sys/dev/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 #include <linux/seq_file.h>
  32 #include <linux/atomic.h>
  33 #include <linux/wait.h>
  34 #include <linux/kref.h>
  35 #include <linux/firmware.h>
  36 #include <drm/drmP.h>
  37 #include "radeon_reg.h"
  38 #include "radeon.h"
  39 #include "radeon_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.  Whether
  48  * we use a scratch register or memory location depends on the asic
  49  * and whether writeback is enabled.
  50  */
  51
  52 /**
  53  * radeon_fence_write - write a fence value
  54  *
  55  * @rdev: radeon_device pointer
  56  * @seq: sequence number to write
  57  * @ring: ring index the fence is associated with
  58  *
  59  * Writes a fence value to memory or a scratch register (all asics).
  60  */
  61 static void radeon_fence_write(struct radeon_device *rdev, u32 seq, int ring)
  62 {
  63         struct radeon_fence_driver *drv = &rdev->fence_drv[ring];
  64         if (likely(rdev->wb.enabled || !drv->scratch_reg)) {
  65                 if (drv->cpu_addr) {
  66                         *drv->cpu_addr = cpu_to_le32(seq);
  67                 }
  68         } else {
  69                 WREG32(drv->scratch_reg, seq);
  70         }
  71 }
  72
  73 /**
  74  * radeon_fence_read - read a fence value
  75  *
  76  * @rdev: radeon_device pointer
  77  * @ring: ring index the fence is associated with
  78  *
  79  * Reads a fence value from memory or a scratch register (all asics).
  80  * Returns the value of the fence read from memory or register.
  81  */
  82 static u32 radeon_fence_read(struct radeon_device *rdev, int ring)
  83 {
  84         struct radeon_fence_driver *drv = &rdev->fence_drv[ring];
  85         u32 seq = 0;
  86
  87         if (likely(rdev->wb.enabled || !drv->scratch_reg)) {
  88                 if (drv->cpu_addr) {
  89                         seq = le32_to_cpu(*drv->cpu_addr);
  90                 } else {
  91                         seq = lower_32_bits(atomic64_read(&drv->last_seq));
  92                 }
  93         } else {
  94                 seq = RREG32(drv->scratch_reg);
  95         }
  96         return seq;
  97 }
  98
  99 /**
 100  * radeon_fence_schedule_check - schedule lockup check
 101  *
 102  * @rdev: radeon_device pointer
 103  * @ring: ring index we should work with
 104  *
 105  * Queues a delayed work item to check for lockups.
 106  */
 107 static void radeon_fence_schedule_check(struct radeon_device *rdev, int ring)
 108 {
 109         /*
 110          * Do not reset the timer here with mod_delayed_work,
 111          * this can livelock in an interaction with TTM delayed destroy.
 112          */
 113         queue_delayed_work(system_power_efficient_wq,
 114                            &rdev->fence_drv[ring].lockup_work,
 115                            RADEON_FENCE_JIFFIES_TIMEOUT);
 116 }
 117
 118 /**
 119  * radeon_fence_emit - emit a fence on the requested ring
 120  *
 121  * @rdev: radeon_device pointer
 122  * @fence: radeon fence object
 123  * @ring: ring index the fence is associated with
 124  *
 125  * Emits a fence command on the requested ring (all asics).
 126  * Returns 0 on success, -ENOMEM on failure.
 127  */
 128 int radeon_fence_emit(struct radeon_device *rdev,
 129                       struct radeon_fence **fence,
 130                       int ring)
 131 {
 132         u64 seq;
 133
 134         /* we are protected by the ring emission mutex */
 135         *fence = kmalloc(sizeof(struct radeon_fence), M_DRM, M_WAITOK);
 136         if ((*fence) == NULL) {
 137                 return -ENOMEM;
 138         }
 139         (*fence)->rdev = rdev;
 140         (*fence)->seq = seq = ++rdev->fence_drv[ring].sync_seq[ring];
 141         (*fence)->ring = ring;
 142         (*fence)->is_vm_update = false;
 143         fence_init(&(*fence)->base, &radeon_fence_ops,
 144                    &rdev->fence_queue.lock, rdev->fence_context + ring, seq);
 145         radeon_fence_ring_emit(rdev, ring, *fence);
 146         trace_radeon_fence_emit(rdev->ddev, ring, (*fence)->seq);
 147         radeon_fence_schedule_check(rdev, ring);
 148         return 0;
 149 }
 150
 151 /**
 152  * radeon_fence_check_signaled - callback from fence_queue
 153  *
 154  * this function is called with fence_queue lock held, which is also used
 155  * for the fence locking itself, so unlocked variants are used for
 156  * fence_signal, and remove_wait_queue.
 157  */
 158 static int radeon_fence_check_signaled(wait_queue_t *wait, unsigned mode, int flags, void *key)
 159 {
 160         struct radeon_fence *fence;
 161         u64 seq;
 162
 163         fence = container_of(wait, struct radeon_fence, fence_wake);
 164
 165         /*
 166          * We cannot use radeon_fence_process here because we're already
 167          * in the waitqueue, in a call from wake_up_all.
