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 /* max num of queues for KV.TODO should be a dynamic value */ 37 .max_no_of_hqd = 24, 38 .ih_ring_entry_size = 4 * sizeof(uint32_t), 39 .event_interrupt_class = &event_interrupt_class_cik, 40 .num_of_watch_points = 4, 41 .mqd_size_aligned = MQD_SIZE_ALIGNED 42 }; 43 44 static const struct kfd_device_info carrizo_device_info = { 45 .asic_family = CHIP_CARRIZO, 46 .max_pasid_bits = 16, 47 .ih_ring_entry_size = 4 * sizeof(uint32_t), 48 .num_of_watch_points = 4, 49 .mqd_size_aligned = MQD_SIZE_ALIGNED 50 }; 51 52 struct kfd_deviceid { 53 unsigned short did; 54 const struct kfd_device_info *device_info; 55 }; 56 57 /* Please keep this sorted by increasing device id. */ 58 static const struct kfd_deviceid supported_devices[] = { 59 { 0x1304, &kaveri_device_info }, /* Kaveri */ 60 { 0x1305, &kaveri_device_info }, /* Kaveri */ 61 { 0x1306, &kaveri_device_info }, /* Kaveri */ 62 { 0x1307, &kaveri_device_info }, /* Kaveri */ 63 { 0x1309, &kaveri_device_info }, /* Kaveri */ 64 { 0x130A, &kaveri_device_info }, /* Kaveri */ 65 { 0x130B, &kaveri_device_info }, /* Kaveri */ 66 { 0x130C, &kaveri_device_info }, /* Kaveri */ 67 { 0x130D, &kaveri_device_info }, /* Kaveri */ 68 { 0x130E, &kaveri_device_info }, /* Kaveri */ 69 { 0x130F, &kaveri_device_info }, /* Kaveri */ 70 { 0x1310, &kaveri_device_info }, /* Kaveri */ 71 { 0x1311, &kaveri_device_info }, /* Kaveri */ 72 { 0x1312, &kaveri_device_info }, /* Kaveri */ 73 { 0x1313, &kaveri_device_info }, /* Kaveri */ 74 { 0x1315, &kaveri_device_info }, /* Kaveri */ 75 { 0x1316, &kaveri_device_info }, /* Kaveri */ 76 { 0x1317, &kaveri_device_info }, /* Kaveri */ 77 { 0x1318, &kaveri_device_info }, /* Kaveri */ 78 { 0x131B, &kaveri_device_info }, /* Kaveri */ 79 { 0x131C, &kaveri_device_info }, /* Kaveri */ 80 { 0x131D, &kaveri_device_info } /* Kaveri */ 81 }; 82 83 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size, 84 unsigned int chunk_size); 85 static void kfd_gtt_sa_fini(struct kfd_dev *kfd); 86 87 static const struct kfd_device_info *lookup_device_info(unsigned short did) 88 { 89 size_t i; 90 91 for (i = 0; i < ARRAY_SIZE(supported_devices); i++) { 92 if (supported_devices[i].did == did) { 93 BUG_ON(supported_devices[i].device_info == NULL); 94 return supported_devices[i].device_info; 95 } 96 } 97 98 return NULL; 99 } 100 101 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, 102 struct pci_dev *pdev, const struct kfd2kgd_calls *f2g) 103 { 104 struct kfd_dev *kfd; 105 106 const struct kfd_device_info *device_info = 107 lookup_device_info(pdev->device); 108 109 if (!device_info) 110 return NULL; 111 112 kfd = kzalloc(sizeof(*kfd), GFP_KERNEL); 113 if (!kfd) 114 return NULL; 115 116 kfd->kgd = kgd; 117 kfd->device_info = device_info; 118 kfd->pdev = pdev; 119 kfd->init_complete = false; 120 kfd->kfd2kgd = f2g; 121 122 mutex_init(&kfd->doorbell_mutex); 123 memset(&kfd->doorbell_available_index, 0, 124 sizeof(kfd->doorbell_available_index)); 125 126 return kfd; 127 } 128 129 static bool device_iommu_pasid_init(struct kfd_dev *kfd) 130 { 131 const u32 required_iommu_flags = AMD_IOMMU_DEVICE_FLAG_ATS_SUP | 132 AMD_IOMMU_DEVICE_FLAG_PRI_SUP | 133 AMD_IOMMU_DEVICE_FLAG_PASID_SUP; 134 135 struct amd_iommu_device_info iommu_info; 136 unsigned int pasid_limit; 137 int err; 138 139 err = amd_iommu_device_info(kfd->pdev, &iommu_info); 140 if (err < 0) { 141 dev_err(kfd_device, 142 "error getting iommu info. is the iommu enabled?