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 24 #include <linux/slab.h> 25 #include <linux/mutex.h> 26 #include "kfd_device_queue_manager.h" 27 #include "kfd_kernel_queue.h" 28 #include "kfd_priv.h" 29 30 static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes, 31 unsigned int buffer_size_bytes) 32 { 33 unsigned int temp = *wptr + increment_bytes / sizeof(uint32_t); 34 35 WARN((temp * sizeof(uint32_t)) > buffer_size_bytes, 36 "Runlist IB overflow"); 37 *wptr = temp; 38 } 39 40 static void pm_calc_rlib_size(struct packet_manager *pm, 41 unsigned int *rlib_size, 42 bool *over_subscription) 43 { 44 unsigned int process_count, queue_count, compute_queue_count, gws_queue_count; 45 unsigned int map_queue_size; 46 unsigned int max_proc_per_quantum = 1; 47 struct kfd_dev *dev = pm->dqm->dev; 48 49 process_count = pm->dqm->processes_count; 50 queue_count = pm->dqm->active_queue_count; 51 compute_queue_count = pm->dqm->active_cp_queue_count; 52 gws_queue_count = pm->dqm->gws_queue_count; 53 54 /* check if there is over subscription 55 * Note: the arbitration between the number of VMIDs and 56 * hws_max_conc_proc has been done in 57 * kgd2kfd_device_init(). 58 */ 59 *over_subscription = false; 60 61 if (dev->max_proc_per_quantum > 1) 62 max_proc_per_quantum = dev->max_proc_per_quantum; 63 64 if ((process_count > max_proc_per_quantum) || 65 compute_queue_count > get_cp_queues_num(pm->dqm) || 66 gws_queue_count > 1) { 67 *over_subscription = true; 68 pr_debug("Over subscribed runlist\n"); 69 } 70 71 map_queue_size = pm->pmf->map_queues_size; 72 /* calculate run list ib allocation size */ 73 *rlib_size = process_count * pm->pmf->map_process_size + 74 queue_count * map_queue_size; 75 76 /* 77 * Increase the allocation size in case we need a chained run list 78 * when over subscription 79 */ 80 if (*over_subscription) 81 *rlib_size += pm->pmf->runlist_size; 82 83 pr_debug("runlist ib size %d\n", *rlib_size); 84 } 85 86 static int pm_allocate_runlist_ib(struct packet_manager *pm, 87 unsigned int **rl_buffer, 88 uint64_t *rl_gpu_buffer, 89 unsigned int *rl_buffer_size, 90 bool *is_over_subscription) 91 { 92 int retval; 93 94 if (WARN_ON(pm->allocated)) 95 return -EINVAL; 96 97 pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription); 98 99 mutex_lock(&pm->lock); 100 101 retval = kfd_gtt_sa_allocate(pm->dqm->dev, *rl_buffer_size, 102 &pm->ib_buffer_obj); 103 104 if (retval) { 105 pr_err("Failed to allocate runlist IB\n"); 106 goto out; 107 } 108 109 *(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr; 110 *rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr; 111 112 memset(*rl_buffer, 0, *rl_buffer_size); 113 pm->allocated = true; 114 115 out: 116 mutex_unlock(&pm->lock); 117 return retval; 118 } 119 120 static int pm_create_runlist_ib(struct packet_manager *pm, 121 struct list_head *queues, 122 uint64_t *rl_gpu_addr, 123 size_t *rl_size_bytes) 124 { 125 unsigned int alloc_size_bytes; 126 unsigned int *rl_buffer, rl_wptr, i; 127 int retval, proccesses_mapped; 128 struct device_process_node *cur; 129 struct qcm_process_device *qpd; 130 struct queue *q; 131 struct kernel_queue *kq; 132 bool is_over_subscription; 133 134 rl_wptr = retval = proccesses_mapped = 0; 135 136 retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr, 137 &alloc_size_bytes, &is_over_subscription); 138 if (retval) 139 return retval; 140 141 *rl_size_bytes = alloc_size_bytes; 142 pm->ib_size_bytes = alloc_size_bytes; 143 144 pr_debug("Building runlist ib process count: %d queues count %d\n", 145 pm->dqm->processes_count, pm->dqm->active_queue_count); 146 147 /* build the run list ib packet */ 148 list_for_each_entry(cur, queues, list) { 149 qpd = cur->qpd; 150 /* build map process packet */ 151 if (proccesses_mapped >= pm->dqm->processes_count) { 152 pr_debug("Not enough space left in runlist IB\n"); 153 