1 /* 2 * Copyright 2016-2018 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 "v9_structs.h" 30 #include "gc/gc_9_0_offset.h" 31 #include "gc/gc_9_0_sh_mask.h" 32 #include "sdma0/sdma0_4_0_sh_mask.h" 33 #include "amdgpu_amdkfd.h" 34 35 static inline struct v9_mqd *get_mqd(void *mqd) 36 { 37 return (struct v9_mqd *)mqd; 38 } 39 40 static inline struct v9_sdma_mqd *get_sdma_mqd(void *mqd) 41 { 42 return (struct v9_sdma_mqd *)mqd; 43 } 44 45 static void update_cu_mask(struct mqd_manager *mm, void *mqd, 46 struct queue_properties *q) 47 { 48 struct v9_mqd *m; 49 uint32_t se_mask[KFD_MAX_NUM_SE] = {0}; 50 51 if (q->cu_mask_count == 0) 52 return; 53 54 mqd_symmetrically_map_cu_mask(mm, 55 q->cu_mask, q->cu_mask_count, se_mask); 56 57 m = get_mqd(mqd); 58 m->compute_static_thread_mgmt_se0 = se_mask[0]; 59 m->compute_static_thread_mgmt_se1 = se_mask[1]; 60 m->compute_static_thread_mgmt_se2 = se_mask[2]; 61 m->compute_static_thread_mgmt_se3 = se_mask[3]; 62 m->compute_static_thread_mgmt_se4 = se_mask[4]; 63 m->compute_static_thread_mgmt_se5 = se_mask[5]; 64 m->compute_static_thread_mgmt_se6 = se_mask[6]; 65 m->compute_static_thread_mgmt_se7 = se_mask[7]; 66 67 pr_debug("update cu mask to %#x %#x %#x %#x %#x %#x %#x %#x\n", 68 m->compute_static_thread_mgmt_se0, 69 m->compute_static_thread_mgmt_se1, 70 m->compute_static_thread_mgmt_se2, 71 m->compute_static_thread_mgmt_se3, 72 m->compute_static_thread_mgmt_se4, 73 m->compute_static_thread_mgmt_se5, 74 m->compute_static_thread_mgmt_se6, 75 m->compute_static_thread_mgmt_se7); 76 } 77 78 static void set_priority(struct v9_mqd *m, struct queue_properties *q) 79 { 80 m->cp_hqd_pipe_priority = pipe_priority_map[q->priority]; 81 m->cp_hqd_queue_priority = q->priority; 82 } 83 84 static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd, 85 struct queue_properties *q) 86 { 87 int retval; 88 struct kfd_mem_obj *mqd_mem_obj = NULL; 89 90 /* For V9 only, due to a HW bug, the control stack of a user mode 91 * compute queue needs to be allocated just behind the page boundary 92 * of its regular MQD buffer. So we allocate an enlarged MQD buffer: 93 * the first page of the buffer serves as the regular MQD buffer 94 * purpose and the remaining is for control stack. Although the two 95 * parts are in the same buffer object, they need different memory 96 * types: MQD part needs UC (uncached) as usual, while control stack 97 * needs NC (non coherent), which is different from the UC type which 98 * is used when control stack is allocated in user space. 99 * 100 * Because of all those, we use the gtt allocation function instead 101 * of sub-allocation function for this enlarged MQD buffer. Moreover, 102 * in order to achieve two memory types in a single buffer object, we 103 * pass a special bo flag AMDGPU_GEM_CREATE_CP_MQD_GFX9 to instruct 104 * amdgpu memory functions to do so. 105 */ 106 if (kfd->cwsr_enabled && (q->type == KFD_QUEUE_TYPE_COMPUTE)) { 107 mqd_mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL); 108 if (!mqd_mem_obj) 109 return NULL; 110 retval = amdgpu_amdkfd_alloc_gtt_mem(kfd->kgd, 111 ALIGN(q->ctl_stack_size, PAGE_SIZE) + 112 ALIGN(sizeof(struct v9_mqd), PAGE_SIZE), 