drivers/gpu/drm/radeon/radeon_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 <glisse@freedesktop.org>
  29  *    Dave Airlie
  30  */
  31
  32 #include <linux/atomic.h>
  33 #include <linux/firmware.h>
  34 #include <linux/kref.h>
  35 #include <linux/sched/signal.h>
  36 #include <linux/seq_file.h>
  37 #include <linux/slab.h>
  38 #include <linux/wait.h>
  39
  40 #include <drm/drm_device.h>
  41 #include <drm/drm_file.h>
  42
  43 #include "radeon.h"
  44 #include "radeon_reg.h"
  45 #include "radeon_trace.h"
  46
  47 /*
  48  * Fences
  49  * Fences mark an event in the GPUs pipeline and are used
  50  * for GPU/CPU synchronization.  When the fence is written,
  51  * it is expected that all buffers associated with that fence
  52  * are no longer in use by the associated ring on the GPU and
  53  * that the the relevant GPU caches have been flushed.  Whether
  54  * we use a scratch register or memory location depends on the asic
  55  * and whether writeback is enabled.
  56  */
  57
  58 /**
  59  * radeon_fence_write - write a fence value
  60  *
  61  * @rdev: radeon_device pointer
  62  * @seq: sequence number to write
  63  * @ring: ring index the fence is associated with
  64  *
  65  * Writes a fence value to memory or a scratch register (all asics).
  66  */
  67 static void radeon_fence_write(struct radeon_device *rdev, u32 seq, int ring)
  68 {
  69         struct radeon_fence_driver *drv = &rdev->fence_drv[ring];
  70         if (likely(rdev->wb.enabled || !drv->scratch_reg)) {
  71                 if (drv->cpu_addr) {
  72                         *drv->cpu_addr = cpu_to_le32(seq);
  73                 }
  74         } else {
  75                 WREG32(drv->scratch_reg, seq);
  76         }
  77 }
  78
  79 /**
  80  * radeon_fence_read - read a fence value
  81  *
  82  * @rdev: radeon_device pointer
  83  * @ring: ring index the fence is associated with
  84  *
  85  * Reads a fence value from memory or a scratch register (all asics).
  86  * Returns the value of the fence read from memory or register.
  87  */
  88 static u32 radeon_fence_read(struct radeon_device *rdev, int ring)
  89 {
  90         struct radeon_fence_driver *drv = &rdev->fence_drv[ring];
  91         u32 seq = 0;
  92
  93         if (likely(rdev->wb.enabled || !drv->scratch_reg)) {
  94                 if (drv->cpu_addr) {
  95                         seq = le32_to_cpu(*drv->cpu_addr);
  96                 } else {
  97                         seq = lower_32_bits(atomic64_read(&drv->last_seq));
  98                 }
  99         } else {
 100                 seq = RREG32(drv->scratch_reg);
 101         }
 102         return seq;
 103 }
 104
 105 /**
 106  * radeon_fence_schedule_check - schedule lockup check
 107  *
 108  * @rdev: radeon_device pointer
 109  * @ring: ring index we should work with
 110  *
 111  * Queues a delayed work item to check for lockups.
 112  */
 113 static void radeon_fence_schedule_check(struct radeon_device *rdev, int ring)
 114 {
 115         /*
 116          * Do not reset the timer here with mod_delayed_work,
 117          * this can livelock in an interaction with TTM delayed destroy.
 118          */
 119         queue_delayed_work(system_power_efficient_wq,
 120                            &rdev->fence_drv[ring].lockup_work,
 121                            RADEON_FENCE_JIFFIES_TIMEOUT);
 122 }
 123
 124 /**
 125  * radeon_fence_emit - emit a fence on the requested ring
 126  *
 127  * @rdev: radeon_device pointer
 128  * @fence: radeon fence object
 129  * @ring: ring index the fence is associated with
 130  *
 131  * Emits a fence command on the requested ring (all asics).
 132  * Returns 0 on success, -ENOMEM on failure.
 133  */
 134 int radeon_fence_emit(struct radeon_device *rdev,
 135                       struct radeon_fence **fence,
 136                       int ring)
 137 {
 138         u64 seq;
 139
 140         /* we are protected by the ring emission mutex */
 141         *fence = kmalloc(sizeof(struct radeon_fence), GFP_KERNEL);
 142         if ((*fence) == NULL) {
 143                 return -ENOMEM;
 144         }
 145         (*fence)->rdev = rdev;
 146         (*fence)->seq = seq = ++rdev->fence_drv[ring].sync_seq[ring];
 147         (*fence)->ring = ring;
 148         (*fence)->is_vm_update = false;
 149         dma_fence_init(&(*fence)->base, &radeon_fence_ops,
 150                        &rdev->fence_queue.lock,
 151                        rdev->fence_context + ring,
 152                        seq);
 153         radeon_fence_ring_emit(rdev, ring, *fence);
 154         trace_radeon_fence_emit(rdev->ddev, ring, (*fence)->seq);
 155         radeon_fence_schedule_check(rdev, ring);
 156         return 0;
 157 }
 158
 159 /*
 160  * radeon_fence_check_signaled - callback from fence_queue
 161  *
 162  * this function is called with fence_queue lock held, which is also used
 163  * for the fence locking itself, so unlocked variants are used for
 164  * fence_signal, and remove_wait_queue.
