drivers/gpu/drm/i915/intel_breadcrumbs.c

   1 /*
   2  * Copyright © 2015 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  */
  24
  25 #include <linux/kthread.h>
  26 #include <trace/events/dma_fence.h>
  27 #include <uapi/linux/sched/types.h>
  28
  29 #include "i915_drv.h"
  30
  31 static void irq_enable(struct intel_engine_cs *engine)
  32 {
  33         if (!engine->irq_enable)
  34                 return;
  35
  36         /* Caller disables interrupts */
  37         spin_lock(&engine->i915->irq_lock);
  38         engine->irq_enable(engine);
  39         spin_unlock(&engine->i915->irq_lock);
  40 }
  41
  42 static void irq_disable(struct intel_engine_cs *engine)
  43 {
  44         if (!engine->irq_disable)
  45                 return;
  46
  47         /* Caller disables interrupts */
  48         spin_lock(&engine->i915->irq_lock);
  49         engine->irq_disable(engine);
  50         spin_unlock(&engine->i915->irq_lock);
  51 }
  52
  53 static void __intel_breadcrumbs_disarm_irq(struct intel_breadcrumbs *b)
  54 {
  55         lockdep_assert_held(&b->irq_lock);
  56
  57         GEM_BUG_ON(!b->irq_enabled);
  58         if (!--b->irq_enabled)
  59                 irq_disable(container_of(b,
  60                                          struct intel_engine_cs,
  61                                          breadcrumbs));
  62
  63         b->irq_armed = false;
  64 }
  65
  66 void intel_engine_disarm_breadcrumbs(struct intel_engine_cs *engine)
  67 {
  68         struct intel_breadcrumbs *b = &engine->breadcrumbs;
  69
  70         if (!b->irq_armed)
  71                 return;
  72
  73         spin_lock_irq(&b->irq_lock);
  74         if (b->irq_armed)
  75                 __intel_breadcrumbs_disarm_irq(b);
  76         spin_unlock_irq(&b->irq_lock);
  77 }
  78
  79 static inline bool __request_completed(const struct i915_request *rq)
  80 {
  81         return i915_seqno_passed(__hwsp_seqno(rq), rq->fence.seqno);
  82 }
  83
  84 static bool
  85 __dma_fence_signal(struct dma_fence *fence)
  86 {
  87         return !test_and_set_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags);
  88 }
  89
  90 static void
  91 __dma_fence_signal__timestamp(struct dma_fence *fence, ktime_t timestamp)
  92 {
  93         fence->timestamp = timestamp;
  94         set_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &fence->flags);
  95         trace_dma_fence_signaled(fence);
  96 }
  97
  98 static void
  99 __dma_fence_signal__notify(struct dma_fence *fence)
 100 {
 101         struct dma_fence_cb *cur, *tmp;
 102
 103         lockdep_assert_held(fence->lock);
 104         lockdep_assert_irqs_disabled();
 105
 106         list_for_each_entry_safe(cur, tmp, &fence->cb_list, node) {
 107                 INIT_LIST_HEAD(&cur->node);
 108                 cur->func(fence, cur);
 109         }
 110         INIT_LIST_HEAD(&fence->cb_list);
 111 }
 112
 113 void intel_engine_breadcrumbs_irq(struct intel_engine_cs *engine)
 114 {
 115         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 116         const ktime_t timestamp = ktime_get();
 117         struct intel_context *ce, *cn;
 118         struct list_head *pos, *next;
 119         LIST_HEAD(signal);
 120
 121         spin_lock(&b->irq_lock);
 122
 123         if (b->irq_armed && list_empty(&b->signalers))
 124                 __intel_breadcrumbs_disarm_irq(b);
 125
 126         list_for_each_entry_safe(ce, cn, &b->signalers, signal_link) {
 127                 GEM_BUG_ON(list_empty(&ce->signals));
 128
 129                 list_for_each_safe(pos, next, &ce->signals) {
 130                         struct i915_request *rq =
 131                                 list_entry(pos, typeof(*rq), signal_link);
 132
 133                         if (!__request_completed(rq))
 134                                 break;
 135
 136                         GEM_BUG_ON(!test_bit(I915_FENCE_FLAG_SIGNAL,
 137                                              &rq->fence.flags));
 138                         clear_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
 139
 140                         if (!__dma_fence_signal(&rq->fence))
 141                                 continue;
 142
 143                         /*
 144                          * Queue for execution after dropping the signaling
 145                          * spinlock as the callback chain may end up adding
 146                          * more signalers to the same context or engine.