\n"); 143 return false; 144 } 145 146 if ((iommu_info.flags & required_iommu_flags) != required_iommu_flags) { 147 dev_err(kfd_device, "error required iommu flags ats(%i), pri(%i), pasid(%i)\n", 148 (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_ATS_SUP) != 0, 149 (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_PRI_SUP) != 0, 150 (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_PASID_SUP) != 0); 151 return false; 152 } 153 154 pasid_limit = min_t(unsigned int, 155 (unsigned int)1 << kfd->device_info->max_pasid_bits, 156 iommu_info.max_pasids); 157 /* 158 * last pasid is used for kernel queues doorbells 159 * in the future the last pasid might be used for a kernel thread. 160 */ 161 pasid_limit = min_t(unsigned int, 162 pasid_limit, 163 kfd->doorbell_process_limit - 1); 164 165 err = amd_iommu_init_device(kfd->pdev, pasid_limit); 166 if (err < 0) { 167 dev_err(kfd_device, "error initializing iommu device\n"); 168 return false; 169 } 170 171 if (!kfd_set_pasid_limit(pasid_limit)) { 172 dev_err(kfd_device, "error setting pasid limit\n"); 173 amd_iommu_free_device(kfd->pdev); 174 return false; 175 } 176 177 return true; 178 } 179 180 static void iommu_pasid_shutdown_callback(struct pci_dev *pdev, int pasid) 181 { 182 struct kfd_dev *dev = kfd_device_by_pci_dev(pdev); 183 184 if (dev) 185 kfd_unbind_process_from_device(dev, pasid); 186 } 187 188 /* 189 * This function called by IOMMU driver on PPR failure 190 */ 191 static int iommu_invalid_ppr_cb(struct pci_dev *pdev, int pasid, 192 unsigned long address, u16 flags) 193 { 194 struct kfd_dev *dev; 195 196 dev_warn(kfd_device, 197 "Invalid PPR device %x:%x.%x pasid %d address 0x%lX flags 0x%X", 198 PCI_BUS_NUM(pdev->devfn), 199 PCI_SLOT(pdev->devfn), 200 PCI_FUNC(pdev->devfn), 201 pasid, 202 address, 203 flags); 204 205 dev = kfd_device_by_pci_dev(pdev); 206 BUG_ON(dev == NULL); 207 208 kfd_signal_iommu_event(dev, pasid, address, 209 flags & PPR_FAULT_WRITE, flags & PPR_FAULT_EXEC); 210 211 return AMD_IOMMU_INV_PRI_RSP_INVALID; 212 } 213 214 bool kgd2kfd_device_init(struct kfd_dev *kfd, 215 const struct kgd2kfd_shared_resources *gpu_resources) 216 { 217 unsigned int size; 218 219 kfd->shared_resources = *gpu_resources; 220 221 /* calculate max size of mqds needed for queues */ 222 size = max_num_of_queues_per_device * 223 kfd->device_info->mqd_size_aligned; 224 225 /* 226 * calculate max size of runlist packet. 