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 <drm/drmP.h>
  38 #include "amdgpu.h"
  39 #include "amdgpu_trace.h"
  40
  41 /*
  42  * Fences
  43  * Fences mark an event in the GPUs pipeline and are used
  44  * for GPU/CPU synchronization.  When the fence is written,
  45  * it is expected that all buffers associated with that fence
  46  * are no longer in use by the associated ring on the GPU and
  47  * that the the relevant GPU caches have been flushed.
  48  */
  49
  50 /**
  51  * amdgpu_fence_write - write a fence value
  52  *
  53  * @ring: ring the fence is associated with
  54  * @seq: sequence number to write
  55  *
  56  * Writes a fence value to memory (all asics).
  57  */
  58 static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq)
  59 {
  60         struct amdgpu_fence_driver *drv = &ring->fence_drv;
  61
  62         if (drv->cpu_addr)
  63                 *drv->cpu_addr = cpu_to_le32(seq);
  64 }
  65
  66 /**
  67  * amdgpu_fence_read - read a fence value
  68  *
  69  * @ring: ring the fence is associated with
  70  *
  71  * Reads a fence value from memory (all asics).
  72  * Returns the value of the fence read from memory.
  73  */
  74 static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
  75 {
  76         struct amdgpu_fence_driver *drv = &ring->fence_drv;
  77         u32 seq = 0;
  78
  79         if (drv->cpu_addr)
  80                 seq = le32_to_cpu(*drv->cpu_addr);
  81         else
  82                 seq = lower_32_bits(atomic64_read(&drv->last_seq));
  83
  84         return seq;
  85 }
  86
  87 /**
  88  * amdgpu_fence_schedule_check - schedule lockup check
  89  *
  90  * @ring: pointer to struct amdgpu_ring
  91  *
  92  * Queues a delayed work item to check for lockups.
  93  */
  94 static void amdgpu_fence_schedule_check(struct amdgpu_ring *ring)
  95 {
  96         /*
  97          * Do not reset the timer here with mod_delayed_work,
  98          * this can livelock in an interaction with TTM delayed destroy.
  99          */
 100         queue_delayed_work(system_power_efficient_wq,
 101                 &ring->fence_drv.lockup_work,
 102                 AMDGPU_FENCE_JIFFIES_TIMEOUT);
 103 }
 104
 105 /**
 106  * amdgpu_fence_emit - emit a fence on the requested ring
 107  *
 108  * @ring: ring the fence is associated with
 109  * @owner: creator of the fence
 110  * @fence: amdgpu fence object
 111  *
 112  * Emits a fence command on the requested ring (all asics).
 113  * Returns 0 on success, -ENOMEM on failure.
 114  */
 115 int amdgpu_fence_emit(struct amdgpu_ring *ring, void *owner,
 116                       struct amdgpu_fence **fence)
 117 {
 118         struct amdgpu_device *adev = ring->adev;
 119
 120         /* we are protected by the ring emission mutex */
 121         *fence = kmalloc(sizeof(struct amdgpu_fence), GFP_KERNEL);
 122         if ((*fence) == NULL) {
 123                 return -ENOMEM;
 124         }
 125         (*fence)->seq = ++ring->fence_drv.sync_seq[ring->idx];
 126         (*fence)->ring = ring;
 127         (*fence)->owner = owner;
 128         fence_init(&(*fence)->base, &amdgpu_fence_ops,
 129                 &adev->fence_queue.lock, adev->fence_context + ring->idx,
 130                 (*fence)->seq);
 131         amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr, (*fence)->seq, false);
 132         trace_amdgpu_fence_emit(ring->adev->ddev, ring->idx, (*fence)->seq);
 133         return 0;
 134 }
 135
 136 /**
 137  * amdgpu_fence_check_signaled - callback from fence_queue
 138  *
 139  * this function is called with fence_queue lock held, which is also used
 140  * for the fence locking itself, so unlocked variants are used for
 141  * fence_signal, and remove_wait_queue.
 142  */
 143 static int amdgpu_fence_check_signaled(wait_queue_t *wait, unsigned mode, int flags, void *key)
 144 {
 145         struct amdgpu_fence *fence;
 146         struct amdgpu_device *adev;
 147         u64 seq;
 148         int ret;
 149
 150         fence = container_of(wait, struct amdgpu_fence, fence_wake);
 151         adev = fence->ring->adev;
 152
 153         /*
 154          * We cannot use amdgpu_fence_process here because we're already
 155          * in the waitqueue, in a call from wake_up_all.
 156          */
 157         seq = atomic64_read(&fence->ring->fence_drv.last_seq);
 158         if (seq >= fence->seq) {
 159                 ret = fence_signal_locked(&fence->base);
 160                 if (!ret)
 161                         FENCE_TRACE(&fence->base, "signaled from irq context\n");
 162                 else
 163                         FENCE_TRACE(&fence->base, "was already signaled\n");
 164
 165                 amdgpu_irq_put(adev, fence->ring->fence_drv.irq_src,
 166                                 fence->ring->fence_drv.irq_type);
 167                 __remove_wait_queue(&adev->fence_queue, &fence->fence_wake);
 168                 fence_put(&fence->base);
 169         } else
 170                 FENCE_TRACE(&fence->base, "pending\n");
 171         return 0;
 172 }
 173
 174 /**
 175  * amdgpu_fence_activity - check for fence activity
 176  *
 177  * @ring: pointer to struct amdgpu_ring
 178  *
 179  * Checks the current fence value and calculates the last
 180  * signalled fence value. Returns true if activity occured
 181  * on the ring, and the fence_queue should be waken up.
