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

   1 /*
   2  * Copyright 2014 Advanced Micro Devices, Inc.
   3  * Copyright 2008 Red Hat Inc.
   4  * Copyright 2009 Jerome Glisse.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22  * OTHER DEALINGS IN THE SOFTWARE.
  23  *
  24  */
  25
  26 #include "amdgpu.h"
  27 #include "amdgpu_gfx.h"
  28 #include "amdgpu_rlc.h"
  29 #include "amdgpu_ras.h"
  30
  31 /* delay 0.1 second to enable gfx off feature */
  32 #define GFX_OFF_DELAY_ENABLE         msecs_to_jiffies(100)
  33
  34 /*
  35  * GPU GFX IP block helpers function.
  36  */
  37
  38 int amdgpu_gfx_mec_queue_to_bit(struct amdgpu_device *adev, int mec,
  39                                 int pipe, int queue)
  40 {
  41         int bit = 0;
  42
  43         bit += mec * adev->gfx.mec.num_pipe_per_mec
  44                 * adev->gfx.mec.num_queue_per_pipe;
  45         bit += pipe * adev->gfx.mec.num_queue_per_pipe;
  46         bit += queue;
  47
  48         return bit;
  49 }
  50
  51 void amdgpu_queue_mask_bit_to_mec_queue(struct amdgpu_device *adev, int bit,
  52                                  int *mec, int *pipe, int *queue)
  53 {
  54         *queue = bit % adev->gfx.mec.num_queue_per_pipe;
  55         *pipe = (bit / adev->gfx.mec.num_queue_per_pipe)
  56                 % adev->gfx.mec.num_pipe_per_mec;
  57         *mec = (bit / adev->gfx.mec.num_queue_per_pipe)
  58                / adev->gfx.mec.num_pipe_per_mec;
  59
  60 }
  61
  62 bool amdgpu_gfx_is_mec_queue_enabled(struct amdgpu_device *adev,
  63                                      int mec, int pipe, int queue)
  64 {
  65         return test_bit(amdgpu_gfx_mec_queue_to_bit(adev, mec, pipe, queue),
  66                         adev->gfx.mec.queue_bitmap);
  67 }
  68
  69 int amdgpu_gfx_me_queue_to_bit(struct amdgpu_device *adev,
  70                                int me, int pipe, int queue)
  71 {
  72         int bit = 0;
  73
  74         bit += me * adev->gfx.me.num_pipe_per_me
  75                 * adev->gfx.me.num_queue_per_pipe;
  76         bit += pipe * adev->gfx.me.num_queue_per_pipe;
  77         bit += queue;
  78
  79         return bit;
  80 }
  81
  82 void amdgpu_gfx_bit_to_me_queue(struct amdgpu_device *adev, int bit,
  83                                 int *me, int *pipe, int *queue)
  84 {
  85         *queue = bit % adev->gfx.me.num_queue_per_pipe;
  86         *pipe = (bit / adev->gfx.me.num_queue_per_pipe)
  87                 % adev->gfx.me.num_pipe_per_me;
  88         *me = (bit / adev->gfx.me.num_queue_per_pipe)
  89                 / adev->gfx.me.num_pipe_per_me;
  90 }
  91
  92 bool amdgpu_gfx_is_me_queue_enabled(struct amdgpu_device *adev,
  93                                     int me, int pipe, int queue)
  94 {
  95         return test_bit(amdgpu_gfx_me_queue_to_bit(adev, me, pipe, queue),
  96                         adev->gfx.me.queue_bitmap);
  97 }
  98
  99 /**
 100  * amdgpu_gfx_scratch_get - Allocate a scratch register
 101  *
 102  * @adev: amdgpu_device pointer
 103  * @reg: scratch register mmio offset
 104  *
 105  * Allocate a CP scratch register for use by the driver (all asics).
 106  * Returns 0 on success or -EINVAL on failure.