pm_release_ib(pm); 154 return -ENOMEM; 155 } 156 157 retval = pm->pmf->map_process(pm, &rl_buffer[rl_wptr], qpd); 158 if (retval) 159 return retval; 160 161 proccesses_mapped++; 162 inc_wptr(&rl_wptr, pm->pmf->map_process_size, 163 alloc_size_bytes); 164 165 list_for_each_entry(kq, &qpd->priv_queue_list, list) { 166 if (!kq->queue->properties.is_active) 167 continue; 168 169 pr_debug("static_queue, mapping kernel q %d, is debug status %d\n", 170 kq->queue->queue, qpd->is_debug); 171 172 retval = pm->pmf->map_queues(pm, 173 &rl_buffer[rl_wptr], 174 kq->queue, 175 qpd->is_debug); 176 if (retval) 177 return retval; 178 179 inc_wptr(&rl_wptr, 180 pm->pmf->map_queues_size, 181 alloc_size_bytes); 182 } 183 184 list_for_each_entry(q, &qpd->queues_list, list) { 185 if (!q->properties.is_active) 186 continue; 187 188 pr_debug("static_queue, mapping user queue %d, is debug status %d\n", 189 q->queue, qpd->is_debug); 190 191 retval = pm->pmf->map_queues(pm, 192 &rl_buffer[rl_wptr], 193 q, 194 qpd->is_debug); 195 196 if (retval) 197 return retval; 198 199 inc_wptr(&rl_wptr, 200 pm->pmf->map_queues_size, 201 alloc_size_bytes); 202 } 203 } 204 205 pr_debug("Finished map process and queues to runlist\n"); 206 207 if (is_over_subscription) { 208 if (!pm->is_over_subscription) 209 pr_warn("Runlist is getting oversubscribed. Expect reduced ROCm performance.\n"); 210 retval = pm->pmf->runlist(pm, &rl_buffer[rl_wptr], 211 *rl_gpu_addr, 212 alloc_size_bytes / sizeof(uint32_t), 213 true); 214 } 215 pm->is_over_subscription = is_over_subscription; 216 217 for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++) 218 pr_debug("0x%2X ", rl_buffer[i]); 219 pr_debug("\n"); 220 221 return retval; 222 } 223 224 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm) 225 { 226 switch (dqm->dev->device_info->asic_family) { 227 case CHIP_KAVERI: 228 case CHIP_HAWAII: 229 /* PM4 packet structures on CIK are the same as on VI */ 230 case CHIP_CARRIZO: 231 case CHIP_TONGA: 232 case CHIP_FIJI: 233 case CHIP_POLARIS10: 234 case CHIP_POLARIS11: 235 case CHIP_POLARIS12: 236 case CHIP_VEGAM: 237 pm->pmf = &kfd_vi_pm_funcs; 238 break; 239 case CHIP_VEGA10: 240 case CHIP_VEGA12: 241 case CHIP_VEGA20: 242 case CHIP_RAVEN: 243 case CHIP_RENOIR: 244 case CHIP_ARCTURUS: 245 case CHIP_NAVI10: 246 case CHIP_NAVI12: 247 case CHIP_NAVI14: 248 case CHIP_SIENNA_CICHLID: 249 case CHIP_NAVY_FLOUNDER: 250 case CHIP_VANGOGH: 251 case CHIP_DIMGREY_CAVEFISH: 252 pm->pmf = &kfd_v9_pm_funcs; 253 break; 254 default: 255 WARN(1, "Unexpected ASIC family %u", 256 dqm->dev->device_info->asic_family); 257 return -EINVAL; 258 } 259 260 pm->dqm = dqm; 261 mutex_init(&pm->lock); 262 pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ); 263 if (!pm->priv_queue) { 264 mutex_destroy(&pm->lock); 265 return -ENOMEM; 266 } 267 pm->allocated = false; 268 269 return 0; 270 } 271 272 void pm_uninit(struct packet_manager *pm, bool hanging) 273 { 274 mutex_destroy(&pm->lock); 275 kernel_queue_uninit(pm->priv_queue, hanging); 276 } 277 278 int pm_send_set_resources(struct packet_manager *pm, 279 struct scheduling_resources *res) 280 { 281 uint32_t *buffer, size; 282 int retval = 0; 283 284 size = pm->pmf->set_resources_size; 285 mutex_lock(&pm->lock); 286 kq_acquire_packet_buffer(pm->priv_queue, 287 size / sizeof(uint32_t), 288 (unsigned int **)&buffer); 289 if (!buffer) { 290 pr_err("Failed to allocate buffer on kernel queue\n"); 291 retval = -ENOMEM; 292 goto out; 293 } 294 295 retval = pm->pmf->set_resources(pm, buffer, res); 296 if (!retval) 297 kq_submit_packet(pm->priv_queue); 298 else 299 kq_rollback_packet(pm->priv_queue); 300 301 out: 302 mutex_unlock(&pm->lock); 303 304 return retval; 305 } 306 307 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues) 308 { 309 uint64_t rl_gpu_ib_addr; 310 uint32_t *rl_buffer; 311 size_t rl_ib_size, packet_size_dwords; 