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/printk.h> 25 #include <linux/slab.h> 26 #include <linux/mm_types.h> 27 28 #include "kfd_priv.h" 29 #include "kfd_mqd_manager.h" 30 #include "vi_structs.h" 31 #include "gca/gfx_8_0_sh_mask.h" 32 #include "gca/gfx_8_0_enum.h" 33 #include "oss/oss_3_0_sh_mask.h" 34 35 #define CP_MQD_CONTROL__PRIV_STATE__SHIFT 0x8 36 37 static inline struct vi_mqd *get_mqd(void *mqd) 38 { 39 return (struct vi_mqd *)mqd; 40 } 41 42 static inline struct vi_sdma_mqd *get_sdma_mqd(void *mqd) 43 { 44 return (struct vi_sdma_mqd *)mqd; 45 } 46 47 static void update_cu_mask(struct mqd_manager *mm, void *mqd, 48 struct mqd_update_info *minfo) 49 { 50 struct vi_mqd *m; 51 uint32_t se_mask[4] = {0}; /* 4 is the max # of SEs */ 52 53 if (!minfo || (minfo->update_flag != UPDATE_FLAG_CU_MASK) || 54 !minfo->cu_mask.ptr) 55 return; 56 57 mqd_symmetrically_map_cu_mask(mm, 58 minfo->cu_mask.ptr, minfo->cu_mask.count, se_mask); 59 60 m = get_mqd(mqd); 61 m->compute_static_thread_mgmt_se0 = se_mask[0]; 62 m->compute_static_thread_mgmt_se1 = se_mask[1]; 63 m->compute_static_thread_mgmt_se2 = se_mask[2]; 64 m->compute_static_thread_mgmt_se3 = se_mask[3]; 65 66 pr_debug("Update cu mask to %#x %#x %#x %#x\n", 67 m->compute_static_thread_mgmt_se0, 68 m->compute_static_thread_mgmt_se1, 69 m->compute_static_thread_mgmt_se2, 70 m->compute_static_thread_mgmt_se3); 71 } 72 73 static void set_priority(struct vi_mqd *m, struct queue_properties *q) 74 { 75 m->cp_hqd_pipe_priority = pipe_priority_map[q->priority]; 76 m->cp_hqd_queue_priority = q->priority; 77 } 78 79 static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd, 80 struct queue_properties *q) 81 { 82 struct kfd_mem_obj *mqd_mem_obj; 83 84 if (kfd_gtt_sa_allocate(kfd, sizeof(struct vi_mqd), 85 &mqd_mem_obj)) 86 return NULL; 87 88 return mqd_mem_obj; 89 } 90 91 static void init_mqd(struct mqd_manager *mm, void **mqd, 92 struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, 93 struct queue_properties *q) 94 { 95 uint64_t addr; 96 struct vi_mqd *m; 97 98 m = (struct vi_mqd *) mqd_mem_obj->cpu_ptr; 99 addr = mqd_mem_obj->gpu_addr; 100 101 memset(m, 0, sizeof(struct vi_mqd)); 102 103 m->header = 0xC0310800; 104 m->compute_pipelinestat_enable = 1; 105 m->compute_static_thread_mgmt_se0 = 0xFFFFFFFF; 106 m->compute_static_thread_mgmt_se1 = 0xFFFFFFFF; 107 m->compute_static_thread_mgmt_se2 = 0xFFFFFFFF; 108 m->compute_static_thread_mgmt_se3 = 0xFFFFFFFF; 109 110 m->cp_hqd_persistent_state = CP_HQD_PERSISTENT_STATE__PRELOAD_REQ_MASK | 111 0x53 << CP_HQD_PERSISTENT_STATE__PRELOAD_SIZE__SHIFT; 112 113 m->cp_mqd_control = 1 << CP_MQD_CONTROL__PRIV_STATE__SHIFT | 114 MTYPE_UC << CP_MQD_CONTROL__MTYPE__SHIFT; 115 116 m->cp_mqd_base_addr_lo = lower_32_bits(addr); 117 m->cp_mqd_base_addr_hi = upper_32_bits(addr); 118 119 m->cp_hqd_quantum = 1 << CP_HQD_QUANTUM__QUANTUM_EN__SHIFT | 120 1 << CP_HQD_QUANTUM__QUANTUM_SCALE__SHIFT | 121 1 << CP_HQD_QUANTUM__QUANTUM_DURATION__SHIFT; 122 123 set_priority(m, q); 124 m->cp_hqd_eop_rptr = 1 << CP_HQD_EOP_RPTR__INIT_FETCHER__SHIFT; 125 126 if (q->format == KFD_QUEUE_FORMAT_AQL) 127 m->cp_hqd_iq_rptr = 1; 128 129 if (q->tba_addr) { 130 m->compute_tba_lo = lower_32_bits(q->tba_addr >> 8); 131 m->compute_tba_hi = upper_32_bits(q->tba_addr >> 8); 132 m->compute_tma_lo = lower_32_bits(q->tma_addr >> 8); 133 m->compute_tma_hi = upper_32_bits(q->tma_addr >> 8); 134 m->compute_pgm_rsrc2 |= 135 (1 << COMPUTE_PGM_RSRC2__TRAP_PRESENT__SHIFT); 136 } 137 138 if (mm->dev->cwsr_enabled && q->ctx_save_restore_area_address) { 139 m->cp_hqd_persistent_state |= 140 (1 << CP_HQD_PERSISTENT_STATE__QSWITCH_MODE__SHIFT); 141 m->cp_hqd_ctx_save_base_addr_lo = 142 lower_32_bits(q->ctx_save_restore_area_address); 143 m->cp_hqd_ctx_save_base_addr_hi = 144 upper_32_bits(q->ctx_save_restore_area_address); 145 m->cp_hqd_ctx_save_size = q->ctx_save_restore_area_size; 146 m->cp_hqd_cntl_stack_size = q->ctl_stack_size; 147 m->cp_hqd_cntl_stack_offset = q->ctl_stack_size; 148 m->cp_hqd_wg_state_offset = q->ctl_stack_size; 149 } 150 151 *mqd = m; 152 if (gart_addr) 153 *gart_addr = addr; 154 mm->update_mqd(mm, m, q, NULL); 155 } 156 157 static int load_mqd(struct mqd_manager *mm, void *mqd, 158 uint32_t pipe_id, uint32_t queue_id, 159 struct queue_properties *p, struct mm_struct *mms) 160 { 161 /* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */ 162 uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0); 163 uint32_t wptr_mask = (uint32_t)((p->queue_size / 4) - 1); 164 165 return mm->dev->kfd2kgd->hqd_load(mm->dev->kgd, mqd, pipe_id, queue_id, 166 (uint32_t __user *)p->write_ptr, 167 wptr_shift, wptr_mask, mms); 168 } 169 170 static void __update_mqd(struct mqd_manager *mm, void *mqd, 171 struct queue_properties *q, struct mqd_update_info *minfo, 172 unsigned int mtype, unsigned int atc_bit) 173 { 174 struct vi_mqd *m; 175 176 m = get_mqd(mqd); 177 178 m->cp_hqd_pq_control = 5 << CP_HQD_PQ_CONTROL__RPTR_BLOCK_SIZE__SHIFT | 179 atc_bit << CP_HQD_PQ_CONTROL__PQ_ATC__SHIFT | 180 mtype << CP_HQD_PQ_CONTROL__MTYPE__SHIFT; 181 m->cp_hqd_pq_control |= order_base_2(q->queue_size / 4) - 1; 182 pr_debug("cp_hqd_pq_control 0x%x\n", m->cp_hqd_pq_control); 183 184 m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8); 185 m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8); 186 187 m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr); 188 m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr); 189 m->cp_hqd_pq_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr); 190 m->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr); 191 192 m->cp_hqd_pq_doorbell_control = 193 q->doorbell_off << 194 CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT; 195 pr_debug("cp_hqd_pq_doorbell_control 0x%x\n", 196 m->cp_hqd_pq_doorbell_control); 197 198 