drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c

   1 /*
   2  * Copyright 2009 Jerome Glisse.
   3  * All Rights Reserved.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the
   7  * "Software"), to deal in the Software without restriction, including
   8  * without limitation the rights to use, copy, modify, merge, publish,
   9  * distribute, sub license, and/or sell copies of the Software, and to
  10  * permit persons to whom the Software is furnished to do so, subject to
  11  * the following conditions:
  12  *
  13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  20  *
  21  * The above copyright notice and this permission notice (including the
  22  * next paragraph) shall be included in all copies or substantial portions
  23  * of the Software.
  24  *
  25  */
  26 /*
  27  * Authors:
  28  *    Jerome Glisse <[email protected]>
  29  *    Dave Airlie
  30  */
  31 #include <linux/seq_file.h>
  32 #include <linux/atomic.h>
  33 #include <linux/wait.h>
  34 #include <linux/kref.h>
  35 #include <linux/slab.h>
  36 #include <linux/firmware.h>
  37 #include <linux/pm_runtime.h>
  38
  39 #include <drm/drm_drv.h>
  40 #include "amdgpu.h"
  41 #include "amdgpu_trace.h"
  42
  43 /*
  44  * Fences
  45  * Fences mark an event in the GPUs pipeline and are used
  46  * for GPU/CPU synchronization.  When the fence is written,
  47  * it is expected that all buffers associated with that fence
  48  * are no longer in use by the associated ring on the GPU and
  49  * that the the relevant GPU caches have been flushed.
  50  */
  51
  52 struct amdgpu_fence {
  53         struct dma_fence base;
  54
  55         /* RB, DMA, etc. */
  56         struct amdgpu_ring              *ring;
  57 };
  58
  59 static struct kmem_cache *amdgpu_fence_slab;
  60
  61 int amdgpu_fence_slab_init(void)
  62 {
  63         amdgpu_fence_slab = kmem_cache_create(
  64                 "amdgpu_fence", sizeof(struct amdgpu_fence), 0,
  65                 SLAB_HWCACHE_ALIGN, NULL);
  66         if (!amdgpu_fence_slab)
  67                 return -ENOMEM;
  68         return 0;
  69 }
  70
  71 void amdgpu_fence_slab_fini(void)
  72 {
  73         rcu_barrier();
  74         kmem_cache_destroy(amdgpu_fence_slab);
  75 }
  76 /*
  77  * Cast helper
  78  */
  79 static const struct dma_fence_ops amdgpu_fence_ops;
  80 static inline struct amdgpu_fence *to_amdgpu_fence(struct dma_fence *f)
  81 {
  82         struct amdgpu_fence *__f = container_of(f, struct amdgpu_fence, base);
  83
  84         if (__f->base.ops == &amdgpu_fence_ops)
  85                 return __f;
  86
  87         return NULL;
  88 }
  89
  90 /**
  91  * amdgpu_fence_write - write a fence value
  92  *
  93  * @ring: ring the fence is associated with
  94  * @seq: sequence number to write
  95  *
  96  * Writes a fence value to memory (all asics).
  97  */
  98 static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq)
  99 {
 100         struct amdgpu_fence_driver *drv = &ring->fence_drv;
 101
 102         if (drv->cpu_addr)
 103                 *drv->cpu_addr = cpu_to_le32(seq);
 104 }
 105
 106 /**
 107  * amdgpu_fence_read - read a fence value
 108  *
 109  * @ring: ring the fence is associated with
 110  *
 111  * Reads a fence value from memory (all asics).
 112  * Returns the value of the fence read from memory.
 113  */
 114 static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
 115 {
 116         struct amdgpu_fence_driver *drv = &ring->fence_drv;
 117         u32 seq = 0;
 118
 119         if (drv->cpu_addr)
 120                 seq = le32_to_cpu(*drv->cpu_addr);
 121         else
 122                 seq = atomic_read(&drv->last_seq);
 123
 124         return seq;
 125 }
 126
 127 /**
 128  * amdgpu_fence_emit - emit a fence on the requested ring
 129  *
 130  * @ring: ring the fence is associated with
 131  * @f: resulting fence object
 132  * @flags: flags to pass into the subordinate .emit_fence() call
 133  *
 134  * Emits a fence command on the requested ring (all asics).