 182  */
 183 static bool amdgpu_fence_activity(struct amdgpu_ring *ring)
 184 {
 185         uint64_t seq, last_seq, last_emitted;
 186         unsigned count_loop = 0;
 187         bool wake = false;
 188
 189         /* Note there is a scenario here for an infinite loop but it's
 190          * very unlikely to happen. For it to happen, the current polling
 191          * process need to be interrupted by another process and another
 192          * process needs to update the last_seq btw the atomic read and
 193          * xchg of the current process.
 194          *
 195          * More over for this to go in infinite loop there need to be
 196          * continuously new fence signaled ie amdgpu_fence_read needs
 197          * to return a different value each time for both the currently
 198          * polling process and the other process that xchg the last_seq
 199          * btw atomic read and xchg of the current process. And the
 200          * value the other process set as last seq must be higher than
 201          * the seq value we just read. Which means that current process
 202          * need to be interrupted after amdgpu_fence_read and before
 203          * atomic xchg.
 204          *
 205          * To be even more safe we count the number of time we loop and
 206          * we bail after 10 loop just accepting the fact that we might
 207          * have temporarly set the last_seq not to the true real last
 208          * seq but to an older one.
 209          */
 210         last_seq = atomic64_read(&ring->fence_drv.last_seq);
 211         do {
 212                 last_emitted = ring->fence_drv.sync_seq[ring->idx];
 213                 seq = amdgpu_fence_read(ring);
 214                 seq |= last_seq & 0xffffffff00000000LL;
 215                 if (seq < last_seq) {
 216                         seq &= 0xffffffff;
 217                         seq |= last_emitted & 0xffffffff00000000LL;
 218                 }
 219
 220                 if (seq <= last_seq || seq > last_emitted) {
 221                         break;
 222                 }
 223                 /* If we loop over we don't want to return without
 224                  * checking if a fence is signaled as it means that the
 225                  * seq we just read is different from the previous on.
 226                  */
 227                 wake = true;
 228                 last_seq = seq;
 229                 if ((count_loop++) > 10) {
 230                         /* We looped over too many time leave with the
 231                          * fact that we might have set an older fence
 232                          * seq then the current real last seq as signaled
 233                          * by the hw.
 234                          */
 235                         break;
 236                 }
 237         } while (atomic64_xchg(&ring->fence_drv.last_seq, seq) > seq);
 238
 239         if (seq < last_emitted)
 240                 amdgpu_fence_schedule_check(ring);
 241
 242         return wake;
 243 }
 244
 245 /**
 246  * amdgpu_fence_check_lockup - check for hardware lockup
 247  *
 248  * @work: delayed work item
 249  *
 250  * Checks for fence activity and if there is none probe
 251  * the hardware if a lockup occured.
 252  */
 253 static void amdgpu_fence_check_lockup(struct work_struct *work)
 254 {
 255         struct amdgpu_fence_driver *fence_drv;
 256         struct amdgpu_ring *ring;
 257
 258         fence_drv = container_of(work, struct amdgpu_fence_driver,
 259                                 lockup_work.work);
 260         ring = fence_drv->ring;
 261
 262         if (!down_read_trylock(&ring->adev->exclusive_lock)) {
 263                 /* just reschedule the check if a reset is going on */
 264                 amdgpu_fence_schedule_check(ring);
 265                 return;
 266         }
 267
 268         if (fence_drv->delayed_irq && ring->adev->ddev->irq_enabled) {
 269                 fence_drv->delayed_irq = false;
 270                 amdgpu_irq_update(ring->adev, fence_drv->irq_src,
 271                                 fence_drv->irq_type);
 272         }
 273
 274         if (amdgpu_fence_activity(ring))
 275                 wake_up_all(&ring->adev->fence_queue);
 276         else if (amdgpu_ring_is_lockup(ring)) {
 277                 /* good news we believe it's a lockup */
 278                 dev_warn(ring->adev->dev, "GPU lockup (current fence id "
 279                         "0x%016llx last fence id 0x%016llx on ring %d)\n",
 280                         (uint64_t)atomic64_read(&fence_drv->last_seq),
 281                         fence_drv->sync_seq[ring->idx], ring->idx);
 282
 283                 /* remember that we need an reset */
 284                 ring->adev->needs_reset = true;
 285                 wake_up_all(&ring->adev->fence_queue);
 286         }
 287         up_read(&ring->adev->exclusive_lock);
 288 }
 289
 290 /**
 291  * amdgpu_fence_process - process a fence
 292  *
 293  * @adev: amdgpu_device pointer
 294  * @ring: ring index the fence is associated with
 295  *
 296  * Checks the current fence value and wakes the fence queue
 297  * if the sequence number has increased (all asics).