 107  */
 108 int amdgpu_gfx_scratch_get(struct amdgpu_device *adev, uint32_t *reg)
 109 {
 110         int i;
 111
 112         i = ffs(adev->gfx.scratch.free_mask);
 113         if (i != 0 && i <= adev->gfx.scratch.num_reg) {
 114                 i--;
 115                 adev->gfx.scratch.free_mask &= ~(1u << i);
 116                 *reg = adev->gfx.scratch.reg_base + i;
 117                 return 0;
 118         }
 119         return -EINVAL;
 120 }
 121
 122 /**
 123  * amdgpu_gfx_scratch_free - Free a scratch register
 124  *
 125  * @adev: amdgpu_device pointer
 126  * @reg: scratch register mmio offset
 127  *
 128  * Free a CP scratch register allocated for use by the driver (all asics)
 129  */
 130 void amdgpu_gfx_scratch_free(struct amdgpu_device *adev, uint32_t reg)
 131 {
 132         adev->gfx.scratch.free_mask |= 1u << (reg - adev->gfx.scratch.reg_base);
 133 }
 134
 135 /**
 136  * amdgpu_gfx_parse_disable_cu - Parse the disable_cu module parameter
 137  *
 138  * @mask: array in which the per-shader array disable masks will be stored
 139  * @max_se: number of SEs
 140  * @max_sh: number of SHs
 141  *
 142  * The bitmask of CUs to be disabled in the shader array determined by se and
 143  * sh is stored in mask[se * max_sh + sh].
 144  */
 145 void amdgpu_gfx_parse_disable_cu(unsigned *mask, unsigned max_se, unsigned max_sh)
 146 {
 147         unsigned se, sh, cu;
 148         const char *p;
 149
 150         memset(mask, 0, sizeof(*mask) * max_se * max_sh);
 151
 152         if (!amdgpu_disable_cu || !*amdgpu_disable_cu)
 153                 return;
 154
 155         p = amdgpu_disable_cu;
 156         for (;;) {
 157                 char *next;
 158                 int ret = sscanf(p, "%u.%u.%u", &se, &sh, &cu);
 159                 if (ret < 3) {
 160                         DRM_ERROR("amdgpu: could not parse disable_cu\n");
 161                         return;
 162                 }
 163
 164                 if (se < max_se && sh < max_sh && cu < 16) {
 165                         DRM_INFO("amdgpu: disabling CU %u.%u.%u\n", se, sh, cu);
 166                         mask[se * max_sh + sh] |= 1u << cu;
 167                 } else {
 168                         DRM_ERROR("amdgpu: disable_cu %u.%u.%u is out of range\n",
 169                                   se, sh, cu);
 170                 }
 171
 172                 next = strchr(p, ',');
 173                 if (!next)
 174                         break;
 175                 p = next + 1;
 176         }
 177 }
 178
 179 static bool amdgpu_gfx_is_multipipe_capable(struct amdgpu_device *adev)
 180 {
 181         if (amdgpu_compute_multipipe != -1) {
 182                 DRM_INFO("amdgpu: forcing compute pipe policy %d\n",
 183                          amdgpu_compute_multipipe);
 184                 return amdgpu_compute_multipipe == 1;
 185         }
 186
 187         /* FIXME: spreading the queues across pipes causes perf regressions
 188          * on POLARIS11 compute workloads */
 189         if (adev->asic_type == CHIP_POLARIS11)
 190                 return false;
 191
 192         return adev->gfx.mec.num_mec > 1;
 193 }
 194
 195 bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
 196                                                struct amdgpu_ring *ring)
 197 {
 198         /* Policy: use 1st queue as high priority compute queue if we
 199          * have more than one compute queue.
