long flags;
+       u32 desc_idx = msg[0];
+
+       if (unlikely(len < 2)) {
+               drm_err(&guc_to_gt(guc)->i915->drm, "Invalid length %u", len);
+               return -EPROTO;
+       }
+
+       ce = g2h_context_lookup(guc, desc_idx);
+       if (unlikely(!ce))
+               return -EPROTO;
+
+       if (unlikely(context_destroyed(ce) ||
+                    (!context_pending_enable(ce) &&
+                    !context_pending_disable(ce)))) {
+               drm_err(&guc_to_gt(guc)->i915->drm,
+                       "Bad context sched_state 0x%x, 0x%x, desc_idx %u",
+                       atomic_read(&ce->guc_sched_state_no_lock),
+                       ce->guc_state.sched_state, desc_idx);
+               return -EPROTO;
+       }
+
+       if (context_pending_enable(ce)) {
+               clr_context_pending_enable(ce);
+       } else if (context_pending_disable(ce)) {
+               intel_context_sched_disable_unpin(ce);
+
+               spin_lock_irqsave(&ce->guc_state.lock, flags);
+               clr_context_pending_disable(ce);
+               spin_unlock_irqrestore(&ce->guc_state.lock, flags);
+       }
+
+       intel_context_put(ce);
+
+       return 0;
+}