 298  */
 299 void amdgpu_fence_process(struct amdgpu_ring *ring)
 300 {
 301         uint64_t seq, last_seq, last_emitted;
 302         unsigned count_loop = 0;
 303         bool wake = false;
 304
 305         /* Note there is a scenario here for an infinite loop but it's
 306          * very unlikely to happen. For it to happen, the current polling
 307          * process need to be interrupted by another process and another
 308          * process needs to update the last_seq btw the atomic read and
 309          * xchg of the current process.
 310          *
 311          * More over for this to go in infinite loop there need to be
 312          * continuously new fence signaled ie amdgpu_fence_read needs
 313          * to return a different value each time for both the currently
 314          * polling process and the other process that xchg the last_seq
 315          * btw atomic read and xchg of the current process. And the
 316          * value the other process set as last seq must be higher than
 317          * the seq value we just read. Which means that current process
 318          * need to be interrupted after amdgpu_fence_read and before
 319          * atomic xchg.
 320          *
 321          * To be even more safe we count the number of time we loop and
 322          * we bail after 10 loop just accepting the fact that we might
 323          * have temporarly set the last_seq not to the true real last
 324          * seq but to an older one.
 325          */
 326         last_seq = atomic64_read(&ring->fence_drv.last_seq);
 327         do {
 328                 last_emitted = ring->fence_drv.sync_seq[ring->idx];
 329                 seq = amdgpu_fence_read(ring);
 330                 seq |= last_seq & 0xffffffff00000000LL;
 331                 if (seq < last_seq) {
 332                         seq &= 0xffffffff;
 333                         seq |= last_emitted & 0xffffffff00000000LL;
 334                 }
 335
 336                 if (seq <= last_seq || seq > last_emitted) {
 337                         break;
 338                 }
 339                 /* If we loop over we don't want to return without
 340                  * checking if a fence is signaled as it means that the
 341                  * seq we just read is different from the previous on.
 342                  */
 343                 wake = true;
 344                 last_seq = seq;
 345                 if ((count_loop++) > 10) {
 346                         /* We looped over too many time leave with the
 347                          * fact that we might have set an older fence
 348                          * seq then the current real last seq as signaled
 349                          * by the hw.
 350                          */
 351                         break;
 352                 }
 353         } while (atomic64_xchg(&ring->fence_drv.last_seq, seq) > seq);
 354
 355         if (wake)
 356                 wake_up_all(&ring->adev->fence_queue);
 357 }
 358
 359 /**
 360  * amdgpu_fence_seq_signaled - check if a fence sequence number has signaled
 361  *
 362  * @ring: ring the fence is associated with
 363  * @seq: sequence number
 364  *
 365  * Check if the last signaled fence sequnce number is >= the requested
 366  * sequence number (all asics).
 367  * Returns true if the fence has signaled (current fence value
 368  * is >= requested value) or false if it has not (current fence
 369  * value is < the requested value.  Helper function for
 370  * amdgpu_fence_signaled().
 371  */
 372 static bool amdgpu_fence_seq_signaled(struct amdgpu_ring *ring, u64 seq)
 373 {
 374         if (atomic64_read(&ring->fence_drv.last_seq) >= seq)
 375                 return true;
 376
 377         /* poll new last sequence at least once */
 378         amdgpu_fence_process(ring);
 379         if (atomic64_read(&ring->fence_drv.last_seq) >= seq)
 380                 return true;
 381
 382         return false;
 383 }
 384
 385 static bool amdgpu_fence_is_signaled(struct fence *f)
 386 {
 387         struct amdgpu_fence *fence = to_amdgpu_fence(f);
 388         struct amdgpu_ring *ring = fence->ring;
 389         struct amdgpu_device *adev = ring->adev;
 390
 391         if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq)
 392                 return true;
 393
 394         if (down_read_trylock(&adev->exclusive_lock)) {
 395                 amdgpu_fence_process(ring);
 396                 up_read(&adev->exclusive_lock);
 397
 398                 if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq)
 399                         return true;
 400         }
 401         return false;
 402 }
 403
 404 /**
 405  * amdgpu_fence_enable_signaling - enable signalling on fence
 406  * @fence: fence
 407  *
 408  * This function is called with fence_queue lock held, and adds a callback
 409  * to fence_queue that checks if this fence is signaled, and if so it
 410  * signals the fence and removes itself.