 165  */
 166 static int radeon_fence_check_signaled(wait_queue_entry_t *wait, unsigned mode, int flags, void *key)
 167 {
 168         struct radeon_fence *fence;
 169         u64 seq;
 170
 171         fence = container_of(wait, struct radeon_fence, fence_wake);
 172
 173         /*
 174          * We cannot use radeon_fence_process here because we're already
 175          * in the waitqueue, in a call from wake_up_all.
 176          */
 177         seq = atomic64_read(&fence->rdev->fence_drv[fence->ring].last_seq);
 178         if (seq >= fence->seq) {
 179                 int ret = dma_fence_signal_locked(&fence->base);
 180
 181                 if (!ret)
 182                         DMA_FENCE_TRACE(&fence->base, "signaled from irq context\n");
 183                 else
 184                         DMA_FENCE_TRACE(&fence->base, "was already signaled\n");
 185
 186                 radeon_irq_kms_sw_irq_put(fence->rdev, fence->ring);
 187                 __remove_wait_queue(&fence->rdev->fence_queue, &fence->fence_wake);
 188                 dma_fence_put(&fence->base);
 189         } else
 190                 DMA_FENCE_TRACE(&fence->base, "pending\n");
 191         return 0;
 192 }
 193
 194 /**
 195  * radeon_fence_activity - check for fence activity
 196  *
 197  * @rdev: radeon_device pointer
 198  * @ring: ring index the fence is associated with
 199  *
 200  * Checks the current fence value and calculates the last
 201  * signalled fence value. Returns true if activity occured
 202  * on the ring, and the fence_queue should be waken up.
 203  */
 204 static bool radeon_fence_activity(struct radeon_device *rdev, int ring)
 205 {
 206         uint64_t seq, last_seq, last_emitted;
 207         unsigned count_loop = 0;
 208         bool wake = false;
 209
 210         /* Note there is a scenario here for an infinite loop but it's
 211          * very unlikely to happen. For it to happen, the current polling
 212          * process need to be interrupted by another process and another
 213          * process needs to update the last_seq btw the atomic read and
 214          * xchg of the current process.
 215          *
 216          * More over for this to go in infinite loop there need to be
 217          * continuously new fence signaled ie radeon_fence_read needs
 218          * to return a different value each time for both the currently
 219          * polling process and the other process that xchg the last_seq
 220          * btw atomic read and xchg of the current process. And the
 221          * value the other process set as last seq must be higher than
 222          * the seq value we just read. Which means that current process
 223          * need to be interrupted after radeon_fence_read and before
 224          * atomic xchg.
 225          *
 226          * To be even more safe we count the number of time we loop and
 227          * we bail after 10 loop just accepting the fact that we might
 228          * have temporarly set the last_seq not to the true real last
 229          * seq but to an older one.
 230          */
 231         last_seq = atomic64_read(&rdev->fence_drv[ring].last_seq);
 232         do {
 233                 last_emitted = rdev->fence_drv[ring].sync_seq[ring];
 234                 seq = radeon_fence_read(rdev, ring);
 235                 seq |= last_seq & 0xffffffff00000000LL;
 236                 if (seq < last_seq) {
 237                         seq &= 0xffffffff;
 238                         seq |= last_emitted & 0xffffffff00000000LL;
 239                 }
 240
 241                 if (seq <= last_seq || seq > last_emitted) {
 242                         break;
 243                 }
 244                 /* If we loop over we don't want to return without
 245                  * checking if a fence is signaled as it means that the
 246                  * seq we just read is different from the previous on.
 247                  */
 248                 wake = true;
 249                 last_seq = seq;
 250                 if ((count_loop++) > 10) {
 251                         /* We looped over too many time leave with the
 252                          * fact that we might have set an older fence
 253                          * seq then the current real last seq as signaled
 254                          * by the hw.
 255                          */
 256                         break;
 257                 }
 258         } while (atomic64_xchg(&rdev->fence_drv[ring].last_seq, seq) > seq);
 259
 260         if (seq < last_emitted)
 261                 radeon_fence_schedule_check(rdev, ring);
 262
 263         return wake;
 264 }
 265
 266 /**
 267  * radeon_fence_check_lockup - check for hardware lockup
 268  *
 269  * @work: delayed work item
 270  *
 271  * Checks for fence activity and if there is none probe
 272  * the hardware if a lockup occured.
 273  */
 274 static void radeon_fence_check_lockup(struct work_struct *work)
 275 {
 276         struct radeon_fence_driver *fence_drv;
 277         struct radeon_device *rdev;
 278         int ring;
 279
 280         fence_drv = container_of(work, struct radeon_fence_driver,
 281                                  lockup_work.work);
 282         rdev = fence_drv->rdev;
 283         ring = fence_drv - &rdev->fence_drv[0];
 284
 285         if (!down_read_trylock(&rdev->exclusive_lock)) {
 286                 /* just reschedule the check if a reset is going on */
 287                 radeon_fence_schedule_check(rdev, ring);
 288                 return;
 289         }
 290
 291         if (fence_drv->delayed_irq && rdev->irq.installed) {
 292                 unsigned long irqflags;
 293
 294                 fence_drv->delayed_irq = false;
 295                 spin_lock_irqsave(&rdev->irq.lock, irqflags);
 296                 radeon_irq_set(rdev);
 297                 spin_unlock_irqrestore(&rdev->irq.lock, irqflags);
 298         }
 299
 300         if (radeon_fence_activity(rdev, ring))
 301                 wake_up_all(&rdev->fence_queue);
 302
 303         else if (radeon_ring_is_lockup(rdev, ring, &rdev->ring[ring])) {
 304
 305                 /* good news we believe it's a lockup */
 306                 dev_warn(rdev->dev, "GPU lockup (current fence id "
 307                          "0x%016llx last fence id 0x%016llx on ring %d)\n",
 308                          (uint64_t)atomic64_read(&fence_drv->last_seq),
 309                          fence_drv->sync_seq[ring], ring);
 310
 311                 /* remember that we need an reset */
 312                 rdev->needs_reset = true;
 313                 wake_up_all(&rdev->fence_queue);
 314         }
 315         up_read(&rdev->exclusive_lock);
 316 }
 317
 318 /**
 319  * radeon_fence_process - process a fence
 320  *
 321  * @rdev: radeon_device pointer
 322  * @ring: ring index the fence is associated with
 323  *
 324  * Checks the current fence value and wakes the fence queue
 325  * if the sequence number has increased (all asics).