 135  * Returns 0 on success, -ENOMEM on failure.
 136  */
 137 int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f,
 138                       unsigned flags)
 139 {
 140         struct amdgpu_device *adev = ring->adev;
 141         struct amdgpu_fence *fence;
 142         struct dma_fence __rcu **ptr;
 143         uint32_t seq;
 144         int r;
 145
 146         fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
 147         if (fence == NULL)
 148                 return -ENOMEM;
 149
 150         seq = ++ring->fence_drv.sync_seq;
 151         fence->ring = ring;
 152         dma_fence_init(&fence->base, &amdgpu_fence_ops,
 153                        &ring->fence_drv.lock,
 154                        adev->fence_context + ring->idx,
 155                        seq);
 156         amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
 157                                seq, flags | AMDGPU_FENCE_FLAG_INT);
 158         pm_runtime_get_noresume(adev_to_drm(adev)->dev);
 159         ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
 160         if (unlikely(rcu_dereference_protected(*ptr, 1))) {
 161                 struct dma_fence *old;
 162
 163                 rcu_read_lock();
 164                 old = dma_fence_get_rcu_safe(ptr);
 165                 rcu_read_unlock();
 166
 167                 if (old) {
 168                         r = dma_fence_wait(old, false);
 169                         dma_fence_put(old);
 170                         if (r)
 171                                 return r;
 172                 }
 173         }
 174
 175         /* This function can't be called concurrently anyway, otherwise
 176          * emitting the fence would mess up the hardware ring buffer.
 177          */
 178         rcu_assign_pointer(*ptr, dma_fence_get(&fence->base));
 179
 180         *f = &fence->base;
 181
 182         return 0;
 183 }
 184
 185 /**
 186  * amdgpu_fence_emit_polling - emit a fence on the requeste ring
 187  *
 188  * @ring: ring the fence is associated with
 189  * @s: resulting sequence number
 190  * @timeout: the timeout for waiting in usecs
 191  *
 192  * Emits a fence command on the requested ring (all asics).
 193  * Used For polling fence.
 194  * Returns 0 on success, -ENOMEM on failure.
 195  */
 196 int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s,
 197                               uint32_t timeout)
 198 {
 199         uint32_t seq;
 200         signed long r;
 201
 202         if (!s)
 203                 return -EINVAL;
 204
 205         seq = ++ring->fence_drv.sync_seq;
 206         r = amdgpu_fence_wait_polling(ring,
 207                                       seq - ring->fence_drv.num_fences_mask,
 208                                       timeout);
 209         if (r < 1)
 210                 return -ETIMEDOUT;
 211
 212         amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
 213                                seq, 0);
 214
 215         *s = seq;
 216
 217         return 0;
 218 }
 219
 220 /**
 221  * amdgpu_fence_schedule_fallback - schedule fallback check
 222  *
 223  * @ring: pointer to struct amdgpu_ring
 224  *
 225  * Start a timer as fallback to our interrupts.
 226  */
 227 static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring)
 228 {
 229         mod_timer(&ring->fence_drv.fallback_timer,
 230                   jiffies + AMDGPU_FENCE_JIFFIES_TIMEOUT);
 231 }
 232
 233 /**
 234  * amdgpu_fence_process - check for fence activity
 235  *
 236  * @ring: pointer to struct amdgpu_ring
 237  *
 238  * Checks the current fence value and calculates the last
 239  * signalled fence value. Wakes the fence queue if the
 240  * sequence number has increased.