 168          */
 169         seq = atomic64_read(&fence->rdev->fence_drv[fence->ring].last_seq);
 170         if (seq >= fence->seq) {
 171                 int ret = fence_signal_locked(&fence->base);
 172
 173                 if (!ret)
 174                         FENCE_TRACE(&fence->base, "signaled from irq context\n");
 175                 else
 176                         FENCE_TRACE(&fence->base, "was already signaled\n");
 177
 178                 radeon_irq_kms_sw_irq_put(fence->rdev, fence->ring);
 179                 __remove_wait_queue(&fence->rdev->fence_queue, &fence->fence_wake);
 180                 fence_put(&fence->base);
 181         } else
 182                 FENCE_TRACE(&fence->base, "pending\n");
 183         return 0;
 184 }
 185
 186 /**
 187  * radeon_fence_activity - check for fence activity
 188  *
 189  * @rdev: radeon_device pointer
 190  * @ring: ring index the fence is associated with
 191  *
 192  * Checks the current fence value and calculates the last
 193  * signalled fence value. Returns true if activity occured
 194  * on the ring, and the fence_queue should be waken up.
 195  */
 196 static bool radeon_fence_activity(struct radeon_device *rdev, int ring)
 197 {
 198         uint64_t seq, last_seq, last_emitted;
 199         unsigned count_loop = 0;
 200         bool wake = false;
 201
 202         /* Note there is a scenario here for an infinite loop but it's
 203          * very unlikely to happen. For it to happen, the current polling
 204          * process need to be interrupted by another process and another
 205          * process needs to update the last_seq btw the atomic read and
 206          * xchg of the current process.
 207          *
 208          * More over for this to go in infinite loop there need to be
 209          * continuously new fence signaled ie radeon_fence_read needs
 210          * to return a different value each time for both the currently
 211          * polling process and the other process that xchg the last_seq
 212          * btw atomic read and xchg of the current process. And the
 213          * value the other process set as last seq must be higher than
 214          * the seq value we just read. Which means that current process
 215          * need to be interrupted after radeon_fence_read and before
 216          * atomic xchg.
 217          *
 218          * To be even more safe we count the number of time we loop and
 219          * we bail after 10 loop just accepting the fact that we might
 220          * have temporarly set the last_seq not to the true real last
 221          * seq but to an older one.
 222          */
 223         last_seq = atomic64_read(&rdev->fence_drv[ring].last_seq);
 224         do {
 225                 last_emitted = rdev->fence_drv[ring].sync_seq[ring];
 226                 seq = radeon_fence_read(rdev, ring);
 227                 seq |= last_seq & 0xffffffff00000000LL;
 228                 if (seq < last_seq) {
 229                         seq &= 0xffffffff;
 230                         seq |= last_emitted & 0xffffffff00000000LL;
 231                 }
 232
 233                 if (seq <= last_seq || seq > last_emitted) {
 234                         break;
 235                 }
 236                 /* If we loop over we don't want to return without
 237                  * checking if a fence is signaled as it means that the
 238                  * seq we just read is different from the previous on.
 239                  */
 240                 wake = true;
 241                 last_seq = seq;
 242                 if ((count_loop++) > 10) {
 243                         /* We looped over too many time leave with the
 244                          * fact that we might have set an older fence
 245                          * seq then the current real last seq as signaled
 246                          * by the hw.
 247                          */
 248                         break;
 249                 }
 250         } while (atomic64_xchg(&rdev->fence_drv[ring].last_seq, seq) > seq);
 251
 252         if (seq < last_emitted)
 253                 radeon_fence_schedule_check(rdev, ring);
 254
 255         return wake;
 256 }
 257
 258 /**
 259  * radeon_fence_check_lockup - check for hardware lockup
 260  *
 261  * @work: delayed work item
 262  *
 263  * Checks for fence activity and if there is none probe
 264  * the hardware if a lockup occured.
 265  */
 266 static void radeon_fence_check_lockup(struct work_struct *work)
 267 {
 268         struct radeon_fence_driver *fence_drv;
 269         struct radeon_device *rdev;
 270         int ring;
 271
 272         fence_drv = container_of(work, struct radeon_fence_driver,
 273                                  lockup_work.work);
 274         rdev = fence_drv->rdev;
 275         ring = fence_drv - &rdev->fence_drv[0];
 276
 277         if (!down_read_trylock(&rdev->exclusive_lock)) {
 278                 /* just reschedule the check if a reset is going on */
 279                 radeon_fence_schedule_check(rdev, ring);
 280                 return;
 281         }
 282
 283         if (fence_drv->delayed_irq && rdev->ddev->irq_enabled) {
 284                 unsigned long irqflags;
 285
 286                 fence_drv->delayed_irq = false;
 287                 spin_lock_irqsave(&rdev->irq.lock, irqflags);
 288                 radeon_irq_set(rdev);
 289                 spin_unlock_irqrestore(&rdev->irq.lock, irqflags);
 290         }
 291
 292         if (radeon_fence_activity(rdev, ring))
 293                 wake_up_all(&rdev->fence_queue);
 294
 295         else if (radeon_ring_is_lockup(rdev, ring, &rdev->ring[ring])) {
 296
 297                 /* good news we believe it's a lockup */
 298                 dev_warn(rdev->dev, "GPU lockup (current fence id "
 299                          "0x%016lx last fence id 0x%016lx on ring %d)\n",
 300                          (uint64_t)atomic64_read(&fence_drv->last_seq),
 301                          fence_drv->sync_seq[ring], ring);
 302
 303                 /* remember that we need an reset */
 304                 rdev->needs_reset = true;
 305                 wake_up_all(&rdev->fence_queue);
 306         }
 307         up_read(&rdev->exclusive_lock);
 308 }
 309
 310 /**
 311  * radeon_fence_process - process a fence
 312  *
 313  * @rdev: radeon_device pointer
 314  * @ring: ring index the fence is associated with
 315  *
 316  * Checks the current fence value and wakes the fence queue
 317  * if the sequence number has increased (all asics).