 200          */
 201         if (adev->gfx.num_compute_rings > 1 &&
 202             ring == &adev->gfx.compute_ring[0])
 203                 return true;
 204
 205         return false;
 206 }
 207
 208 void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
 209 {
 210         int i, queue, pipe;
 211         bool multipipe_policy = amdgpu_gfx_is_multipipe_capable(adev);
 212         int max_queues_per_mec = min(adev->gfx.mec.num_pipe_per_mec *
 213                                      adev->gfx.mec.num_queue_per_pipe,
 214                                      adev->gfx.num_compute_rings);
 215
 216         if (multipipe_policy) {
 217                 /* policy: make queues evenly cross all pipes on MEC1 only */
 218                 for (i = 0; i < max_queues_per_mec; i++) {
 219                         pipe = i % adev->gfx.mec.num_pipe_per_mec;
 220                         queue = (i / adev->gfx.mec.num_pipe_per_mec) %
 221                                 adev->gfx.mec.num_queue_per_pipe;
 222
 223                         set_bit(pipe * adev->gfx.mec.num_queue_per_pipe + queue,
 224                                         adev->gfx.mec.queue_bitmap);
 225                 }
 226         } else {
 227                 /* policy: amdgpu owns all queues in the given pipe */
 228                 for (i = 0; i < max_queues_per_mec; ++i)
 229                         set_bit(i, adev->gfx.mec.queue_bitmap);
 230         }
 231
 232         dev_dbg(adev->dev, "mec queue bitmap weight=%d\n", bitmap_weight(adev->gfx.mec.queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES));
 233 }
 234
 235 void amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device *adev)
 236 {
 237         int i, queue, me;
 238
 239         for (i = 0; i < AMDGPU_MAX_GFX_QUEUES; ++i) {
 240                 queue = i % adev->gfx.me.num_queue_per_pipe;
 241                 me = (i / adev->gfx.me.num_queue_per_pipe)
 242                       / adev->gfx.me.num_pipe_per_me;
 243
 244                 if (me >= adev->gfx.me.num_me)
 245                         break;
 246                 /* policy: amdgpu owns the first queue per pipe at this stage
 247                  * will extend to mulitple queues per pipe later */
 248                 if (me == 0 && queue < 1)
 249                         set_bit(i, adev->gfx.me.queue_bitmap);
 250         }
 251
 252         /* update the number of active graphics rings */
 253         adev->gfx.num_gfx_rings =
 254                 bitmap_weight(adev->gfx.me.queue_bitmap, AMDGPU_MAX_GFX_QUEUES);
 255 }
 256
 257 static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev,
 258                                   struct amdgpu_ring *ring)
 259 {
 260         int queue_bit;
 261         int mec, pipe, queue;
 262
 263         queue_bit = adev->gfx.mec.num_mec
 264                     * adev->gfx.mec.num_pipe_per_mec
 265                     * adev->gfx.mec.num_queue_per_pipe;
 266
 267         while (queue_bit-- >= 0) {
 268                 if (test_bit(queue_bit, adev->gfx.mec.queue_bitmap))
 269                         continue;
 270
 271                 amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
 272
 273                 /*
 274                  * 1. Using pipes 2/3 from MEC 2 seems cause problems.
 275                  * 2. It must use queue id 0, because CGPG_IDLE/SAVE/LOAD/RUN
 276                  * only can be issued on queue 0.