 241  *
 242  * Returns true if fence was processed
 243  */
 244 bool amdgpu_fence_process(struct amdgpu_ring *ring)
 245 {
 246         struct amdgpu_fence_driver *drv = &ring->fence_drv;
 247         struct amdgpu_device *adev = ring->adev;
 248         uint32_t seq, last_seq;
 249         int r;
 250
 251         do {
 252                 last_seq = atomic_read(&ring->fence_drv.last_seq);
 253                 seq = amdgpu_fence_read(ring);
 254
 255         } while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
 256
 257         if (del_timer(&ring->fence_drv.fallback_timer) &&
 258             seq != ring->fence_drv.sync_seq)
 259                 amdgpu_fence_schedule_fallback(ring);
 260
 261         if (unlikely(seq == last_seq))
 262                 return false;
 263
 264         last_seq &= drv->num_fences_mask;
 265         seq &= drv->num_fences_mask;
 266
 267         do {
 268                 struct dma_fence *fence, **ptr;
 269
 270                 ++last_seq;
 271                 last_seq &= drv->num_fences_mask;
 272                 ptr = &drv->fences[last_seq];
 273
 274                 /* There is always exactly one thread signaling this fence slot */
 275                 fence = rcu_dereference_protected(*ptr, 1);
 276                 RCU_INIT_POINTER(*ptr, NULL);
 277
 278                 if (!fence)
 279                         continue;
 280
 281                 r = dma_fence_signal(fence);
 282                 if (!r)
 283                         DMA_FENCE_TRACE(fence, "signaled from irq context\n");
 284                 else
 285                         BUG();
 286
 287                 dma_fence_put(fence);
 288                 pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 289                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 290         } while (last_seq != seq);
 291
 292         return true;
 293 }
 294
 295 /**
 296  * amdgpu_fence_fallback - fallback for hardware interrupts
 297  *
 298  * @t: timer context used to obtain the pointer to ring structure
 299  *
 300  * Checks for fence activity.
 301  */
 302 static void amdgpu_fence_fallback(struct timer_list *t)
 303 {
 304         struct amdgpu_ring *ring = from_timer(ring, t,
 305                                               fence_drv.fallback_timer);
 306
 307         if (amdgpu_fence_process(ring))
 308                 DRM_WARN("Fence fallback timer expired on ring %s\n", ring->name);
 309 }
 310
 311 /**
 312  * amdgpu_fence_wait_empty - wait for all fences to signal
 313  *
 314  * @ring: ring index the fence is associated with
 315  *
 316  * Wait for all fences on the requested ring to signal (all asics).
 317  * Returns 0 if the fences have passed, error for all other cases.
 318  */
 319 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
 320 {
 321         uint64_t seq = READ_ONCE(ring->fence_drv.sync_seq);
 322         struct dma_fence *fence, **ptr;
 323         int r;
 324
 325         if (!seq)
 326                 return 0;
 327
 328         ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
 329         rcu_read_lock();
 330         fence = rcu_dereference(*ptr);
 331         if (!fence || !dma_fence_get_rcu(fence)) {
 332                 rcu_read_unlock();
 333                 return 0;
 334         }
 335         rcu_read_unlock();
 336
 337         r = dma_fence_wait(fence, false);
 338         dma_fence_put(fence);
 339         return r;
 340 }
 341
 342 /**
 343  * amdgpu_fence_wait_polling - busy wait for givn sequence number
 344  *
 345  * @ring: ring index the fence is associated with
 346  * @wait_seq: sequence number to wait
 347  * @timeout: the timeout for waiting in usecs
 348  *
 349  * Wait for all fences on the requested ring to signal (all asics).
 350  * Returns left time if no timeout, 0 or minus if timeout.
 351  */
 352 signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
 353                                       uint32_t wait_seq,
 354                                       signed long timeout)
 355 {
 356         uint32_t seq;
 357
 358         do {
 359                 seq = amdgpu_fence_read(ring);
 360                 udelay(5);
 361                 timeout -= 5;
 362         } while ((int32_t)(wait_seq - seq) > 0 && timeout > 0);
 363
 364         return timeout > 0 ? timeout : 0;
 365 }
 366 /**
 367  * amdgpu_fence_count_emitted - get the count of emitted fences
 368  *
 369  * @ring: ring the fence is associated with
 370  *
 371  * Get the number of fences emitted on the requested ring (all asics).
