1 // SPDX-License-Identifier: GPL-2.0 OR MIT 2 /* 3 * Copyright 2014-2022 Advanced Micro Devices, Inc. 4 * 5 * Permission is hereby granted, free of charge, to any person obtaining a 6 * copy of this software and associated documentation files (the "Software"), 7 * to deal in the Software without restriction, including without limitation 8 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 9 * and/or sell copies of the Software, and to permit persons to whom the 10 * Software is furnished to do so, subject to the following conditions: 11 * 12 * The above copyright notice and this permission notice shall be included in 13 * all copies or substantial portions of the Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 21 * OTHER DEALINGS IN THE SOFTWARE. 22 * 23 */ 24 25 #include <linux/slab.h> 26 #include <linux/mutex.h> 27 #include "kfd_device_queue_manager.h" 28 #include "kfd_kernel_queue.h" 29 #include "kfd_priv.h" 30 31 static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes, 32 unsigned int buffer_size_bytes) 33 { 34 unsigned int temp = *wptr + increment_bytes / sizeof(uint32_t); 35 36 WARN((temp * sizeof(uint32_t)) > buffer_size_bytes, 37 "Runlist IB overflow"); 38 *wptr = temp; 39 } 40 41 static void pm_calc_rlib_size(struct packet_manager *pm, 42 unsigned int *rlib_size, 43 bool *over_subscription) 44 { 45 unsigned int process_count, queue_count, compute_queue_count, gws_queue_count; 46 unsigned int map_queue_size; 47 unsigned int max_proc_per_quantum = 1; 48 struct kfd_node *dev = pm->dqm->dev; 49 50 process_count = pm->dqm->processes_count; 51 queue_count = pm->dqm->active_queue_count; 52 compute_queue_count = pm->dqm->active_cp_queue_count; 53 gws_queue_count = pm->dqm->gws_queue_count; 54 55 /* check if there is over subscription 56 * Note: the arbitration between the number of VMIDs and 57 * hws_max_conc_proc has been done in 58 * kgd2kfd_device_init(). 59 */ 60 *over_subscription = false; 61 62 if (dev->max_proc_per_quantum > 1) 63 max_proc_per_quantum = dev->max_proc_per_quantum; 64 65 if ((process_count > max_proc_per_quantum) || 66 compute_queue_count > get_cp_queues_num(pm->dqm) || 67 gws_queue_count > 1) { 68 *over_subscription = true; 69 pr_debug("Over subscribed runlist\n"); 70 } 71 72 map_queue_size = pm->pmf->map_queues_size; 73 /* calculate run list ib allocation size */ 74 *rlib_size = process_count * pm->pmf->map_process_size + 75 queue_count * map_queue_size; 76 77 /* 78 * Increase the allocation size in case we need a chained run list 79 * when over subscription 80 */ 81 if (*over_subscription) 82 *rlib_size += pm->pmf->runlist_size; 83 84 pr_debug("runlist ib size %d\n", *rlib_size); 85 } 86 87 static int pm_allocate_runlist_ib(struct packet_manager *pm, 88 unsigned int **rl_buffer, 89 uint64_t *rl_gpu_buffer, 90 unsigned int *rl_buffer_size, 91 bool *is_over_subscription) 92 { 93 int retval; 94 95 if (WARN_ON(pm->allocated)) 96 return -EINVAL; 97 98 pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription); 99 100 mutex_lock(&pm->lock); 101 102 retval = kfd_gtt_sa_allocate(pm->dqm->dev, *rl_buffer_size, 103 &pm->ib_buffer_obj); 104 105 if (retval) { 106 pr_err("Failed to allocate runlist IB\n"); 107 goto out; 108 } 109 110 *(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr; 