 147                          */
 148                         i915_request_get(rq);
 149                         list_add_tail(&rq->signal_link, &signal);
 150                 }
 151
 152                 /*
 153                  * We process the list deletion in bulk, only using a list_add
 154                  * (not list_move) above but keeping the status of
 155                  * rq->signal_link known with the I915_FENCE_FLAG_SIGNAL bit.
 156                  */
 157                 if (!list_is_first(pos, &ce->signals)) {
 158                         /* Advance the list to the first incomplete request */
 159                         __list_del_many(&ce->signals, pos);
 160                         if (&ce->signals == pos) /* now empty */
 161                                 list_del_init(&ce->signal_link);
 162                 }
 163         }
 164
 165         spin_unlock(&b->irq_lock);
 166
 167         list_for_each_safe(pos, next, &signal) {
 168                 struct i915_request *rq =
 169                         list_entry(pos, typeof(*rq), signal_link);
 170
 171                 __dma_fence_signal__timestamp(&rq->fence, timestamp);
 172
 173                 spin_lock(&rq->lock);
 174                 __dma_fence_signal__notify(&rq->fence);
 175                 spin_unlock(&rq->lock);
 176
 177                 i915_request_put(rq);
 178         }
 179 }
 180
 181 void intel_engine_signal_breadcrumbs(struct intel_engine_cs *engine)
 182 {
 183         local_irq_disable();
 184         intel_engine_breadcrumbs_irq(engine);
 185         local_irq_enable();
 186 }
 187
 188 static void signal_irq_work(struct irq_work *work)
 189 {
 190         struct intel_engine_cs *engine =
 191                 container_of(work, typeof(*engine), breadcrumbs.irq_work);
 192
 193         intel_engine_breadcrumbs_irq(engine);
 194 }
 195
 196 void intel_engine_pin_breadcrumbs_irq(struct intel_engine_cs *engine)
 197 {
 198         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 199
 200         spin_lock_irq(&b->irq_lock);
 201         if (!b->irq_enabled++)
 202                 irq_enable(engine);
 203         GEM_BUG_ON(!b->irq_enabled); /* no overflow! */
 204         spin_unlock_irq(&b->irq_lock);
 205 }
 206
 207 void intel_engine_unpin_breadcrumbs_irq(struct intel_engine_cs *engine)
 208 {
 209         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 210
 211         spin_lock_irq(&b->irq_lock);
 212         GEM_BUG_ON(!b->irq_enabled); /* no underflow! */
 213         if (!--b->irq_enabled)
 214                 irq_disable(engine);
 215         spin_unlock_irq(&b->irq_lock);
 216 }
 217
 218 static void __intel_breadcrumbs_arm_irq(struct intel_breadcrumbs *b)
 219 {
 220         struct intel_engine_cs *engine =
 221                 container_of(b, struct intel_engine_cs, breadcrumbs);
 222
 223         lockdep_assert_held(&b->irq_lock);
 224         if (b->irq_armed)
 225                 return;
 226
 227         /*
 228          * The breadcrumb irq will be disarmed on the interrupt after the
 229          * waiters are signaled. This gives us a single interrupt window in
 230          * which we can add a new waiter and avoid the cost of re-enabling
 231          * the irq.
 232          */
 233         b->irq_armed = true;
 234
 235         /*
 236          * Since we are waiting on a request, the GPU should be busy
 237          * and should have its own rpm reference. This is tracked
 238          * by i915->gt.awake, we can forgo holding our own wakref
 239          * for the interrupt as before i915->gt.awake is released (when
 240          * the driver is idle) we disarm the breadcrumbs.
