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_queues_per_device * 187 kfd->device_info->mqd_size_aligned; 188 189 /* 190 * calculate max size of runlist packet. 191 * There can be only 2 packets at once 192 */ 193 size += (KFD_MAX_NUM_OF_PROCESSES * sizeof(struct pm4_map_process) + 194 max_num_of_queues_per_device * 195 sizeof(struct pm4_map_queues) + sizeof(struct pm4_runlist)) * 2; 196 197 /* Add size of HIQ & DIQ */ 198 size += KFD_KERNEL_QUEUE_SIZE * 2; 199 200 /* add another 512KB for all other allocations on gart (HPD, fences) */ 201 size += 512 * 1024; 202 203 if (kfd2kgd->init_gtt_mem_allocation(kfd->kgd, size, &kfd->gtt_mem, 204 &kfd->gtt_start_gpu_addr, &kfd->gtt_start_cpu_ptr)) { 205 dev_err(kfd_device, 206 "Could not allocate %d bytes for device (%x:%x)\n", 207 size, kfd->pdev->vendor, kfd->pdev->device); 208 goto out; 209 } 210 211 dev_info(kfd_device, 212 "Allocated %d bytes on gart for device(%x:%x)\n", 213 size, kfd->pdev->vendor, kfd->pdev->device); 214 215 /* Initialize GTT sa with 512 byte chunk size */ 216 if (kfd_gtt_sa_init(kfd, size, 512) != 0) { 217 dev_err(kfd_device, 218 "Error initializing gtt sub-allocator\n"); 219 goto kfd_gtt_sa_init_error; 220 } 221 222 kfd_doorbell_init(kfd); 223 224 if (kfd_topology_add_device(kfd) != 0) { 225 dev_err(kfd_device, 226 "Error adding device (%x:%x) to topology\n", 227 kfd->pdev->vendor, kfd->pdev->device); 228 goto kfd_topology_add_device_error; 229 } 230 231 if (!device_iommu_pasid_init(kfd)) { 232 dev_err(kfd_device, 233 "Error initializing iommuv2 for device (%x:%x)\n", 234 kfd->pdev->vendor, kfd->pdev->device); 235 goto device_iommu_pasid_error; 236 } 237 amd_iommu_set_invalidate_ctx_cb(kfd->pdev, 238 iommu_pasid_shutdown_callback); 239 240 kfd->dqm = device_queue_manager_init(kfd); 241 if (!kfd->dqm) { 242 dev_err(kfd_device, 243 "Error initializing queue manager for device (%x:%x)\n", 244 kfd->pdev->vendor, kfd->pdev->device); 245 goto device_queue_manager_error; 246 } 247 248 if (kfd->dqm->ops.start(kfd->dqm) != 0) { 249 dev_err(kfd_device, 250 "Error starting queuen manager for device (%x:%x)\n", 251 kfd->pdev->vendor, kfd->pdev->device); 252 goto dqm_start_error; 253 } 254 255 kfd->init_complete = true; 256 dev_info(kfd_device, "added device (%x:%x)\n", kfd->pdev->vendor, 257 kfd->pdev->device); 258 259 pr_debug("kfd: Starting kfd with the following scheduling policy %d\n", 260 sched_policy); 261 262 goto out; 263 264 dqm_start_error: 265 device_queue_manager_uninit(kfd->dqm); 266 device_queue_manager_error: 267 amd_iommu_free_device(kfd->pdev); 268 device_iommu_pasid_error: 269 kfd_topology_remove_device(kfd); 270 kfd_topology_add_device_error: 271 kfd_gtt_sa_fini(kfd); 272 kfd_gtt_sa_init_error: 273 kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem); 274 dev_err(kfd_device, 275 "device (%x:%x) NOT added due to errors\n", 276 kfd->pdev->vendor, kfd->pdev->device); 277 out: 278 return kfd->init_complete; 279 } 280 281 void kgd2kfd_device_exit(struct kfd_dev *kfd) 282 { 283 if (kfd->init_complete) { 284 