 318  */
 319 void radeon_fence_process(struct radeon_device *rdev, int ring)
 320 {
 321         if (radeon_fence_activity(rdev, ring))
 322                 wake_up_all(&rdev->fence_queue);
 323 }
 324
 325 /**
 326  * radeon_fence_seq_signaled - check if a fence sequence number has signaled
 327  *
 328  * @rdev: radeon device pointer
 329  * @seq: sequence number
 330  * @ring: ring index the fence is associated with
 331  *
 332  * Check if the last signaled fence sequnce number is >= the requested
 333  * sequence number (all asics).
 334  * Returns true if the fence has signaled (current fence value
 335  * is >= requested value) or false if it has not (current fence
 336  * value is < the requested value.  Helper function for
 337  * radeon_fence_signaled().
 338  */
 339 static bool radeon_fence_seq_signaled(struct radeon_device *rdev,
 340                                       u64 seq, unsigned ring)
 341 {
 342         if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) {
 343                 return true;
 344         }
 345         /* poll new last sequence at least once */
 346         radeon_fence_process(rdev, ring);
 347         if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) {
 348                 return true;
 349         }
 350         return false;
 351 }
 352
 353 static bool radeon_fence_is_signaled(struct fence *f)
 354 {
 355         struct radeon_fence *fence = to_radeon_fence(f);
 356         struct radeon_device *rdev = fence->rdev;
 357         unsigned ring = fence->ring;
 358         u64 seq = fence->seq;
 359
 360         if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) {
 361                 return true;
 362         }
 363
 364         if (down_read_trylock(&rdev->exclusive_lock)) {
 365                 radeon_fence_process(rdev, ring);
 366                 up_read(&rdev->exclusive_lock);
 367
 368                 if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) {
 369                         return true;
 370                 }
 371         }
 372         return false;
 373 }
 374
 375 /**
 376  * radeon_fence_enable_signaling - enable signalling on fence
 377  * @fence: fence
 378  *
 379  * This function is called with fence_queue lock held, and adds a callback
 380  * to fence_queue that checks if this fence is signaled, and if so it
 381  * signals the fence and removes itself.
 382  */
 383 static bool radeon_fence_enable_signaling(struct fence *f)
 384 {
 385         struct radeon_fence *fence = to_radeon_fence(f);
 386         struct radeon_device *rdev = fence->rdev;
 387
 388         if (atomic64_read(&rdev->fence_drv[fence->ring].last_seq) >= fence->seq)
 389                 return false;
 390
 391         if (down_read_trylock(&rdev->exclusive_lock)) {
 392                 radeon_irq_kms_sw_irq_get(rdev, fence->ring);
 393
 394                 if (radeon_fence_activity(rdev, fence->ring))
 395                         wake_up_all_locked(&rdev->fence_queue);
 396
 397                 /* did fence get signaled after we enabled the sw irq? */
 398                 if (atomic64_read(&rdev->fence_drv[fence->ring].last_seq) >= fence->seq) {
 399                         radeon_irq_kms_sw_irq_put(rdev, fence->ring);
 400                         up_read(&rdev->exclusive_lock);
 401                         return false;
 402                 }
 403
 404                 up_read(&rdev->exclusive_lock);
 405         } else {
 406                 /* we're probably in a lockup, lets not fiddle too much */
 407                 if (radeon_irq_kms_sw_irq_get_delayed(rdev, fence->ring))
 408                         rdev->fence_drv[fence->ring].delayed_irq = true;
 409                 radeon_fence_schedule_check(rdev, fence->ring);
 410         }
 411
 412         fence->fence_wake.flags = 0;
 413         fence->fence_wake.private = NULL;
 414         fence->fence_wake.func = radeon_fence_check_signaled;
 415         __add_wait_queue(&rdev->fence_queue, &fence->fence_wake);
 416         fence_get(f);
 417
 418         FENCE_TRACE(&fence->base, "armed on ring %i!\n", fence->ring);
 419         return true;
 420 }
 421
 422 /**
 423  * radeon_fence_signaled - check if a fence has signaled
 424  *
 425  * @fence: radeon fence object
 426  *
 427  * Check if the requested fence has signaled (all asics).
 428  * Returns true if the fence has signaled or false if it has not.
 429  */
 430 bool radeon_fence_signaled(struct radeon_fence *fence)
 431 {
 432         if (!fence)
 433                 return true;
 434
 435         if (radeon_fence_seq_signaled(fence->rdev, fence->seq, fence->ring)) {
 436                 int ret;
 437
 438                 ret = fence_signal(&fence->base);
 439                 if (!ret)
 440                         FENCE_TRACE(&fence->base, "signaled from radeon_fence_signaled\n");
 441                 return true;
 442         }
 443         return false;
 444 }
 445
 446 /**
 447  * radeon_fence_any_seq_signaled - check if any sequence number is signaled
 448  *
 449  * @rdev: radeon device pointer
 450  * @seq: sequence numbers
 451  *
 452  * Check if the last signaled fence sequnce number is >= the requested
 453  * sequence number (all asics).
