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/printk.h> 26 #include <linux/slab.h> 27 #include <linux/mm_types.h> 28 29 #include "kfd_priv.h" 30 #include "kfd_mqd_manager.h" 31 #include "vi_structs.h" 32 #include "gca/gfx_8_0_sh_mask.h" 33 #include "gca/gfx_8_0_enum.h" 34 #include "oss/oss_3_0_sh_mask.h" 35 36 #define CP_MQD_CONTROL__PRIV_STATE__SHIFT 0x8 37 38 static inline struct vi_mqd *get_mqd(void *mqd) 39 { 40 return (struct vi_mqd *)mqd; 41 } 42 43 static inline struct vi_sdma_mqd *get_sdma_mqd(void *mqd) 44 { 45 return (struct vi_sdma_mqd *)mqd; 46 } 47 48 static void update_cu_mask(struct mqd_manager *mm, void *mqd, 49 struct mqd_update_info *minfo) 50 { 51 struct vi_mqd *m; 52 uint32_t se_mask[4] = {0}; /* 4 is the max # of SEs */ 53 54 if (!minfo || !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_node *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->kfd->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->adev, mqd, pipe_id, queue_id, 166 (uint32_t __user *)p->write_ptr, 167 wptr_shift, wptr_mask, mms, 0); 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->kfd->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 static uint32_t read_doorbell_id(void *mqd) 241 { 242 struct vi_mqd *m = (struct vi_mqd *)mqd; 243 244 return m->queue_doorbell_id0; 245 } 246 247 static void update_mqd(struct mqd_manager *mm, void *mqd, 248 struct queue_properties *q, 249 struct mqd_update_info *minfo) 250 { 251 __update_mqd(mm, mqd, q, minfo, MTYPE_UC, 0); 252 } 253 254 static int get_wave_state(struct mqd_manager *mm, void *mqd, 255 struct queue_properties *q, 256 void __user *ctl_stack, 257 u32 *ctl_stack_used_size, 258 u32 *save_area_used_size) 259 { 260 struct vi_mqd *m; 261 262 m = get_mqd(mqd); 263 264 *ctl_stack_used_size = m->cp_hqd_cntl_stack_size - 265 m->cp_hqd_cntl_stack_offset; 266 *save_area_used_size = m->cp_hqd_wg_state_offset - 267 m->cp_hqd_cntl_stack_size; 268 269 /* Control stack is not copied to user mode for GFXv8 because 270 * it's part of the context save area that is already 271 * accessible to user mode 272 */ 273 274 return 0; 275 } 276 277 static void get_checkpoint_info(struct mqd_manager *mm, void *mqd, u32 *ctl_stack_size) 278 { 279 /* Control stack is stored in user mode */ 280 *ctl_stack_size = 0; 281 } 282 283 static void checkpoint_mqd(struct mqd_manager *mm, void *mqd, void *mqd_dst, void *ctl_stack_dst) 284 { 285 struct vi_mqd *m; 286 287 m = get_mqd(mqd); 288 289 memcpy(mqd_dst, m, sizeof(struct vi_mqd)); 290 } 291 292 static void restore_mqd(struct mqd_manager *mm, void **mqd, 293 struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, 294 struct queue_properties *qp, 295 const void *mqd_src, 296 const void *ctl_stack_src, const u32 ctl_stack_size) 297 { 298 uint64_t addr; 299 struct vi_mqd *m; 300 301 m = (struct vi_mqd *) mqd_mem_obj->cpu_ptr; 302 addr = mqd_mem_obj->gpu_addr; 303 304 memcpy(m, mqd_src, sizeof(*m)); 305 306 *mqd = m; 307 if (gart_addr) 308 *gart_addr = addr; 309 310 m->cp_hqd_pq_doorbell_control = 311 qp->doorbell_off << 312 CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT; 313 pr_debug("cp_hqd_pq_doorbell_control 0x%x\n", 314 m->cp_hqd_pq_doorbell_control); 315 316 qp->is_active = 0; 317 } 318 319 static void init_mqd_hiq(struct mqd_manager *mm, void **mqd, 320 struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, 321 struct queue_properties *q) 322 { 323 struct vi_mqd *m; 324 325 init_mqd(mm, mqd, mqd_mem_obj, gart_addr, q); 326 327 m = get_mqd(*mqd); 328 329 m->cp_hqd_pq_control |= 1 << CP_HQD_PQ_CONTROL__PRIV_STATE__SHIFT | 330 1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT; 331 } 332 333 static void update_mqd_hiq(struct mqd_manager *mm, void *mqd, 334 struct queue_properties *q, 335 struct mqd_update_info *minfo) 336 { 337 __update_mqd(mm, mqd, q, minfo, MTYPE_UC, 0); 338 } 339 340 static void init_mqd_sdma(struct mqd_manager *mm, void **mqd, 341 struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, 342 struct queue_properties *q) 343 { 344 struct vi_sdma_mqd *m; 345 346 m = (struct vi_sdma_mqd *) mqd_mem_obj->cpu_ptr; 347 348 memset(m, 0, sizeof(struct vi_sdma_mqd)); 349 350 *mqd = m; 351 if (gart_addr) 352 *gart_addr = mqd_mem_obj->gpu_addr; 353 354 mm->update_mqd(mm, m, q, NULL); 355 } 356 357 static void update_mqd_sdma(struct mqd_manager *mm, void *mqd, 358 struct queue_properties *q, 359 struct mqd_update_info *minfo) 360 { 361 struct vi_sdma_mqd *m; 362 363 m = get_sdma_mqd(mqd); 364 m->sdmax_rlcx_rb_cntl = order_base_2(q->queue_size / 4) 365 << SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT | 366 q->vmid << SDMA0_RLC0_RB_CNTL__RB_VMID__SHIFT | 367 1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT | 368 6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT; 369 370 m->sdmax_rlcx_rb_base = lower_32_bits(q->queue_address >> 8); 371 m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8); 372 m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr); 373 m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr); 374 m->sdmax_rlcx_doorbell = 375 q->doorbell_off << SDMA0_RLC0_DOORBELL__OFFSET__SHIFT; 376 377 m->sdmax_rlcx_virtual_addr = q->sdma_vm_addr; 378 379 m->sdma_engine_id = q->sdma_engine_id; 380 m->sdma_queue_id = q->sdma_queue_id; 381 382 q->is_active = QUEUE_IS_ACTIVE(*q); 383 } 384 385 static void checkpoint_mqd_sdma(struct mqd_manager *mm, 386 void *mqd, 387 void *mqd_dst, 388 void *ctl_stack_dst) 389 { 390 struct vi_sdma_mqd *m; 391 392 m = get_sdma_mqd(mqd); 393 394 memcpy(mqd_dst, m, sizeof(struct vi_sdma_mqd)); 395 } 396 397 static void restore_mqd_sdma(struct mqd_manager *mm, void **mqd, 398 struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, 399 struct queue_properties *qp, 400 const void *mqd_src, 401 const void *ctl_stack_src, const u32 ctl_stack_size) 402 { 403 uint64_t addr; 404 struct vi_sdma_mqd *m; 405 406 m = (struct vi_sdma_mqd *) mqd_mem_obj->cpu_ptr; 407 addr = mqd_mem_obj->gpu_addr; 408 409 memcpy(m, mqd_src, sizeof(*m)); 410 411 m->sdmax_rlcx_doorbell = 412 qp->doorbell_off << SDMA0_RLC0_DOORBELL__OFFSET__SHIFT; 413 414 *mqd = m; 415 if (gart_addr) 416 *gart_addr = addr; 