 372  * Returns the number of emitted fences on the ring.  Used by the
 373  * dynpm code to ring track activity.
 374  */
 375 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
 376 {
 377         uint64_t emitted;
 378
 379         /* We are not protected by ring lock when reading the last sequence
 380          * but it's ok to report slightly wrong fence count here.
 381          */
 382         amdgpu_fence_process(ring);
 383         emitted = 0x100000000ull;
 384         emitted -= atomic_read(&ring->fence_drv.last_seq);
 385         emitted += READ_ONCE(ring->fence_drv.sync_seq);
 386         return lower_32_bits(emitted);
 387 }
 388
 389 /**
 390  * amdgpu_fence_driver_start_ring - make the fence driver
 391  * ready for use on the requested ring.
 392  *
 393  * @ring: ring to start the fence driver on
 394  * @irq_src: interrupt source to use for this ring
 395  * @irq_type: interrupt type to use for this ring
 396  *
 397  * Make the fence driver ready for processing (all asics).
 398  * Not all asics have all rings, so each asic will only
 399  * start the fence driver on the rings it has.
 400  * Returns 0 for success, errors for failure.
 401  */
 402 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
 403                                    struct amdgpu_irq_src *irq_src,
 404                                    unsigned irq_type)
 405 {
 406         struct amdgpu_device *adev = ring->adev;
 407         uint64_t index;
 408
 409         if (ring->funcs->type != AMDGPU_RING_TYPE_UVD) {
 410                 ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs];
 411                 ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4);
 412         } else {
 413                 /* put fence directly behind firmware */
 414                 index = ALIGN(adev->uvd.fw->size, 8);
 415                 ring->fence_drv.cpu_addr = adev->uvd.inst[ring->me].cpu_addr + index;
 416                 ring->fence_drv.gpu_addr = adev->uvd.inst[ring->me].gpu_addr + index;
 417         }
 418         amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
 419
 420         if (irq_src)
 421                 amdgpu_irq_get(adev, irq_src, irq_type);
 422
 423         ring->fence_drv.irq_src = irq_src;
 424         ring->fence_drv.irq_type = irq_type;
 425         ring->fence_drv.initialized = true;
 426
 427         DRM_DEV_DEBUG(adev->dev, "fence driver on ring %s use gpu addr 0x%016llx\n",
 428                       ring->name, ring->fence_drv.gpu_addr);
 429         return 0;
 430 }
 431
 432 /**
 433  * amdgpu_fence_driver_init_ring - init the fence driver
 434  * for the requested ring.
 435  *
 436  * @ring: ring to init the fence driver on
 437  * @num_hw_submission: number of entries on the hardware queue
 438  * @sched_score: optional score atomic shared with other schedulers
 439  *
 440  * Init the fence driver for the requested ring (all asics).
 441  * Helper function for amdgpu_fence_driver_init().