 411  */
 412 static bool amdgpu_fence_enable_signaling(struct fence *f)
 413 {
 414         struct amdgpu_fence *fence = to_amdgpu_fence(f);
 415         struct amdgpu_ring *ring = fence->ring;
 416         struct amdgpu_device *adev = ring->adev;
 417
 418         if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq)
 419                 return false;
 420
 421         if (down_read_trylock(&adev->exclusive_lock)) {
 422                 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
 423                         ring->fence_drv.irq_type);
 424                 if (amdgpu_fence_activity(ring))
 425                         wake_up_all_locked(&adev->fence_queue);
 426
 427                 /* did fence get signaled after we enabled the sw irq? */
 428                 if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq) {
 429                         amdgpu_irq_put(adev, ring->fence_drv.irq_src,
 430                                 ring->fence_drv.irq_type);
 431                         up_read(&adev->exclusive_lock);
 432                         return false;
 433                 }
 434
 435                 up_read(&adev->exclusive_lock);
 436         } else {
 437                 /* we're probably in a lockup, lets not fiddle too much */
 438                 if (amdgpu_irq_get_delayed(adev, ring->fence_drv.irq_src,
 439                         ring->fence_drv.irq_type))
 440                         ring->fence_drv.delayed_irq = true;
 441                 amdgpu_fence_schedule_check(ring);
 442         }
 443
 444         fence->fence_wake.flags = 0;
 445         fence->fence_wake.private = NULL;
 446         fence->fence_wake.func = amdgpu_fence_check_signaled;
 447         __add_wait_queue(&adev->fence_queue, &fence->fence_wake);
 448         fence_get(f);
 449         FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx);
 450         return true;
 451 }
 452
 453 /**
 454  * amdgpu_fence_signaled - check if a fence has signaled
 455  *
 456  * @fence: amdgpu fence object
 457  *
 458  * Check if the requested fence has signaled (all asics).
 459  * Returns true if the fence has signaled or false if it has not.
 460  */
 461 bool amdgpu_fence_signaled(struct amdgpu_fence *fence)
 462 {
 463         if (!fence)
 464                 return true;
 465
 466         if (amdgpu_fence_seq_signaled(fence->ring, fence->seq)) {
 467                 if (!fence_signal(&fence->base))
 468                         FENCE_TRACE(&fence->base, "signaled from amdgpu_fence_signaled\n");
 469                 return true;
 470         }
 471
 472         return false;
 473 }
 474
 475 /**
 476  * amdgpu_fence_any_seq_signaled - check if any sequence number is signaled
 477  *
 478  * @adev: amdgpu device pointer
 479  * @seq: sequence numbers
 480  *
 481  * Check if the last signaled fence sequnce number is >= the requested
 482  * sequence number (all asics).
 483  * Returns true if any has signaled (current value is >= requested value)
 484  * or false if it has not. Helper function for amdgpu_fence_wait_seq.
 485  */
 486 static bool amdgpu_fence_any_seq_signaled(struct amdgpu_device *adev, u64 *seq)
 487 {
 488         unsigned i;
 489
 490         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
 491                 if (!adev->rings[i] || !seq[i])
 492                         continue;
 493
 494                 if (amdgpu_fence_seq_signaled(adev->rings[i], seq[i]))
 495                         return true;
 496         }
 497
 498         return false;
 499 }
 500
 501 /**
 502  * amdgpu_fence_wait_seq_timeout - wait for a specific sequence numbers
 503  *
 504  * @adev: amdgpu device pointer
 505  * @target_seq: sequence number(s) we want to wait for
 506  * @intr: use interruptable sleep
 507  * @timeout: maximum time to wait, or MAX_SCHEDULE_TIMEOUT for infinite wait
 508  *
 509  * Wait for the requested sequence number(s) to be written by any ring
 510  * (all asics).  Sequnce number array is indexed by ring id.
 511  * @intr selects whether to use interruptable (true) or non-interruptable
 512  * (false) sleep when waiting for the sequence number.  Helper function
 513  * for amdgpu_fence_wait_*().
 514  * Returns remaining time if the sequence number has passed, 0 when
 515  * the wait timeout, or an error for all other cases.
 516  * -EDEADLK is returned when a GPU lockup has been detected.
 517  */
 518 long amdgpu_fence_wait_seq_timeout(struct amdgpu_device *adev, u64 *target_seq,
 519                                    bool intr, long timeout)
 520 {
 521         uint64_t last_seq[AMDGPU_MAX_RINGS];
 522         bool signaled;
 523         int i, r;
 524
 525         if (timeout == 0) {
 526                 return amdgpu_fence_any_seq_signaled(adev, target_seq);
 527         }
 528
 529         while (!amdgpu_fence_any_seq_signaled(adev, target_seq)) {
 530
 531                 /* Save current sequence values, used to check for GPU lockups */
 532                 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
 533                         struct amdgpu_ring *ring = adev->rings[i];
 534
 535                         if (!ring || !target_seq[i])
 536                                 continue;
 537
 538                         last_seq[i] = atomic64_read(&ring->fence_drv.last_seq);
 539                         trace_amdgpu_fence_wait_begin(adev->ddev, i, target_seq[i]);
 540                         amdgpu_irq_get(adev, ring->fence_drv.irq_src,
 541                                        ring->fence_drv.irq_type);
 542                 }
 543
 544                 if (intr) {
 545                         r = wait_event_interruptible_timeout(adev->fence_queue, (
 546                                 (signaled = amdgpu_fence_any_seq_signaled(adev, target_seq))
 547                                  || adev->needs_reset), AMDGPU_FENCE_JIFFIES_TIMEOUT);
 548                 } else {
 549                         r = wait_event_timeout(adev->fence_queue, (
 550                                 (signaled = amdgpu_fence_any_seq_signaled(adev, target_seq))
