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 "cik_regs.h" 31 #include "cik_structs.h" 32 #include "oss/oss_2_4_sh_mask.h" 33 34 static inline struct cik_mqd *get_mqd(void *mqd) 35 { 36 return (struct cik_mqd *)mqd; 37 } 38 39 static inline struct cik_sdma_rlc_registers *get_sdma_mqd(void *mqd) 40 { 41 return (struct cik_sdma_rlc_registers *)mqd; 42 } 43 44 static void update_cu_mask(struct mqd_manager *mm, void *mqd, 45 struct queue_properties *q) 46 { 47 struct cik_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 int init_mqd(struct mqd_manager *mm, void **mqd, 70 struct kfd_mem_obj **mqd_mem_obj, uint64_t *gart_addr, 71 struct queue_properties *q) 72 { 73 uint64_t addr; 74 struct cik_mqd *m; 75 int retval; 76 77 retval = kfd_gtt_sa_allocate(mm->dev, sizeof(struct cik_mqd), 78 mqd_mem_obj); 79 80 if (retval != 0) 81 return -ENOMEM; 82 83 m = (struct cik_mqd *) (*mqd_mem_obj)->cpu_ptr; 84 addr = (*mqd_mem_obj)->gpu_addr; 85 86 memset(m, 0, ALIGN(sizeof(struct cik_mqd), 256)); 87 88 m->header = 0xC0310800; 89 m->compute_pipelinestat_enable = 1; 90 m->compute_static_thread_mgmt_se0 = 0xFFFFFFFF; 91 m->compute_static_thread_mgmt_se1 = 0xFFFFFFFF; 92 m->compute_static_thread_mgmt_se2 = 0xFFFFFFFF; 93 m->compute_static_thread_mgmt_se3 = 0xFFFFFFFF; 94 95 /* 96 * Make sure to use the last queue state saved on mqd when the cp 97 * reassigns the queue, so when queue is switched on/off (e.g over 98 * subscription or quantum timeout) the context will be consistent 99 */ 100 m->cp_hqd_persistent_state = 101 DEFAULT_CP_HQD_PERSISTENT_STATE | PRELOAD_REQ; 102 103 m->cp_mqd_control = MQD_CONTROL_PRIV_STATE_EN; 104 m->cp_mqd_base_addr_lo = lower_32_bits(addr); 105 m->cp_mqd_base_addr_hi = upper_32_bits(addr); 106 107 m->cp_hqd_quantum = QUANTUM_EN | QUANTUM_SCALE_1MS | 108 QUANTUM_DURATION(10); 109 110 /* 111 * Pipe Priority 112 * Identifies the pipe relative priority when this queue is connected 113 * to the pipeline. The pipe priority is against the GFX pipe and HP3D. 114 * In KFD we are using a fixed pipe priority set to CS_MEDIUM. 115 * 0 = CS_LOW (typically below GFX) 116 * 1 = CS_MEDIUM (typically between HP3D and GFX 117 * 2 = CS_HIGH (typically above HP3D) 118 */ 119 m->cp_hqd_pipe_priority = 1; 120 m->cp_hqd_queue_priority = 15; 121 122 if (q->format == KFD_QUEUE_FORMAT_AQL) 123 m->cp_hqd_iq_rptr = AQL_ENABLE; 124 125 *mqd = m; 126 if (gart_addr) 127 *gart_addr = addr; 128 retval = mm->update_mqd(mm, m, q); 129 130 return retval; 131 } 132 133 static int init_mqd_sdma(struct mqd_manager *mm, void **mqd, 134 struct kfd_mem_obj **mqd_mem_obj, uint64_t *gart_addr, 135 struct queue_properties *q) 136 { 137 int retval; 138 struct cik_sdma_rlc_registers *m; 139 140 retval = kfd_gtt_sa_allocate(mm->dev, 141 sizeof(struct cik_sdma_rlc_registers), 142 mqd_mem_obj); 143 144 if (retval != 0) 145 return -ENOMEM; 146 147 m = (struct cik_sdma_rlc_registers *) (*mqd_mem_obj)->cpu_ptr; 148 149 memset(m, 0, sizeof(struct cik_sdma_rlc_registers)); 150 151 *mqd = m; 152 if (gart_addr) 153 *gart_addr = (*mqd_mem_obj)->gpu_addr; 154 155 retval = mm->update_mqd(mm, m, q); 