 241          */
 242
 243         if (!b->irq_enabled++)
 244                 irq_enable(engine);
 245 }
 246
 247 void intel_engine_init_breadcrumbs(struct intel_engine_cs *engine)
 248 {
 249         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 250
 251         spin_lock_init(&b->irq_lock);
 252         INIT_LIST_HEAD(&b->signalers);
 253
 254         init_irq_work(&b->irq_work, signal_irq_work);
 255 }
 256
 257 void intel_engine_reset_breadcrumbs(struct intel_engine_cs *engine)
 258 {
 259         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 260         unsigned long flags;
 261
 262         spin_lock_irqsave(&b->irq_lock, flags);
 263
 264         if (b->irq_enabled)
 265                 irq_enable(engine);
 266         else
 267                 irq_disable(engine);
 268
 269         spin_unlock_irqrestore(&b->irq_lock, flags);
 270 }
 271
 272 void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine)
 273 {
 274 }
 275
 276 bool i915_request_enable_breadcrumb(struct i915_request *rq)
 277 {
 278         lockdep_assert_held(&rq->lock);
 279         lockdep_assert_irqs_disabled();
 280
 281         if (test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags)) {
 282                 struct intel_breadcrumbs *b = &rq->engine->breadcrumbs;
 283                 struct intel_context *ce = rq->hw_context;
 284                 struct list_head *pos;
 285
 286                 spin_lock(&b->irq_lock);
 287                 GEM_BUG_ON(test_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags));
 288
 289                 __intel_breadcrumbs_arm_irq(b);
 290
 291                 /*
 292                  * We keep the seqno in retirement order, so we can break
 293                  * inside intel_engine_breadcrumbs_irq as soon as we've passed
 294                  * the last completed request (or seen a request that hasn't
 295                  * event started). We could iterate the timeline->requests list,
 296                  * but keeping a separate signalers_list has the advantage of
 297                  * hopefully being much smaller than the full list and so
 298                  * provides faster iteration and detection when there are no
 299                  * more interrupts required for this context.
 300                  *
 301                  * We typically expect to add new signalers in order, so we
 302                  * start looking for our insertion point from the tail of
 303                  * the list.
 304                  */
 305                 list_for_each_prev(pos, &ce->signals) {
 306                         struct i915_request *it =
 307                                 list_entry(pos, typeof(*it), signal_link);
 308
 309                         if (i915_seqno_passed(rq->fence.seqno, it->fence.seqno))
 310                                 break;
 311                 }
 312                 list_add(&rq->signal_link, pos);
 313                 if (pos == &ce->signals) /* catch transitions from empty list */
 314                         list_move_tail(&ce->signal_link, &b->signalers);
 315
 316                 set_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
 317                 spin_unlock(&b->irq_lock);
 318         }
 319
 320         return !__request_completed(rq);
 321 }
 322
 323 void i915_request_cancel_breadcrumb(struct i915_request *rq)
 324 {
 325         struct intel_breadcrumbs *b = &rq->engine->breadcrumbs;
 326
 327         lockdep_assert_held(&rq->lock);
 328         lockdep_assert_irqs_disabled();
 329
 330         /*
 331          * We must wait for b->irq_lock so that we know the interrupt handler
 332          * has released its reference to the intel_context and has completed
 333          * the DMA_FENCE_FLAG_SIGNALED_BIT/I915_FENCE_FLAG_SIGNAL dance (if
 334          * required).
 335          */
 336         spin_lock(&b->irq_lock);
 337         if (test_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags)) {
 338                 struct intel_context *ce = rq->hw_context;
 339
 340                 list_del(&rq->signal_link);
 341                 if (list_empty(&ce->signals))
 342                         list_del_init(&ce->signal_link);
 343
 344                 clear_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
 345         }
 346         spin_unlock(&b->irq_lock);
 347 }
 348
 349 void intel_engine_print_breadcrumbs(struct intel_engine_cs *engine,
 350                                     struct drm_printer *p)
 351 {
 352         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 353         struct intel_context *ce;
 354         struct i915_request *rq;
 355
 356         if (list_empty(&b->signalers))
 357                 return;
 358
 359         drm_printf(p, "Signals:\n");
 360
 361         spin_lock_irq(&b->irq_lock);
 362         list_for_each_entry(ce, &b->signalers, signal_link) {
 363                 list_for_each_entry(rq, &ce->signals, signal_link) {
 364                         drm_printf(p, "\t[%llx:%llx%s] @ %dms\n",
 365                                    rq->fence.context, rq->fence.seqno,
 366                                    i915_request_completed(rq) ? "!" :
 367                                    i915_request_started(rq) ? "*" :
 368                                    "",
 369                                    jiffies_to_msecs(jiffies - rq->emitted_jiffies));
 370                 }
 371         }
 372         spin_unlock_irq(&b->irq_lock);
 373 }