 326  */
 327 void radeon_fence_process(struct radeon_device *rdev, int ring)
 328 {
 329         if (radeon_fence_activity(rdev, ring))
 330                 wake_up_all(&rdev->fence_queue);
 331 }
 332
 333 /**
 334  * radeon_fence_seq_signaled - check if a fence sequence number has signaled
 335  *
 336  * @rdev: radeon device pointer
 337  * @seq: sequence number
 338  * @ring: ring index the fence is associated with
 339  *
 340  * Check if the last signaled fence sequnce number is >= the requested
 341  * sequence number (all asics).
 342  * Returns true if the fence has signaled (current fence value
 343  * is >= requested value) or false if it has not (current fence
 344  * value is < the requested value.  Helper function for
 345  * radeon_fence_signaled().
 346  */
 347 static bool radeon_fence_seq_signaled(struct radeon_device *rdev,
 348                                       u64 seq, unsigned ring)
 349 {
 350         if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) {
 351                 return true;
 352         }
 353         /* poll new last sequence at least once */
 354         radeon_fence_process(rdev, ring);
 355         if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) {
 356                 return true;
 357         }
 358         return false;
 359 }
 360
 361 static bool radeon_fence_is_signaled(struct dma_fence *f)
 362 {
 363         struct radeon_fence *fence = to_radeon_fence(f);
 364         struct radeon_device *rdev = fence->rdev;
 365         unsigned ring = fence->ring;
 366         u64 seq = fence->seq;
 367
 368         if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) {
 369                 return true;
 370         }
 371
 372         if (down_read_trylock(&rdev->exclusive_lock)) {
 373                 radeon_fence_process(rdev, ring);
 374                 up_read(&rdev->exclusive_lock);
 375
 376                 if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) {
 377                         return true;
 378                 }
 379         }
 380         return false;
 381 }
 382
 383 /**
 384  * radeon_fence_enable_signaling - enable signalling on fence
 385  * @f: fence
 386  *
 387  * This function is called with fence_queue lock held, and adds a callback
 388  * to fence_queue that checks if this fence is signaled, and if so it
 389  * signals the fence and removes itself.
 390  */
 391 static bool radeon_fence_enable_signaling(struct dma_fence *f)
 392 {
 393         struct radeon_fence *fence = to_radeon_fence(f);
 394         struct radeon_device *rdev = fence->rdev;
 395
 396         if (atomic64_read(&rdev->fence_drv[fence->ring].last_seq) >= fence->seq)
 397                 return false;
 398
 399         if (down_read_trylock(&rdev->exclusive_lock)) {
 400                 radeon_irq_kms_sw_irq_get(rdev, fence->ring);
 401
 402                 if (radeon_fence_activity(rdev, fence->ring))
 403                         wake_up_all_locked(&rdev->fence_queue);
 404
 405                 /* did fence get signaled after we enabled the sw irq? */
 406                 if (atomic64_read(&rdev->fence_drv[fence->ring].last_seq) >= fence->seq) {
 407                         radeon_irq_kms_sw_irq_put(rdev, fence->ring);
 408                         up_read(&rdev->exclusive_lock);
 409                         return false;
 410                 }
 411
 412                 up_read(&rdev->exclusive_lock);
 413         } else {
 414                 /* we're probably in a lockup, lets not fiddle too much */
 415                 if (radeon_irq_kms_sw_irq_get_delayed(rdev, fence->ring))
 416                         rdev->fence_drv[fence->ring].delayed_irq = true;
 417                 radeon_fence_schedule_check(rdev, fence->ring);
 418         }
 419
 420         fence->fence_wake.flags = 0;
 421         fence->fence_wake.private = NULL;
 422         fence->fence_wake.func = radeon_fence_check_signaled;
 423         __add_wait_queue(&rdev->fence_queue, &fence->fence_wake);
 424         dma_fence_get(f);
 425
 426         DMA_FENCE_TRACE(&fence->base, "armed on ring %i!\n", fence->ring);
 427         return true;
 428 }
 429
 430 /**
 431  * radeon_fence_signaled - check if a fence has signaled
 432  *
 433  * @fence: radeon fence object
 434  *
 435  * Check if the requested fence has signaled (all asics).
 436  * Returns true if the fence has signaled or false if it has not.