111 *rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr; 112 113 memset(*rl_buffer, 0, *rl_buffer_size); 114 pm->allocated = true; 115 116 out: 117 mutex_unlock(&pm->lock); 118 return retval; 119 } 120 121 static int pm_create_runlist_ib(struct packet_manager *pm, 122 struct list_head *queues, 123 uint64_t *rl_gpu_addr, 124 size_t *rl_size_bytes) 125 { 126 unsigned int alloc_size_bytes; 127 unsigned int *rl_buffer, rl_wptr, i; 128 int retval, processes_mapped; 129 struct device_process_node *cur; 130 struct qcm_process_device *qpd; 131 struct queue *q; 132 struct kernel_queue *kq; 133 bool is_over_subscription; 134 135 rl_wptr = retval = processes_mapped = 0; 136 137 retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr, 138 &alloc_size_bytes, &is_over_subscription); 139 if (retval) 140 return retval; 141 142 *rl_size_bytes = alloc_size_bytes; 143 pm->ib_size_bytes = alloc_size_bytes; 144 145 pr_debug("Building runlist ib process count: %d queues count %d\n", 146 pm->dqm->processes_count, pm->dqm->active_queue_count); 147 148 /* build the run list ib packet */ 149 list_for_each_entry(cur, queues, list) { 150 qpd = cur->qpd; 151 /* build map process packet */ 152 if (processes_mapped >= pm->dqm->processes_count) { 153 pr_debug("Not enough space left in runlist IB\n"); 154 pm_release_ib(pm); 155 return -ENOMEM; 156 } 157 158 retval = pm->pmf->map_process(pm, &rl_buffer[rl_wptr], qpd); 159 if (retval) 160 return retval; 161 162 processes_mapped++; 163 inc_wptr(&rl_wptr, pm->pmf->map_process_size, 164 alloc_size_bytes); 165 166 list_for_each_entry(kq, &qpd->priv_queue_list, list) { 167 if (!kq->queue->properties.is_active) 168 continue; 169 170 pr_debug("static_queue, mapping kernel q %d, is debug status %d\n", 171 kq->queue->queue, qpd->is_debug); 172 173 retval = pm->pmf->map_queues(pm, 174 &rl_buffer[rl_wptr], 175 kq->queue, 176 qpd->is_debug); 177 if (retval) 178 return retval; 179 180 inc_wptr(&rl_wptr, 181 pm->pmf->map_queues_size, 182 alloc_size_bytes); 183 } 184 185 list_for_each_entry(q, &qpd->queues_list, list) { 186 if (!q->properties.is_active) 187 continue; 188 189 pr_debug("static_queue, mapping user queue %d, is debug status %d\n", 190 q->queue, qpd->is_debug); 191 192 retval = pm->pmf->map_queues(pm, 193 &rl_buffer[rl_wptr], 194 q, 195 qpd->is_debug); 196 197 if (retval) 198 return retval; 199 200 inc_wptr(&rl_wptr, 201 pm->pmf->map_queues_size, 202 alloc_size_bytes); 203 } 204 } 205 206 pr_debug("Finished map process and queues to runlist\n"); 207 208 if (is_over_subscription) { 209 if (!pm->is_over_subscription) 210 pr_warn("Runlist is getting oversubscribed. Expect reduced ROCm performance.\n"); 211 retval = pm->pmf->runlist(pm, &rl_buffer[rl_wptr], 212 *rl_gpu_addr, 213 alloc_size_bytes / sizeof(uint32_t), 214 true); 215 } 216 pm->is_over_subscription = is_over_subscription; 217 218 for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++) 219 pr_debug("0x%2X ", rl_buffer[i]); 220 pr_debug("\n"); 221 222 return retval; 223 } 224 225 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm) 226 { 227 switch (dqm->dev->adev->asic_type) { 228 case CHIP_KAVERI: 229 case CHIP_HAWAII: 230 /* PM4 packet structures on CIK are the same as on VI */ 231 