227 * There can be only 2 packets at once 228 */ 229 size += (KFD_MAX_NUM_OF_PROCESSES * sizeof(struct pm4_map_process) + 230 max_num_of_queues_per_device * 231 sizeof(struct pm4_map_queues) + sizeof(struct pm4_runlist)) * 2; 232 233 /* Add size of HIQ & DIQ */ 234 size += KFD_KERNEL_QUEUE_SIZE * 2; 235 236 /* add another 512KB for all other allocations on gart (HPD, fences) */ 237 size += 512 * 1024; 238 239 if (kfd->kfd2kgd->init_gtt_mem_allocation( 240 kfd->kgd, size, &kfd->gtt_mem, 241 &kfd->gtt_start_gpu_addr, &kfd->gtt_start_cpu_ptr)){ 242 dev_err(kfd_device, 243 "Could not allocate %d bytes for device (%x:%x)\n", 244 size, kfd->pdev->vendor, kfd->pdev->device); 245 goto out; 246 } 247 248 dev_info(kfd_device, 249 "Allocated %d bytes on gart for device(%x:%x)\n", 250 size, kfd->pdev->vendor, kfd->pdev->device); 251 252 /* Initialize GTT sa with 512 byte chunk size */ 253 if (kfd_gtt_sa_init(kfd, size, 512) != 0) { 254 dev_err(kfd_device, 255 "Error initializing gtt sub-allocator\n"); 256 goto kfd_gtt_sa_init_error; 257 } 258 259 kfd_doorbell_init(kfd); 260 261 if (kfd_topology_add_device(kfd) != 0) { 262 dev_err(kfd_device, 263 "Error adding device (%x:%x) to topology\n", 264 kfd->pdev->vendor, kfd->pdev->device); 265 goto kfd_topology_add_device_error; 266 } 267 268 if (kfd_interrupt_init(kfd)) { 269 dev_err(kfd_device, 270 "Error initializing interrupts for device (%x:%x)\n", 271 kfd->pdev->vendor, kfd->pdev->device); 272 goto kfd_interrupt_error; 273 } 274 275 if (!device_iommu_pasid_init(kfd)) { 276 dev_err(kfd_device, 277 "Error initializing iommuv2 for device (%x:%x)\n", 278 kfd->pdev->vendor, kfd->pdev->device); 279 goto device_iommu_pasid_error; 280 } 281 amd_iommu_set_invalidate_ctx_cb(kfd->pdev, 282 iommu_pasid_shutdown_callback); 283 amd_iommu_set_invalid_ppr_cb(kfd->pdev, iommu_invalid_ppr_cb); 284 285 kfd->dqm = device_queue_manager_init(kfd); 286 if (!kfd->dqm) { 287 dev_err(kfd_device, 288 "Error initializing queue manager for device (%x:%x)\n", 289 kfd->pdev->vendor, kfd->pdev->device); 290 goto device_queue_manager_error; 291 } 292 293 if (kfd->dqm->ops.start(kfd->dqm) != 0) { 294 dev_err(kfd_device, 295 "Error starting queuen manager for device (%x:%x)\n", 296 kfd->pdev->vendor, kfd->pdev->device); 297 goto dqm_start_error; 298 } 299 300 kfd->dbgmgr = NULL; 301 302 kfd->init_complete = true; 303 dev_info(kfd_device, "added device (%x:%x)\n", kfd->pdev->vendor, 304 kfd->pdev->device); 305 306 pr_debug("kfd: Starting kfd with the following scheduling policy %d\n", 307 sched_policy); 308 309 goto out; 310 311 dqm_start_error: 312 device_queue_manager_uninit(kfd->dqm); 313 device_queue_manager_error: 314 amd_iommu_free_device(kfd->pdev); 315 device_iommu_pasid_error: 316 kfd_interrupt_exit(kfd); 317 kfd_interrupt_error: 318 kfd_topology_remove_device(kfd); 319 kfd_topology_add_device_error: 320 kfd_gtt_sa_fini(kfd); 321 kfd_gtt_sa_init_error: 322 kfd->kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem); 323 dev_err(kfd_device, 324 "device (%x:%x) NOT added due to errors\n", 325 kfd->pdev->vendor, kfd->pdev->device); 326 out: 327 return kfd->init_complete; 328 } 329 330 void kgd2kfd_device_exit(struct kfd_dev *kfd) 331 { 332 if (kfd->init_complete) { 333 device_queue_manager_uninit(kfd->dqm); 334 amd_iommu_free_device(kfd->pdev); 335 kfd_interrupt_exit(kfd); 336 kfd_topology_remove_device(kfd); 337 kfd_gtt_sa_fini(kfd); 338 kfd->kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem); 339 } 340 