 437  */
 438 bool radeon_fence_signaled(struct radeon_fence *fence)
 439 {
 440         if (!fence)
 441                 return true;
 442
 443         if (radeon_fence_seq_signaled(fence->rdev, fence->seq, fence->ring)) {
 444                 int ret;
 445
 446                 ret = dma_fence_signal(&fence->base);
 447                 if (!ret)
 448                         DMA_FENCE_TRACE(&fence->base, "signaled from radeon_fence_signaled\n");
 449                 return true;
 450         }
 451         return false;
 452 }
 453
 454 /**
 455  * radeon_fence_any_seq_signaled - check if any sequence number is signaled
 456  *
 457  * @rdev: radeon device pointer
 458  * @seq: sequence numbers
 459  *
 460  * Check if the last signaled fence sequnce number is >= the requested
 461  * sequence number (all asics).
 462  * Returns true if any has signaled (current value is >= requested value)
 463  * or false if it has not. Helper function for radeon_fence_wait_seq.
 464  */
 465 static bool radeon_fence_any_seq_signaled(struct radeon_device *rdev, u64 *seq)
 466 {
 467         unsigned i;
 468
 469         for (i = 0; i < RADEON_NUM_RINGS; ++i) {
 470                 if (seq[i] && radeon_fence_seq_signaled(rdev, seq[i], i))
 471                         return true;
 472         }
 473         return false;
 474 }
 475
 476 /**
 477  * radeon_fence_wait_seq_timeout - wait for a specific sequence numbers
 478  *
 479  * @rdev: radeon device pointer
 480  * @target_seq: sequence number(s) we want to wait for
 481  * @intr: use interruptable sleep
 482  * @timeout: maximum time to wait, or MAX_SCHEDULE_TIMEOUT for infinite wait
 483  *
 484  * Wait for the requested sequence number(s) to be written by any ring
 485  * (all asics).  Sequnce number array is indexed by ring id.
 486  * @intr selects whether to use interruptable (true) or non-interruptable
 487  * (false) sleep when waiting for the sequence number.  Helper function
 488  * for radeon_fence_wait_*().
 489  * Returns remaining time if the sequence number has passed, 0 when
 490  * the wait timeout, or an error for all other cases.
 491  * -EDEADLK is returned when a GPU lockup has been detected.
 492  */
 493 static long radeon_fence_wait_seq_timeout(struct radeon_device *rdev,
 494                                           u64 *target_seq, bool intr,
 495                                           long timeout)
 496 {
 497         long r;
 498         int i;
 499
 500         if (radeon_fence_any_seq_signaled(rdev, target_seq))
 501                 return timeout;
 502
 503         /* enable IRQs and tracing */
 504         for (i = 0; i < RADEON_NUM_RINGS; ++i) {
 505                 if (!target_seq[i])
 506                         continue;
 507
 508                 trace_radeon_fence_wait_begin(rdev->ddev, i, target_seq[i]);
 509                 radeon_irq_kms_sw_irq_get(rdev, i);
 510         }
 511
 512         if (intr) {
 513                 r = wait_event_interruptible_timeout(rdev->fence_queue, (
 514                         radeon_fence_any_seq_signaled(rdev, target_seq)
 515                          || rdev->needs_reset), timeout);
 516         } else {
 517                 r = wait_event_timeout(rdev->fence_queue, (
 518                         radeon_fence_any_seq_signaled(rdev, target_seq)
 519                          || rdev->needs_reset), timeout);
 520         }
 521
 522         if (rdev->needs_reset)
 523                 r = -EDEADLK;
 524
 525         for (i = 0; i < RADEON_NUM_RINGS; ++i) {
 526                 if (!target_seq[i])
 527                         continue;
 528
 529                 radeon_irq_kms_sw_irq_put(rdev, i);
 530                 trace_radeon_fence_wait_end(rdev->ddev, i, target_seq[i]);
 531         }
 532
 533         return r;
 534 }
 535
 536 /**
 537  * radeon_fence_wait_timeout - wait for a fence to signal with timeout
 538  *
 539  * @fence: radeon fence object
 540  * @intr: use interruptible sleep
 541  *
 542  * Wait for the requested fence to signal (all asics).
 543  * @intr selects whether to use interruptable (true) or non-interruptable
 544  * (false) sleep when waiting for the fence.
 545  * @timeout: maximum time to wait, or MAX_SCHEDULE_TIMEOUT for infinite wait
 546  * Returns remaining time if the sequence number has passed, 0 when
 547  * the wait timeout, or an error for all other cases.
 548  */
 549 long radeon_fence_wait_timeout(struct radeon_fence *fence, bool intr, long timeout)
 550 {
 551         uint64_t seq[RADEON_NUM_RINGS] = {};
 552         long r;
 553         int r_sig;
 554
 555         /*
 556          * This function should not be called on !radeon fences.
 557          * If this is the case, it would mean this function can
 558          * also be called on radeon fences belonging to another card.
 559          * exclusive_lock is not held in that case.
 560          */
 561         if (WARN_ON_ONCE(!to_radeon_fence(&fence->base)))
 562                 return dma_fence_wait(&fence->base, intr);
 563
 564         seq[fence->ring] = fence->seq;
 565         r = radeon_fence_wait_seq_timeout(fence->rdev, seq, intr, timeout);
 566         if (r <= 0) {
 567                 return r;
 568         }
 569
 570         r_sig = dma_fence_signal(&fence->base);
 571         if (!r_sig)
 572                 DMA_FENCE_TRACE(&fence->base, "signaled from fence_wait\n");
 573         return r;
 574 }
 575
 576 /**
 577  * radeon_fence_wait - wait for a fence to signal
 578  *
 579  * @fence: radeon fence object
 580  * @intr: use interruptible sleep
 581  *
 582  * Wait for the requested fence to signal (all asics).