device_queue_manager_uninit(kfd->dqm); 285 amd_iommu_free_device(kfd->pdev); 286 kfd_topology_remove_device(kfd); 287 kfd_gtt_sa_fini(kfd); 288 kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem); 289 } 290 291 kfree(kfd); 292 } 293 294 void kgd2kfd_suspend(struct kfd_dev *kfd) 295 { 296 BUG_ON(kfd == NULL); 297 298 if (kfd->init_complete) { 299 kfd->dqm->ops.stop(kfd->dqm); 300 amd_iommu_set_invalidate_ctx_cb(kfd->pdev, NULL); 301 amd_iommu_free_device(kfd->pdev); 302 } 303 } 304 305 int kgd2kfd_resume(struct kfd_dev *kfd) 306 { 307 unsigned int pasid_limit; 308 int err; 309 310 BUG_ON(kfd == NULL); 311 312 pasid_limit = kfd_get_pasid_limit(); 313 314 if (kfd->init_complete) { 315 err = amd_iommu_init_device(kfd->pdev, pasid_limit); 316 if (err < 0) 317 return -ENXIO; 318 amd_iommu_set_invalidate_ctx_cb(kfd->pdev, 319 iommu_pasid_shutdown_callback); 320 kfd->dqm->ops.start(kfd->dqm); 321 } 322 323 return 0; 324 } 325 326 /* This is called directly from KGD at ISR. */ 327 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry) 328 { 329 /* Process interrupts / schedule work as necessary */ 330 } 331 332 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size, 333 unsigned int chunk_size) 334 { 335 unsigned int num_of_bits; 336 337 BUG_ON(!kfd); 338 BUG_ON(!kfd->gtt_mem); 339 BUG_ON(buf_size < chunk_size); 340 BUG_ON(buf_size == 0); 341 BUG_ON(chunk_size == 0); 342 343 kfd->gtt_sa_chunk_size = chunk_size; 344 kfd->gtt_sa_num_of_chunks = buf_size / chunk_size; 345 346 num_of_bits = kfd->gtt_sa_num_of_chunks / BITS_PER_BYTE; 347 BUG_ON(num_of_bits == 0); 348 349 kfd->gtt_sa_bitmap = kzalloc(num_of_bits, GFP_KERNEL); 350 351 if (!kfd->gtt_sa_bitmap) 352 return -ENOMEM; 353 354 pr_debug("kfd: gtt_sa_num_of_chunks = %d, gtt_sa_bitmap = %p\n", 355 kfd->gtt_sa_num_of_chunks, kfd->gtt_sa_bitmap); 356 357 mutex_init(&kfd->gtt_sa_lock); 358 359 return 0; 360 361 } 362 363 static void kfd_gtt_sa_fini(struct kfd_dev *kfd) 364 { 365 mutex_destroy(&kfd->gtt_sa_lock); 366 kfree(kfd->gtt_sa_bitmap); 367 } 368 369 static inline uint64_t kfd_gtt_sa_calc_gpu_addr(uint64_t start_addr, 370 unsigned int bit_num, 371 unsigned int chunk_size) 372 { 373 return start_addr + bit_num * chunk_size; 374 } 375 376 static inline uint32_t *kfd_gtt_sa_calc_cpu_addr(void *start_addr, 377 unsigned int bit_num, 378 unsigned int chunk_size) 379 { 380 return (uint32_t *) ((uint64_t) start_addr + bit_num * chunk_size); 381 } 382 383 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, 384 struct kfd_mem_obj **mem_obj) 385 { 386 unsigned int found, start_search, cur_size; 387 388 BUG_ON(!kfd); 389 390 if (size == 0) 391 return -EINVAL; 392 393 if (size > kfd->gtt_sa_num_of_chunks * kfd->gtt_sa_chunk_size) 394 return -ENOMEM; 395 396 *mem_obj = kmalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL); 397 if ((*mem_obj) == NULL) 398 return -ENOMEM; 399 400 pr_debug("kfd: allocated mem_obj = %p for size = %d\n", *mem_obj, size); 401 402 start_search = 0; 403 404 mutex_lock(&kfd->gtt_sa_lock); 405 406 kfd_gtt_restart_search: 407 /* Find the first chunk that is free */ 408 found = find_next_zero_bit(kfd->gtt_sa_bitmap, 409 kfd->gtt_sa_num_of_chunks, 410 start_search); 411 412 pr_debug("kfd: found = %d\n", found); 413 414 /* If there wasn't any free chunk, bail out */ 415 if (found == kfd->gtt_sa_num_of_chunks) 416 goto kfd_gtt_no_free_chunk; 417 418 /* Update fields of mem_obj */ 419 (*mem_obj)->range_start = found; 420 (*mem_obj)->range_end = found; 421 (*mem_obj)->gpu_addr = kfd_gtt_sa_calc_gpu_addr( 422 kfd->gtt_start_gpu_addr, 423 found, 424 kfd->gtt_sa_chunk_size); 425 (*mem_obj)->cpu_ptr = kfd_gtt_sa_calc_cpu_addr( 426 kfd->gtt_start_cpu_ptr, 427 found, 428 kfd->gtt_sa_chunk_size); 429 430 pr_debug("kfd: gpu_addr = %p, cpu_addr = %p\n", 431 (uint64_t *) (*mem_obj)->gpu_addr, (*mem_obj)->cpu_ptr); 432 433 /* If we need only one chunk, mark it as allocated and get out */ 434 if (size <= kfd->gtt_sa_chunk_size) { 435 pr_debug("kfd: single bit\n"); 436 set_bit(found, kfd->gtt_sa_bitmap); 437 goto kfd_gtt_out; 438 } 439 440 /* Otherwise, try to see if we have enough contiguous chunks */ 441 cur_size = size - kfd->gtt_sa_chunk_size; 442 do { 443 (*mem_obj)->range_end = 444 find_next_zero_bit(kfd->gtt_sa_bitmap, 445 kfd->gtt_sa_num_of_chunks, ++found); 446 /* 447 * If next free chunk is not contiguous than we need to 448 * restart our search from the last free chunk we found (which 449 * wasn't contiguous to the previous ones 450 */ 451 if ((*mem_obj)->range_end != found) { 452 start_search = found; 453 goto kfd_gtt_restart_search; 454 } 455 456 /* 457 * If we reached end of buffer, bail out with error 458 */ 459 if (found == kfd->gtt_sa_num_of_chunks) 460 goto kfd_gtt_no_free_chunk; 461 462 /* Check if we don't need another chunk */ 463 if (cur_size <= kfd->gtt_sa_chunk_size) 464 cur_size = 0; 465 else 466 cur_size -= kfd->gtt_sa_chunk_size; 467 468 } while (cur_size > 0); 469 470 pr_debug("kfd: range_start = %d, range_end = %d\n", 471 (*mem_obj)->range_start, (*mem_obj)->range_end); 472 473 /* Mark the chunks as allocated */ 474 for (found = (*mem_obj)->range_start; 475 found <= (*mem_obj)->range_end; 476 found++) 477 set_bit(found, kfd->gtt_sa_bitmap); 478 479 kfd_gtt_out: 480 mutex_unlock(&kfd->gtt_sa_lock); 481 return 0; 482 483 kfd_gtt_no_free_chunk: 484 pr_debug("kfd: allocation failed with mem_obj = %p\n", mem_obj); 485 mutex_unlock(&kfd->gtt_sa_lock); 486 kfree(mem_obj); 487 return -ENOMEM; 488 } 489 490 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj) 491 { 492 unsigned int bit; 493 494 BUG_ON(!kfd); 495 496 /* Act like kfree when trying to free a NULL object */ 497 if (!mem_obj) 498 return 0; 499 500 pr_debug("kfd: free mem_obj = %p, range_start = %d, range_end = %d\n", 501 mem_obj, mem_obj->range_start, mem_obj->range_end); 502 503 mutex_lock(&kfd->gtt_sa_lock); 504 505 /* Mark the chunks as free */ 506 for (bit = mem_obj->range_start; 507 bit <= mem_obj->range_end; 508 bit++) 509 clear_bit(bit, kfd->gtt_sa_bitmap); 510 511 mutex_unlock(&kfd->gtt_sa_lock); 512 513 kfree(mem_obj); 514 return 0; 515 } 516