m->cp_hqd_eop_control = atc_bit << CP_HQD_EOP_CONTROL__EOP_ATC__SHIFT | 199 mtype << CP_HQD_EOP_CONTROL__MTYPE__SHIFT; 200 201 m->cp_hqd_ib_control = atc_bit << CP_HQD_IB_CONTROL__IB_ATC__SHIFT | 202 3 << CP_HQD_IB_CONTROL__MIN_IB_AVAIL_SIZE__SHIFT | 203 mtype << CP_HQD_IB_CONTROL__MTYPE__SHIFT; 204 205 /* 206 * HW does not clamp this field correctly. Maximum EOP queue size 207 * is constrained by per-SE EOP done signal count, which is 8-bit. 208 * Limit is 0xFF EOP entries (= 0x7F8 dwords). CP will not submit 209 * more than (EOP entry count - 1) so a queue size of 0x800 dwords 210 * is safe, giving a maximum field value of 0xA. 211 */ 212 m->cp_hqd_eop_control |= min(0xA, 213 order_base_2(q->eop_ring_buffer_size / 4) - 1); 214 m->cp_hqd_eop_base_addr_lo = 215 lower_32_bits(q->eop_ring_buffer_address >> 8); 216 m->cp_hqd_eop_base_addr_hi = 217 upper_32_bits(q->eop_ring_buffer_address >> 8); 218 219 m->cp_hqd_iq_timer = atc_bit << CP_HQD_IQ_TIMER__IQ_ATC__SHIFT | 220 mtype << CP_HQD_IQ_TIMER__MTYPE__SHIFT; 221 222 m->cp_hqd_vmid = q->vmid; 223 224 if (q->format == KFD_QUEUE_FORMAT_AQL) { 225 m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK | 226 2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT; 227 } 228 229 if (mm->dev->cwsr_enabled && q->ctx_save_restore_area_address) 230 m->cp_hqd_ctx_save_control = 231 atc_bit << CP_HQD_CTX_SAVE_CONTROL__ATC__SHIFT | 232 mtype << CP_HQD_CTX_SAVE_CONTROL__MTYPE__SHIFT; 233 234 update_cu_mask(mm, mqd, minfo); 235 set_priority(m, q); 236 237 q->is_active = QUEUE_IS_ACTIVE(*q); 238 } 239 240 241 static void update_mqd(struct mqd_manager *mm, void *mqd, 242 struct queue_properties *q, 243 struct mqd_update_info *minfo) 244 { 245 __update_mqd(mm, mqd, q, minfo, MTYPE_CC, 1); 246 } 247 248 static uint32_t read_doorbell_id(void *mqd) 249 { 250 struct vi_mqd *m = (struct vi_mqd *)mqd; 251 252 return m->queue_doorbell_id0; 253 } 254 255 static void update_mqd_tonga(struct mqd_manager *mm, void *mqd, 256 struct queue_properties *q, 257 struct mqd_update_info *minfo) 258 { 259 __update_mqd(mm, mqd, q, minfo, MTYPE_UC, 0); 260 } 261 262 static int destroy_mqd(struct mqd_manager *mm, void *mqd, 263 enum kfd_preempt_type type, 264 unsigned int timeout, uint32_t pipe_id, 265 uint32_t queue_id) 266 { 267 return mm->dev->kfd2kgd->hqd_destroy 268 (mm->dev->kgd, mqd, type, timeout, 269 pipe_id, queue_id); 270 } 271 272 static void free_mqd(struct mqd_manager *mm, void *mqd, 273 struct kfd_mem_obj *mqd_mem_obj) 274 { 275 kfd_gtt_sa_free(mm->dev, mqd_mem_obj); 276 } 277 278 static bool is_occupied(struct mqd_manager *mm, void *mqd, 279 uint64_t queue_address, uint32_t pipe_id, 280 uint32_t queue_id) 281 { 282 return mm->dev->kfd2kgd->hqd_is_occupied( 283 mm->dev->kgd, queue_address, 284 pipe_id, queue_id); 285 } 286 287 static int get_wave_state(struct mqd_manager *mm, void *mqd, 288 void __user *ctl_stack, 289 u32 *ctl_stack_used_size, 290 