1 /* 2 * Copyright 2018 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/printk.h> 25 #include <linux/slab.h> 26 #include <linux/uaccess.h> 27 #include "kfd_priv.h" 28 #include "kfd_mqd_manager.h" 29 #include "v10_structs.h" 30 #include "gc/gc_10_1_0_offset.h" 31 #include "gc/gc_10_1_0_sh_mask.h" 32 #include "amdgpu_amdkfd.h" 33 34 static inline struct v10_compute_mqd *get_mqd(void *mqd) 35 { 36 return (struct v10_compute_mqd *)mqd; 37 } 38 39 static inline struct v10_sdma_mqd *get_sdma_mqd(void *mqd) 40 { 41 return (struct v10_sdma_mqd *)mqd; 42 } 43 44 static void update_cu_mask(struct mqd_manager *mm, void *mqd, 45 struct queue_properties *q) 46 { 47 struct v10_compute_mqd *m; 48 uint32_t se_mask[4] = {0}; /* 4 is the max # of SEs */ 49 50 if (q->cu_mask_count == 0) 51 return; 52 53 mqd_symmetrically_map_cu_mask(mm, 54 q->cu_mask, q->cu_mask_count, se_mask); 55 56 m = get_mqd(mqd); 57 m->compute_static_thread_mgmt_se0 = se_mask[0]; 58 m->compute_static_thread_mgmt_se1 = se_mask[1]; 59 m->compute_static_thread_mgmt_se2 = se_mask[2]; 60 m->compute_static_thread_mgmt_se3 = se_mask[3]; 61 62 pr_debug("update cu mask to %#x %#x %#x %#x\n", 63 m->compute_static_thread_mgmt_se0, 64 m->compute_static_thread_mgmt_se1, 65 m->compute_static_thread_mgmt_se2, 66 m->compute_static_thread_mgmt_se3); 67 } 68 69 static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd, 70 struct queue_properties *q) 71 { 72 int retval; 73 struct kfd_mem_obj *mqd_mem_obj = NULL; 74 75 /* From V9, for CWSR, the control stack is located on the next page 76 * boundary after the mqd, we will use the gtt allocation function 77 * instead of sub-allocation function. 78 */ 79 if (kfd->cwsr_enabled && (q->type == KFD_QUEUE_TYPE_COMPUTE)) { 80 mqd_mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_NOIO); 81 if (!mqd_mem_obj) 82 return NULL; 83 retval = amdgpu_amdkfd_alloc_gtt_mem(kfd->kgd, 84 ALIGN(q->ctl_stack_size, PAGE_SIZE) + 85 ALIGN(sizeof(struct v10_compute_mqd), PAGE_SIZE), 86 &(mqd_mem_obj->gtt_mem), 87 &(mqd_mem_obj->gpu_addr), 88 (void *)&(mqd_mem_obj->cpu_ptr), true); 89 } else { 90 retval = kfd_gtt_sa_allocate(kfd, sizeof(struct v10_compute_mqd), 91 &mqd_mem_obj); 92 } 93 94 if (retval) { 95 kfree(mqd_mem_obj); 96 return NULL; 97 } 98 99 return mqd_mem_obj; 100 101 } 102 103 static void init_mqd(struct mqd_manager *mm, void **mqd, 104 struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, 105 struct queue_properties *q) 106 { 107 uint64_t addr; 108 struct v10_compute_mqd *m; 109 110 m = (struct v10_compute_mqd *) mqd_mem_obj->cpu_ptr; 111 addr = mqd_mem_obj->gpu_addr; 112 113 memset(m, 0, sizeof(struct v10_compute_mqd)); 114 115 m->header = 0xC0310800; 116 m->compute_pipelinestat_enable = 1; 117 m->compute_static_thread_mgmt_se0 = 0xFFFFFFFF; 118 m->compute_static_thread_mgmt_se1 = 0xFFFFFFFF; 119 m->compute_static_thread_mgmt_se2 = 0xFFFFFFFF; 120 m->compute_static_thread_mgmt_se3 = 0xFFFFFFFF; 121 122 