 277                  */
 278                 if ((mec == 1 && pipe > 1) || queue != 0)
 279                         continue;
 280
 281                 ring->me = mec + 1;
 282                 ring->pipe = pipe;
 283                 ring->queue = queue;
 284
 285                 return 0;
 286         }
 287
 288         dev_err(adev->dev, "Failed to find a queue for KIQ\n");
 289         return -EINVAL;
 290 }
 291
 292 int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev,
 293                              struct amdgpu_ring *ring,
 294                              struct amdgpu_irq_src *irq)
 295 {
 296         struct amdgpu_kiq *kiq = &adev->gfx.kiq;
 297         int r = 0;
 298
 299         spin_lock_init(&kiq->ring_lock);
 300
 301         ring->adev = NULL;
 302         ring->ring_obj = NULL;
 303         ring->use_doorbell = true;
 304         ring->doorbell_index = adev->doorbell_index.kiq;
 305
 306         r = amdgpu_gfx_kiq_acquire(adev, ring);
 307         if (r)
 308                 return r;
 309
 310         ring->eop_gpu_addr = kiq->eop_gpu_addr;
 311         ring->no_scheduler = true;
 312         sprintf(ring->name, "kiq_%d.%d.%d", ring->me, ring->pipe, ring->queue);
 313         r = amdgpu_ring_init(adev, ring, 1024,
 314                              irq, AMDGPU_CP_KIQ_IRQ_DRIVER0,
 315                              AMDGPU_RING_PRIO_DEFAULT);
 316         if (r)
 317                 dev_warn(adev->dev, "(%d) failed to init kiq ring\n", r);
 318
 319         return r;
 320 }
 321
 322 void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring)
 323 {
 324         amdgpu_ring_fini(ring);
 325 }
 326
 327 void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev)
 328 {
 329         struct amdgpu_kiq *kiq = &adev->gfx.kiq;
 330
 331         amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL);
 332 }
 333
 334 int amdgpu_gfx_kiq_init(struct amdgpu_device *adev,
 335                         unsigned hpd_size)
 336 {
 337         int r;
 338         u32 *hpd;
 339         struct amdgpu_kiq *kiq = &adev->gfx.kiq;
 340
 341         r = amdgpu_bo_create_kernel(adev, hpd_size, PAGE_SIZE,
 342                                     AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj,
 343                                     &kiq->eop_gpu_addr, (void **)&hpd);
 344         if (r) {
 345                 dev_warn(adev->dev, "failed to create KIQ bo (%d).\n", r);
 346                 return r;
 347         }
 348
 349         memset(hpd, 0, hpd_size);
 350
 351         r = amdgpu_bo_reserve(kiq->eop_obj, true);
 352         if (unlikely(r != 0))
 353                 dev_warn(adev->dev, "(%d) reserve kiq eop bo failed\n", r);
 354         amdgpu_bo_kunmap(kiq->eop_obj);
 355         amdgpu_bo_unreserve(kiq->eop_obj);
 356
 357         return 0;
 358 }
 359
 360 /* create MQD for each compute/gfx queue */
 361 int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
 362                            unsigned mqd_size)
 363 {
 364         struct amdgpu_ring *ring = NULL;
 365         int r, i;
 366
 367         /* create MQD for KIQ */
 368         ring = &adev->gfx.kiq.ring;
 369         if (!ring->mqd_obj) {
 370                 /* originaly the KIQ MQD is put in GTT domain, but for SRIOV VRAM domain is a must
 371                  * otherwise hypervisor trigger SAVE_VF fail after driver unloaded which mean MQD
 372                  * deallocated and gart_unbind, to strict diverage we decide to use VRAM domain for
 373                  * KIQ MQD no matter SRIOV or Bare-metal
 374                  */
 375                 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
 376                                             AMDGPU_GEM_DOMAIN_VRAM, &ring->mqd_obj,
 377                                             &ring->mqd_gpu_addr, &ring->mqd_ptr);
 378                 if (r) {
 379                         dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
 380                         return r;
 381                 }
 382
 383                 /* prepare MQD backup */
 384                 adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS] = kmalloc(mqd_size, GFP_KERNEL);
 385                 if (!adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS])
 386                                 dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
 387         }
 388
 389         if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
 390                 /* create MQD for each KGQ */
 391                 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
 392                         ring = &adev->gfx.gfx_ring[i];
 393                         if (!ring->mqd_obj) {
 394                                 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
 395                                                             AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
 396                                                             &ring->mqd_gpu_addr, &ring->mqd_ptr);
 397                                 if (r) {
 398                                         dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
 399                                         return r;
 400                                 }
 401
 402                                 /* prepare MQD backup */
 403                                 adev->gfx.me.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
 404                                 if (!adev->gfx.me.mqd_backup[i])
 405                                         dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
 406                         }
 407                 }
 408         }
 409
 410         /* create MQD for each KCQ */
 411         for (i = 0; i < adev->gfx.num_compute_rings; i++) {