 583  * @intr selects whether to use interruptable (true) or non-interruptable
 584  * (false) sleep when waiting for the fence.
 585  * Returns 0 if the fence has passed, error for all other cases.
 586  */
 587 int radeon_fence_wait(struct radeon_fence *fence, bool intr)
 588 {
 589         long r = radeon_fence_wait_timeout(fence, intr, MAX_SCHEDULE_TIMEOUT);
 590         if (r > 0) {
 591                 return 0;
 592         } else {
 593                 return r;
 594         }
 595 }
 596
 597 /**
 598  * radeon_fence_wait_any - wait for a fence to signal on any ring
 599  *
 600  * @rdev: radeon device pointer
 601  * @fences: radeon fence object(s)
 602  * @intr: use interruptable sleep
 603  *
 604  * Wait for any requested fence to signal (all asics).  Fence
 605  * array is indexed by ring id.  @intr selects whether to use
 606  * interruptable (true) or non-interruptable (false) sleep when
 607  * waiting for the fences. Used by the suballocator.
 608  * Returns 0 if any fence has passed, error for all other cases.
 609  */
 610 int radeon_fence_wait_any(struct radeon_device *rdev,
 611                           struct radeon_fence **fences,
 612                           bool intr)
 613 {
 614         uint64_t seq[RADEON_NUM_RINGS];
 615         unsigned i, num_rings = 0;
 616         long r;
 617
 618         for (i = 0; i < RADEON_NUM_RINGS; ++i) {
 619                 seq[i] = 0;
 620
 621                 if (!fences[i]) {
 622                         continue;
 623                 }
 624
 625                 seq[i] = fences[i]->seq;
 626                 ++num_rings;
 627         }
 628
 629         /* nothing to wait for ? */
 630         if (num_rings == 0)
 631                 return -ENOENT;
 632
 633         r = radeon_fence_wait_seq_timeout(rdev, seq, intr, MAX_SCHEDULE_TIMEOUT);
 634         if (r < 0) {
 635                 return r;
 636         }
 637         return 0;
 638 }
 639
 640 /**
 641  * radeon_fence_wait_next - wait for the next fence to signal
 642  *
 643  * @rdev: radeon device pointer
 644  * @ring: ring index the fence is associated with
 645  *
 646  * Wait for the next fence on the requested ring to signal (all asics).
 647  * Returns 0 if the next fence has passed, error for all other cases.
 648  * Caller must hold ring lock.
 649  */
 650 int radeon_fence_wait_next(struct radeon_device *rdev, int ring)
 651 {
 652         uint64_t seq[RADEON_NUM_RINGS] = {};
 653         long r;
 654
 655         seq[ring] = atomic64_read(&rdev->fence_drv[ring].last_seq) + 1ULL;
 656         if (seq[ring] >= rdev->fence_drv[ring].sync_seq[ring]) {
 657                 /* nothing to wait for, last_seq is
 658                    already the last emited fence */
 659                 return -ENOENT;
 660         }
 661         r = radeon_fence_wait_seq_timeout(rdev, seq, false, MAX_SCHEDULE_TIMEOUT);
 662         if (r < 0)
 663                 return r;
 664         return 0;
 665 }
 666
 667 /**
 668  * radeon_fence_wait_empty - wait for all fences to signal
 669  *
 670  * @rdev: radeon device pointer
 671  * @ring: ring index the fence is associated with
 672  *
 673  * Wait for all fences on the requested ring to signal (all asics).
 674  * Returns 0 if the fences have passed, error for all other cases.
 675  * Caller must hold ring lock.
 676  */
 677 int radeon_fence_wait_empty(struct radeon_device *rdev, int ring)
 678 {
 679         uint64_t seq[RADEON_NUM_RINGS] = {};
 680         long r;
 681
 682         seq[ring] = rdev->fence_drv[ring].sync_seq[ring];
 683         if (!seq[ring])
 684                 return 0;
 685
 686         r = radeon_fence_wait_seq_timeout(rdev, seq, false, MAX_SCHEDULE_TIMEOUT);
 687         if (r < 0) {
 688                 if (r == -EDEADLK)
 689                         return -EDEADLK;
 690
 691                 dev_err(rdev->dev, "error waiting for ring[%d] to become idle (%ld)\n",
 692                         ring, r);
 693         }
 694         return 0;
 695 }
 696
 697 /**
 698  * radeon_fence_ref - take a ref on a fence
 699  *
 700  * @fence: radeon fence object
 701  *
 702  * Take a reference on a fence (all asics).
 703  * Returns the fence.
 704  */
 705 struct radeon_fence *radeon_fence_ref(struct radeon_fence *fence)
 706 {
 707         dma_fence_get(&fence->base);
 708         return fence;
 709 }
 710
 711 /**
 712  * radeon_fence_unref - remove a ref on a fence
 713  *
 714  * @fence: radeon fence object
 715  *
 716  * Remove a reference on a fence (all asics).
