drivers/gpu/drm/amd/amdkfd/kfd_device.c

   1 /*
   2  * Copyright 2014 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  */
  22
  23 #include <linux/amd-iommu.h>
  24 #include <linux/bsearch.h>
  25 #include <linux/pci.h>
  26 #include <linux/slab.h>
  27 #include "kfd_priv.h"
  28 #include "kfd_device_queue_manager.h"
  29 #include "kfd_pm4_headers.h"
  30
  31 #define MQD_SIZE_ALIGNED 768
  32
  33 static const struct kfd_device_info kaveri_device_info = {
  34         .asic_family = CHIP_KAVERI,
  35         .max_pasid_bits = 16,
  36         .ih_ring_entry_size = 4 * sizeof(uint32_t),
  37         .mqd_size_aligned = MQD_SIZE_ALIGNED
  38 };
  39
  40 static const struct kfd_device_info carrizo_device_info = {
  41         .asic_family = CHIP_CARRIZO,
  42         .max_pasid_bits = 16,
  43         .ih_ring_entry_size = 4 * sizeof(uint32_t),
  44         .num_of_watch_points = 4,
  45         .mqd_size_aligned = MQD_SIZE_ALIGNED
  46 };
  47
  48 struct kfd_deviceid {
  49         unsigned short did;
  50         const struct kfd_device_info *device_info;
  51 };
  52
  53 /* Please keep this sorted by increasing device id. */
  54 static const struct kfd_deviceid supported_devices[] = {
  55         { 0x1304, &kaveri_device_info },        /* Kaveri */
  56         { 0x1305, &kaveri_device_info },        /* Kaveri */
  57         { 0x1306, &kaveri_device_info },        /* Kaveri */
  58         { 0x1307, &kaveri_device_info },        /* Kaveri */
  59         { 0x1309, &kaveri_device_info },        /* Kaveri */
  60         { 0x130A, &kaveri_device_info },        /* Kaveri */
  61         { 0x130B, &kaveri_device_info },        /* Kaveri */
  62         { 0x130C, &kaveri_device_info },        /* Kaveri */
  63         { 0x130D, &kaveri_device_info },        /* Kaveri */
  64         { 0x130E, &kaveri_device_info },        /* Kaveri */
  65         { 0x130F, &kaveri_device_info },        /* Kaveri */
  66         { 0x1310, &kaveri_device_info },        /* Kaveri */
  67         { 0x1311, &kaveri_device_info },        /* Kaveri */
  68         { 0x1312, &kaveri_device_info },        /* Kaveri */
  69         { 0x1313, &kaveri_device_info },        /* Kaveri */
  70         { 0x1315, &kaveri_device_info },        /* Kaveri */
  71         { 0x1316, &kaveri_device_info },        /* Kaveri */
  72         { 0x1317, &kaveri_device_info },        /* Kaveri */
  73         { 0x1318, &kaveri_device_info },        /* Kaveri */
  74         { 0x131B, &kaveri_device_info },        /* Kaveri */
  75         { 0x131C, &kaveri_device_info },        /* Kaveri */
  76         { 0x131D, &kaveri_device_info }         /* Kaveri */
  77 };
  78
  79 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
  80                                 unsigned int chunk_size);
  81 static void kfd_gtt_sa_fini(struct kfd_dev *kfd);
  82
  83 static const struct kfd_device_info *lookup_device_info(unsigned short did)
  84 {
  85         size_t i;
  86
  87         for (i = 0; i < ARRAY_SIZE(supported_devices); i++) {
  88                 if (supported_devices[i].did == did) {
  89                         BUG_ON(supported_devices[i].device_info == NULL);
  90                         return supported_devices[i].device_info;
  91                 }
  92         }
  93
  94         return NULL;
  95 }
  96
  97 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev)
  98 {
  99         struct kfd_dev *kfd;
 100
 101         const struct kfd_device_info *device_info =