 442  */
 443 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
 444                                   unsigned num_hw_submission,
 445                                   atomic_t *sched_score)
 446 {
 447         struct amdgpu_device *adev = ring->adev;
 448         long timeout;
 449         int r;
 450
 451         if (!adev)
 452                 return -EINVAL;
 453
 454         if (!is_power_of_2(num_hw_submission))
 455                 return -EINVAL;
 456
 457         ring->fence_drv.cpu_addr = NULL;
 458         ring->fence_drv.gpu_addr = 0;
 459         ring->fence_drv.sync_seq = 0;
 460         atomic_set(&ring->fence_drv.last_seq, 0);
 461         ring->fence_drv.initialized = false;
 462
 463         timer_setup(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, 0);
 464
 465         ring->fence_drv.num_fences_mask = num_hw_submission * 2 - 1;
 466         spin_lock_init(&ring->fence_drv.lock);
 467         ring->fence_drv.fences = kcalloc(num_hw_submission * 2, sizeof(void *),
 468                                          GFP_KERNEL);
 469         if (!ring->fence_drv.fences)
 470                 return -ENOMEM;
 471
 472         /* No need to setup the GPU scheduler for rings that don't need it */
 473         if (ring->no_scheduler)
 474                 return 0;
 475
 476         switch (ring->funcs->type) {
 477         case AMDGPU_RING_TYPE_GFX:
 478                 timeout = adev->gfx_timeout;
 479                 break;
 480         case AMDGPU_RING_TYPE_COMPUTE:
 481                 timeout = adev->compute_timeout;
 482                 break;
 483         case AMDGPU_RING_TYPE_SDMA:
 484                 timeout = adev->sdma_timeout;
 485                 break;
 486         default:
 487                 timeout = adev->video_timeout;
 488                 break;
 489         }
 490
 491         r = drm_sched_init(&ring->sched, &amdgpu_sched_ops,
 492                            num_hw_submission, amdgpu_job_hang_limit,
 493                            timeout, sched_score, ring->name);
 494         if (r) {
 495                 DRM_ERROR("Failed to create scheduler on ring %s.\n",
 496                           ring->name);
 497                 return r;
 498         }
 499
 500         return 0;
 501 }
 502
 503 /**
 504  * amdgpu_fence_driver_init - init the fence driver
 505  * for all possible rings.
 506  *
 507  * @adev: amdgpu device pointer
 508  *
 509  * Init the fence driver for all possible rings (all asics).
 510  * Not all asics have all rings, so each asic will only
 511  * start the fence driver on the rings it has using
 512  * amdgpu_fence_driver_start_ring().
 513  * Returns 0 for success.
 514  */
 515 int amdgpu_fence_driver_init(struct amdgpu_device *adev)
 516 {
 517         return 0;
 518 }
 519
 520 /**
 521  * amdgpu_fence_driver_fini - tear down the fence driver
 522  * for all possible rings.
 523  *
 524  * @adev: amdgpu device pointer
 525  *
 526  * Tear down the fence driver for all possible rings (all asics).
 527  */
 528 void amdgpu_fence_driver_fini_hw(struct amdgpu_device *adev)
 529 {
 530         int i, r;
 531
 532         for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
 533                 struct amdgpu_ring *ring = adev->rings[i];
 534
 535                 if (!ring || !ring->fence_drv.initialized)
 536                         continue;
 537                 if (!ring->no_scheduler)
 538                         drm_sched_fini(&ring->sched);
 539                 /* You can't wait for HW to signal if it's gone */
 540                 if (!drm_dev_is_unplugged(&adev->ddev))
 541                         r = amdgpu_fence_wait_empty(ring);
 542                 else
 543                         r = -ENODEV;
 544                 /* no need to trigger GPU reset as we are unloading */
 545                 if (r)
 546                         amdgpu_fence_driver_force_completion(ring);
 547
 548                 if (ring->fence_drv.irq_src)
 549                         amdgpu_irq_put(adev, ring->fence_drv.irq_src,
 550                                        ring->fence_drv.irq_type);
 551
 552                 del_timer_sync(&ring->fence_drv.fallback_timer);
 553         }
 554 }
 555
 556 void amdgpu_fence_driver_fini_sw(struct amdgpu_device *adev)
 557 {
 558         unsigned int i, j;
 559
 560         for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
 561                 struct amdgpu_ring *ring = adev->rings[i];
 562
 563                 if (!ring || !ring->fence_drv.initialized)
 564                         continue;
 565
 566                 for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
 567                         dma_fence_put(ring->fence_drv.fences[j]);
 568                 kfree(ring->fence_drv.fences);
 569                 ring->fence_drv.fences = NULL;
 570                 ring->fence_drv.initialized = false;
 571         }
 572 }
 573
 574 /**
 575  * amdgpu_fence_driver_suspend - suspend the fence driver
 576  * for all possible rings.
 577  *
 578  * @adev: amdgpu device pointer
 579  *
 580  * Suspend the fence driver for all possible rings (all asics).
