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