case CHIP_CARRIZO: 232 case CHIP_TONGA: 233 case CHIP_FIJI: 234 case CHIP_POLARIS10: 235 case CHIP_POLARIS11: 236 case CHIP_POLARIS12: 237 case CHIP_VEGAM: 238 pm->pmf = &kfd_vi_pm_funcs; 239 break; 240 default: 241 if (KFD_GC_VERSION(dqm->dev) == IP_VERSION(9, 4, 2) || 242 KFD_GC_VERSION(dqm->dev) == IP_VERSION(9, 4, 3) || 243 KFD_GC_VERSION(dqm->dev) == IP_VERSION(9, 4, 4)) 244 pm->pmf = &kfd_aldebaran_pm_funcs; 245 else if (KFD_GC_VERSION(dqm->dev) >= IP_VERSION(9, 0, 1)) 246 pm->pmf = &kfd_v9_pm_funcs; 247 else { 248 WARN(1, "Unexpected ASIC family %u", 249 dqm->dev->adev->asic_type); 250 return -EINVAL; 251 } 252 } 253 254 pm->dqm = dqm; 255 mutex_init(&pm->lock); 256 pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ); 257 if (!pm->priv_queue) { 258 mutex_destroy(&pm->lock); 259 return -ENOMEM; 260 } 261 pm->allocated = false; 262 263 return 0; 264 } 265 266 void pm_uninit(struct packet_manager *pm, bool hanging) 267 { 268 mutex_destroy(&pm->lock); 269 kernel_queue_uninit(pm->priv_queue, hanging); 270 pm->priv_queue = NULL; 271 } 272 273 int pm_send_set_resources(struct packet_manager *pm, 274 struct scheduling_resources *res) 275 { 276 uint32_t *buffer, size; 277 int retval = 0; 278 279 size = pm->pmf->set_resources_size; 280 mutex_lock(&pm->lock); 281 kq_acquire_packet_buffer(pm->priv_queue, 282 size / sizeof(uint32_t), 283 (unsigned int **)&buffer); 284 if (!buffer) { 285 pr_err("Failed to allocate buffer on kernel queue\n"); 286 retval = -ENOMEM; 287 goto out; 288 } 289 290 retval = pm->pmf->set_resources(pm, buffer, res); 291 if (!retval) 292 retval = kq_submit_packet(pm->priv_queue); 293 else 294 kq_rollback_packet(pm->priv_queue); 295 296 out: 297 mutex_unlock(&pm->lock); 298 299 return retval; 300 } 301 302 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues) 303 { 304 uint64_t rl_gpu_ib_addr; 305 uint32_t *rl_buffer; 306 size_t rl_ib_size, packet_size_dwords; 307 int retval; 308 309 retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr, 310 &rl_ib_size); 311 if (retval) 312 goto fail_create_runlist_ib; 313 314 pr_debug("runlist IB address: 0x%llX\n", rl_gpu_ib_addr); 315 316 packet_size_dwords = pm->pmf->runlist_size / sizeof(uint32_t); 317 mutex_lock(&pm->lock); 318 319 retval = kq_acquire_packet_buffer(pm->priv_queue, 320 packet_size_dwords, &rl_buffer); 321 if (retval) 322 goto fail_acquire_packet_buffer; 323 324 retval = pm->pmf->runlist(pm, rl_buffer, rl_gpu_ib_addr, 325 rl_ib_size / sizeof(uint32_t), false); 326 if (retval) 327 goto fail_create_runlist; 328 329 retval = kq_submit_packet(pm->priv_queue); 330 331 mutex_unlock(&pm->lock); 332 333 return retval; 334 335 fail_create_runlist: 336 kq_rollback_packet(pm->priv_queue); 337 fail_acquire_packet_buffer: 338 mutex_unlock(&pm->lock); 339 fail_create_runlist_ib: 340 pm_release_ib(pm); 341 return retval; 342 } 343 344 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 345 uint64_t fence_value) 346 { 347 uint32_t *buffer, size; 348 int retval = 0; 349 350 if (WARN_ON(!fence_address)) 351 return -EFAULT; 352 353 size = pm->pmf->query_status_size; 354 mutex_lock(&pm->lock); 355 kq_acquire_packet_buffer(pm->priv_queue, 356 size / sizeof(uint32_t), (unsigned int **)&buffer); 357 if (!buffer) { 358 pr_err("Failed to