 551                                  || adev->needs_reset), AMDGPU_FENCE_JIFFIES_TIMEOUT);
 552                 }
 553
 554                 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
 555                         struct amdgpu_ring *ring = adev->rings[i];
 556
 557                         if (!ring || !target_seq[i])
 558                                 continue;
 559
 560                         amdgpu_irq_put(adev, ring->fence_drv.irq_src,
 561                                        ring->fence_drv.irq_type);
 562                         trace_amdgpu_fence_wait_end(adev->ddev, i, target_seq[i]);
 563                 }
 564
 565                 if (unlikely(r < 0))
 566                         return r;
 567
 568                 if (unlikely(!signaled)) {
 569
 570                         if (adev->needs_reset)
 571                                 return -EDEADLK;
 572
 573                         /* we were interrupted for some reason and fence
 574                          * isn't signaled yet, resume waiting */
 575                         if (r)
 576                                 continue;
 577
 578                         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
 579                                 struct amdgpu_ring *ring = adev->rings[i];
 580
 581                                 if (!ring || !target_seq[i])
 582                                         continue;
 583
 584                                 if (last_seq[i] != atomic64_read(&ring->fence_drv.last_seq))
 585                                         break;
 586                         }
 587
 588                         if (i != AMDGPU_MAX_RINGS)
 589                                 continue;
 590
 591                         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
 592                                 if (!adev->rings[i] || !target_seq[i])
 593                                         continue;
 594
 595                                 if (amdgpu_ring_is_lockup(adev->rings[i]))
 596                                         break;
 597                         }
 598
 599                         if (i < AMDGPU_MAX_RINGS) {
 600                                 /* good news we believe it's a lockup */
 601                                 dev_warn(adev->dev, "GPU lockup (waiting for "
 602                                          "0x%016llx last fence id 0x%016llx on"
 603                                          " ring %d)\n",
 604                                          target_seq[i], last_seq[i], i);
 605
 606                                 /* remember that we need an reset */
 607                                 adev->needs_reset = true;
 608                                 wake_up_all(&adev->fence_queue);
 609                                 return -EDEADLK;
 610                         }
 611
 612                         if (timeout < MAX_SCHEDULE_TIMEOUT) {
 613                                 timeout -= AMDGPU_FENCE_JIFFIES_TIMEOUT;
 614                                 if (timeout <= 0) {
 615                                         return 0;
 616                                 }
 617                         }
 618                 }
 619         }
 620         return timeout;
 621 }
 622
 623 /**
 624  * amdgpu_fence_wait - wait for a fence to signal
 625  *
 626  * @fence: amdgpu fence object
 627  * @intr: use interruptable sleep
 628  *
 629  * Wait for the requested fence to signal (all asics).
 630  * @intr selects whether to use interruptable (true) or non-interruptable
 631  * (false) sleep when waiting for the fence.
 632  * Returns 0 if the fence has passed, error for all other cases.
 633  */
 634 int amdgpu_fence_wait(struct amdgpu_fence *fence, bool intr)
 635 {
 636         uint64_t seq[AMDGPU_MAX_RINGS] = {};
 637         long r;
 638
 639         seq[fence->ring->idx] = fence->seq;
 640         r = amdgpu_fence_wait_seq_timeout(fence->ring->adev, seq, intr, MAX_SCHEDULE_TIMEOUT);
 641         if (r < 0) {
 642                 return r;
 643         }
 644
 645         r = fence_signal(&fence->base);
 646         if (!r)
 647                 FENCE_TRACE(&fence->base, "signaled from fence_wait\n");
 648         return 0;
 649 }
 650
 651 /**
 652  * amdgpu_fence_wait_any - wait for a fence to signal on any ring
 653  *
 654  * @adev: amdgpu device pointer
 655  * @fences: amdgpu fence object(s)
 656  * @intr: use interruptable sleep
 657  *
 658  * Wait for any requested fence to signal (all asics).  Fence
 659  * array is indexed by ring id.  @intr selects whether to use
 660  * interruptable (true) or non-interruptable (false) sleep when
 661  * waiting for the fences. Used by the suballocator.
 662  * Returns 0 if any fence has passed, error for all other cases.
 663  */
 664 int amdgpu_fence_wait_any(struct amdgpu_device *adev,
 665                           struct amdgpu_fence **fences,
 666                           bool intr)
 667 {
 668         uint64_t seq[AMDGPU_MAX_RINGS];
 669         unsigned i, num_rings = 0;
 670         long r;
 671
 672         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
 673                 seq[i] = 0;
 674
 675                 if (!fences[i]) {
 676                         continue;
 677                 }
 678
 679                 seq[i] = fences[i]->seq;
 680                 ++num_rings;
 681         }
 682
 683         /* nothing to wait for ? */
 684         if (num_rings == 0)
 685                 return -ENOENT;
 686
 687         r = amdgpu_fence_wait_seq_timeout(adev, seq, intr, MAX_SCHEDULE_TIMEOUT);
 688         if (r < 0) {
 689                 return r;
 690         }
 691         return 0;
 692 }
 693
 694 /**
 695  * amdgpu_fence_wait_next - wait for the next fence to signal
 696  *
 697  * @adev: amdgpu device pointer
 698  * @ring: ring index the fence is associated with
 699  *
 700  * Wait for the next fence on the requested ring to signal (all asics).
 701  * Returns 0 if the next fence has passed, error for all other cases.