 581  */
 582 void amdgpu_fence_driver_suspend(struct amdgpu_device *adev)
 583 {
 584         int i, r;
 585
 586         for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
 587                 struct amdgpu_ring *ring = adev->rings[i];
 588                 if (!ring || !ring->fence_drv.initialized)
 589                         continue;
 590
 591                 /* wait for gpu to finish processing current batch */
 592                 r = amdgpu_fence_wait_empty(ring);
 593                 if (r) {
 594                         /* delay GPU reset to resume */
 595                         amdgpu_fence_driver_force_completion(ring);
 596                 }
 597
 598                 /* disable the interrupt */
 599                 if (ring->fence_drv.irq_src)
 600                         amdgpu_irq_put(adev, ring->fence_drv.irq_src,
 601                                        ring->fence_drv.irq_type);
 602         }
 603 }
 604
 605 /**
 606  * amdgpu_fence_driver_resume - resume the fence driver
 607  * for all possible rings.
 608  *
 609  * @adev: amdgpu device pointer
 610  *
 611  * Resume the fence driver for all possible rings (all asics).
 612  * Not all asics have all rings, so each asic will only
 613  * start the fence driver on the rings it has using
 614  * amdgpu_fence_driver_start_ring().
 615  * Returns 0 for success.
 616  */
 617 void amdgpu_fence_driver_resume(struct amdgpu_device *adev)
 618 {
 619         int i;
 620
 621         for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
 622                 struct amdgpu_ring *ring = adev->rings[i];
 623                 if (!ring || !ring->fence_drv.initialized)
 624                         continue;
 625
 626                 /* enable the interrupt */
 627                 if (ring->fence_drv.irq_src)
 628                         amdgpu_irq_get(adev, ring->fence_drv.irq_src,
 629                                        ring->fence_drv.irq_type);
 630         }
 631 }
 632
 633 /**
 634  * amdgpu_fence_driver_force_completion - force signal latest fence of ring
 635  *
 636  * @ring: fence of the ring to signal
 637  *
 638  */
 639 void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
 640 {
 641         amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
 642         amdgpu_fence_process(ring);
 643 }
 644
 645 /*
 646  * Common fence implementation
 647  */
 648
 649 static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
 650 {
 651         return "amdgpu";
 652 }
 653
 654 static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
 655 {
 656         struct amdgpu_fence *fence = to_amdgpu_fence(f);
 657         return (const char *)fence->ring->name;
 658 }
 659
 660 /**
 661  * amdgpu_fence_enable_signaling - enable signalling on fence
 662  * @f: fence
 663  *
 664  * This function is called with fence_queue lock held, and adds a callback
 665  * to fence_queue that checks if this fence is signaled, and if so it
 666  * signals the fence and removes itself.
 667  */
 668 static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
 669 {
 670         struct amdgpu_fence *fence = to_amdgpu_fence(f);
 671         struct amdgpu_ring *ring = fence->ring;
 672
 673         if (!timer_pending(&ring->fence_drv.fallback_timer))
 674                 amdgpu_fence_schedule_fallback(ring);
 675
 676         DMA_FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx);
 677
 678         return true;
 679 }
 680
 681 /**
 682  * amdgpu_fence_free - free up the fence memory
 683  *
 684  * @rcu: RCU callback head
 685  *
 686  * Free up the fence memory after the RCU grace period.
 687  */
 688 static void amdgpu_fence_free(struct rcu_head *rcu)
 689 {
 690         struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
 691         struct amdgpu_fence *fence = to_amdgpu_fence(f);
 692         kmem_cache_free(amdgpu_fence_slab, fence);
 693 }
 694
 695 /**
 696  * amdgpu_fence_release - callback that fence can be freed
 697  *
 698  * @f: fence
 699  *
 700  * This function is called when the reference count becomes zero.
 701  * It just RCU schedules freeing up the fence.
