drivers/gpu/drm/i915/gt/intel_engine_pm.c

   1 /*
   2  * SPDX-License-Identifier: MIT
   3  *
   4  * Copyright © 2019 Intel Corporation
   5  */
   6
   7 #include "i915_drv.h"
   8
   9 #include "intel_context.h"
  10 #include "intel_engine.h"
  11 #include "intel_engine_heartbeat.h"
  12 #include "intel_engine_pm.h"
  13 #include "intel_engine_pool.h"
  14 #include "intel_gt.h"
  15 #include "intel_gt_pm.h"
  16 #include "intel_rc6.h"
  17 #include "intel_ring.h"
  18
  19 static int __engine_unpark(struct intel_wakeref *wf)
  20 {
  21         struct intel_engine_cs *engine =
  22                 container_of(wf, typeof(*engine), wakeref);
  23         struct intel_context *ce;
  24         void *map;
  25
  26         ENGINE_TRACE(engine, "\n");
  27
  28         intel_gt_pm_get(engine->gt);
  29
  30         /* Pin the default state for fast resets from atomic context. */
  31         map = NULL;
  32         if (engine->default_state)
  33                 map = i915_gem_object_pin_map(engine->default_state,
  34                                               I915_MAP_WB);
  35         if (!IS_ERR_OR_NULL(map))
  36                 engine->pinned_default_state = map;
  37
  38         /* Discard stale context state from across idling */
  39         ce = engine->kernel_context;
  40         if (ce) {
  41                 GEM_BUG_ON(test_bit(CONTEXT_VALID_BIT, &ce->flags));
  42
  43                 /* First poison the image to verify we never fully trust it */
  44                 if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) && ce->state) {
  45                         struct drm_i915_gem_object *obj = ce->state->obj;
  46                         int type = i915_coherent_map_type(engine->i915);
  47
  48                         map = i915_gem_object_pin_map(obj, type);
  49                         if (!IS_ERR(map)) {
  50                                 memset(map, CONTEXT_REDZONE, obj->base.size);
  51                                 i915_gem_object_flush_map(obj);
  52                                 i915_gem_object_unpin_map(obj);
  53                         }
  54                 }
  55
  56                 ce->ops->reset(ce);
  57         }
  58
  59         if (engine->unpark)
  60                 engine->unpark(engine);
  61
  62         intel_engine_unpark_heartbeat(engine);
  63         return 0;
  64 }
  65
  66 #if IS_ENABLED(CONFIG_LOCKDEP)
  67
  68 static inline unsigned long __timeline_mark_lock(struct intel_context *ce)
  69 {
  70         unsigned long flags;
  71
  72         local_irq_save(flags);
  73         mutex_acquire(&ce->timeline->mutex.dep_map, 2, 0, _THIS_IP_);
  74
  75         return flags;
  76 }
  77
  78 static inline void __timeline_mark_unlock(struct intel_context *ce,
  79                                           unsigned long flags)
  80 {
  81         mutex_release(&ce->timeline->mutex.dep_map, _THIS_IP_);
  82         local_irq_restore(flags);
  83 }
  84
  85 #else
  86
  87 static inline unsigned long __timeline_mark_lock(struct intel_context *ce)
  88 {
  89         return 0;
  90 }
  91
  92 static inline void __timeline_mark_unlock(struct intel_context *ce,
  93                                           unsigned long flags)
  94 {
  95 }
  96
  97 #endif /* !IS_ENABLED(CONFIG_LOCKDEP) */
  98
  99 static void duration(struct dma_fence *fence, struct dma_fence_cb *cb)
 100 {
 101         struct i915_request *rq = to_request(fence);
 102
 103         ewma__engine_latency_add(&rq->engine->latency,
 104                                  ktime_us_delta(rq->fence.timestamp,
 105                                                 rq->duration.emitted));
 106 }
 107
 108 static void
 109 __queue_and_release_pm(struct i915_request *rq,
 110                        struct intel_timeline *tl,
 111                        struct intel_engine_cs *engine)
 112 {
 113         struct intel_gt_timelines *timelines = &engine->gt->timelines;
 114
 115         ENGINE_TRACE(engine, "parking\n");
 116
 117         /*
 118          * We have to serialise all potential retirement paths with our
 119          * submission, as we don't want to underflow either the
 120          * engine->wakeref.counter or our timeline->active_count.
 121          *
 122          * Equally, we cannot allow a new submission to start until
 123          * after we finish queueing, nor could we allow that submitter
 124          * to retire us before we are ready!
 125          */
 126         spin_lock(&timelines->lock);
 127
 128         /* Let intel_gt_retire_requests() retire us (acquired under lock) */
 129         if (!atomic_fetch_inc(&tl->active_count))
 130                 list_add_tail(&tl->link, &timelines->active_list);
 131
 132         /* Hand the request over to HW and so engine_retire() */
 133         __i915_request_queue(rq, NULL);
 134
 135         /* Let new submissions commence (and maybe retire this timeline) */
 136         __intel_wakeref_defer_park(&engine->wakeref);
 137
 138         spin_unlock(&timelines->lock);
 139 }
 140
 141 static bool switch_to_kernel_context(struct intel_engine_cs *engine)
 142 {
 143         struct intel_context *ce = engine->kernel_context;
 144         struct i915_request *rq;
 145         unsigned long flags;
 146         bool result = true;
 147
 148         /* GPU is pointing to the void, as good as in the kernel context. */
 149         if (intel_gt_is_wedged(engine->gt))
 150                 return true;
 151
 152         GEM_BUG_ON(!intel_context_is_barrier(ce));
 153
 154         /* Already inside the kernel context, safe to power down. */
 155         if (engine->wakeref_serial == engine->serial)
 156                 return true;
 157
 158         /*
 159          * Note, we do this without taking the timeline->mutex. We cannot
 160          * as we may be called while retiring the kernel context and so
 161          * already underneath the timeline->mutex. Instead we rely on the
 162          * exclusive property of the __engine_park that prevents anyone
 163          * else from creating a request on this engine. This also requires
 164          * that the ring is empty and we avoid any waits while constructing
 165          * the context, as they assume protection by the timeline->mutex.