156 157 return retval; 158 } 159 160 static void uninit_mqd(struct mqd_manager *mm, void *mqd, 161 struct kfd_mem_obj *mqd_mem_obj) 162 { 163 kfd_gtt_sa_free(mm->dev, mqd_mem_obj); 164 } 165 166 static void uninit_mqd_sdma(struct mqd_manager *mm, void *mqd, 167 struct kfd_mem_obj *mqd_mem_obj) 168 { 169 kfd_gtt_sa_free(mm->dev, mqd_mem_obj); 170 } 171 172 static int load_mqd(struct mqd_manager *mm, void *mqd, uint32_t pipe_id, 173 uint32_t queue_id, struct queue_properties *p, 174 struct mm_struct *mms) 175 { 176 /* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */ 177 uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0); 178 uint32_t wptr_mask = (uint32_t)((p->queue_size / 4) - 1); 179 180 return mm->dev->kfd2kgd->hqd_load(mm->dev->kgd, mqd, pipe_id, queue_id, 181 (uint32_t __user *)p->write_ptr, 182 wptr_shift, wptr_mask, mms); 183 } 184 185 static int load_mqd_sdma(struct mqd_manager *mm, void *mqd, 186 uint32_t pipe_id, uint32_t queue_id, 187 struct queue_properties *p, struct mm_struct *mms) 188 { 189 return mm->dev->kfd2kgd->hqd_sdma_load(mm->dev->kgd, mqd, 190 (uint32_t __user *)p->write_ptr, 191 mms); 192 } 193 194 static int __update_mqd(struct mqd_manager *mm, void *mqd, 195 struct queue_properties *q, unsigned int atc_bit) 196 { 197 struct cik_mqd *m; 198 199 m = get_mqd(mqd); 200 m->cp_hqd_pq_control = DEFAULT_RPTR_BLOCK_SIZE | 201 DEFAULT_MIN_AVAIL_SIZE; 202 m->cp_hqd_ib_control = DEFAULT_MIN_IB_AVAIL_SIZE; 203 if (atc_bit) { 204 m->cp_hqd_pq_control |= PQ_ATC_EN; 205 m->cp_hqd_ib_control |= IB_ATC_EN; 206 } 207 208 /* 209 * Calculating queue size which is log base 2 of actual queue size -1 210 * dwords and another -1 for ffs 211 */ 212 m->cp_hqd_pq_control |= order_base_2(q->queue_size / 4) - 1; 213 m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8); 214 m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8); 215 m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr); 216 m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr); 217 m->cp_hqd_pq_doorbell_control = DOORBELL_OFFSET(q->doorbell_off); 218 219 m->cp_hqd_vmid = q->vmid; 220 221 if (q->format == KFD_QUEUE_FORMAT_AQL) 222 m->cp_hqd_pq_control |= NO_UPDATE_RPTR; 223 224 update_cu_mask(mm, mqd, q); 225 226 q->is_active = (q->queue_size > 0 && 227 q->queue_address != 0 && 228 q->queue_percent > 0 && 229 !q->is_evicted); 230 231 return 0; 232 } 233 234 static int update_mqd(struct mqd_manager *mm, void *mqd, 235 struct queue_properties *q) 236 { 237 return __update_mqd(mm, mqd, q, 1); 238 } 239 240 static int update_mqd_hawaii(struct mqd_manager *mm, void *mqd, 241 struct queue_properties *q) 242 { 243 return __update_mqd(mm, mqd, q, 0); 244 } 245 246 static int update_mqd_sdma(struct mqd_manager *mm, void *mqd, 247 struct queue_properties *q) 248 { 249 struct cik_sdma_rlc_registers *m; 250 251 m = get_sdma_mqd(mqd); 252 m->sdma_rlc_rb_cntl = order_base_2(q->queue_size / 4) 253 << SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT | 254 q->vmid << SDMA0_RLC0_RB_CNTL__RB_VMID__SHIFT | 255 1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT | 256 