 454  * Returns true if any has signaled (current value is >= requested value)
 455  * or false if it has not. Helper function for radeon_fence_wait_seq.
 456  */
 457 static bool radeon_fence_any_seq_signaled(struct radeon_device *rdev, u64 *seq)
 458 {
 459         unsigned i;
 460
 461         for (i = 0; i < RADEON_NUM_RINGS; ++i) {
 462                 if (seq[i] && radeon_fence_seq_signaled(rdev, seq[i], i))
 463                         return true;
 464         }
 465         return false;
 466 }
 467
 468 /**
 469  * radeon_fence_wait_seq_timeout - wait for a specific sequence numbers
 470  *
 471  * @rdev: radeon device pointer
 472  * @target_seq: sequence number(s) we want to wait for
 473  * @intr: use interruptable sleep
 474  * @timeout: maximum time to wait, or MAX_SCHEDULE_TIMEOUT for infinite wait
 475  *
 476  * Wait for the requested sequence number(s) to be written by any ring
 477  * (all asics).  Sequnce number array is indexed by ring id.
 478  * @intr selects whether to use interruptable (true) or non-interruptable
 479  * (false) sleep when waiting for the sequence number.  Helper function
 480  * for radeon_fence_wait_*().
 481  * Returns remaining time if the sequence number has passed, 0 when
 482  * the wait timeout, or an error for all other cases.
 483  * -EDEADLK is returned when a GPU lockup has been detected.
 484  */
 485 static long radeon_fence_wait_seq_timeout(struct radeon_device *rdev,
 486                                           u64 *target_seq, bool intr,
 487                                           long timeout)
 488 {
 489         long r;
 490         int i;
 491
 492         if (radeon_fence_any_seq_signaled(rdev, target_seq))
 493                 return timeout;
 494
 495         /* enable IRQs and tracing */
 496         for (i = 0; i < RADEON_NUM_RINGS; ++i) {
 497                 if (!target_seq[i])
 498                         continue;
 499
 500                 trace_radeon_fence_wait_begin(rdev->ddev, i, target_seq[i]);
 501                 radeon_irq_kms_sw_irq_get(rdev, i);
 502         }
 503
 504         if (intr) {
 505                 r = wait_event_interruptible_timeout(rdev->fence_queue, (
 506                         radeon_fence_any_seq_signaled(rdev, target_seq)
 507                          || rdev->needs_reset), timeout);
 508         } else {
 509                 r = wait_event_timeout(rdev->fence_queue, (
 510                         radeon_fence_any_seq_signaled(rdev, target_seq)
 511                          || rdev->needs_reset), timeout);
 512         }
 513
 514         if (rdev->needs_reset)
 515                 r = -EDEADLK;
 516
 517         for (i = 0; i < RADEON_NUM_RINGS; ++i) {
 518                 if (!target_seq[i])
 519                         continue;
 520
 521                 radeon_irq_kms_sw_irq_put(rdev, i);
 522                 trace_radeon_fence_wait_end(rdev->ddev, i, target_seq[i]);
 523         }
 524
 525         return r;
 526 }
 527
 528 /**
 529  * radeon_fence_wait_timeout - wait for a fence to signal with timeout
 530  *
 531  * @fence: radeon fence object
 532  * @intr: use interruptible sleep
 533  *
 534  * Wait for the requested fence to signal (all asics).
 535  * @intr selects whether to use interruptable (true) or non-interruptable
 536  * (false) sleep when waiting for the fence.
 537  * @timeout: maximum time to wait, or MAX_SCHEDULE_TIMEOUT for infinite wait
 538  * Returns remaining time if the sequence number has passed, 0 when
 539  * the wait timeout, or an error for all other cases.
 540  */
 541 long radeon_fence_wait_timeout(struct radeon_fence *fence, bool intr, long timeout)
 542 {
 543         u64 seq[RADEON_NUM_RINGS] = {};
 544         long r;
 545         int r_sig;
 546
 547         /*
 548          * This function should not be called on !radeon fences.
 549          * If this is the case, it would mean this function can
 550          * also be called on radeon fences belonging to another card.
 551          * exclusive_lock is not held in that case.
 552          */
 553         if (WARN_ON_ONCE(!to_radeon_fence(&fence->base)))
 554                 return fence_wait(&fence->base, intr);
 555
 556         seq[fence->ring] = fence->seq;
 557         r = radeon_fence_wait_seq_timeout(fence->rdev, seq, intr, timeout);
 558         if (r <= 0) {
 559                 return r;
 560         }
 561
 562         r_sig = fence_signal(&fence->base);
 563         if (!r_sig)
 564                 FENCE_TRACE(&fence->base, "signaled from fence_wait\n");
 565         return r;
 566 }
 567
 568 /**
 569  * radeon_fence_wait - wait for a fence to signal
 570  *
 571  * @fence: radeon fence object
 572  * @intr: use interruptible sleep
 573  *
 574  * Wait for the requested fence to signal (all asics).
 575  * @intr selects whether to use interruptable (true) or non-interruptable
 576  * (false) sleep when waiting for the fence.
 577  * Returns 0 if the fence has passed, error for all other cases.