417 418 qp->is_active = 0; 419 } 420 421 #if defined(CONFIG_DEBUG_FS) 422 423 424 static int debugfs_show_mqd(struct seq_file *m, void *data) 425 { 426 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4, 427 data, sizeof(struct vi_mqd), false); 428 return 0; 429 } 430 431 static int debugfs_show_mqd_sdma(struct seq_file *m, void *data) 432 { 433 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4, 434 data, sizeof(struct vi_sdma_mqd), false); 435 return 0; 436 } 437 438 #endif 439 440 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, 441 struct kfd_node *dev) 442 { 443 struct mqd_manager *mqd; 444 445 if (WARN_ON(type >= KFD_MQD_TYPE_MAX)) 446 return NULL; 447 448 mqd = kzalloc(sizeof(*mqd), GFP_KERNEL); 449 if (!mqd) 450 return NULL; 451 452 mqd->dev = dev; 453 454 switch (type) { 455 case KFD_MQD_TYPE_CP: 456 mqd->allocate_mqd = allocate_mqd; 457 mqd->init_mqd = init_mqd; 458 mqd->free_mqd = kfd_free_mqd_cp; 459 mqd->load_mqd = load_mqd; 460 mqd->update_mqd = update_mqd; 461 mqd->destroy_mqd = kfd_destroy_mqd_cp; 462 mqd->is_occupied = kfd_is_occupied_cp; 463 mqd->get_wave_state = get_wave_state; 464 mqd->get_checkpoint_info = get_checkpoint_info; 465 mqd->checkpoint_mqd = checkpoint_mqd; 466 mqd->restore_mqd = restore_mqd; 467 mqd->mqd_size = sizeof(struct vi_mqd); 468 #if defined(CONFIG_DEBUG_FS) 469 mqd->debugfs_show_mqd = debugfs_show_mqd; 470 #endif 471 break; 472 case KFD_MQD_TYPE_HIQ: 473 mqd->allocate_mqd = allocate_hiq_mqd; 474 mqd->init_mqd = init_mqd_hiq; 475 mqd->free_mqd = free_mqd_hiq_sdma; 476 mqd->load_mqd = load_mqd; 477 mqd->update_mqd = update_mqd_hiq; 478 mqd->destroy_mqd = kfd_destroy_mqd_cp; 479 mqd->is_occupied = kfd_is_occupied_cp; 480 mqd->mqd_size = sizeof(struct vi_mqd); 481 mqd->mqd_stride = kfd_mqd_stride; 482 #if defined(CONFIG_DEBUG_FS) 483 mqd->debugfs_show_mqd = debugfs_show_mqd; 484 #endif 485 mqd->read_doorbell_id = read_doorbell_id; 486 break; 487 case KFD_MQD_TYPE_DIQ: 488 mqd->allocate_mqd = allocate_mqd; 489 mqd->init_mqd = init_mqd_hiq; 490 mqd->free_mqd = kfd_free_mqd_cp; 491 mqd->load_mqd = load_mqd; 492 mqd->update_mqd = update_mqd_hiq; 493 mqd->destroy_mqd = kfd_destroy_mqd_cp; 494 mqd->is_occupied = kfd_is_occupied_cp; 495 mqd->mqd_size = sizeof(struct vi_mqd); 496 mqd->mqd_stride = kfd_mqd_stride; 497 #if defined(CONFIG_DEBUG_FS) 498 mqd->debugfs_show_mqd = debugfs_show_mqd; 499 #endif 500 break; 501 case KFD_MQD_TYPE_SDMA: 502 mqd->allocate_mqd = allocate_sdma_mqd; 503 mqd->init_mqd = init_mqd_sdma; 504 mqd->free_mqd = free_mqd_hiq_sdma; 505 mqd->load_mqd = kfd_load_mqd_sdma; 506 mqd->update_mqd = update_mqd_sdma; 507 mqd->destroy_mqd = kfd_destroy_mqd_sdma; 508 mqd->is_occupied = kfd_is_occupied_sdma; 509 mqd->checkpoint_mqd = checkpoint_mqd_sdma; 510 mqd->restore_mqd = restore_mqd_sdma; 511 mqd->mqd_size = sizeof(struct vi_sdma_mqd); 512 mqd->mqd_stride = kfd_mqd_stride; 513 #if defined(CONFIG_DEBUG_FS) 514 mqd->debugfs_show_mqd = debugfs_show_mqd_sdma; 515 #endif 516 break; 517 default: 518 kfree(mqd); 519 return NULL; 520 } 521 522 return mqd; 523 } 524