m->cp_hqd_persistent_state = CP_HQD_PERSISTENT_STATE__PRELOAD_REQ_MASK | 123 0x53 << CP_HQD_PERSISTENT_STATE__PRELOAD_SIZE__SHIFT; 124 125 m->cp_mqd_control = 1 << CP_MQD_CONTROL__PRIV_STATE__SHIFT; 126 127 m->cp_mqd_base_addr_lo = lower_32_bits(addr); 128 m->cp_mqd_base_addr_hi = upper_32_bits(addr); 129 130 m->cp_hqd_quantum = 1 << CP_HQD_QUANTUM__QUANTUM_EN__SHIFT | 131 1 << CP_HQD_QUANTUM__QUANTUM_SCALE__SHIFT | 132 10 << CP_HQD_QUANTUM__QUANTUM_DURATION__SHIFT; 133 134 m->cp_hqd_pipe_priority = 1; 135 m->cp_hqd_queue_priority = 15; 136 137 if (q->format == KFD_QUEUE_FORMAT_AQL) { 138 m->cp_hqd_aql_control = 139 1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT; 140 } 141 142 if (mm->dev->cwsr_enabled) { 143 m->cp_hqd_persistent_state |= 144 (1 << CP_HQD_PERSISTENT_STATE__QSWITCH_MODE__SHIFT); 145 m->cp_hqd_ctx_save_base_addr_lo = 146 lower_32_bits(q->ctx_save_restore_area_address); 147 m->cp_hqd_ctx_save_base_addr_hi = 148 upper_32_bits(q->ctx_save_restore_area_address); 149 m->cp_hqd_ctx_save_size = q->ctx_save_restore_area_size; 150 m->cp_hqd_cntl_stack_size = q->ctl_stack_size; 151 m->cp_hqd_cntl_stack_offset = q->ctl_stack_size; 152 m->cp_hqd_wg_state_offset = q->ctl_stack_size; 153 } 154 155 *mqd = m; 156 if (gart_addr) 157 *gart_addr = addr; 158 mm->update_mqd(mm, m, q); 159 } 160 161 static int load_mqd(struct mqd_manager *mm, void *mqd, 162 uint32_t pipe_id, uint32_t queue_id, 163 struct queue_properties *p, struct mm_struct *mms) 164 { 165 int r = 0; 166 /* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */ 167 uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0); 168 169 r = mm->dev->kfd2kgd->hqd_load(mm->dev->kgd, mqd, pipe_id, queue_id, 170 (uint32_t __user *)p->write_ptr, 171 wptr_shift, 0, mms); 172 return r; 173 } 174 175 static void update_mqd(struct mqd_manager *mm, void *mqd, 176 struct queue_properties *q) 177 { 178 struct v10_compute_mqd *m; 179 180 m = get_mqd(mqd); 181 182 m->cp_hqd_pq_control = 5 << CP_HQD_PQ_CONTROL__RPTR_BLOCK_SIZE__SHIFT; 183 m->cp_hqd_pq_control |= 184 ffs(q->queue_size / sizeof(unsigned int)) - 1 - 1; 185 pr_debug("cp_hqd_pq_control 0x%x\n", m->cp_hqd_pq_control); 186 187 m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8); 188 m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8); 189 190 m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr); 191 m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr); 192 m->cp_hqd_pq_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr); 193 m->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr); 194 195 m->cp_hqd_pq_doorbell_control = 196 q->doorbell_off << 197 CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT; 198 pr_debug("cp_hqd_pq_doorbell_control 0x%x\n", 199 m->cp_hqd_pq_doorbell_control); 200 201 m->cp_hqd_ib_control = 3 << CP_HQD_IB_CONTROL__MIN_IB_AVAIL_SIZE__SHIFT; 202 203 /* 204 * HW does not clamp this field correctly. Maximum EOP queue size 205 * is constrained by