 578  */
 579 int radeon_fence_wait(struct radeon_fence *fence, bool intr)
 580 {
 581         long r = radeon_fence_wait_timeout(fence, intr, MAX_SCHEDULE_TIMEOUT);
 582         if (r > 0) {
 583                 return 0;
 584         } else {
 585                 return r;
 586         }
 587 }
 588
 589 /**
 590  * radeon_fence_wait_any - wait for a fence to signal on any ring
 591  *
 592  * @rdev: radeon device pointer
 593  * @fences: radeon fence object(s)
 594  * @intr: use interruptable sleep
 595  *
 596  * Wait for any requested fence to signal (all asics).  Fence
 597  * array is indexed by ring id.  @intr selects whether to use
 598  * interruptable (true) or non-interruptable (false) sleep when
 599  * waiting for the fences. Used by the suballocator.
 600  * Returns 0 if any fence has passed, error for all other cases.
 601  */
 602 int radeon_fence_wait_any(struct radeon_device *rdev,
 603                           struct radeon_fence **fences,
 604                           bool intr)
 605 {
 606         u64 seq[RADEON_NUM_RINGS];
 607         unsigned i, num_rings = 0;
 608         long r;
 609
 610         for (i = 0; i < RADEON_NUM_RINGS; ++i) {
 611                 seq[i] = 0;
 612
 613                 if (!fences[i]) {
 614                         continue;
 615                 }
 616
 617                 seq[i] = fences[i]->seq;
 618                 ++num_rings;
 619         }
 620
 621         /* nothing to wait for ? */
 622         if (num_rings == 0)
 623                 return -ENOENT;
 624
 625         r = radeon_fence_wait_seq_timeout(rdev, seq, intr, MAX_SCHEDULE_TIMEOUT);
 626         if (r < 0) {
 627                 return r;
 628         }
 629         return 0;
 630 }
 631
 632 /**
 633  * radeon_fence_wait_next - wait for the next fence to signal
 634  *
 635  * @rdev: radeon device pointer
 636  * @ring: ring index the fence is associated with
 637  *
 638  * Wait for the next fence on the requested ring to signal (all asics).
 639  * Returns 0 if the next fence has passed, error for all other cases.
 640  * Caller must hold ring lock.
 641  */
 642 int radeon_fence_wait_next(struct radeon_device *rdev, int ring)
 643 {
 644         u64 seq[RADEON_NUM_RINGS] = {};
 645         long r;
 646
 647         seq[ring] = atomic64_read(&rdev->fence_drv[ring].last_seq) + 1ULL;
 648         if (seq[ring] >= rdev->fence_drv[ring].sync_seq[ring]) {
 649                 /* nothing to wait for, last_seq is
 650                    already the last emited fence */
 651                 return -ENOENT;
 652         }
 653         r = radeon_fence_wait_seq_timeout(rdev, seq, false, MAX_SCHEDULE_TIMEOUT);
 654         if (r < 0)
 655                 return r;
 656         return 0;
 657 }
 658
 659 /**
 660  * radeon_fence_wait_empty - wait for all fences to signal
 661  *
 662  * @rdev: radeon device pointer
 663  * @ring: ring index the fence is associated with
 664  *
 665  * Wait for all fences on the requested ring to signal (all asics).
 666  * Returns 0 if the fences have passed, error for all other cases.
 667  * Caller must hold ring lock.
 668  */
 669 int radeon_fence_wait_empty(struct radeon_device *rdev, int ring)
 670 {
 671         u64 seq[RADEON_NUM_RINGS] = {};
 672         long r;
 673
 674         seq[ring] = rdev->fence_drv[ring].sync_seq[ring];
 675         if (!seq[ring])
 676                 return 0;
 677
 678         r = radeon_fence_wait_seq_timeout(rdev, seq, false, MAX_SCHEDULE_TIMEOUT);
 679         if (r < 0) {
 680                 if (r == -EDEADLK)
 681                         return -EDEADLK;
 682
 683                 dev_err(rdev->dev, "error waiting for ring[%d] to become idle (%ld)\n",
 684                         ring, r);
 685         }
 686         return 0;
 687 }
 688
 689 /**
 690  * radeon_fence_ref - take a ref on a fence
 691  *
 692  * @fence: radeon fence object
 693  *
 694  * Take a reference on a fence (all asics).
 695  * Returns the fence.
 696  */
 697 struct radeon_fence *radeon_fence_ref(struct radeon_fence *fence)
 698 {
 699         fence_get(&fence->base);
 700         return fence;
 701 }
 702
 703 /**
 704  * radeon_fence_unref - remove a ref on a fence
 705  *
 706  * @fence: radeon fence object
 707  *
 708  * Remove a reference on a fence (all asics).
 709  */
 710 void radeon_fence_unref(struct radeon_fence **fence)
 711 {
 712         struct radeon_fence *tmp = *fence;
 713
 714         *fence = NULL;
 715         if (tmp) {
 716                 fence_put(&tmp->base);
 717         }
 718 }
 719
 720 /**
 721  * radeon_fence_count_emitted - get the count of emitted fences
 722  *
 723  * @rdev: radeon device pointer
 724  * @ring: ring index the fence is associated with
 725  *
 726  * Get the number of fences emitted on the requested ring (all asics).
 727  * Returns the number of emitted fences on the ring.  Used by the
 728  * dynpm code to ring track activity.