 166          * This should hold true as we can only park the engine after
 167          * retiring the last request, thus all rings should be empty and
 168          * all timelines idle.
 169          *
 170          * For unlocking, there are 2 other parties and the GPU who have a
 171          * stake here.
 172          *
 173          * A new gpu user will be waiting on the engine-pm to start their
 174          * engine_unpark. New waiters are predicated on engine->wakeref.count
 175          * and so intel_wakeref_defer_park() acts like a mutex_unlock of the
 176          * engine->wakeref.
 177          *
 178          * The other party is intel_gt_retire_requests(), which is walking the
 179          * list of active timelines looking for completions. Meanwhile as soon
 180          * as we call __i915_request_queue(), the GPU may complete our request.
 181          * Ergo, if we put ourselves on the timelines.active_list
 182          * (se intel_timeline_enter()) before we increment the
 183          * engine->wakeref.count, we may see the request completion and retire
 184          * it causing an undeflow of the engine->wakeref.
 185          */
 186         flags = __timeline_mark_lock(ce);
 187         GEM_BUG_ON(atomic_read(&ce->timeline->active_count) < 0);
 188
 189         rq = __i915_request_create(ce, GFP_NOWAIT);
 190         if (IS_ERR(rq))
 191                 /* Context switch failed, hope for the best! Maybe reset? */
 192                 goto out_unlock;
 193
 194         /* Check again on the next retirement. */
 195         engine->wakeref_serial = engine->serial + 1;
 196         i915_request_add_active_barriers(rq);
 197
 198         /* Install ourselves as a preemption barrier */
 199         rq->sched.attr.priority = I915_PRIORITY_BARRIER;
 200         if (likely(!__i915_request_commit(rq))) { /* engine should be idle! */
 201                 /*
 202                  * Use an interrupt for precise measurement of duration,
 203                  * otherwise we rely on someone else retiring all the requests
 204                  * which may delay the signaling (i.e. we will likely wait
 205                  * until the background request retirement running every
 206                  * second or two).
 207                  */
 208                 BUILD_BUG_ON(sizeof(rq->duration) > sizeof(rq->submitq));
 209                 dma_fence_add_callback(&rq->fence, &rq->duration.cb, duration);
 210                 rq->duration.emitted = ktime_get();
 211         }
 212
 213         /* Expose ourselves to the world */
 214         __queue_and_release_pm(rq, ce->timeline, engine);
 215
 216         result = false;
 217 out_unlock:
 218         __timeline_mark_unlock(ce, flags);
 219         return result;
 220 }
 221
 222 static void call_idle_barriers(struct intel_engine_cs *engine)
 223 {
 224         struct llist_node *node, *next;
 225
 226         llist_for_each_safe(node, next, llist_del_all(&engine->barrier_tasks)) {
 227                 struct dma_fence_cb *cb =
 228                         container_of((struct list_head *)node,
 229                                      typeof(*cb), node);
 230
 231                 cb->func(ERR_PTR(-EAGAIN), cb);
 232         }
 233 }
 234
 235 static int __engine_park(struct intel_wakeref *wf)
 236 {
 237         struct intel_engine_cs *engine =
 238                 container_of(wf, typeof(*engine), wakeref);
 239
 240         engine->saturated = 0;
 241
 242         /*
 243          * If one and only one request is completed between pm events,
 244          * we know that we are inside the kernel context and it is
 245          * safe to power down. (We are paranoid in case that runtime
 246          * suspend causes corruption to the active context image, and
 247          * want to avoid that impacting userspace.)
 248          */
 249         if (!switch_to_kernel_context(engine))
 250                 return -EBUSY;
 251
 252         ENGINE_TRACE(engine, "parked\n");
 253
 254         call_idle_barriers(engine); /* cleanup after wedging */
 255
 256         intel_engine_park_heartbeat(engine);
 257         intel_engine_disarm_breadcrumbs(engine);
 258         intel_engine_pool_park(&engine->pool);
 259
 260         /* Must be reset upon idling, or we may miss the busy wakeup. */
 261         GEM_BUG_ON(engine->execlists.queue_priority_hint != INT_MIN);
 262
 263         if (engine->park)
 264                 engine->park(engine);
 265
 266         if (engine->pinned_default_state) {
 267                 i915_gem_object_unpin_map(engine->default_state);
 268                 engine->pinned_default_state = NULL;
 269         }
 270
 271         engine->execlists.no_priolist = false;
 272
 273         /* While gt calls i915_vma_parked(), we have to break the lock cycle */
 274         intel_gt_pm_put_async(engine->gt);
 275         return 0;
 276 }
 277
 278 static const struct intel_wakeref_ops wf_ops = {
 279         .get = __engine_unpark,
 280         .put = __engine_park,
 281 };
 282
 283 void intel_engine_init__pm(struct intel_engine_cs *engine)
 284 {
 285         struct intel_runtime_pm *rpm = engine->uncore->rpm;
 286
 287         intel_wakeref_init(&engine->wakeref, rpm, &wf_ops);
 288         intel_engine_init_heartbeat(engine);
 289 }
 290
 291 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
 292 #include "selftest_engine_pm.c"
 293 #endif