 102                                         lookup_device_info(pdev->device);
 103
 104         if (!device_info)
 105                 return NULL;
 106
 107         kfd = kzalloc(sizeof(*kfd), GFP_KERNEL);
 108         if (!kfd)
 109                 return NULL;
 110
 111         kfd->kgd = kgd;
 112         kfd->device_info = device_info;
 113         kfd->pdev = pdev;
 114         kfd->init_complete = false;
 115
 116         return kfd;
 117 }
 118
 119 static bool device_iommu_pasid_init(struct kfd_dev *kfd)
 120 {
 121         const u32 required_iommu_flags = AMD_IOMMU_DEVICE_FLAG_ATS_SUP |
 122                                         AMD_IOMMU_DEVICE_FLAG_PRI_SUP |
 123                                         AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
 124
 125         struct amd_iommu_device_info iommu_info;
 126         unsigned int pasid_limit;
 127         int err;
 128
 129         err = amd_iommu_device_info(kfd->pdev, &iommu_info);
 130         if (err < 0) {
 131                 dev_err(kfd_device,
 132                         "error getting iommu info. is the iommu enabled?\n");
 133                 return false;
 134         }
 135
 136         if ((iommu_info.flags & required_iommu_flags) != required_iommu_flags) {
 137                 dev_err(kfd_device, "error required iommu flags ats(%i), pri(%i), pasid(%i)\n",
 138                        (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_ATS_SUP) != 0,
 139                        (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_PRI_SUP) != 0,
 140                        (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_PASID_SUP) != 0);
 141                 return false;
 142         }
 143
 144         pasid_limit = min_t(unsigned int,
 145                         (unsigned int)1 << kfd->device_info->max_pasid_bits,
 146                         iommu_info.max_pasids);
 147         /*
 148          * last pasid is used for kernel queues doorbells
 149          * in the future the last pasid might be used for a kernel thread.
 150          */
 151         pasid_limit = min_t(unsigned int,
 152                                 pasid_limit,
 153                                 kfd->doorbell_process_limit - 1);
 154
 155         err = amd_iommu_init_device(kfd->pdev, pasid_limit);
 156         if (err < 0) {
 157                 dev_err(kfd_device, "error initializing iommu device\n");
 158                 return false;
 159         }
 160
 161         if (!kfd_set_pasid_limit(pasid_limit)) {
 162                 dev_err(kfd_device, "error setting pasid limit\n");
 163                 amd_iommu_free_device(kfd->pdev);
 164                 return false;
 165         }
 166
 167         return true;
 168 }
 169
 170 static void iommu_pasid_shutdown_callback(struct pci_dev *pdev, int pasid)
 171 {
 172         struct kfd_dev *dev = kfd_device_by_pci_dev(pdev);
 173
 174         if (dev)
 175                 kfd_unbind_process_from_device(dev, pasid);
 176 }
 177
 178 bool kgd2kfd_device_init(struct kfd_dev *kfd,
 179                          const struct kgd2kfd_shared_resources *gpu_resources)
 180 {
 181         unsigned int size;
 182
 183         kfd->shared_resources = *gpu_resources;
 184
 185         /* calculate max size of mqds needed for queues */
 186         size = max_num_of_processes *
 187                 max_num_of_queues_per_process *
 188                 kfd->device_info->mqd_size_aligned;
 189
 190         /*
 191          * calculate max size of runlist packet.