allocate buffer on kernel queue\n"); 359 retval = -ENOMEM; 360 goto out; 361 } 362 363 retval = pm->pmf->query_status(pm, buffer, fence_address, fence_value); 364 if (!retval) 365 retval = kq_submit_packet(pm->priv_queue); 366 else 367 kq_rollback_packet(pm->priv_queue); 368 369 out: 370 mutex_unlock(&pm->lock); 371 return retval; 372 } 373 374 int pm_update_grace_period(struct packet_manager *pm, uint32_t grace_period) 375 { 376 int retval = 0; 377 uint32_t *buffer, size; 378 379 size = pm->pmf->set_grace_period_size; 380 381 mutex_lock(&pm->lock); 382 383 if (size) { 384 kq_acquire_packet_buffer(pm->priv_queue, 385 size / sizeof(uint32_t), 386 (unsigned int **)&buffer); 387 388 if (!buffer) { 389 pr_err("Failed to allocate buffer on kernel queue\n"); 390 retval = -ENOMEM; 391 goto out; 392 } 393 394 retval = pm->pmf->set_grace_period(pm, buffer, grace_period); 395 if (!retval) 396 retval = kq_submit_packet(pm->priv_queue); 397 else 398 kq_rollback_packet(pm->priv_queue); 399 } 400 401 out: 402 mutex_unlock(&pm->lock); 403 return retval; 404 } 405 406 int pm_send_unmap_queue(struct packet_manager *pm, 407 enum kfd_unmap_queues_filter filter, 408 uint32_t filter_param, bool reset) 409 { 410 uint32_t *buffer, size; 411 int retval = 0; 412 413 size = pm->pmf->unmap_queues_size; 414 mutex_lock(&pm->lock); 415 kq_acquire_packet_buffer(pm->priv_queue, 416 size / sizeof(uint32_t), (unsigned int **)&buffer); 417 if (!buffer) { 418 pr_err("Failed to allocate buffer on kernel queue\n"); 419 retval = -ENOMEM; 420 goto out; 421 } 422 423 retval = pm->pmf->unmap_queues(pm, buffer, filter, filter_param, reset); 424 if (!retval) 425 retval = kq_submit_packet(pm->priv_queue); 426 else 427 kq_rollback_packet(pm->priv_queue); 428 429 out: 430 mutex_unlock(&pm->lock); 431 return retval; 432 } 433 434 void pm_release_ib(struct packet_manager *pm) 435 { 436 mutex_lock(&pm->lock); 437 if (pm->allocated) { 438 kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj); 439 pm->allocated = false; 440 } 441 mutex_unlock(&pm->lock); 442 } 443 444 #if defined(CONFIG_DEBUG_FS) 445 446 int pm_debugfs_runlist(struct seq_file *m, void *data) 447 { 448 struct packet_manager *pm = data; 449 450 mutex_lock(&pm->lock); 451 452 if (!pm->allocated) { 453 seq_puts(m, " No active runlist\n"); 454 goto out; 455 } 456 457 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4, 458 pm->ib_buffer_obj->cpu_ptr, pm->ib_size_bytes, false); 459 460 out: 461 mutex_unlock(&pm->lock); 462 return 0; 463 } 464 465 int pm_debugfs_hang_hws(struct packet_manager *pm) 466 { 467 uint32_t *buffer, size; 468 int r = 0; 469 470 if (!pm->priv_queue) 471 return -EAGAIN; 472 473 size = pm->pmf->query_status_size; 474 mutex_lock(&pm->lock); 475 kq_acquire_packet_buffer(pm->priv_queue, 476 size / sizeof(uint32_t), (unsigned int **)&buffer); 477 if (!buffer) { 478 pr_err("Failed to allocate buffer on kernel queue\n"); 479 r = -ENOMEM; 480 goto out; 481 } 482 memset(buffer, 0x55, size); 483 kq_submit_packet(pm->priv_queue); 484 485 pr_info("Submitting %x %x %x %x %x %x %x to HIQ to hang the HWS.", 486 buffer[0], buffer[1], buffer[2], buffer[3], 487 buffer[4], buffer[5], buffer[6]); 488 out: 489 mutex_unlock(&pm->lock); 490 return r; 491 } 492 493 494 #endif 495