 717  */
 718 void radeon_fence_unref(struct radeon_fence **fence)
 719 {
 720         struct radeon_fence *tmp = *fence;
 721
 722         *fence = NULL;
 723         if (tmp) {
 724                 dma_fence_put(&tmp->base);
 725         }
 726 }
 727
 728 /**
 729  * radeon_fence_count_emitted - get the count of emitted fences
 730  *
 731  * @rdev: radeon device pointer
 732  * @ring: ring index the fence is associated with
 733  *
 734  * Get the number of fences emitted on the requested ring (all asics).
 735  * Returns the number of emitted fences on the ring.  Used by the
 736  * dynpm code to ring track activity.
 737  */
 738 unsigned radeon_fence_count_emitted(struct radeon_device *rdev, int ring)
 739 {
 740         uint64_t emitted;
 741
 742         /* We are not protected by ring lock when reading the last sequence
 743          * but it's ok to report slightly wrong fence count here.
 744          */
 745         radeon_fence_process(rdev, ring);
 746         emitted = rdev->fence_drv[ring].sync_seq[ring]
 747                 - atomic64_read(&rdev->fence_drv[ring].last_seq);
 748         /* to avoid 32bits warp around */
 749         if (emitted > 0x10000000) {
 750                 emitted = 0x10000000;
 751         }
 752         return (unsigned)emitted;
 753 }
 754
 755 /**
 756  * radeon_fence_need_sync - do we need a semaphore
 757  *
 758  * @fence: radeon fence object
 759  * @dst_ring: which ring to check against
 760  *
 761  * Check if the fence needs to be synced against another ring
 762  * (all asics).  If so, we need to emit a semaphore.
 763  * Returns true if we need to sync with another ring, false if
 764  * not.
 765  */
 766 bool radeon_fence_need_sync(struct radeon_fence *fence, int dst_ring)
 767 {
 768         struct radeon_fence_driver *fdrv;
 769
 770         if (!fence) {
 771                 return false;
 772         }
 773
 774         if (fence->ring == dst_ring) {
 775                 return false;
 776         }
 777
 778         /* we are protected by the ring mutex */
 779         fdrv = &fence->rdev->fence_drv[dst_ring];
 780         if (fence->seq <= fdrv->sync_seq[fence->ring]) {
 781                 return false;
 782         }
 783
 784         return true;
 785 }
 786
 787 /**
 788  * radeon_fence_note_sync - record the sync point
 789  *
 790  * @fence: radeon fence object
 791  * @dst_ring: which ring to check against
 792  *
 793  * Note the sequence number at which point the fence will
 794  * be synced with the requested ring (all asics).
 795  */
 796 void radeon_fence_note_sync(struct radeon_fence *fence, int dst_ring)
 797 {
 798         struct radeon_fence_driver *dst, *src;
 799         unsigned i;
 800
 801         if (!fence) {
 802                 return;
 803         }
 804
 805         if (fence->ring == dst_ring) {
 806                 return;
 807         }
 808
 809         /* we are protected by the ring mutex */
 810         src = &fence->rdev->fence_drv[fence->ring];
 811         dst = &fence->rdev->fence_drv[dst_ring];
 812         for (i = 0; i < RADEON_NUM_RINGS; ++i) {
 813                 if (i == dst_ring) {
 814                         continue;
 815                 }
 816                 dst->sync_seq[i] = max(dst->sync_seq[i], src->sync_seq[i]);
 817         }
 818 }
 819
 820 /**
 821  * radeon_fence_driver_start_ring - make the fence driver
 822  * ready for use on the requested ring.
 823  *
 824  * @rdev: radeon device pointer
 825  * @ring: ring index to start the fence driver on
 826  *
 827  * Make the fence driver ready for processing (all asics).
 828  * Not all asics have all rings, so each asic will only
 829  * start the fence driver on the rings it has.
 830  * Returns 0 for success, errors for failure.
 831  */
 832 int radeon_fence_driver_start_ring(struct radeon_device *rdev, int ring)
 833 {
 834         uint64_t index;
 835         int r;
 836
 837         radeon_scratch_free(rdev, rdev->fence_drv[ring].scratch_reg);
 838         if (rdev->wb.use_event || !radeon_ring_supports_scratch_reg(rdev, &rdev->ring[ring])) {
 839                 rdev->fence_drv[ring].scratch_reg = 0;
 840                 if (ring != R600_RING_TYPE_UVD_INDEX) {
 841                         index = R600_WB_EVENT_OFFSET + ring * 4;
 842                         rdev->fence_drv[ring].cpu_addr = &rdev->wb.wb[index/4];
 843                         rdev->fence_drv[ring].gpu_addr = rdev->wb.gpu_addr +
 844                                                          index;
 845
 846                 } else {
 847                         /* put fence directly behind firmware */
 848                         index = ALIGN(rdev->uvd_fw->size, 8);
 849                         rdev->fence_drv[ring].cpu_addr = rdev->uvd.cpu_addr + index;
 850                         rdev->fence_drv[ring].gpu_addr = rdev->uvd.gpu_addr + index;
 851                 }
 852
 853         } else {
 854                 r = radeon_scratch_get(rdev, &rdev->fence_drv[ring].scratch_reg);
 855                 if (r) {
 856                         dev_err(rdev->dev, "fence failed to get scratch register\n");
 857                         return r;
 858                 }
 859                 index = RADEON_WB_SCRATCH_OFFSET +
 860                         rdev->fence_drv[ring].scratch_reg -
 861                         rdev->scratch.reg_base;
 862                 rdev->fence_drv[ring].cpu_addr = &rdev->wb.wb[index/4];
 863                 rdev->fence_drv[ring].gpu_addr = rdev->wb.gpu_addr + index;
 864         }
 865         radeon_fence_write(rdev, atomic64_read(&rdev->fence_drv[ring].last_seq), ring);
 866         rdev->fence_drv[ring].initialized = true;
 867         dev_info(rdev->dev, "fence driver on ring %d use gpu addr 0x%016llx\n",
 868                  ring, rdev->fence_drv[ring].gpu_addr);
 869         return 0;
 870 }
 871
 872 /**
 873  * radeon_fence_driver_init_ring - init the fence driver
 874  * for the requested ring.