 412                 ring = &adev->gfx.compute_ring[i];
 413                 if (!ring->mqd_obj) {
 414                         r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
 415                                                     AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
 416                                                     &ring->mqd_gpu_addr, &ring->mqd_ptr);
 417                         if (r) {
 418                                 dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
 419                                 return r;
 420                         }
 421
 422                         /* prepare MQD backup */
 423                         adev->gfx.mec.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
 424                         if (!adev->gfx.mec.mqd_backup[i])
 425                                 dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
 426                 }
 427         }
 428
 429         return 0;
 430 }
 431
 432 void amdgpu_gfx_mqd_sw_fini(struct amdgpu_device *adev)
 433 {
 434         struct amdgpu_ring *ring = NULL;
 435         int i;
 436
 437         if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
 438                 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
 439                         ring = &adev->gfx.gfx_ring[i];
 440                         kfree(adev->gfx.me.mqd_backup[i]);
 441                         amdgpu_bo_free_kernel(&ring->mqd_obj,
 442                                               &ring->mqd_gpu_addr,
 443                                               &ring->mqd_ptr);
 444                 }
 445         }
 446
 447         for (i = 0; i < adev->gfx.num_compute_rings; i++) {
 448                 ring = &adev->gfx.compute_ring[i];
 449                 kfree(adev->gfx.mec.mqd_backup[i]);
 450                 amdgpu_bo_free_kernel(&ring->mqd_obj,
 451                                       &ring->mqd_gpu_addr,
 452                                       &ring->mqd_ptr);
 453         }
 454
 455         ring = &adev->gfx.kiq.ring;
 456         kfree(adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS]);
 457         amdgpu_bo_free_kernel(&ring->mqd_obj,
 458                               &ring->mqd_gpu_addr,
 459                               &ring->mqd_ptr);
 460 }
 461
 462 int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev)
 463 {
 464         struct amdgpu_kiq *kiq = &adev->gfx.kiq;
 465         struct amdgpu_ring *kiq_ring = &kiq->ring;
 466         int i;
 467
 468         if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
 469                 return -EINVAL;
 470
 471         if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
 472                                         adev->gfx.num_compute_rings))
 473                 return -ENOMEM;
 474
 475         for (i = 0; i < adev->gfx.num_compute_rings; i++)
 476                 kiq->pmf->kiq_unmap_queues(kiq_ring, &adev->gfx.compute_ring[i],
 477                                            RESET_QUEUES, 0, 0);
 478
 479         return amdgpu_ring_test_helper(kiq_ring);
 480 }
 481
 482 int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
 483                                         int queue_bit)
 484 {
 485         int mec, pipe, queue;
 486         int set_resource_bit = 0;
 487
 488         amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
 489
 490         set_resource_bit = mec * 4 * 8 + pipe * 8 + queue;
 491
 492         return set_resource_bit;
 493 }
 494
 495 int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev)
 496 {
 497         struct amdgpu_kiq *kiq = &adev->gfx.kiq;
 498         struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
 499         uint64_t queue_mask = 0;
 500         int r, i;
 501
 502         if (!kiq->pmf || !kiq->pmf->kiq_map_queues || !kiq->pmf->kiq_set_resources)
 503                 return -EINVAL;
 504
 505         for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
 506                 if (!test_bit(i, adev->gfx.mec.queue_bitmap))
 507                         continue;
 508
 509                 /* This situation may be hit in the future if a new HW
 510                  * generation exposes more than 64 queues. If so, the
 511                  * definition of queue_mask needs updating */
 512                 if (WARN_ON(i > (sizeof(queue_mask)*8))) {
 513                         DRM_ERROR("Invalid KCQ enabled: %d\n", i);
 514                         break;
 515                 }
 516
 517                 queue_mask |= (1ull << amdgpu_queue_mask_bit_to_set_resource_bit(adev, i));
 518         }
 519
 520         DRM_INFO("kiq ring mec %d pipe %d q %d\n", kiq_ring->me, kiq_ring->pipe,
 521                                                         kiq_ring->queue);
 522
 523         r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
 524                                         adev->gfx.num_compute_rings +
 525                                         kiq->pmf->set_resources_size);
 526         if (r) {
 527                 DRM_ERROR("Failed to lock KIQ (%d).\n", r);
 528                 return r;
 529         }
 530
 531         kiq->pmf->kiq_set_resources(kiq_ring, queue_mask);
 532         for (i = 0; i < adev->gfx.num_compute_rings; i++)
 533                 kiq->pmf->kiq_map_queues(kiq_ring, &adev->gfx.compute_ring[i]);
 534
 535         r = amdgpu_ring_test_helper(kiq_ring);
 536         if (r)
 537                 DRM_ERROR("KCQ enable failed\n");
 538
 539         return r;
 540 }
 541
 542 /* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
 543  *
 544  * @adev: amdgpu_device pointer
 545  * @bool enable true: enable gfx off feature, false: disable gfx off feature
 546  *
 547  * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled.