 192          * There can be only 2 packets at once
 193          */
 194         size += (max_num_of_processes * sizeof(struct pm4_map_process) +
 195                 max_num_of_processes * max_num_of_queues_per_process *
 196                 sizeof(struct pm4_map_queues) + sizeof(struct pm4_runlist)) * 2;
 197
 198         /* Add size of HIQ & DIQ */
 199         size += KFD_KERNEL_QUEUE_SIZE * 2;
 200
 201         /* add another 512KB for all other allocations on gart (HPD, fences) */
 202         size += 512 * 1024;
 203
 204         if (kfd2kgd->init_gtt_mem_allocation(kfd->kgd, size, &kfd->gtt_mem,
 205                         &kfd->gtt_start_gpu_addr, &kfd->gtt_start_cpu_ptr)) {
 206                 dev_err(kfd_device,
 207                         "Could not allocate %d bytes for device (%x:%x)\n",
 208                         size, kfd->pdev->vendor, kfd->pdev->device);
 209                 goto out;
 210         }
 211
 212         dev_info(kfd_device,
 213                 "Allocated %d bytes on gart for device(%x:%x)\n",
 214                 size, kfd->pdev->vendor, kfd->pdev->device);
 215
 216         /* Initialize GTT sa with 512 byte chunk size */
 217         if (kfd_gtt_sa_init(kfd, size, 512) != 0) {
 218                 dev_err(kfd_device,
 219                         "Error initializing gtt sub-allocator\n");
 220                 goto kfd_gtt_sa_init_error;
 221         }
 222
 223         kfd_doorbell_init(kfd);
 224
 225         if (kfd_topology_add_device(kfd) != 0) {
 226                 dev_err(kfd_device,
 227                         "Error adding device (%x:%x) to topology\n",
 228                         kfd->pdev->vendor, kfd->pdev->device);
 229                 goto kfd_topology_add_device_error;
 230         }
 231
 232         if (kfd_interrupt_init(kfd)) {
 233                 dev_err(kfd_device,
 234                         "Error initializing interrupts for device (%x:%x)\n",
 235                         kfd->pdev->vendor, kfd->pdev->device);
 236                 goto kfd_interrupt_error;
 237         }
 238
 239         if (!device_iommu_pasid_init(kfd)) {
 240                 dev_err(kfd_device,
 241                         "Error initializing iommuv2 for device (%x:%x)\n",
 242                         kfd->pdev->vendor, kfd->pdev->device);
 243                 goto device_iommu_pasid_error;
 244         }
 245         amd_iommu_set_invalidate_ctx_cb(kfd->pdev,
 246                                                 iommu_pasid_shutdown_callback);
 247
 248         kfd->dqm = device_queue_manager_init(kfd);
 249         if (!kfd->dqm) {
 250                 dev_err(kfd_device,
 251                         "Error initializing queue manager for device (%x:%x)\n",
 252                         kfd->pdev->vendor, kfd->pdev->device);
 253                 goto device_queue_manager_error;
 254         }
 255
 256         if (kfd->dqm->ops.start(kfd->dqm) != 0) {
 257                 dev_err(kfd_device,
 258                         "Error starting queuen manager for device (%x:%x)\n",
 259                         kfd->pdev->vendor, kfd->pdev->device);
 260                 goto dqm_start_error;
 261         }
 262
 263         kfd->init_complete = true;
 264         dev_info(kfd_device, "added device (%x:%x)\n", kfd->pdev->vendor,
 265                  kfd->pdev->device);
 266
 267         pr_debug("kfd: Starting kfd with the following scheduling policy %d\n",
 268                 sched_policy);
 269
 270         goto out;
 271
 272 dqm_start_error:
 273         device_queue_manager_uninit(kfd->dqm);
 274 device_queue_manager_error:
 275         amd_iommu_free_device(kfd->pdev);
 276 device_iommu_pasid_error:
 277         kfd_interrupt_exit(kfd);
 278 kfd_interrupt_error:
 279         kfd_topology_remove_device(kfd);
 280 kfd_topology_add_device_error:
 281         kfd_gtt_sa_fini(kfd);
 282 kfd_gtt_sa_init_error:
 283         kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem);
 284         dev_err(kfd_device,
 285                 "device (%x:%x) NOT added due to errors\n",
 286                 kfd->pdev->vendor, kfd->pdev->device);
 287 out:
 288         return kfd->init_complete;
 289 }
 290
 291 void kgd2kfd_device_exit(struct kfd_dev *kfd)
 292 {
 293         if (kfd->init_complete) {
 294                 device_queue_manager_uninit(kfd->dqm);
 295                 amd_iommu_free_device(kfd->pdev);