341 kfree(kfd); 342 } 343 344 void kgd2kfd_suspend(struct kfd_dev *kfd) 345 { 346 BUG_ON(kfd == NULL); 347 348 if (kfd->init_complete) { 349 kfd->dqm->ops.stop(kfd->dqm); 350 amd_iommu_set_invalidate_ctx_cb(kfd->pdev, NULL); 351 amd_iommu_set_invalid_ppr_cb(kfd->pdev, NULL); 352 amd_iommu_free_device(kfd->pdev); 353 } 354 } 355 356 int kgd2kfd_resume(struct kfd_dev *kfd) 357 { 358 unsigned int pasid_limit; 359 int err; 360 361 BUG_ON(kfd == NULL); 362 363 pasid_limit = kfd_get_pasid_limit(); 364 365 if (kfd->init_complete) { 366 err = amd_iommu_init_device(kfd->pdev, pasid_limit); 367 if (err < 0) 368 return -ENXIO; 369 amd_iommu_set_invalidate_ctx_cb(kfd->pdev, 370 iommu_pasid_shutdown_callback); 371 amd_iommu_set_invalid_ppr_cb(kfd->pdev, iommu_invalid_ppr_cb); 372 kfd->dqm->ops.start(kfd->dqm); 373 } 374 375 return 0; 376 } 377 378 /* This is called directly from KGD at ISR. */ 379 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry) 380 { 381 if (!kfd->init_complete) 382 return; 383 384 spin_lock(&kfd->interrupt_lock); 385 386 if (kfd->interrupts_active 387 && interrupt_is_wanted(kfd, ih_ring_entry) 388 && enqueue_ih_ring_entry(kfd, ih_ring_entry)) 389 schedule_work(&kfd->interrupt_work); 390 391 spin_unlock(&kfd->interrupt_lock); 392 } 393 394 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size, 395 unsigned int chunk_size) 396 { 397 unsigned int num_of_bits; 398 399 BUG_ON(!kfd); 400 BUG_ON(!kfd->gtt_mem); 401 BUG_ON(buf_size < chunk_size); 402 BUG_ON(buf_size == 0); 403 BUG_ON(chunk_size == 0); 404 405 kfd->gtt_sa_chunk_size = chunk_size; 406 kfd->gtt_sa_num_of_chunks = buf_size / chunk_size; 407 408 num_of_bits = kfd->gtt_sa_num_of_chunks / BITS_PER_BYTE; 409 BUG_ON(num_of_bits == 0); 410 411 kfd->gtt_sa_bitmap = kzalloc(num_of_bits, GFP_KERNEL); 412 413 if (!kfd->gtt_sa_bitmap) 414 return -ENOMEM; 415 416 pr_debug("kfd: gtt_sa_num_of_chunks = %d, gtt_sa_bitmap = %p\n", 417 kfd->gtt_sa_num_of_chunks, kfd->gtt_sa_bitmap); 418 419 mutex_init(&kfd->gtt_sa_lock); 420 421 return 0; 422 423 } 424 425 static void kfd_gtt_sa_fini(struct kfd_dev *kfd) 426 { 427 mutex_destroy(&kfd->gtt_sa_lock); 428 kfree(kfd->gtt_sa_bitmap); 429 } 430 431 static inline uint64_t kfd_gtt_sa_calc_gpu_addr(uint64_t start_addr, 432 unsigned int bit_num, 433 unsigned int chunk_size) 434 { 435 return start_addr + bit_num * chunk_size; 436 } 437 438 static inline uint32_t *kfd_gtt_sa_calc_cpu_addr(void *start_addr, 439 unsigned int bit_num, 440 unsigned int chunk_size) 441 { 442 return (uint32_t *) ((uint64_t) start_addr + bit_num * chunk_size); 443 } 444 445 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, 446 struct kfd_mem_obj **mem_obj) 447 { 448 unsigned int found, start_search, cur_size; 449 450 BUG_ON(!kfd); 451 452 if (size == 0) 453 return -EINVAL; 454 455 if (size > kfd->gtt_sa_num_of_chunks * kfd->gtt_sa_chunk_size) 456 return -ENOMEM; 457 458 *mem_obj = kmalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL); 459 if ((*mem_obj) == NULL) 460 return -ENOMEM; 461 462 pr_debug("kfd: allocated mem_obj = %p for size = %d\n", *mem_obj, size); 463 464 start_search = 0; 