6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT; 257 258 m->sdma_rlc_rb_base = lower_32_bits(q->queue_address >> 8); 259 m->sdma_rlc_rb_base_hi = upper_32_bits(q->queue_address >> 8); 260 m->sdma_rlc_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr); 261 m->sdma_rlc_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr); 262 m->sdma_rlc_doorbell = 263 q->doorbell_off << SDMA0_RLC0_DOORBELL__OFFSET__SHIFT; 264 265 m->sdma_rlc_virtual_addr = q->sdma_vm_addr; 266 267 m->sdma_engine_id = q->sdma_engine_id; 268 m->sdma_queue_id = q->sdma_queue_id; 269 270 q->is_active = (q->queue_size > 0 && 271 q->queue_address != 0 && 272 q->queue_percent > 0 && 273 !q->is_evicted); 274 275 return 0; 276 } 277 278 static int destroy_mqd(struct mqd_manager *mm, void *mqd, 279 enum kfd_preempt_type type, 280 unsigned int timeout, uint32_t pipe_id, 281 uint32_t queue_id) 282 { 283 return mm->dev->kfd2kgd->hqd_destroy(mm->dev->kgd, mqd, type, timeout, 284 pipe_id, queue_id); 285 } 286 287 /* 288 * preempt type here is ignored because there is only one way 289 * to preempt sdma queue 290 */ 291 static int destroy_mqd_sdma(struct mqd_manager *mm, void *mqd, 292 enum kfd_preempt_type type, 293 unsigned int timeout, uint32_t pipe_id, 294 uint32_t queue_id) 295 { 296 return mm->dev->kfd2kgd->hqd_sdma_destroy(mm->dev->kgd, mqd, timeout); 297 } 298 299 static bool is_occupied(struct mqd_manager *mm, void *mqd, 300 uint64_t queue_address, uint32_t pipe_id, 301 uint32_t queue_id) 302 { 303 304 return mm->dev->kfd2kgd->hqd_is_occupied(mm->dev->kgd, queue_address, 305 pipe_id, queue_id); 306 307 } 308 309 static bool is_occupied_sdma(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_sdma_is_occupied(mm->dev->kgd, mqd); 314 } 315 316 /* 317 * HIQ MQD Implementation, concrete implementation for HIQ MQD implementation. 318 * The HIQ queue in Kaveri is using the same MQD structure as all the user mode 319 * queues but with different initial values. 320 */ 321 322 static int init_mqd_hiq(struct mqd_manager *mm, void **mqd, 323 struct kfd_mem_obj **mqd_mem_obj, uint64_t *gart_addr, 324 struct queue_properties *q) 325 { 326 uint64_t addr; 327 struct cik_mqd *m; 328 int retval; 329 330 retval = kfd_gtt_sa_allocate(mm->dev, sizeof(struct cik_mqd), 331 mqd_mem_obj); 332 333 if (retval != 0) 334 return -ENOMEM; 335 336 m = (struct cik_mqd *) (*mqd_mem_obj)->cpu_ptr; 337 addr = (*mqd_mem_obj)->gpu_addr; 338 339 memset(m, 0, ALIGN(sizeof(struct cik_mqd), 256)); 340 341 m->header = 0xC0310800; 342 m->compute_pipelinestat_enable = 1; 343 m->compute_static_thread_mgmt_se0 = 0xFFFFFFFF; 344 m->compute_static_thread_mgmt_se1 = 0xFFFFFFFF; 345 m->compute_static_thread_mgmt_se2 = 0xFFFFFFFF; 346 m->compute_static_thread_mgmt_se3 = 0xFFFFFFFF; 347 348 m->cp_hqd_persistent_state = DEFAULT_CP_HQD_PERSISTENT_STATE | 349 PRELOAD_REQ; 350 m->cp_hqd_quantum = QUANTUM_EN | QUANTUM_SCALE_1MS | 351 QUANTUM_DURATION(10); 352 353 m->cp_mqd_control = MQD_CONTROL_PRIV_STATE_EN; 354 m->cp_mqd_base_addr_lo = lower_32_bits(addr); 355 m->cp_mqd_base_addr_hi = upper_32_bits(addr); 356 357 m->cp_hqd_ib_control = DEFAULT_MIN_IB_AVAIL_SIZE; 358 359 /* 360 * Pipe Priority 361 * Identifies the pipe relative priority when this queue is connected 362 * to the pipeline. The pipe priority is against the GFX pipe and HP3D. 363 * In KFD we are using a fixed pipe priority set to CS_MEDIUM. 