113 &(mqd_mem_obj->gtt_mem), 114 &(mqd_mem_obj->gpu_addr), 115 (void *)&(mqd_mem_obj->cpu_ptr), true); 116 } else { 117 retval = kfd_gtt_sa_allocate(kfd, sizeof(struct v9_mqd), 118 &mqd_mem_obj); 119 } 120 121 if (retval) { 122 kfree(mqd_mem_obj); 123 return NULL; 124 } 125 126 return mqd_mem_obj; 127 128 } 129 130 static void init_mqd(struct mqd_manager *mm, void **mqd, 131 struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, 132 struct queue_properties *q) 133 { 134 uint64_t addr; 135 struct v9_mqd *m; 136 137 m = (struct v9_mqd *) mqd_mem_obj->cpu_ptr; 138 addr = mqd_mem_obj->gpu_addr; 139 140 memset(m, 0, sizeof(struct v9_mqd)); 141 142 m->header = 0xC0310800; 143 m->compute_pipelinestat_enable = 1; 144 m->compute_static_thread_mgmt_se0 = 0xFFFFFFFF; 145 m->compute_static_thread_mgmt_se1 = 0xFFFFFFFF; 146 m->compute_static_thread_mgmt_se2 = 0xFFFFFFFF; 147 m->compute_static_thread_mgmt_se3 = 0xFFFFFFFF; 148 m->compute_static_thread_mgmt_se4 = 0xFFFFFFFF; 149 m->compute_static_thread_mgmt_se5 = 0xFFFFFFFF; 150 m->compute_static_thread_mgmt_se6 = 0xFFFFFFFF; 151 m->compute_static_thread_mgmt_se7 = 0xFFFFFFFF; 152 153 m->cp_hqd_persistent_state = CP_HQD_PERSISTENT_STATE__PRELOAD_REQ_MASK | 154 0x53 << CP_HQD_PERSISTENT_STATE__PRELOAD_SIZE__SHIFT; 155 156 m->cp_mqd_control = 1 << CP_MQD_CONTROL__PRIV_STATE__SHIFT; 157 158 m->cp_mqd_base_addr_lo = lower_32_bits(addr); 159 m->cp_mqd_base_addr_hi = upper_32_bits(addr); 160 161 m->cp_hqd_quantum = 1 << CP_HQD_QUANTUM__QUANTUM_EN__SHIFT | 162 1 << CP_HQD_QUANTUM__QUANTUM_SCALE__SHIFT | 163 1 << CP_HQD_QUANTUM__QUANTUM_DURATION__SHIFT; 164 165 if (q->format == KFD_QUEUE_FORMAT_AQL) { 166 m->cp_hqd_aql_control = 167 1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT; 168 } 169 170 if (q->tba_addr) { 171 m->compute_pgm_rsrc2 |= 172 (1 << COMPUTE_PGM_RSRC2__TRAP_PRESENT__SHIFT); 173 } 174 175 if (mm->dev->cwsr_enabled && q->ctx_save_restore_area_address) { 176 m->cp_hqd_persistent_state |= 177 (1 << CP_HQD_PERSISTENT_STATE__QSWITCH_MODE__SHIFT); 178 m->cp_hqd_ctx_save_base_addr_lo = 179 lower_32_bits(q->ctx_save_restore_area_address); 180 m->cp_hqd_ctx_save_base_addr_hi = 181 upper_32_bits(q->ctx_save_restore_area_address); 182 m->cp_hqd_ctx_save_size = q->ctx_save_restore_area_size; 183 m->cp_hqd_cntl_stack_size = q->ctl_stack_size; 184 m->cp_hqd_cntl_stack_offset = q->ctl_stack_size; 185 m->cp_hqd_wg_state_offset = q->ctl_stack_size; 186 } 187 188 *mqd = m; 189 if (gart_addr) 190 *gart_addr = addr; 191 mm->update_mqd(mm, m, q); 192 } 193 194 static int load_mqd(struct mqd_manager *mm, void *mqd, 195 uint32_t pipe_id, uint32_t queue_id, 196 struct queue_properties *p, struct mm_struct *mms) 197 { 198 /* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */ 199 uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0); 200 201 return mm->dev->kfd2kgd->hqd_load(mm->dev->kgd, mqd, pipe_id, queue_id, 202 (uint32_t __user *)p->write_ptr, 203 wptr_shift, 0, mms); 204 } 205 206 static int hiq_load_mqd_kiq(struct mqd_manager *mm, void *mqd, 207 uint32_t pipe_id, uint32_t queue_id, 208 struct queue_properties *p, struct mm_struct *mms) 209 { 210 return