465 466 mutex_lock(&kfd->gtt_sa_lock); 467 468 kfd_gtt_restart_search: 469 /* Find the first chunk that is free */ 470 found = find_next_zero_bit(kfd->gtt_sa_bitmap, 471 kfd->gtt_sa_num_of_chunks, 472 start_search); 473 474 pr_debug("kfd: found = %d\n", found); 475 476 /* If there wasn't any free chunk, bail out */ 477 if (found == kfd->gtt_sa_num_of_chunks) 478 goto kfd_gtt_no_free_chunk; 479 480 /* Update fields of mem_obj */ 481 (*mem_obj)->range_start = found; 482 (*mem_obj)->range_end = found; 483 (*mem_obj)->gpu_addr = kfd_gtt_sa_calc_gpu_addr( 484 kfd->gtt_start_gpu_addr, 485 found, 486 kfd->gtt_sa_chunk_size); 487 (*mem_obj)->cpu_ptr = kfd_gtt_sa_calc_cpu_addr( 488 kfd->gtt_start_cpu_ptr, 489 found, 490 kfd->gtt_sa_chunk_size); 491 492 pr_debug("kfd: gpu_addr = %p, cpu_addr = %p\n", 493 (uint64_t *) (*mem_obj)->gpu_addr, (*mem_obj)->cpu_ptr); 494 495 /* If we need only one chunk, mark it as allocated and get out */ 496 if (size <= kfd->gtt_sa_chunk_size) { 497 pr_debug("kfd: single bit\n"); 498 set_bit(found, kfd->gtt_sa_bitmap); 499 goto kfd_gtt_out; 500 } 501 502 /* Otherwise, try to see if we have enough contiguous chunks */ 503 cur_size = size - kfd->gtt_sa_chunk_size; 504 do { 505 (*mem_obj)->range_end = 506 find_next_zero_bit(kfd->gtt_sa_bitmap, 507 kfd->gtt_sa_num_of_chunks, ++found); 508 /* 509 * If next free chunk is not contiguous than we need to 510 * restart our search from the last free chunk we found (which 511 * wasn't contiguous to the previous ones 512 */ 513 if ((*mem_obj)->range_end != found) { 514 start_search = found; 515 goto kfd_gtt_restart_search; 516 } 517 518 /* 519 * If we reached end of buffer, bail out with error 520 */ 521 if (found == kfd->gtt_sa_num_of_chunks) 522 goto kfd_gtt_no_free_chunk; 523 524 /* Check if we don't need another chunk */ 525 if (cur_size <= kfd->gtt_sa_chunk_size) 526 cur_size = 0; 527 else 528 cur_size -= kfd->gtt_sa_chunk_size; 529 530 } while (cur_size > 0); 531 532 pr_debug("kfd: range_start = %d, range_end = %d\n", 533 (*mem_obj)->range_start, (*mem_obj)->range_end); 534 535 /* Mark the chunks as allocated */ 536 for (found = (*mem_obj)->range_start; 537 found <= (*mem_obj)->range_end; 538 found++) 539 set_bit(found, kfd->gtt_sa_bitmap); 540 541 kfd_gtt_out: 542 mutex_unlock(&kfd->gtt_sa_lock); 543 return 0; 544 545 kfd_gtt_no_free_chunk: 546 pr_debug("kfd: allocation failed with mem_obj = %p\n", mem_obj); 547 mutex_unlock(&kfd->gtt_sa_lock); 548 kfree(mem_obj); 549 return -ENOMEM; 550 } 551 552 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj) 553 { 554 unsigned int bit; 555 556 BUG_ON(!kfd); 557 558 /* Act like kfree when trying to free a NULL object */ 559 if (!mem_obj) 560 return 0; 561 562 pr_debug("kfd: free mem_obj = %p, range_start = %d, range_end = %d\n", 563 mem_obj, mem_obj->range_start, mem_obj->range_end); 564 565 mutex_lock(&kfd->gtt_sa_lock); 566 567 /* Mark the chunks as free */ 568 for (bit = mem_obj->range_start; 569 bit <= mem_obj->range_end; 570 bit++) 571 clear_bit(bit, kfd->gtt_sa_bitmap); 572 573 mutex_unlock(&kfd->gtt_sa_lock); 574 575 kfree(mem_obj); 576 return 0; 577 } 578