per-SE EOP done signal count, which is 8-bit. 206 * Limit is 0xFF EOP entries (= 0x7F8 dwords). CP will not submit 207 * more than (EOP entry count - 1) so a queue size of 0x800 dwords 208 * is safe, giving a maximum field value of 0xA. 209 */ 210 m->cp_hqd_eop_control = min(0xA, 211 ffs(q->eop_ring_buffer_size / sizeof(unsigned int)) - 1 - 1); 212 m->cp_hqd_eop_base_addr_lo = 213 lower_32_bits(q->eop_ring_buffer_address >> 8); 214 m->cp_hqd_eop_base_addr_hi = 215 upper_32_bits(q->eop_ring_buffer_address >> 8); 216 217 m->cp_hqd_iq_timer = 0; 218 219 m->cp_hqd_vmid = q->vmid; 220 221 if (q->format == KFD_QUEUE_FORMAT_AQL) { 222 /* GC 10 removed WPP_CLAMP from PQ Control */ 223 m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK | 224 2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT | 225 1 << CP_HQD_PQ_CONTROL__QUEUE_FULL_EN__SHIFT ; 226 m->cp_hqd_pq_doorbell_control |= 227 1 << CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_BIF_DROP__SHIFT; 228 } 229 if (mm->dev->cwsr_enabled) 230 m->cp_hqd_ctx_save_control = 0; 231 232 update_cu_mask(mm, mqd, q); 233 234 q->is_active = (q->queue_size > 0 && 235 q->queue_address != 0 && 236 q->queue_percent > 0 && 237 !q->is_evicted); 238 } 239 240 static int destroy_mqd(struct mqd_manager *mm, void *mqd, 241 enum kfd_preempt_type type, 242 unsigned int timeout, uint32_t pipe_id, 243 uint32_t queue_id) 244 { 245 return mm->dev->kfd2kgd->hqd_destroy 246 (mm->dev->kgd, mqd, type, timeout, 247 pipe_id, queue_id); 248 } 249 250 static void free_mqd(struct mqd_manager *mm, void *mqd, 251 struct kfd_mem_obj *mqd_mem_obj) 252 { 253 struct kfd_dev *kfd = mm->dev; 254 255 if (mqd_mem_obj->gtt_mem) { 256 amdgpu_amdkfd_free_gtt_mem(kfd->kgd, mqd_mem_obj->gtt_mem); 257 kfree(mqd_mem_obj); 258 } else { 259 kfd_gtt_sa_free(mm->dev, mqd_mem_obj); 260 } 261 } 262 263 static bool is_occupied(struct mqd_manager *mm, void *mqd, 264 uint64_t queue_address, uint32_t pipe_id, 265 uint32_t queue_id) 266 { 267 return mm->dev->kfd2kgd->hqd_is_occupied( 268 mm->dev->kgd, queue_address, 269 pipe_id, queue_id); 270 } 271 272 static int get_wave_state(struct mqd_manager *mm, void *mqd, 273 void __user *ctl_stack, 274 u32 *ctl_stack_used_size, 275 u32 *save_area_used_size) 276 { 277 struct v10_compute_mqd *m; 278 279 /* Control stack is located one page after MQD. */ 280 void *mqd_ctl_stack = (void *)((uintptr_t)mqd + PAGE_SIZE); 281 282 m = get_mqd(mqd); 283 284 *ctl_stack_used_size = m->cp_hqd_cntl_stack_size - 285 m->cp_hqd_cntl_stack_offset; 286 *save_area_used_size = m->cp_hqd_wg_state_offset - 287 m->cp_hqd_cntl_stack_size; 288 289 if (copy_to_user(ctl_stack, mqd_ctl_stack, m->cp_hqd_cntl_stack_size)) 290 return -EFAULT; 291 292 return 0; 293 } 294 295 static void init_mqd_hiq(struct mqd_manager *mm, void **mqd, 296 struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, 297 struct queue_properties *q) 298 { 299 struct v10_compute_mqd *m; 300 301 