mm->dev->kfd2kgd->hiq_mqd_load(mm->dev->kgd, mqd, pipe_id, 211 queue_id, p->doorbell_off); 212 } 213 214 static void update_mqd(struct mqd_manager *mm, void *mqd, 215 struct queue_properties *q) 216 { 217 struct v9_mqd *m; 218 219 m = get_mqd(mqd); 220 221 m->cp_hqd_pq_control = 5 << CP_HQD_PQ_CONTROL__RPTR_BLOCK_SIZE__SHIFT; 222 m->cp_hqd_pq_control |= order_base_2(q->queue_size / 4) - 1; 223 pr_debug("cp_hqd_pq_control 0x%x\n", m->cp_hqd_pq_control); 224 225 m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8); 226 m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8); 227 228 m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr); 229 m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr); 230 m->cp_hqd_pq_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr); 231 m->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr); 232 233 m->cp_hqd_pq_doorbell_control = 234 q->doorbell_off << 235 CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT; 236 pr_debug("cp_hqd_pq_doorbell_control 0x%x\n", 237 m->cp_hqd_pq_doorbell_control); 238 239 m->cp_hqd_ib_control = 240 3 << CP_HQD_IB_CONTROL__MIN_IB_AVAIL_SIZE__SHIFT | 241 1 << CP_HQD_IB_CONTROL__IB_EXE_DISABLE__SHIFT; 242 243 /* 244 * HW does not clamp this field correctly. Maximum EOP queue size 245 * is constrained by per-SE EOP done signal count, which is 8-bit. 246 * Limit is 0xFF EOP entries (= 0x7F8 dwords). CP will not submit 247 * more than (EOP entry count - 1) so a queue size of 0x800 dwords 248 * is safe, giving a maximum field value of 0xA. 249 */ 250 m->cp_hqd_eop_control = min(0xA, 251 order_base_2(q->eop_ring_buffer_size / 4) - 1); 252 m->cp_hqd_eop_base_addr_lo = 253 lower_32_bits(q->eop_ring_buffer_address >> 8); 254 m->cp_hqd_eop_base_addr_hi = 255 upper_32_bits(q->eop_ring_buffer_address >> 8); 256 257 m->cp_hqd_iq_timer = 0; 258 259 m->cp_hqd_vmid = q->vmid; 260 261 if (q->format == KFD_QUEUE_FORMAT_AQL) { 262 m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK | 263 2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT | 264 1 << CP_HQD_PQ_CONTROL__QUEUE_FULL_EN__SHIFT | 265 1 << CP_HQD_PQ_CONTROL__WPP_CLAMP_EN__SHIFT; 266 m->cp_hqd_pq_doorbell_control |= 1 << 267 CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_BIF_DROP__SHIFT; 268 } 269 if (mm->dev->cwsr_enabled && q->ctx_save_restore_area_address) 270 m->cp_hqd_ctx_save_control = 0; 271 272 update_cu_mask(mm, mqd, q); 273 set_priority(m, q); 274 275 q->is_active = QUEUE_IS_ACTIVE(*q); 276 } 277 278 279 static uint32_t read_doorbell_id(void *mqd) 280 { 281 struct v9_mqd *m = (struct v9_mqd *)mqd; 282 283 return m->queue_doorbell_id0; 284 } 285 286 static int destroy_mqd(struct mqd_manager *mm, void *mqd, 287 enum kfd_preempt_type type, 288 unsigned int timeout, uint32_t pipe_id, 289 uint32_t queue_id) 290 { 291 return mm->dev->kfd2kgd->hqd_destroy 292 (mm->dev->kgd, mqd, type, timeout, 293 pipe_id, queue_id); 294 } 295 296 static void free_mqd(struct mqd_manager *mm, void *mqd, 297 struct kfd_mem_obj *mqd_mem_obj) 298 { 299 struct kfd_dev *kfd = mm->dev; 300 301 if (mqd_mem_obj->gtt_mem) { 302 amdgpu_amdkfd_free_gtt_mem(kfd->kgd, mqd_mem_obj->gtt_mem); 303 kfree(mqd_mem_obj); 304 } else { 305 