 548  * 2. other client can send request to disable gfx off feature, the request should be honored.
 549  * 3. other client can cancel their request of disable gfx off feature
 550  * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
 551  */
 552
 553 void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
 554 {
 555         if (!(adev->pm.pp_feature & PP_GFXOFF_MASK))
 556                 return;
 557
 558         mutex_lock(&adev->gfx.gfx_off_mutex);
 559
 560         if (!enable)
 561                 adev->gfx.gfx_off_req_count++;
 562         else if (adev->gfx.gfx_off_req_count > 0)
 563                 adev->gfx.gfx_off_req_count--;
 564
 565         if (enable && !adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
 566                 schedule_delayed_work(&adev->gfx.gfx_off_delay_work, GFX_OFF_DELAY_ENABLE);
 567         } else if (!enable && adev->gfx.gfx_off_state) {
 568                 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false)) {
 569                         adev->gfx.gfx_off_state = false;
 570
 571                         if (adev->gfx.funcs->init_spm_golden) {
 572                                 dev_dbg(adev->dev, "GFXOFF is disabled, re-init SPM golden settings\n");
 573                                 amdgpu_gfx_init_spm_golden(adev);
 574                         }
 575                 }
 576         }
 577
 578         mutex_unlock(&adev->gfx.gfx_off_mutex);
 579 }
 580
 581 int amdgpu_get_gfx_off_status(struct amdgpu_device *adev, uint32_t *value)
 582 {
 583
 584         int r = 0;
 585
 586         mutex_lock(&adev->gfx.gfx_off_mutex);
 587
 588         r = smu_get_status_gfxoff(adev, value);
 589
 590         mutex_unlock(&adev->gfx.gfx_off_mutex);
 591
 592         return r;
 593 }
 594
 595 int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev)
 596 {
 597         int r;
 598         struct ras_fs_if fs_info = {
 599                 .sysfs_name = "gfx_err_count",
 600         };
 601         struct ras_ih_if ih_info = {
 602                 .cb = amdgpu_gfx_process_ras_data_cb,
 603         };
 604
 605         if (!adev->gfx.ras_if) {
 606                 adev->gfx.ras_if = kmalloc(sizeof(struct ras_common_if), GFP_KERNEL);
 607                 if (!adev->gfx.ras_if)
 608                         return -ENOMEM;
 609                 adev->gfx.ras_if->block = AMDGPU_RAS_BLOCK__GFX;
 610                 adev->gfx.ras_if->type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
 611                 adev->gfx.ras_if->sub_block_index = 0;
 612                 strcpy(adev->gfx.ras_if->name, "gfx");
 613         }
 614         fs_info.head = ih_info.head = *adev->gfx.ras_if;
 615
 616         r = amdgpu_ras_late_init(adev, adev->gfx.ras_if,
 617                                  &fs_info, &ih_info);
 618         if (r)
 619                 goto free;
 620
 621         if (amdgpu_ras_is_supported(adev, adev->gfx.ras_if->block)) {
 622                 r = amdgpu_irq_get(adev, &adev->gfx.cp_ecc_error_irq, 0);
 623                 if (r)
 624                         goto late_fini;
 625         } else {
 626                 /* free gfx ras_if if ras is not supported */
 627                 r = 0;
 628                 goto free;
 629         }
 630
 631         return 0;
 632 late_fini:
 633         amdgpu_ras_late_fini(adev, adev->gfx.ras_if, &ih_info);
 634 free:
 635         kfree(adev->gfx.ras_if);
 636         adev->gfx.ras_if = NULL;
 637         return r;
 638 }
 639