u32 *save_area_used_size) 291 { 292 struct vi_mqd *m; 293 294 m = get_mqd(mqd); 295 296 *ctl_stack_used_size = m->cp_hqd_cntl_stack_size - 297 m->cp_hqd_cntl_stack_offset; 298 *save_area_used_size = m->cp_hqd_wg_state_offset - 299 m->cp_hqd_cntl_stack_size; 300 301 /* Control stack is not copied to user mode for GFXv8 because 302 * it's part of the context save area that is already 303 * accessible to user mode 304 */ 305 306 return 0; 307 } 308 309 static void init_mqd_hiq(struct mqd_manager *mm, void **mqd, 310 struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, 311 struct queue_properties *q) 312 { 313 struct vi_mqd *m; 314 init_mqd(mm, mqd, mqd_mem_obj, gart_addr, q); 315 316 m = get_mqd(*mqd); 317 318 m->cp_hqd_pq_control |= 1 << CP_HQD_PQ_CONTROL__PRIV_STATE__SHIFT | 319 1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT; 320 } 321 322 static void update_mqd_hiq(struct mqd_manager *mm, void *mqd, 323 struct queue_properties *q, 324 struct mqd_update_info *minfo) 325 { 326 __update_mqd(mm, mqd, q, minfo, MTYPE_UC, 0); 327 } 328 329 static void init_mqd_sdma(struct mqd_manager *mm, void **mqd, 330 struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, 331 struct queue_properties *q) 332 { 333 struct vi_sdma_mqd *m; 334 335 m = (struct vi_sdma_mqd *) mqd_mem_obj->cpu_ptr; 336 337 memset(m, 0, sizeof(struct vi_sdma_mqd)); 338 339 *mqd = m; 340 if (gart_addr) 341 *gart_addr = mqd_mem_obj->gpu_addr; 342 343 mm->update_mqd(mm, m, q, NULL); 344 } 345 346 static int load_mqd_sdma(struct mqd_manager *mm, void *mqd, 347 uint32_t pipe_id, uint32_t queue_id, 348 struct queue_properties *p, struct mm_struct *mms) 349 { 350 return mm->dev->kfd2kgd->hqd_sdma_load(mm->dev->kgd, mqd, 351 (uint32_t __user *)p->write_ptr, 352 mms); 353 } 354 355 static void update_mqd_sdma(struct mqd_manager *mm, void *mqd, 356 struct queue_properties *q, 357 struct mqd_update_info *minfo) 358 { 359 struct vi_sdma_mqd *m; 360 361 m = get_sdma_mqd(mqd); 362 m->sdmax_rlcx_rb_cntl = order_base_2(q->queue_size / 4) 363 << SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT | 364 q->vmid << SDMA0_RLC0_RB_CNTL__RB_VMID__SHIFT | 365 1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT | 366 6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT; 367 368 m->sdmax_rlcx_rb_base = lower_32_bits(q->queue_address >> 8); 369 m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8); 370 m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr); 371 m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr); 372 m->sdmax_rlcx_doorbell = 373 q->doorbell_off << SDMA0_RLC0_DOORBELL__OFFSET__SHIFT; 374 375 m->sdmax_rlcx_virtual_addr = q->sdma_vm_addr; 376 377 m->sdma_engine_id = q->sdma_engine_id; 378 m->sdma_queue_id = q->sdma_queue_id; 379 380 q->is_active = QUEUE_IS_ACTIVE(*q); 381 } 382 383 /* 384 * * preempt type here is ignored because there is only one way 385 * * to preempt sdma queue 