kfd_gtt_sa_free(mm->dev, mqd_mem_obj); 306 } 307 } 308 309 static bool is_occupied(struct mqd_manager *mm, void *mqd, 310 uint64_t queue_address, uint32_t pipe_id, 311 uint32_t queue_id) 312 { 313 return mm->dev->kfd2kgd->hqd_is_occupied( 314 mm->dev->kgd, queue_address, 315 pipe_id, queue_id); 316 } 317 318 static int get_wave_state(struct mqd_manager *mm, void *mqd, 319 void __user *ctl_stack, 320 u32 *ctl_stack_used_size, 321 u32 *save_area_used_size) 322 { 323 struct v9_mqd *m; 324 325 /* Control stack is located one page after MQD. */ 326 void *mqd_ctl_stack = (void *)((uintptr_t)mqd + PAGE_SIZE); 327 328 m = get_mqd(mqd); 329 330 *ctl_stack_used_size = m->cp_hqd_cntl_stack_size - 331 m->cp_hqd_cntl_stack_offset; 332 *save_area_used_size = m->cp_hqd_wg_state_offset - 333 m->cp_hqd_cntl_stack_size; 334 335 if (copy_to_user(ctl_stack, mqd_ctl_stack, m->cp_hqd_cntl_stack_size)) 336 return -EFAULT; 337 338 return 0; 339 } 340 341 static void init_mqd_hiq(struct mqd_manager *mm, void **mqd, 342 struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, 343 struct queue_properties *q) 344 { 345 struct v9_mqd *m; 346 347 init_mqd(mm, mqd, mqd_mem_obj, gart_addr, q); 348 349 m = get_mqd(*mqd); 350 351 m->cp_hqd_pq_control |= 1 << CP_HQD_PQ_CONTROL__PRIV_STATE__SHIFT | 352 1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT; 353 } 354 355 static void init_mqd_sdma(struct mqd_manager *mm, void **mqd, 356 struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, 357 struct queue_properties *q) 358 { 359 struct v9_sdma_mqd *m; 360 361 m = (struct v9_sdma_mqd *) mqd_mem_obj->cpu_ptr; 362 363 memset(m, 0, sizeof(struct v9_sdma_mqd)); 364 365 *mqd = m; 366 if (gart_addr) 367 *gart_addr = mqd_mem_obj->gpu_addr; 368 369 mm->update_mqd(mm, m, q); 370 } 371 372 static int load_mqd_sdma(struct mqd_manager *mm, void *mqd, 373 uint32_t pipe_id, uint32_t queue_id, 374 struct queue_properties *p, struct mm_struct *mms) 375 { 376 return mm->dev->kfd2kgd->hqd_sdma_load(mm->dev->kgd, mqd, 377 (uint32_t __user *)p->write_ptr, 378 mms); 379 } 380 381 #define SDMA_RLC_DUMMY_DEFAULT 0xf 382 383 static void update_mqd_sdma(struct mqd_manager *mm, void *mqd, 384 struct queue_properties *q) 385 { 386 struct v9_sdma_mqd *m; 387 388 m = get_sdma_mqd(mqd); 389 m->sdmax_rlcx_rb_cntl = order_base_2(q->queue_size / 4) 390 << SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT | 391 q->vmid << SDMA0_RLC0_RB_CNTL__RB_VMID__SHIFT | 392 1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT | 393 6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT; 394 395 m->sdmax_rlcx_rb_base = lower_32_bits(q->queue_address >> 8); 396 m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8); 397 m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr); 398 m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr); 399 m->sdmax_rlcx_doorbell_offset = 400 q->doorbell_off << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT; 401 402 m->sdma_engine_id = q->sdma_engine_id; 403 m->sdma_queue_id = q->sdma_queue_id; 404 m->sdmax_rlcx_dummy_reg = SDMA_RLC_DUMMY_DEFAULT; 405 406 q->is_active = QUEUE_IS_ACTIVE(*q); 407 } 408 409 /* 410 * * preempt type here is ignored