 640 void amdgpu_gfx_ras_fini(struct amdgpu_device *adev)
 641 {
 642         if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX) &&
 643                         adev->gfx.ras_if) {
 644                 struct ras_common_if *ras_if = adev->gfx.ras_if;
 645                 struct ras_ih_if ih_info = {
 646                         .head = *ras_if,
 647                         .cb = amdgpu_gfx_process_ras_data_cb,
 648                 };
 649
 650                 amdgpu_ras_late_fini(adev, ras_if, &ih_info);
 651                 kfree(ras_if);
 652         }
 653 }
 654
 655 int amdgpu_gfx_process_ras_data_cb(struct amdgpu_device *adev,
 656                 void *err_data,
 657                 struct amdgpu_iv_entry *entry)
 658 {
 659         /* TODO ue will trigger an interrupt.
 660          *
 661          * When “Full RAS” is enabled, the per-IP interrupt sources should
 662          * be disabled and the driver should only look for the aggregated
 663          * interrupt via sync flood
 664          */
 665         if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX)) {
 666                 kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
 667                 if (adev->gfx.funcs->query_ras_error_count)
 668                         adev->gfx.funcs->query_ras_error_count(adev, err_data);
 669                 amdgpu_ras_reset_gpu(adev);
 670         }
 671         return AMDGPU_RAS_SUCCESS;
 672 }
 673
 674 int amdgpu_gfx_cp_ecc_error_irq(struct amdgpu_device *adev,
 675                                   struct amdgpu_irq_src *source,
 676                                   struct amdgpu_iv_entry *entry)
 677 {
 678         struct ras_common_if *ras_if = adev->gfx.ras_if;
 679         struct ras_dispatch_if ih_data = {
 680                 .entry = entry,
 681         };
 682
 683         if (!ras_if)
 684                 return 0;
 685
 686         ih_data.head = *ras_if;
 687
 688         DRM_ERROR("CP ECC ERROR IRQ\n");
 689         amdgpu_ras_interrupt_dispatch(adev, &ih_data);
 690         return 0;
 691 }
 692
 693 uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg)
 694 {
 695         signed long r, cnt = 0;
 696         unsigned long flags;
 697         uint32_t seq, reg_val_offs = 0, value = 0;
 698         struct amdgpu_kiq *kiq = &adev->gfx.kiq;
 699         struct amdgpu_ring *ring = &kiq->ring;
 700
 701         if (adev->in_pci_err_recovery)
 702                 return 0;
 703
 704         BUG_ON(!ring->funcs->emit_rreg);
 705
 706         spin_lock_irqsave(&kiq->ring_lock, flags);
 707         if (amdgpu_device_wb_get(adev, &reg_val_offs)) {
 708                 pr_err("critical bug! too many kiq readers\n");
 709                 goto failed_unlock;
 710         }
 711         amdgpu_ring_alloc(ring, 32);
 712         amdgpu_ring_emit_rreg(ring, reg, reg_val_offs);
 713         r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
 714         if (r)
 715                 goto failed_undo;
 716
 717         amdgpu_ring_commit(ring);
 718         spin_unlock_irqrestore(&kiq->ring_lock, flags);
 719
 720         r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
 721
 722         /* don't wait anymore for gpu reset case because this way may
 723          * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
 724          * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
 725          * never return if we keep waiting in virt_kiq_rreg, which cause
 726          * gpu_recover() hang there.