 729  */
 730 unsigned radeon_fence_count_emitted(struct radeon_device *rdev, int ring)
 731 {
 732         uint64_t emitted;
 733
 734         /* We are not protected by ring lock when reading the last sequence
 735          * but it's ok to report slightly wrong fence count here.
 736          */
 737         radeon_fence_process(rdev, ring);
 738         emitted = rdev->fence_drv[ring].sync_seq[ring]
 739                 - atomic64_read(&rdev->fence_drv[ring].last_seq);
 740         /* to avoid 32bits warp around */
 741         if (emitted > 0x10000000) {
 742                 emitted = 0x10000000;
 743         }
 744         return (unsigned)emitted;
 745 }
 746
 747 /**
 748  * radeon_fence_need_sync - do we need a semaphore
 749  *
 750  * @fence: radeon fence object
 751  * @dst_ring: which ring to check against
 752  *
 753  * Check if the fence needs to be synced against another ring
 754  * (all asics).  If so, we need to emit a semaphore.
 755  * Returns true if we need to sync with another ring, false if
 756  * not.
 757  */
 758 bool radeon_fence_need_sync(struct radeon_fence *fence, int dst_ring)
 759 {
 760         struct radeon_fence_driver *fdrv;
 761
 762         if (!fence) {
 763                 return false;
 764         }
 765
 766         if (fence->ring == dst_ring) {
 767                 return false;
 768         }
 769
 770         /* we are protected by the ring mutex */
 771         fdrv = &fence->rdev->fence_drv[dst_ring];
 772         if (fence->seq <= fdrv->sync_seq[fence->ring]) {
 773                 return false;
 774         }
 775
 776         return true;
 777 }
 778
 779 /**
 780  * radeon_fence_note_sync - record the sync point
 781  *
 782  * @fence: radeon fence object
 783  * @dst_ring: which ring to check against
 784  *
 785  * Note the sequence number at which point the fence will
 786  * be synced with the requested ring (all asics).
 787  */
 788 void radeon_fence_note_sync(struct radeon_fence *fence, int dst_ring)
 789 {
 790         struct radeon_fence_driver *dst, *src;
 791         unsigned i;
 792
 793         if (!fence) {
 794                 return;
 795         }
 796
 797         if (fence->ring == dst_ring) {
 798                 return;
 799         }
 800
 801         /* we are protected by the ring mutex */
 802         src = &fence->rdev->fence_drv[fence->ring];
 803         dst = &fence->rdev->fence_drv[dst_ring];
 804         for (i = 0; i < RADEON_NUM_RINGS; ++i) {
 805                 if (i == dst_ring) {
 806                         continue;
 807                 }
 808                 dst->sync_seq[i] = max(dst->sync_seq[i], src->sync_seq[i]);
 809         }
 810 }
 811
 812 /**
 813  * radeon_fence_driver_start_ring - make the fence driver
 814  * ready for use on the requested ring.
 815  *
 816  * @rdev: radeon device pointer
 817  * @ring: ring index to start the fence driver on
 818  *
 819  * Make the fence driver ready for processing (all asics).
 820  * Not all asics have all rings, so each asic will only
 821  * start the fence driver on the rings it has.
 822  * Returns 0 for success, errors for failure.
 823  */
 824 int radeon_fence_driver_start_ring(struct radeon_device *rdev, int ring)
 825 {
 826         uint64_t index;
 827         int r;
 828
 829         radeon_scratch_free(rdev, rdev->fence_drv[ring].scratch_reg);
 830         if (rdev->wb.use_event || !radeon_ring_supports_scratch_reg(rdev, &rdev->ring[ring])) {
 831                 rdev->fence_drv[ring].scratch_reg = 0;
 832                 if (ring != R600_RING_TYPE_UVD_INDEX) {
 833                         index = R600_WB_EVENT_OFFSET + ring * 4;
 834                         rdev->fence_drv[ring].cpu_addr = &rdev->wb.wb[index/4];
 835                         rdev->fence_drv[ring].gpu_addr = rdev->wb.gpu_addr +
 836                                                          index;
 837
 838                 } else {
 839                         /* put fence directly behind firmware */
 840                         index = ALIGN(rdev->uvd_fw->datasize, 8);
 841                         rdev->fence_drv[ring].cpu_addr = (void*)((uint8_t*)rdev->uvd.cpu_addr + index);
 842                         rdev->fence_drv[ring].gpu_addr = rdev->uvd.gpu_addr + index;
 843                 }
 844
 845         } else {
 846                 r = radeon_scratch_get(rdev, &rdev->fence_drv[ring].scratch_reg);
 847                 if (r) {
 848                         dev_err(rdev->dev, "fence failed to get scratch register\n");
 849                         return r;
 850                 }
 851                 index = RADEON_WB_SCRATCH_OFFSET +
 852                         rdev->fence_drv[ring].scratch_reg -
 853                         rdev->scratch.reg_base;
 854                 rdev->fence_drv[ring].cpu_addr = &rdev->wb.wb[index/4];
 855                 rdev->fence_drv[ring].gpu_addr = rdev->wb.gpu_addr + index;
 856         }
 857         radeon_fence_write(rdev, atomic64_read(&rdev->fence_drv[ring].last_seq), ring);
 858         rdev->fence_drv[ring].initialized = true;
 859         dev_info(rdev->dev, "fence driver on ring %d use gpu addr 0x%016lx and cpu addr 0x%p\n",
 860                  ring, rdev->fence_drv[ring].gpu_addr, rdev->fence_drv[ring].cpu_addr);
 861         return 0;
 862 }
 863
 864 /**
 865  * radeon_fence_driver_init_ring - init the fence driver
 866  * for the requested ring.