 296                 kfd_interrupt_exit(kfd);
 297                 kfd_topology_remove_device(kfd);
 298                 kfd_gtt_sa_fini(kfd);
 299                 kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem);
 300         }
 301
 302         kfree(kfd);
 303 }
 304
 305 void kgd2kfd_suspend(struct kfd_dev *kfd)
 306 {
 307         BUG_ON(kfd == NULL);
 308
 309         if (kfd->init_complete) {
 310                 kfd->dqm->ops.stop(kfd->dqm);
 311                 amd_iommu_set_invalidate_ctx_cb(kfd->pdev, NULL);
 312                 amd_iommu_free_device(kfd->pdev);
 313         }
 314 }
 315
 316 int kgd2kfd_resume(struct kfd_dev *kfd)
 317 {
 318         unsigned int pasid_limit;
 319         int err;
 320
 321         BUG_ON(kfd == NULL);
 322
 323         pasid_limit = kfd_get_pasid_limit();
 324
 325         if (kfd->init_complete) {
 326                 err = amd_iommu_init_device(kfd->pdev, pasid_limit);
 327                 if (err < 0)
 328                         return -ENXIO;
 329                 amd_iommu_set_invalidate_ctx_cb(kfd->pdev,
 330                                                 iommu_pasid_shutdown_callback);
 331                 kfd->dqm->ops.start(kfd->dqm);
 332         }
 333
 334         return 0;
 335 }
 336
 337 /* This is called directly from KGD at ISR. */
 338 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
 339 {
 340         if (kfd->init_complete) {
 341                 spin_lock(&kfd->interrupt_lock);
 342
 343                 if (kfd->interrupts_active
 344                     && enqueue_ih_ring_entry(kfd, ih_ring_entry))
 345                         schedule_work(&kfd->interrupt_work);
 346
 347                 spin_unlock(&kfd->interrupt_lock);
 348         }
 349 }
 350
 351 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
 352                                 unsigned int chunk_size)
 353 {
 354         unsigned int num_of_bits;
 355
 356         BUG_ON(!kfd);
 357         BUG_ON(!kfd->gtt_mem);
 358         BUG_ON(buf_size < chunk_size);
 359         BUG_ON(buf_size == 0);
 360         BUG_ON(chunk_size == 0);
 361
 362         kfd->gtt_sa_chunk_size = chunk_size;
 363         kfd->gtt_sa_num_of_chunks = buf_size / chunk_size;
 364
 365         num_of_bits = kfd->gtt_sa_num_of_chunks / BITS_PER_BYTE;
 366         BUG_ON(num_of_bits == 0);
 367
 368         kfd->gtt_sa_bitmap = kzalloc(num_of_bits, GFP_KERNEL);
 369
 370         if (!kfd->gtt_sa_bitmap)
 371                 return -ENOMEM;
 372
 373         pr_debug("kfd: gtt_sa_num_of_chunks = %d, gtt_sa_bitmap = %p\n",
 374                         kfd->gtt_sa_num_of_chunks, kfd->gtt_sa_bitmap);
 375
 376         mutex_init(&kfd->gtt_sa_lock);
 377
 378         return 0;
 379
 380 }
 381
 382 static void kfd_gtt_sa_fini(struct kfd_dev *kfd)
 383 {
 384         mutex_destroy(&kfd->gtt_sa_lock);
 385         kfree(kfd->gtt_sa_bitmap);
 386 }
 387
 388 static inline uint64_t kfd_gtt_sa_calc_gpu_addr(uint64_t start_addr,
 389                                                 unsigned int bit_num,
 390                                                 unsigned int chunk_size)
 391 {
 392         return start_addr + bit_num * chunk_size;
 393 }
 394
 395 static inline uint32_t *kfd_gtt_sa_calc_cpu_addr(void *start_addr,
 396                                                 unsigned int bit_num,
 397                                                 unsigned int chunk_size)
 398 {
 399         return (uint32_t *) ((uint64_t) start_addr + bit_num * chunk_size);
 400 }
 401
 402 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
 403                         struct kfd_mem_obj **mem_obj)
 404 {
 405         unsigned int found, start_search, cur_size;
 406
 407         BUG_ON(!kfd);
 408
 409         if (size == 0)
 410                 return -EINVAL;
 411
 412         if (size > kfd->gtt_sa_num_of_chunks * kfd->gtt_sa_chunk_size)
 413                 return -ENOMEM;
 414
 415         *mem_obj = kmalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
 416         if ((*mem_obj) == NULL)
 417                 return -ENOMEM;
 418
 419         pr_debug("kfd: allocated mem_obj = %p for size = %d\n", *mem_obj, size);
 420
 421         start_search = 0;
 422
 423         mutex_lock(&kfd->gtt_sa_lock);
 424
 425 kfd_gtt_restart_search:
 426         /* Find the first chunk that is free */