 727          *
 728          * also don't wait anymore for IRQ context
 729          * */
 730         if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
 731                 goto failed_kiq_read;
 732
 733         might_sleep();
 734         while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
 735                 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
 736                 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
 737         }
 738
 739         if (cnt > MAX_KIQ_REG_TRY)
 740                 goto failed_kiq_read;
 741
 742         mb();
 743         value = adev->wb.wb[reg_val_offs];
 744         amdgpu_device_wb_free(adev, reg_val_offs);
 745         return value;
 746
 747 failed_undo:
 748         amdgpu_ring_undo(ring);
 749 failed_unlock:
 750         spin_unlock_irqrestore(&kiq->ring_lock, flags);
 751 failed_kiq_read:
 752         if (reg_val_offs)
 753                 amdgpu_device_wb_free(adev, reg_val_offs);
 754         dev_err(adev->dev, "failed to read reg:%x\n", reg);
 755         return ~0;
 756 }
 757
 758 void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
 759 {
 760         signed long r, cnt = 0;
 761         unsigned long flags;
 762         uint32_t seq;
 763         struct amdgpu_kiq *kiq = &adev->gfx.kiq;
 764         struct amdgpu_ring *ring = &kiq->ring;
 765
 766         BUG_ON(!ring->funcs->emit_wreg);
 767
 768         if (adev->in_pci_err_recovery)
 769                 return;
 770
 771         spin_lock_irqsave(&kiq->ring_lock, flags);
 772         amdgpu_ring_alloc(ring, 32);
 773         amdgpu_ring_emit_wreg(ring, reg, v);
 774         r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
 775         if (r)
 776                 goto failed_undo;
 777
 778         amdgpu_ring_commit(ring);
 779         spin_unlock_irqrestore(&kiq->ring_lock, flags);
 780
 781         r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
 782
 783         /* don't wait anymore for gpu reset case because this way may
 784          * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
 785          * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
 786          * never return if we keep waiting in virt_kiq_rreg, which cause
 787          * gpu_recover() hang there.
 788          *
 789          * also don't wait anymore for IRQ context
 790          * */
 791         if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
 792                 goto failed_kiq_write;
 793
 794         might_sleep();
 795         while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
 796
 797                 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
 798                 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
 799         }
 800
 801         if (cnt > MAX_KIQ_REG_TRY)
 802                 goto failed_kiq_write;
 803
 804         return;
 805
 806 failed_undo:
 807         amdgpu_ring_undo(ring);
 808         spin_unlock_irqrestore(&kiq->ring_lock, flags);
 809 failed_kiq_write:
 810         dev_err(adev->dev, "failed to write reg:%x\n", reg);
 811 }
 812
 813 int amdgpu_gfx_get_num_kcq(struct amdgpu_device *adev)
 814 {
 815         if (amdgpu_num_kcq == -1) {
 816                 return 8;
 817         } else if (amdgpu_num_kcq > 8 || amdgpu_num_kcq < 0) {
 818                 dev_warn(adev->dev, "set kernel compute queue number to 8 due to invalid parameter provided by user\n");
 819                 return 8;
 820         }
 821         return amdgpu_num_kcq;
 822 }
 823
 824 /* amdgpu_gfx_state_change_set - Handle gfx power state change set
 825  * @adev: amdgpu_device pointer
 826  * @state: gfx power state(1 -sGpuChangeState_D0Entry and 2 -sGpuChangeState_D3Entry)
 827  *
 828  */
 829
 830 void amdgpu_gfx_state_change_set(struct amdgpu_device *adev, enum gfx_change_state state)
 831 {
 832         if (is_support_sw_smu(adev)) {
 833                 smu_gfx_state_change_set(&adev->smu, state);
 834         } else {
 835                 mutex_lock(&adev->pm.mutex);
 836                 if (adev->powerplay.pp_funcs &&
 837                     adev->powerplay.pp_funcs->gfx_state_change_set)
 838                         ((adev)->powerplay.pp_funcs->gfx_state_change_set(
 839                                 (adev)->powerplay.pp_handle, state));
 840                 mutex_unlock(&adev->pm.mutex);
 841         }
 842 }