 702  * Caller must hold ring lock.
 703  */
 704 int amdgpu_fence_wait_next(struct amdgpu_ring *ring)
 705 {
 706         uint64_t seq[AMDGPU_MAX_RINGS] = {};
 707         long r;
 708
 709         seq[ring->idx] = atomic64_read(&ring->fence_drv.last_seq) + 1ULL;
 710         if (seq[ring->idx] >= ring->fence_drv.sync_seq[ring->idx]) {
 711                 /* nothing to wait for, last_seq is
 712                    already the last emited fence */
 713                 return -ENOENT;
 714         }
 715         r = amdgpu_fence_wait_seq_timeout(ring->adev, seq, false, MAX_SCHEDULE_TIMEOUT);
 716         if (r < 0)
 717                 return r;
 718         return 0;
 719 }
 720
 721 /**
 722  * amdgpu_fence_wait_empty - wait for all fences to signal
 723  *
 724  * @adev: amdgpu device pointer
 725  * @ring: ring index the fence is associated with
 726  *
 727  * Wait for all fences on the requested ring to signal (all asics).
 728  * Returns 0 if the fences have passed, error for all other cases.
 729  * Caller must hold ring lock.
 730  */
 731 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
 732 {
 733         struct amdgpu_device *adev = ring->adev;
 734         uint64_t seq[AMDGPU_MAX_RINGS] = {};
 735         long r;
 736
 737         seq[ring->idx] = ring->fence_drv.sync_seq[ring->idx];
 738         if (!seq[ring->idx])
 739                 return 0;
 740
 741         r = amdgpu_fence_wait_seq_timeout(adev, seq, false, MAX_SCHEDULE_TIMEOUT);
 742         if (r < 0) {
 743                 if (r == -EDEADLK)
 744                         return -EDEADLK;
 745
 746                 dev_err(adev->dev, "error waiting for ring[%d] to become idle (%ld)\n",
 747                         ring->idx, r);
 748         }
 749         return 0;
 750 }
 751
 752 /**
 753  * amdgpu_fence_ref - take a ref on a fence
 754  *
 755  * @fence: amdgpu fence object
 756  *
 757  * Take a reference on a fence (all asics).
 758  * Returns the fence.
 759  */
 760 struct amdgpu_fence *amdgpu_fence_ref(struct amdgpu_fence *fence)
 761 {
 762         fence_get(&fence->base);
 763         return fence;
 764 }
 765
 766 /**
 767  * amdgpu_fence_unref - remove a ref on a fence
 768  *
 769  * @fence: amdgpu fence object
 770  *
 771  * Remove a reference on a fence (all asics).
 772  */
 773 void amdgpu_fence_unref(struct amdgpu_fence **fence)
 774 {
 775         struct amdgpu_fence *tmp = *fence;
 776
 777         *fence = NULL;
 778         if (tmp)
 779                 fence_put(&tmp->base);
 780 }
 781
 782 /**
 783  * amdgpu_fence_count_emitted - get the count of emitted fences
 784  *
 785  * @ring: ring the fence is associated with
 786  *
 787  * Get the number of fences emitted on the requested ring (all asics).
 788  * Returns the number of emitted fences on the ring.  Used by the
 789  * dynpm code to ring track activity.
 790  */
 791 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
 792 {
 793         uint64_t emitted;
 794
 795         /* We are not protected by ring lock when reading the last sequence
 796          * but it's ok to report slightly wrong fence count here.
 797          */
 798         amdgpu_fence_process(ring);
 799         emitted = ring->fence_drv.sync_seq[ring->idx]
 800                 - atomic64_read(&ring->fence_drv.last_seq);
 801         /* to avoid 32bits warp around */
 802         if (emitted > 0x10000000)
 803                 emitted = 0x10000000;
 804
 805         return (unsigned)emitted;
 806 }
 807
 808 /**
 809  * amdgpu_fence_need_sync - do we need a semaphore
 810  *
 811  * @fence: amdgpu fence object
 812  * @dst_ring: which ring to check against
 813  *
 814  * Check if the fence needs to be synced against another ring
 815  * (all asics).  If so, we need to emit a semaphore.
 816  * Returns true if we need to sync with another ring, false if
 817  * not.
 818  */
 819 bool amdgpu_fence_need_sync(struct amdgpu_fence *fence,
 820                             struct amdgpu_ring *dst_ring)
 821 {
 822         struct amdgpu_fence_driver *fdrv;
 823
 824         if (!fence)
 825                 return false;
 826
 827         if (fence->ring == dst_ring)
 828                 return false;
 829
 830         /* we are protected by the ring mutex */
 831         fdrv = &dst_ring->fence_drv;
 832         if (fence->seq <= fdrv->sync_seq[fence->ring->idx])
 833                 return false;
 834
 835         return true;
 836 }
 837
 838 /**
 839  * amdgpu_fence_note_sync - record the sync point
 840  *
 841  * @fence: amdgpu fence object
 842  * @dst_ring: which ring to check against
 843  *
 844  * Note the sequence number at which point the fence will
 845  * be synced with the requested ring (all asics).