 875  *
 876  * @rdev: radeon device pointer
 877  * @ring: ring index to start the fence driver on
 878  *
 879  * Init the fence driver for the requested ring (all asics).
 880  * Helper function for radeon_fence_driver_init().
 881  */
 882 static void radeon_fence_driver_init_ring(struct radeon_device *rdev, int ring)
 883 {
 884         int i;
 885
 886         rdev->fence_drv[ring].scratch_reg = -1;
 887         rdev->fence_drv[ring].cpu_addr = NULL;
 888         rdev->fence_drv[ring].gpu_addr = 0;
 889         for (i = 0; i < RADEON_NUM_RINGS; ++i)
 890                 rdev->fence_drv[ring].sync_seq[i] = 0;
 891         atomic64_set(&rdev->fence_drv[ring].last_seq, 0);
 892         rdev->fence_drv[ring].initialized = false;
 893         INIT_DELAYED_WORK(&rdev->fence_drv[ring].lockup_work,
 894                           radeon_fence_check_lockup);
 895         rdev->fence_drv[ring].rdev = rdev;
 896 }
 897
 898 /**
 899  * radeon_fence_driver_init - init the fence driver
 900  * for all possible rings.
 901  *
 902  * @rdev: radeon device pointer
 903  *
 904  * Init the fence driver for all possible rings (all asics).
 905  * Not all asics have all rings, so each asic will only
 906  * start the fence driver on the rings it has using
 907  * radeon_fence_driver_start_ring().
 908  */
 909 void radeon_fence_driver_init(struct radeon_device *rdev)
 910 {
 911         int ring;
 912
 913         init_waitqueue_head(&rdev->fence_queue);
 914         for (ring = 0; ring < RADEON_NUM_RINGS; ring++) {
 915                 radeon_fence_driver_init_ring(rdev, ring);
 916         }
 917
 918         radeon_debugfs_fence_init(rdev);
 919 }
 920
 921 /**
 922  * radeon_fence_driver_fini - tear down the fence driver
 923  * for all possible rings.
 924  *
 925  * @rdev: radeon device pointer
 926  *
 927  * Tear down the fence driver for all possible rings (all asics).
 928  */
 929 void radeon_fence_driver_fini(struct radeon_device *rdev)
 930 {
 931         int ring, r;
 932
 933         mutex_lock(&rdev->ring_lock);
 934         for (ring = 0; ring < RADEON_NUM_RINGS; ring++) {
 935                 if (!rdev->fence_drv[ring].initialized)
 936                         continue;
 937                 r = radeon_fence_wait_empty(rdev, ring);
 938                 if (r) {
 939                         /* no need to trigger GPU reset as we are unloading */
 940                         radeon_fence_driver_force_completion(rdev, ring);
 941                 }
 942                 cancel_delayed_work_sync(&rdev->fence_drv[ring].lockup_work);
 943                 wake_up_all(&rdev->fence_queue);
 944                 radeon_scratch_free(rdev, rdev->fence_drv[ring].scratch_reg);
 945                 rdev->fence_drv[ring].initialized = false;
 946         }
 947         mutex_unlock(&rdev->ring_lock);
 948 }
 949
 950 /**
 951  * radeon_fence_driver_force_completion - force all fence waiter to complete
 952  *
 953  * @rdev: radeon device pointer
 954  * @ring: the ring to complete
 955  *
 956  * In case of GPU reset failure make sure no process keep waiting on fence
 957  * that will never complete.
 958  */
 959 void radeon_fence_driver_force_completion(struct radeon_device *rdev, int ring)
 960 {
 961         if (rdev->fence_drv[ring].initialized) {
 962                 radeon_fence_write(rdev, rdev->fence_drv[ring].sync_seq[ring], ring);
 963                 cancel_delayed_work_sync(&rdev->fence_drv[ring].lockup_work);
 964         }
 965 }
 966
 967
 968 /*
 969  * Fence debugfs
 970  */
 971 #if defined(CONFIG_DEBUG_FS)
 972 static int radeon_debugfs_fence_info_show(struct seq_file *m, void *data)
 973 {
 974         struct radeon_device *rdev = (struct radeon_device *)m->private;
 975         int i, j;
 976
 977         for (i = 0; i < RADEON_NUM_RINGS; ++i) {
 978                 if (!rdev->fence_drv[i].initialized)
 979                         continue;
 980
 981                 radeon_fence_process(rdev, i);
 982
 983                 seq_printf(m, "--- ring %d ---\n", i);
 984                 seq_printf(m, "Last signaled fence 0x%016llx\n",
 985                            (unsigned long long)atomic64_read(&rdev->fence_drv[i].last_seq));
 986                 seq_printf(m, "Last emitted        0x%016llx\n",
 987                            rdev->fence_drv[i].sync_seq[i]);
 988
 989                 for (j = 0; j < RADEON_NUM_RINGS; ++j) {
 990                         if (i != j && rdev->fence_drv[j].initialized)
 991                                 seq_printf(m, "Last sync to ring %d 0x%016llx\n",
 992                                            j, rdev->fence_drv[i].sync_seq[j]);
 993                 }
 994         }
 995         return 0;
 996 }
 997
 998 /*
 999  * radeon_debugfs_gpu_reset - manually trigger a gpu reset
1000  *
1001  * Manually trigger a gpu reset at the next fence wait.