 867  *
 868  * @rdev: radeon device pointer
 869  * @ring: ring index to start the fence driver on
 870  *
 871  * Init the fence driver for the requested ring (all asics).
 872  * Helper function for radeon_fence_driver_init().
 873  */
 874 static void radeon_fence_driver_init_ring(struct radeon_device *rdev, int ring)
 875 {
 876         int i;
 877
 878         rdev->fence_drv[ring].scratch_reg = -1;
 879         rdev->fence_drv[ring].cpu_addr = NULL;
 880         rdev->fence_drv[ring].gpu_addr = 0;
 881         for (i = 0; i < RADEON_NUM_RINGS; ++i)
 882                 rdev->fence_drv[ring].sync_seq[i] = 0;
 883         atomic64_set(&rdev->fence_drv[ring].last_seq, 0);
 884         rdev->fence_drv[ring].initialized = false;
 885         INIT_DELAYED_WORK(&rdev->fence_drv[ring].lockup_work,
 886                           radeon_fence_check_lockup);
 887         rdev->fence_drv[ring].rdev = rdev;
 888 }
 889
 890 /**
 891  * radeon_fence_driver_init - init the fence driver
 892  * for all possible rings.
 893  *
 894  * @rdev: radeon device pointer
 895  *
 896  * Init the fence driver for all possible rings (all asics).
 897  * Not all asics have all rings, so each asic will only
 898  * start the fence driver on the rings it has using
 899  * radeon_fence_driver_start_ring().
 900  * Returns 0 for success.
 901  */
 902 int radeon_fence_driver_init(struct radeon_device *rdev)
 903 {
 904         int ring;
 905
 906         init_waitqueue_head(&rdev->fence_queue);
 907         for (ring = 0; ring < RADEON_NUM_RINGS; ring++) {
 908                 radeon_fence_driver_init_ring(rdev, ring);
 909         }
 910         if (radeon_debugfs_fence_init(rdev)) {
 911                 dev_err(rdev->dev, "fence debugfs file creation failed\n");
 912         }
 913         return 0;
 914 }
 915
 916 /**
 917  * radeon_fence_driver_fini - tear down the fence driver
 918  * for all possible rings.
 919  *
 920  * @rdev: radeon device pointer
 921  *
 922  * Tear down the fence driver for all possible rings (all asics).
 923  */
 924 void radeon_fence_driver_fini(struct radeon_device *rdev)
 925 {
 926         int ring, r;
 927
 928         mutex_lock(&rdev->ring_lock);
 929         for (ring = 0; ring < RADEON_NUM_RINGS; ring++) {
 930                 if (!rdev->fence_drv[ring].initialized)
 931                         continue;
 932                 r = radeon_fence_wait_empty(rdev, ring);
 933                 if (r) {
 934                         /* no need to trigger GPU reset as we are unloading */
 935                         radeon_fence_driver_force_completion(rdev, ring);
 936                 }
 937                 cancel_delayed_work_sync(&rdev->fence_drv[ring].lockup_work);
 938                 wake_up_all(&rdev->fence_queue);
 939                 radeon_scratch_free(rdev, rdev->fence_drv[ring].scratch_reg);
 940                 rdev->fence_drv[ring].initialized = false;
 941         }
 942         mutex_unlock(&rdev->ring_lock);
 943 }
 944
 945 /**
 946  * radeon_fence_driver_force_completion - force all fence waiter to complete
 947  *
 948  * @rdev: radeon device pointer
 949  * @ring: the ring to complete
 950  *
 951  * In case of GPU reset failure make sure no process keep waiting on fence
 952  * that will never complete.
 953  */
 954 void radeon_fence_driver_force_completion(struct radeon_device *rdev, int ring)
 955 {
 956         if (rdev->fence_drv[ring].initialized) {
 957                 radeon_fence_write(rdev, rdev->fence_drv[ring].sync_seq[ring], ring);
 958                 cancel_delayed_work_sync(&rdev->fence_drv[ring].lockup_work);
 959         }
 960 }
 961
 962
 963 /*
 964  * Fence debugfs
 965  */
 966 #if defined(CONFIG_DEBUG_FS)
 967 static int radeon_debugfs_fence_info(struct seq_file *m, void *data)
 968 {
 969         struct drm_info_node *node = (struct drm_info_node *)m->private;
 970         struct drm_device *dev = node->minor->dev;
 971         struct radeon_device *rdev = dev->dev_private;
 972         int i, j;
 973
 974         for (i = 0; i < RADEON_NUM_RINGS; ++i) {
 975                 if (!rdev->fence_drv[i].initialized)
 976                         continue;
 977
 978                 radeon_fence_process(rdev, i);
 979
 980                 seq_printf(m, "--- ring %d ---\n", i);
 981                 seq_printf(m, "Last signaled fence 0x%016llx\n",
 982                            (unsigned long long)atomic64_read(&rdev->fence_drv[i].last_seq));
 983                 seq_printf(m, "Last emitted        0x%016llx\n",
 984                            rdev->fence_drv[i].sync_seq[i]);
 985
 986                 for (j = 0; j < RADEON_NUM_RINGS; ++j) {
 987                         if (i != j && rdev->fence_drv[j].initialized)
 988                                 seq_printf(m, "Last sync to ring %d 0x%016llx\n",
 989                                            j, rdev->fence_drv[i].sync_seq[j]);
 990                 }
 991         }
 992         return 0;
 993 }
 994
 995 /**
 996  * radeon_debugfs_gpu_reset - manually trigger a gpu reset
 997  *
 998  * Manually trigger a gpu reset at the next fence wait.