because there is only one way 411 * * to preempt sdma queue 412 */ 413 static int destroy_mqd_sdma(struct mqd_manager *mm, void *mqd, 414 enum kfd_preempt_type type, 415 unsigned int timeout, uint32_t pipe_id, 416 uint32_t queue_id) 417 { 418 return mm->dev->kfd2kgd->hqd_sdma_destroy(mm->dev->kgd, mqd, timeout); 419 } 420 421 static bool is_occupied_sdma(struct mqd_manager *mm, void *mqd, 422 uint64_t queue_address, uint32_t pipe_id, 423 uint32_t queue_id) 424 { 425 return mm->dev->kfd2kgd->hqd_sdma_is_occupied(mm->dev->kgd, mqd); 426 } 427 428 #if defined(CONFIG_DEBUG_FS) 429 430 static int debugfs_show_mqd(struct seq_file *m, void *data) 431 { 432 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4, 433 data, sizeof(struct v9_mqd), false); 434 return 0; 435 } 436 437 static int debugfs_show_mqd_sdma(struct seq_file *m, void *data) 438 { 439 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4, 440 data, sizeof(struct v9_sdma_mqd), false); 441 return 0; 442 } 443 444 #endif 445 446 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, 447 struct kfd_dev *dev) 448 { 449 struct mqd_manager *mqd; 450 451 if (WARN_ON(type >= KFD_MQD_TYPE_MAX)) 452 return NULL; 453 454 mqd = kzalloc(sizeof(*mqd), GFP_KERNEL); 455 if (!mqd) 456 return NULL; 457 458 mqd->dev = dev; 459 460 switch (type) { 461 case KFD_MQD_TYPE_CP: 462 mqd->allocate_mqd = allocate_mqd; 463 mqd->init_mqd = init_mqd; 464 mqd->free_mqd = free_mqd; 465 mqd->load_mqd = load_mqd; 466 mqd->update_mqd = update_mqd; 467 mqd->destroy_mqd = destroy_mqd; 468 mqd->is_occupied = is_occupied; 469 mqd->get_wave_state = get_wave_state; 470 mqd->mqd_size = sizeof(struct v9_mqd); 471 #if defined(CONFIG_DEBUG_FS) 472 mqd->debugfs_show_mqd = debugfs_show_mqd; 473 #endif 474 break; 475 case KFD_MQD_TYPE_HIQ: 476 mqd->allocate_mqd = allocate_hiq_mqd; 477 mqd->init_mqd = init_mqd_hiq; 478 mqd->free_mqd = free_mqd_hiq_sdma; 479 mqd->load_mqd = hiq_load_mqd_kiq; 480 mqd->update_mqd = update_mqd; 481 mqd->destroy_mqd = destroy_mqd; 482 mqd->is_occupied = is_occupied; 483 mqd->mqd_size = sizeof(struct v9_mqd); 484 #if defined(CONFIG_DEBUG_FS) 485 mqd->debugfs_show_mqd = debugfs_show_mqd; 486 #endif 487 mqd->read_doorbell_id = read_doorbell_id; 488 break; 489 case KFD_MQD_TYPE_DIQ: 490 mqd->allocate_mqd = allocate_mqd; 491 mqd->init_mqd = init_mqd_hiq; 492 mqd->free_mqd = free_mqd; 493 mqd->load_mqd = load_mqd; 494 mqd->update_mqd = update_mqd; 495 mqd->destroy_mqd = destroy_mqd; 496 mqd->is_occupied = is_occupied; 497 mqd->mqd_size = sizeof(struct v9_mqd); 498 #if defined(CONFIG_DEBUG_FS) 499 mqd->debugfs_show_mqd = debugfs_show_mqd; 500 #endif 501 break; 502 case KFD_MQD_TYPE_SDMA: 503 mqd->allocate_mqd = allocate_sdma_mqd; 504 mqd->init_mqd = init_mqd_sdma; 505 mqd->free_mqd = free_mqd_hiq_sdma; 506 mqd->load_mqd = load_mqd_sdma; 507 mqd->update_mqd = update_mqd_sdma; 508 mqd->destroy_mqd = destroy_mqd_sdma; 509 mqd->is_occupied = is_occupied_sdma; 510 mqd->mqd_size = sizeof(struct v9_sdma_mqd); 511 #if defined(CONFIG_DEBUG_FS) 512 mqd->debugfs_show_mqd = debugfs_show_mqd_sdma; 513 #endif 514 break; 515 default: 516 kfree(mqd); 517 return NULL; 518 } 519 520 return mqd; 521 } 522