1002  */
1003 static int radeon_debugfs_gpu_reset(void *data, u64 *val)
1004 {
1005         struct radeon_device *rdev = (struct radeon_device *)data;
1006
1007         down_read(&rdev->exclusive_lock);
1008         *val = rdev->needs_reset;
1009         rdev->needs_reset = true;
1010         wake_up_all(&rdev->fence_queue);
1011         up_read(&rdev->exclusive_lock);
1012
1013         return 0;
1014 }
1015 DEFINE_SHOW_ATTRIBUTE(radeon_debugfs_fence_info);
1016 DEFINE_DEBUGFS_ATTRIBUTE(radeon_debugfs_gpu_reset_fops,
1017                          radeon_debugfs_gpu_reset, NULL, "%lld\n");
1018 #endif
1019
1020 void radeon_debugfs_fence_init(struct radeon_device *rdev)
1021 {
1022 #if defined(CONFIG_DEBUG_FS)
1023         struct dentry *root = rdev->ddev->primary->debugfs_root;
1024
1025         debugfs_create_file("radeon_gpu_reset", 0444, root, rdev,
1026                             &radeon_debugfs_gpu_reset_fops);
1027         debugfs_create_file("radeon_fence_info", 0444, root, rdev,
1028                             &radeon_debugfs_fence_info_fops);
1029
1030
1031 #endif
1032 }
1033
1034 static const char *radeon_fence_get_driver_name(struct dma_fence *fence)
1035 {
1036         return "radeon";
1037 }
1038
1039 static const char *radeon_fence_get_timeline_name(struct dma_fence *f)
1040 {
1041         struct radeon_fence *fence = to_radeon_fence(f);
1042         switch (fence->ring) {
1043         case RADEON_RING_TYPE_GFX_INDEX: return "radeon.gfx";
1044         case CAYMAN_RING_TYPE_CP1_INDEX: return "radeon.cp1";
1045         case CAYMAN_RING_TYPE_CP2_INDEX: return "radeon.cp2";
1046         case R600_RING_TYPE_DMA_INDEX: return "radeon.dma";
1047         case CAYMAN_RING_TYPE_DMA1_INDEX: return "radeon.dma1";
1048         case R600_RING_TYPE_UVD_INDEX: return "radeon.uvd";
1049         case TN_RING_TYPE_VCE1_INDEX: return "radeon.vce1";
1050         case TN_RING_TYPE_VCE2_INDEX: return "radeon.vce2";
1051         default: WARN_ON_ONCE(1); return "radeon.unk";
1052         }
1053 }
1054
1055 static inline bool radeon_test_signaled(struct radeon_fence *fence)
1056 {
1057         return test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->base.flags);
1058 }
1059
1060 struct radeon_wait_cb {
1061         struct dma_fence_cb base;
1062         struct task_struct *task;
1063 };
1064
1065 static void
1066 radeon_fence_wait_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
1067 {
1068         struct radeon_wait_cb *wait =
1069                 container_of(cb, struct radeon_wait_cb, base);
1070
1071         wake_up_process(wait->task);
1072 }
1073
1074 static signed long radeon_fence_default_wait(struct dma_fence *f, bool intr,
1075                                              signed long t)
1076 {
1077         struct radeon_fence *fence = to_radeon_fence(f);
1078         struct radeon_device *rdev = fence->rdev;
1079         struct radeon_wait_cb cb;
1080
1081         cb.task = current;
1082
1083         if (dma_fence_add_callback(f, &cb.base, radeon_fence_wait_cb))
1084                 return t;
1085
1086         while (t > 0) {
1087                 if (intr)
1088                         set_current_state(TASK_INTERRUPTIBLE);
1089                 else
1090                         set_current_state(TASK_UNINTERRUPTIBLE);
1091
1092                 /*
1093                  * radeon_test_signaled must be called after
1094                  * set_current_state to prevent a race with wake_up_process
1095                  */
1096                 if (radeon_test_signaled(fence))
1097                         break;
1098
1099                 if (rdev->needs_reset) {
1100                         t = -EDEADLK;
1101                         break;
1102                 }
1103
1104                 t = schedule_timeout(t);
1105
1106                 if (t > 0 && intr && signal_pending(current))
1107                         t = -ERESTARTSYS;
1108         }
1109
1110         __set_current_state(TASK_RUNNING);
1111         dma_fence_remove_callback(f, &cb.base);
1112
1113         return t;
1114 }
1115
1116 const struct dma_fence_ops radeon_fence_ops = {
1117         .get_driver_name = radeon_fence_get_driver_name,
1118         .get_timeline_name = radeon_fence_get_timeline_name,
1119         .enable_signaling = radeon_fence_enable_signaling,
1120         .signaled = radeon_fence_is_signaled,
1121         .wait = radeon_fence_default_wait,
1122         .release = NULL,
1123 };