 846  */
 847 void amdgpu_fence_note_sync(struct amdgpu_fence *fence,
 848                             struct amdgpu_ring *dst_ring)
 849 {
 850         struct amdgpu_fence_driver *dst, *src;
 851         unsigned i;
 852
 853         if (!fence)
 854                 return;
 855
 856         if (fence->ring == dst_ring)
 857                 return;
 858
 859         /* we are protected by the ring mutex */
 860         src = &fence->ring->fence_drv;
 861         dst = &dst_ring->fence_drv;
 862         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
 863                 if (i == dst_ring->idx)
 864                         continue;
 865
 866                 dst->sync_seq[i] = max(dst->sync_seq[i], src->sync_seq[i]);
 867         }
 868 }
 869
 870 /**
 871  * amdgpu_fence_driver_start_ring - make the fence driver
 872  * ready for use on the requested ring.
 873  *
 874  * @ring: ring to start the fence driver on
 875  * @irq_src: interrupt source to use for this ring
 876  * @irq_type: interrupt type to use for this ring
 877  *
 878  * Make the fence driver ready for processing (all asics).
 879  * Not all asics have all rings, so each asic will only
 880  * start the fence driver on the rings it has.
 881  * Returns 0 for success, errors for failure.
 882  */
 883 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
 884                                    struct amdgpu_irq_src *irq_src,
 885                                    unsigned irq_type)
 886 {
 887         struct amdgpu_device *adev = ring->adev;
 888         uint64_t index;
 889
 890         if (ring != &adev->uvd.ring) {
 891                 ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs];
 892                 ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4);
 893         } else {
 894                 /* put fence directly behind firmware */
 895                 index = ALIGN(adev->uvd.fw->size, 8);
 896                 ring->fence_drv.cpu_addr = adev->uvd.cpu_addr + index;
 897                 ring->fence_drv.gpu_addr = adev->uvd.gpu_addr + index;
 898         }
 899         amdgpu_fence_write(ring, atomic64_read(&ring->fence_drv.last_seq));
 900         ring->fence_drv.initialized = true;
 901         ring->fence_drv.irq_src = irq_src;
 902         ring->fence_drv.irq_type = irq_type;
 903         dev_info(adev->dev, "fence driver on ring %d use gpu addr 0x%016llx, "
 904                  "cpu addr 0x%p\n", ring->idx,
 905                  ring->fence_drv.gpu_addr, ring->fence_drv.cpu_addr);
 906         return 0;
 907 }
 908
 909 /**
 910  * amdgpu_fence_driver_init_ring - init the fence driver
 911  * for the requested ring.
 912  *
 913  * @ring: ring to init the fence driver on
 914  *
 915  * Init the fence driver for the requested ring (all asics).
 916  * Helper function for amdgpu_fence_driver_init().
 917  */
 918 void amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring)
 919 {
 920         int i;
 921
 922         ring->fence_drv.cpu_addr = NULL;
 923         ring->fence_drv.gpu_addr = 0;
 924         for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
 925                 ring->fence_drv.sync_seq[i] = 0;
 926
 927         atomic64_set(&ring->fence_drv.last_seq, 0);
 928         ring->fence_drv.initialized = false;
 929
 930         INIT_DELAYED_WORK(&ring->fence_drv.lockup_work,
 931                         amdgpu_fence_check_lockup);
 932         ring->fence_drv.ring = ring;
 933 }
 934
 935 /**
 936  * amdgpu_fence_driver_init - init the fence driver
 937  * for all possible rings.
 938  *
 939  * @adev: amdgpu device pointer
 940  *
 941  * Init the fence driver for all possible rings (all asics).
 942  * Not all asics have all rings, so each asic will only
 943  * start the fence driver on the rings it has using
 944  * amdgpu_fence_driver_start_ring().
 945  * Returns 0 for success.
 946  */
 947 int amdgpu_fence_driver_init(struct amdgpu_device *adev)
 948 {
 949         init_waitqueue_head(&adev->fence_queue);
 950         if (amdgpu_debugfs_fence_init(adev))
 951                 dev_err(adev->dev, "fence debugfs file creation failed\n");
 952
 953         return 0;
 954 }
 955
 956 /**
 957  * amdgpu_fence_driver_fini - tear down the fence driver
 958  * for all possible rings.
 959  *
 960  * @adev: amdgpu device pointer
 961  *
 962  * Tear down the fence driver for all possible rings (all asics).
 963  */
 964 void amdgpu_fence_driver_fini(struct amdgpu_device *adev)
 965 {
 966         int i, r;
 967
 968         mutex_lock(&adev->ring_lock);
 969         for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
 970                 struct amdgpu_ring *ring = adev->rings[i];
 971                 if (!ring || !ring->fence_drv.initialized)
 972                         continue;
 973                 r = amdgpu_fence_wait_empty(ring);
 974                 if (r) {
 975                         /* no need to trigger GPU reset as we are unloading */
 976                         amdgpu_fence_driver_force_completion(adev);
 977                 }
 978                 wake_up_all(&adev->fence_queue);
 979                 ring->fence_drv.initialized = false;
 980         }
 981         mutex_unlock(&adev->ring_lock);
 982 }
 983
 984 /**
 985  * amdgpu_fence_driver_force_completion - force all fence waiter to complete
 986  *
 987  * @adev: amdgpu device pointer
 988  *
 989  * In case of GPU reset failure make sure no process keep waiting on fence
 990  * that will never complete.