init_mqd(mm, mqd, mqd_mem_obj, gart_addr, q); 302 303 m = get_mqd(*mqd); 304 305 m->cp_hqd_pq_control |= 1 << CP_HQD_PQ_CONTROL__PRIV_STATE__SHIFT | 306 1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT; 307 } 308 309 static void update_mqd_hiq(struct mqd_manager *mm, void *mqd, 310 struct queue_properties *q) 311 { 312 struct v10_compute_mqd *m; 313 314 update_mqd(mm, mqd, q); 315 316 /* TODO: what's the point? update_mqd already does this. */ 317 m = get_mqd(mqd); 318 m->cp_hqd_vmid = q->vmid; 319 } 320 321 static void init_mqd_sdma(struct mqd_manager *mm, void **mqd, 322 struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, 323 struct queue_properties *q) 324 { 325 struct v10_sdma_mqd *m; 326 327 m = (struct v10_sdma_mqd *) mqd_mem_obj->cpu_ptr; 328 329 memset(m, 0, sizeof(struct v10_sdma_mqd)); 330 331 *mqd = m; 332 if (gart_addr) 333 *gart_addr = mqd_mem_obj->gpu_addr; 334 335 mm->update_mqd(mm, m, q); 336 } 337 338 static int load_mqd_sdma(struct mqd_manager *mm, void *mqd, 339 uint32_t pipe_id, uint32_t queue_id, 340 struct queue_properties *p, struct mm_struct *mms) 341 { 342 return mm->dev->kfd2kgd->hqd_sdma_load(mm->dev->kgd, mqd, 343 (uint32_t __user *)p->write_ptr, 344 mms); 345 } 346 347 #define SDMA_RLC_DUMMY_DEFAULT 0xf 348 349 static void update_mqd_sdma(struct mqd_manager *mm, void *mqd, 350 struct queue_properties *q) 351 { 352 struct v10_sdma_mqd *m; 353 354 m = get_sdma_mqd(mqd); 355 m->sdmax_rlcx_rb_cntl = (ffs(q->queue_size / sizeof(unsigned int)) - 1) 356 << SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT | 357 q->vmid << SDMA0_RLC0_RB_CNTL__RB_VMID__SHIFT | 358 1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT | 359 6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT; 360 361 m->sdmax_rlcx_rb_base = lower_32_bits(q->queue_address >> 8); 362 m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8); 363 m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr); 364 m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr); 365 m->sdmax_rlcx_doorbell_offset = 366 q->doorbell_off << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT; 367 368 m->sdma_engine_id = q->sdma_engine_id; 369 m->sdma_queue_id = q->sdma_queue_id; 370 m->sdmax_rlcx_dummy_reg = SDMA_RLC_DUMMY_DEFAULT; 371 372 373 q->is_active = (q->queue_size > 0 && 374 q->queue_address != 0 && 375 q->queue_percent > 0 && 376 !q->is_evicted); 377 } 378 379 /* 380 * * preempt type here is ignored because there is only one way 381 * * to preempt sdma queue 382 */ 383 static int destroy_mqd_sdma(struct mqd_manager *mm, void *mqd, 384 enum kfd_preempt_type type, 385 unsigned int timeout, uint32_t pipe_id, 386 uint32_t queue_id) 387 { 388 return mm->dev->kfd2kgd->hqd_sdma_destroy(mm->dev->kgd, mqd, timeout); 389 } 390 391 static bool is_occupied_sdma(struct mqd_manager *mm, void *mqd, 392 uint64_t queue_address, uint32_t pipe_id, 393 uint32_t queue_id) 394 { 395 return mm->dev->kfd2kgd->hqd_sdma_is_occupied(mm->dev->kgd, mqd); 