 999  */
1000 static int radeon_debugfs_gpu_reset(struct seq_file *m, void *data)
1001 {
1002         struct drm_info_node *node = (struct drm_info_node *) m->private;
1003         struct drm_device *dev = node->minor->dev;
1004         struct radeon_device *rdev = dev->dev_private;
1005
1006         down_read(&rdev->exclusive_lock);
1007         seq_printf(m, "%d\n", rdev->needs_reset);
1008         rdev->needs_reset = true;
1009         wake_up_all(&rdev->fence_queue);
1010         up_read(&rdev->exclusive_lock);
1011
1012         return 0;
1013 }
1014
1015 static struct drm_info_list radeon_debugfs_fence_list[] = {
1016         {"radeon_fence_info", &radeon_debugfs_fence_info, 0, NULL},
1017         {"radeon_gpu_reset", &radeon_debugfs_gpu_reset, 0, NULL}
1018 };
1019 #endif
1020
1021 int radeon_debugfs_fence_init(struct radeon_device *rdev)
1022 {
1023 #if defined(CONFIG_DEBUG_FS)
1024         return radeon_debugfs_add_files(rdev, radeon_debugfs_fence_list, 2);
1025 #else
1026         return 0;
1027 #endif
1028 }
1029
1030 static const char *radeon_fence_get_driver_name(struct fence *fence)
1031 {
1032         return "radeon";
1033 }
1034
1035 static const char *radeon_fence_get_timeline_name(struct fence *f)
1036 {
1037         struct radeon_fence *fence = to_radeon_fence(f);
1038         switch (fence->ring) {
1039         case RADEON_RING_TYPE_GFX_INDEX: return "radeon.gfx";
1040         case CAYMAN_RING_TYPE_CP1_INDEX: return "radeon.cp1";
1041         case CAYMAN_RING_TYPE_CP2_INDEX: return "radeon.cp2";
1042         case R600_RING_TYPE_DMA_INDEX: return "radeon.dma";
1043         case CAYMAN_RING_TYPE_DMA1_INDEX: return "radeon.dma1";
1044         case R600_RING_TYPE_UVD_INDEX: return "radeon.uvd";
1045         case TN_RING_TYPE_VCE1_INDEX: return "radeon.vce1";
1046         case TN_RING_TYPE_VCE2_INDEX: return "radeon.vce2";
1047         default: WARN_ON_ONCE(1); return "radeon.unk";
1048         }
1049 }
1050
1051 static inline bool radeon_test_signaled(struct radeon_fence *fence)
1052 {
1053         return test_bit(FENCE_FLAG_SIGNALED_BIT, &fence->base.flags);
1054 }
1055
1056 struct radeon_wait_cb {
1057         struct fence_cb base;
1058         struct task_struct *task;
1059 };
1060
1061 static void
1062 radeon_fence_wait_cb(struct fence *fence, struct fence_cb *cb)
1063 {
1064         struct radeon_wait_cb *wait =
1065                 container_of(cb, struct radeon_wait_cb, base);
1066
1067         wake_up_process(wait->task);
1068 }
1069
1070 static signed long radeon_fence_default_wait(struct fence *f, bool intr,
1071                                              signed long t)
1072 {
1073         struct radeon_fence *fence = to_radeon_fence(f);
1074         struct radeon_device *rdev = fence->rdev;
1075         struct radeon_wait_cb cb;
1076
1077         cb.task = current;
1078
1079         if (fence_add_callback(f, &cb.base, radeon_fence_wait_cb))
1080                 return t;
1081
1082         while (t > 0) {
1083                 if (intr)
1084                         set_current_state(TASK_INTERRUPTIBLE);
1085                 else
1086                         set_current_state(TASK_UNINTERRUPTIBLE);
1087
1088                 /*
1089                  * radeon_test_signaled must be called after
1090                  * set_current_state to prevent a race with wake_up_process
1091                  */
1092                 if (radeon_test_signaled(fence))
1093                         break;
1094
1095                 if (rdev->needs_reset) {
1096                         t = -EDEADLK;
1097                         break;
1098                 }
1099
1100                 t = schedule_timeout(t);
1101
1102                 if (t > 0 && intr && signal_pending(current))
1103                         t = -ERESTARTSYS;
1104         }
1105
1106         __set_current_state(TASK_RUNNING);
1107         fence_remove_callback(f, &cb.base);
1108
1109         return t;
1110 }
1111
1112 const struct fence_ops radeon_fence_ops = {
1113         .get_driver_name = radeon_fence_get_driver_name,
1114         .get_timeline_name = radeon_fence_get_timeline_name,
1115         .enable_signaling = radeon_fence_enable_signaling,
1116         .signaled = radeon_fence_is_signaled,
1117         .wait = radeon_fence_default_wait,
1118         .release = NULL,
1119 };