 991  */
 992 void amdgpu_fence_driver_force_completion(struct amdgpu_device *adev)
 993 {
 994         int i;
 995
 996         for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
 997                 struct amdgpu_ring *ring = adev->rings[i];
 998                 if (!ring || !ring->fence_drv.initialized)
 999                         continue;
1000
1001                 amdgpu_fence_write(ring, ring->fence_drv.sync_seq[i]);
1002         }
1003 }
1004
1005
1006 /*
1007  * Fence debugfs
1008  */
1009 #if defined(CONFIG_DEBUG_FS)
1010 static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data)
1011 {
1012         struct drm_info_node *node = (struct drm_info_node *)m->private;
1013         struct drm_device *dev = node->minor->dev;
1014         struct amdgpu_device *adev = dev->dev_private;
1015         int i, j;
1016
1017         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
1018                 struct amdgpu_ring *ring = adev->rings[i];
1019                 if (!ring || !ring->fence_drv.initialized)
1020                         continue;
1021
1022                 amdgpu_fence_process(ring);
1023
1024                 seq_printf(m, "--- ring %d ---\n", i);
1025                 seq_printf(m, "Last signaled fence 0x%016llx\n",
1026                            (unsigned long long)atomic64_read(&ring->fence_drv.last_seq));
1027                 seq_printf(m, "Last emitted        0x%016llx\n",
1028                            ring->fence_drv.sync_seq[i]);
1029
1030                 for (j = 0; j < AMDGPU_MAX_RINGS; ++j) {
1031                         struct amdgpu_ring *other = adev->rings[j];
1032                         if (i != j && other && other->fence_drv.initialized)
1033                                 seq_printf(m, "Last sync to ring %d 0x%016llx\n",
1034                                            j, ring->fence_drv.sync_seq[j]);
1035                 }
1036         }
1037         return 0;
1038 }
1039
1040 static struct drm_info_list amdgpu_debugfs_fence_list[] = {
1041         {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
1042 };
1043 #endif
1044
1045 int amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
1046 {
1047 #if defined(CONFIG_DEBUG_FS)
1048         return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 1);
1049 #else
1050         return 0;
1051 #endif
1052 }
1053
1054 static const char *amdgpu_fence_get_driver_name(struct fence *fence)
1055 {
1056         return "amdgpu";
1057 }
1058
1059 static const char *amdgpu_fence_get_timeline_name(struct fence *f)
1060 {
1061         struct amdgpu_fence *fence = to_amdgpu_fence(f);
1062         return (const char *)fence->ring->name;
1063 }
1064
1065 static inline bool amdgpu_test_signaled(struct amdgpu_fence *fence)
1066 {
1067         return test_bit(FENCE_FLAG_SIGNALED_BIT, &fence->base.flags);
1068 }
1069
1070 struct amdgpu_wait_cb {
1071         struct fence_cb base;
1072         struct task_struct *task;
1073 };
1074
1075 static void amdgpu_fence_wait_cb(struct fence *fence, struct fence_cb *cb)
1076 {
1077         struct amdgpu_wait_cb *wait =
1078                 container_of(cb, struct amdgpu_wait_cb, base);
1079         wake_up_process(wait->task);
1080 }
1081
1082 static signed long amdgpu_fence_default_wait(struct fence *f, bool intr,
1083                                              signed long t)
1084 {
1085         struct amdgpu_fence *fence = to_amdgpu_fence(f);
1086         struct amdgpu_device *adev = fence->ring->adev;
1087         struct amdgpu_wait_cb cb;
1088
1089         cb.task = current;
1090
1091         if (fence_add_callback(f, &cb.base, amdgpu_fence_wait_cb))
1092                 return t;
1093
1094         while (t > 0) {
1095                 if (intr)
1096                         set_current_state(TASK_INTERRUPTIBLE);
1097                 else
1098                         set_current_state(TASK_UNINTERRUPTIBLE);
1099
1100                 /*
1101                  * amdgpu_test_signaled must be called after
1102                  * set_current_state to prevent a race with wake_up_process
1103                  */
1104                 if (amdgpu_test_signaled(fence))
1105                         break;
1106
1107                 if (adev->needs_reset) {
1108                         t = -EDEADLK;
1109                         break;
1110                 }
1111
1112                 t = schedule_timeout(t);
1113
1114                 if (t > 0 && intr && signal_pending(current))
1115                         t = -ERESTARTSYS;
1116         }
1117
1118         __set_current_state(TASK_RUNNING);
1119         fence_remove_callback(f, &cb.base);
1120
1121         return t;
1122 }
1123
1124 const struct fence_ops amdgpu_fence_ops = {
1125         .get_driver_name = amdgpu_fence_get_driver_name,
1126         .get_timeline_name = amdgpu_fence_get_timeline_name,
1127         .enable_signaling = amdgpu_fence_enable_signaling,
1128         .signaled = amdgpu_fence_is_signaled,
1129         .wait = amdgpu_fence_default_wait,
1130         .release = NULL,
1131 };