312 int retval; 313 314 retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr, 315 &rl_ib_size); 316 if (retval) 317 goto fail_create_runlist_ib; 318 319 pr_debug("runlist IB address: 0x%llX\n", rl_gpu_ib_addr); 320 321 packet_size_dwords = pm->pmf->runlist_size / sizeof(uint32_t); 322 mutex_lock(&pm->lock); 323 324 retval = kq_acquire_packet_buffer(pm->priv_queue, 325 packet_size_dwords, &rl_buffer); 326 if (retval) 327 goto fail_acquire_packet_buffer; 328 329 retval = pm->pmf->runlist(pm, rl_buffer, rl_gpu_ib_addr, 330 rl_ib_size / sizeof(uint32_t), false); 331 if (retval) 332 goto fail_create_runlist; 333 334 kq_submit_packet(pm->priv_queue); 335 336 mutex_unlock(&pm->lock); 337 338 return retval; 339 340 fail_create_runlist: 341 kq_rollback_packet(pm->priv_queue); 342 fail_acquire_packet_buffer: 343 mutex_unlock(&pm->lock); 344 fail_create_runlist_ib: 345 pm_release_ib(pm); 346 return retval; 347 } 348 349 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 350 uint32_t fence_value) 351 { 352 uint32_t *buffer, size; 353 int retval = 0; 354 355 if (WARN_ON(!fence_address)) 356 return -EFAULT; 357 358 size = pm->pmf->query_status_size; 359 mutex_lock(&pm->lock); 360 kq_acquire_packet_buffer(pm->priv_queue, 361 size / sizeof(uint32_t), (unsigned int **)&buffer); 362 if (!buffer) { 363 pr_err("Failed to allocate buffer on kernel queue\n"); 364 retval = -ENOMEM; 365 goto out; 366 } 367 368 retval = pm->pmf->query_status(pm, buffer, fence_address, fence_value); 369 if (!retval) 370 kq_submit_packet(pm->priv_queue); 371 else 372 kq_rollback_packet(pm->priv_queue); 373 374 out: 375 mutex_unlock(&pm->lock); 376 return retval; 377 } 378 379 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, 380 enum kfd_unmap_queues_filter filter, 381 uint32_t filter_param, bool reset, 382 unsigned int sdma_engine) 383 { 384 uint32_t *buffer, size; 385 int retval = 0; 386 387 size = pm->pmf->unmap_queues_size; 388 mutex_lock(&pm->lock); 389 kq_acquire_packet_buffer(pm->priv_queue, 390 size / sizeof(uint32_t), (unsigned int **)&buffer); 391 if (!buffer) { 392 pr_err("Failed to allocate buffer on kernel queue\n"); 393 retval = -ENOMEM; 394 goto out; 395 } 396 397 retval = pm->pmf->unmap_queues(pm, buffer, type, filter, filter_param, 398 reset, sdma_engine); 399 if (!retval) 400 kq_submit_packet(pm->priv_queue); 401 else 402 kq_rollback_packet(pm->priv_queue); 403 404 out: 405 mutex_unlock(&pm->lock); 406 return retval; 407 } 408 409 void pm_release_ib(struct packet_manager *pm) 410 { 411 mutex_lock(&pm->lock); 412 if (pm->allocated) { 413 kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj); 414 pm->allocated = false; 415 } 416 mutex_unlock(&pm->lock); 417 } 418 419 #if defined(CONFIG_DEBUG_FS) 420 421 int pm_debugfs_runlist(struct seq_file *m, void *data) 422 { 423 struct packet_manager *pm = data; 424 425 mutex_lock(&pm->lock); 426 427 if (!pm->allocated) { 428 seq_puts(m, " No active runlist\n"); 429 goto out; 430 } 431 432 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4, 433 pm->ib_buffer_obj->cpu_ptr, pm->ib_size_bytes, false); 434 435 out: 436 mutex_unlock(&pm->lock); 437 return 0; 438 } 439 440 int pm_debugfs_hang_hws(struct packet_manager *pm) 441 { 442 uint32_t *buffer, size; 443 int r = 0; 444 445 size = pm->pmf->query_status_size; 446 mutex_lock(&pm->lock); 447 kq_acquire_packet_buffer(pm->priv_queue, 448 size / sizeof(uint32_t), (unsigned int **)&buffer); 449 if (!buffer) { 450 pr_err("Failed to allocate buffer on kernel queue\n"); 451 r = -ENOMEM; 452 goto out; 453 } 454 memset(buffer, 0x55, size); 455 kq_submit_packet(pm->priv_queue); 456 457 pr_info("Submitting %x %x %x %x %x %x %x to HIQ to hang the HWS.", 458 buffer[0], buffer[1], buffer[2], buffer[3], 459 buffer[4], buffer[5], buffer[6]); 460 out: 461 mutex_unlock(&pm->lock); 462 return r; 463 } 464 465 466 #endif 467