 702  */
 703 static void amdgpu_fence_release(struct dma_fence *f)
 704 {
 705         call_rcu(&f->rcu, amdgpu_fence_free);
 706 }
 707
 708 static const struct dma_fence_ops amdgpu_fence_ops = {
 709         .get_driver_name = amdgpu_fence_get_driver_name,
 710         .get_timeline_name = amdgpu_fence_get_timeline_name,
 711         .enable_signaling = amdgpu_fence_enable_signaling,
 712         .release = amdgpu_fence_release,
 713 };
 714
 715 /*
 716  * Fence debugfs
 717  */
 718 #if defined(CONFIG_DEBUG_FS)
 719 static int amdgpu_debugfs_fence_info_show(struct seq_file *m, void *unused)
 720 {
 721         struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
 722         int i;
 723
 724         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
 725                 struct amdgpu_ring *ring = adev->rings[i];
 726                 if (!ring || !ring->fence_drv.initialized)
 727                         continue;
 728
 729                 amdgpu_fence_process(ring);
 730
 731                 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
 732                 seq_printf(m, "Last signaled fence          0x%08x\n",
 733                            atomic_read(&ring->fence_drv.last_seq));
 734                 seq_printf(m, "Last emitted                 0x%08x\n",
 735                            ring->fence_drv.sync_seq);
 736
 737                 if (ring->funcs->type == AMDGPU_RING_TYPE_GFX ||
 738                     ring->funcs->type == AMDGPU_RING_TYPE_SDMA) {
 739                         seq_printf(m, "Last signaled trailing fence 0x%08x\n",
 740                                    le32_to_cpu(*ring->trail_fence_cpu_addr));
 741                         seq_printf(m, "Last emitted                 0x%08x\n",
 742                                    ring->trail_seq);
 743                 }
 744
 745                 if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
 746                         continue;
 747
 748                 /* set in CP_VMID_PREEMPT and preemption occurred */
 749                 seq_printf(m, "Last preempted               0x%08x\n",
 750                            le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
 751                 /* set in CP_VMID_RESET and reset occurred */
 752                 seq_printf(m, "Last reset                   0x%08x\n",
 753                            le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
 754                 /* Both preemption and reset occurred */
 755                 seq_printf(m, "Last both                    0x%08x\n",
 756                            le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
 757         }
 758         return 0;
 759 }
 760
 761 /*
 762  * amdgpu_debugfs_gpu_recover - manually trigger a gpu reset & recover
 763  *
 764  * Manually trigger a gpu reset at the next fence wait.
 765  */
 766 static int gpu_recover_get(void *data, u64 *val)
 767 {
 768         struct amdgpu_device *adev = (struct amdgpu_device *)data;
 769         struct drm_device *dev = adev_to_drm(adev);
 770         int r;
 771
 772         r = pm_runtime_get_sync(dev->dev);
 773         if (r < 0) {
 774                 pm_runtime_put_autosuspend(dev->dev);
 775                 return 0;
 776         }
 777
 778         *val = amdgpu_device_gpu_recover(adev, NULL);
 779
 780         pm_runtime_mark_last_busy(dev->dev);
 781         pm_runtime_put_autosuspend(dev->dev);
 782
 783         return 0;
 784 }
 785
 786 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_fence_info);
 787 DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_gpu_recover_fops, gpu_recover_get, NULL,
 788                          "%lld\n");
 789
 790 #endif
 791
 792 void amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
 793 {
 794 #if defined(CONFIG_DEBUG_FS)
 795         struct drm_minor *minor = adev_to_drm(adev)->primary;
 796         struct dentry *root = minor->debugfs_root;
 797
 798         debugfs_create_file("amdgpu_fence_info", 0444, root, adev,
 799                             &amdgpu_debugfs_fence_info_fops);
 800
 801         if (!amdgpu_sriov_vf(adev))
 802                 debugfs_create_file("amdgpu_gpu_recover", 0444, root, adev,
 803                                     &amdgpu_debugfs_gpu_recover_fops);
 804 #endif
 805 }
 806