396 } 397 398 #if defined(CONFIG_DEBUG_FS) 399 400 static int debugfs_show_mqd(struct seq_file *m, void *data) 401 { 402 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4, 403 data, sizeof(struct v10_compute_mqd), false); 404 return 0; 405 } 406 407 static int debugfs_show_mqd_sdma(struct seq_file *m, void *data) 408 { 409 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4, 410 data, sizeof(struct v10_sdma_mqd), false); 411 return 0; 412 } 413 414 #endif 415 416 struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, 417 struct kfd_dev *dev) 418 { 419 struct mqd_manager *mqd; 420 421 if (WARN_ON(type >= KFD_MQD_TYPE_MAX)) 422 return NULL; 423 424 mqd = kzalloc(sizeof(*mqd), GFP_NOIO); 425 if (!mqd) 426 return NULL; 427 428 mqd->dev = dev; 429 430 switch (type) { 431 case KFD_MQD_TYPE_CP: 432 case KFD_MQD_TYPE_COMPUTE: 433 pr_debug("%s@%i\n", __func__, __LINE__); 434 mqd->allocate_mqd = allocate_mqd; 435 mqd->init_mqd = init_mqd; 436 mqd->free_mqd = free_mqd; 437 mqd->load_mqd = load_mqd; 438 mqd->update_mqd = update_mqd; 439 mqd->destroy_mqd = destroy_mqd; 440 mqd->is_occupied = is_occupied; 441 mqd->mqd_size = sizeof(struct v10_compute_mqd); 442 mqd->get_wave_state = get_wave_state; 443 #if defined(CONFIG_DEBUG_FS) 444 mqd->debugfs_show_mqd = debugfs_show_mqd; 445 #endif 446 pr_debug("%s@%i\n", __func__, __LINE__); 447 break; 448 case KFD_MQD_TYPE_HIQ: 449 pr_debug("%s@%i\n", __func__, __LINE__); 450 mqd->allocate_mqd = allocate_hiq_mqd; 451 mqd->init_mqd = init_mqd_hiq; 452 mqd->free_mqd = free_mqd_hiq_sdma; 453 mqd->load_mqd = load_mqd; 454 mqd->update_mqd = update_mqd_hiq; 455 mqd->destroy_mqd = destroy_mqd; 456 mqd->is_occupied = is_occupied; 457 mqd->mqd_size = sizeof(struct v10_compute_mqd); 458 #if defined(CONFIG_DEBUG_FS) 459 mqd->debugfs_show_mqd = debugfs_show_mqd; 460 #endif 461 pr_debug("%s@%i\n", __func__, __LINE__); 462 break; 463 case KFD_MQD_TYPE_DIQ: 464 mqd->allocate_mqd = allocate_hiq_mqd; 465 mqd->init_mqd = init_mqd_hiq; 466 mqd->free_mqd = free_mqd; 467 mqd->load_mqd = load_mqd; 468 mqd->update_mqd = update_mqd_hiq; 469 mqd->destroy_mqd = destroy_mqd; 470 mqd->is_occupied = is_occupied; 471 mqd->mqd_size = sizeof(struct v10_compute_mqd); 472 #if defined(CONFIG_DEBUG_FS) 473 mqd->debugfs_show_mqd = debugfs_show_mqd; 474 #endif 475 break; 476 case KFD_MQD_TYPE_SDMA: 477 pr_debug("%s@%i\n", __func__, __LINE__); 478 mqd->allocate_mqd = allocate_sdma_mqd; 479 mqd->init_mqd = init_mqd_sdma; 480 mqd->free_mqd = free_mqd_hiq_sdma; 481 mqd->load_mqd = load_mqd_sdma; 482 mqd->update_mqd = update_mqd_sdma; 483 mqd->destroy_mqd = destroy_mqd_sdma; 484 mqd->is_occupied = is_occupied_sdma; 485 mqd->mqd_size = sizeof(struct v10_sdma_mqd); 486 #if defined(CONFIG_DEBUG_FS) 487 mqd->debugfs_show_mqd = debugfs_show_mqd_sdma; 488 #endif 489 pr_debug("%s@%i\n", __func__, __LINE__); 490 break; 491 default: 492 kfree(mqd); 493 return NULL; 494 } 495 496 return mqd; 497 } 498