386 */ 387 static int destroy_mqd_sdma(struct mqd_manager *mm, void *mqd, 388 enum kfd_preempt_type type, 389 unsigned int timeout, uint32_t pipe_id, 390 uint32_t queue_id) 391 { 392 return mm->dev->kfd2kgd->hqd_sdma_destroy(mm->dev->kgd, mqd, timeout); 393 } 394 395 static bool is_occupied_sdma(struct mqd_manager *mm, void *mqd, 396 uint64_t queue_address, uint32_t pipe_id, 397 uint32_t queue_id) 398 { 399 return mm->dev->kfd2kgd->hqd_sdma_is_occupied(mm->dev->kgd, mqd); 400 } 401 402 #if defined(CONFIG_DEBUG_FS) 403 404 static int debugfs_show_mqd(struct seq_file *m, void *data) 405 { 406 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4, 407 data, sizeof(struct vi_mqd), false); 408 return 0; 409 } 410 411 static int debugfs_show_mqd_sdma(struct seq_file *m, void *data) 412 { 413 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4, 414 data, sizeof(struct vi_sdma_mqd), false); 415 return 0; 416 } 417 418 #endif 419 420 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, 421 struct kfd_dev *dev) 422 { 423 struct mqd_manager *mqd; 424 425 if (WARN_ON(type >= KFD_MQD_TYPE_MAX)) 426 return NULL; 427 428 mqd = kzalloc(sizeof(*mqd), GFP_KERNEL); 429 if (!mqd) 430 return NULL; 431 432 mqd->dev = dev; 433 434 switch (type) { 435 case KFD_MQD_TYPE_CP: 436 mqd->allocate_mqd = allocate_mqd; 437 mqd->init_mqd = init_mqd; 438 mqd->free_mqd = free_mqd; 439 mqd->load_mqd = load_mqd; 440 mqd->update_mqd = update_mqd; 441 mqd->destroy_mqd = destroy_mqd; 442 mqd->is_occupied = is_occupied; 443 mqd->get_wave_state = get_wave_state; 444 mqd->mqd_size = sizeof(struct vi_mqd); 445 #if defined(CONFIG_DEBUG_FS) 446 mqd->debugfs_show_mqd = debugfs_show_mqd; 447 #endif 448 break; 449 case KFD_MQD_TYPE_HIQ: 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 vi_mqd); 458 #if defined(CONFIG_DEBUG_FS) 459 mqd->debugfs_show_mqd = debugfs_show_mqd; 460 #endif 461 mqd->read_doorbell_id = read_doorbell_id; 462 break; 463 case KFD_MQD_TYPE_DIQ: 464 mqd->allocate_mqd = allocate_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 vi_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 mqd->allocate_mqd = allocate_sdma_mqd; 478 mqd->init_mqd = init_mqd_sdma; 479 mqd->free_mqd = free_mqd_hiq_sdma; 480 mqd->load_mqd = load_mqd_sdma; 481 mqd->update_mqd = update_mqd_sdma; 482 mqd->destroy_mqd = destroy_mqd_sdma; 483 mqd->is_occupied = is_occupied_sdma; 484 mqd->mqd_size = sizeof(struct vi_sdma_mqd); 485 #if defined(CONFIG_DEBUG_FS) 486 mqd->debugfs_show_mqd = debugfs_show_mqd_sdma; 487 #endif 488 break; 489 default: 490 kfree(mqd); 491 return NULL; 492 } 493 494 return mqd; 495 } 496 497 struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type, 498 struct kfd_dev *dev) 499 { 500 struct mqd_manager *mqd; 501 502 mqd = mqd_manager_init_vi(type, dev); 503 if (!mqd) 504 return NULL; 505 if (type == KFD_MQD_TYPE_CP) 506 mqd->update_mqd = update_mqd_tonga; 507 return mqd; 508 } 509