 427         found = find_next_zero_bit(kfd->gtt_sa_bitmap,
 428                                         kfd->gtt_sa_num_of_chunks,
 429                                         start_search);
 430
 431         pr_debug("kfd: found = %d\n", found);
 432
 433         /* If there wasn't any free chunk, bail out */
 434         if (found == kfd->gtt_sa_num_of_chunks)
 435                 goto kfd_gtt_no_free_chunk;
 436
 437         /* Update fields of mem_obj */
 438         (*mem_obj)->range_start = found;
 439         (*mem_obj)->range_end = found;
 440         (*mem_obj)->gpu_addr = kfd_gtt_sa_calc_gpu_addr(
 441                                         kfd->gtt_start_gpu_addr,
 442                                         found,
 443                                         kfd->gtt_sa_chunk_size);
 444         (*mem_obj)->cpu_ptr = kfd_gtt_sa_calc_cpu_addr(
 445                                         kfd->gtt_start_cpu_ptr,
 446                                         found,
 447                                         kfd->gtt_sa_chunk_size);
 448
 449         pr_debug("kfd: gpu_addr = %p, cpu_addr = %p\n",
 450                         (uint64_t *) (*mem_obj)->gpu_addr, (*mem_obj)->cpu_ptr);
 451
 452         /* If we need only one chunk, mark it as allocated and get out */
 453         if (size <= kfd->gtt_sa_chunk_size) {
 454                 pr_debug("kfd: single bit\n");
 455                 set_bit(found, kfd->gtt_sa_bitmap);
 456                 goto kfd_gtt_out;
 457         }
 458
 459         /* Otherwise, try to see if we have enough contiguous chunks */
 460         cur_size = size - kfd->gtt_sa_chunk_size;
 461         do {
 462                 (*mem_obj)->range_end =
 463                         find_next_zero_bit(kfd->gtt_sa_bitmap,
 464                                         kfd->gtt_sa_num_of_chunks, ++found);
 465                 /*
 466                  * If next free chunk is not contiguous than we need to
 467                  * restart our search from the last free chunk we found (which
 468                  * wasn't contiguous to the previous ones
 469                  */
 470                 if ((*mem_obj)->range_end != found) {
 471                         start_search = found;
 472                         goto kfd_gtt_restart_search;
 473                 }
 474
 475                 /*
 476                  * If we reached end of buffer, bail out with error
 477                  */
 478                 if (found == kfd->gtt_sa_num_of_chunks)
 479                         goto kfd_gtt_no_free_chunk;
 480
 481                 /* Check if we don't need another chunk */
 482                 if (cur_size <= kfd->gtt_sa_chunk_size)
 483                         cur_size = 0;
 484                 else
 485                         cur_size -= kfd->gtt_sa_chunk_size;
 486
 487         } while (cur_size > 0);
 488
 489         pr_debug("kfd: range_start = %d, range_end = %d\n",
 490                 (*mem_obj)->range_start, (*mem_obj)->range_end);
 491
 492         /* Mark the chunks as allocated */
 493         for (found = (*mem_obj)->range_start;
 494                 found <= (*mem_obj)->range_end;
 495                 found++)
 496                 set_bit(found, kfd->gtt_sa_bitmap);
 497
 498 kfd_gtt_out:
 499         mutex_unlock(&kfd->gtt_sa_lock);
 500         return 0;
 501
 502 kfd_gtt_no_free_chunk:
 503         pr_debug("kfd: allocation failed with mem_obj = %p\n", mem_obj);
 504         mutex_unlock(&kfd->gtt_sa_lock);
 505         kfree(mem_obj);
 506         return -ENOMEM;
 507 }
 508
 509 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj)
 510 {
 511         unsigned int bit;
 512
 513         BUG_ON(!kfd);
 514
 515         /* Act like kfree when trying to free a NULL object */
 516         if (!mem_obj)
 517                 return 0;
 518
 519         pr_debug("kfd: free mem_obj = %p, range_start = %d, range_end = %d\n",
 520                         mem_obj, mem_obj->range_start, mem_obj->range_end);
 521
 522         mutex_lock(&kfd->gtt_sa_lock);
 523
 524         /* Mark the chunks as free */
 525         for (bit = mem_obj->range_start;
 526                 bit <= mem_obj->range_end;
 527                 bit++)
 528                 clear_bit(bit, kfd->gtt_sa_bitmap);
 529
 530         mutex_unlock(&kfd->gtt_sa_lock);
 531
 532         kfree(mem_obj);
 533         return 0;
 534 }