364 * 0 = CS_LOW (typically below GFX) 365 * 1 = CS_MEDIUM (typically between HP3D and GFX 366 * 2 = CS_HIGH (typically above HP3D) 367 */ 368 m->cp_hqd_pipe_priority = 1; 369 m->cp_hqd_queue_priority = 15; 370 371 *mqd = m; 372 if (gart_addr) 373 *gart_addr = addr; 374 retval = mm->update_mqd(mm, m, q); 375 376 return retval; 377 } 378 379 static int update_mqd_hiq(struct mqd_manager *mm, void *mqd, 380 struct queue_properties *q) 381 { 382 struct cik_mqd *m; 383 384 m = get_mqd(mqd); 385 m->cp_hqd_pq_control = DEFAULT_RPTR_BLOCK_SIZE | 386 DEFAULT_MIN_AVAIL_SIZE | 387 PRIV_STATE | 388 KMD_QUEUE; 389 390 /* 391 * Calculating queue size which is log base 2 of actual queue 392 * size -1 dwords 393 */ 394 m->cp_hqd_pq_control |= order_base_2(q->queue_size / 4) - 1; 395 m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8); 396 m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8); 397 m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr); 398 m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr); 399 m->cp_hqd_pq_doorbell_control = DOORBELL_OFFSET(q->doorbell_off); 400 401 m->cp_hqd_vmid = q->vmid; 402 403 q->is_active = (q->queue_size > 0 && 404 q->queue_address != 0 && 405 q->queue_percent > 0 && 406 !q->is_evicted); 407 408 return 0; 409 } 410 411 #if defined(CONFIG_DEBUG_FS) 412 413 static int debugfs_show_mqd(struct seq_file *m, void *data) 414 { 415 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4, 416 data, sizeof(struct cik_mqd), false); 417 return 0; 418 } 419 420 static int debugfs_show_mqd_sdma(struct seq_file *m, void *data) 421 { 422 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4, 423 data, sizeof(struct cik_sdma_rlc_registers), false); 424 return 0; 425 } 426 427 #endif 428 429 430 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, 431 struct kfd_dev *dev) 432 { 433 struct mqd_manager *mqd; 434 435 if (WARN_ON(type >= KFD_MQD_TYPE_MAX)) 436 return NULL; 437 438 mqd = kzalloc(sizeof(*mqd), GFP_KERNEL); 439 if (!mqd) 440 return NULL; 441 442 mqd->dev = dev; 443 444 switch (type) { 445 case KFD_MQD_TYPE_CP: 446 case KFD_MQD_TYPE_COMPUTE: 447 mqd->init_mqd = init_mqd; 448 mqd->uninit_mqd = uninit_mqd; 449 mqd->load_mqd = load_mqd; 450 mqd->update_mqd = update_mqd; 451 mqd->destroy_mqd = destroy_mqd; 452 mqd->is_occupied = is_occupied; 453 #if defined(CONFIG_DEBUG_FS) 454 mqd->debugfs_show_mqd = debugfs_show_mqd; 455 #endif 456 break; 457 case KFD_MQD_TYPE_HIQ: 458 mqd->init_mqd = init_mqd_hiq; 459 mqd->uninit_mqd = uninit_mqd; 460 mqd->load_mqd = load_mqd; 461 mqd->update_mqd = update_mqd_hiq; 462 mqd->destroy_mqd = destroy_mqd; 463 mqd->is_occupied = is_occupied; 464 #if defined(CONFIG_DEBUG_FS) 465 mqd->debugfs_show_mqd = debugfs_show_mqd; 466 #endif 467 break; 468 case KFD_MQD_TYPE_SDMA: 469 mqd->init_mqd = init_mqd_sdma; 470 mqd->uninit_mqd = uninit_mqd_sdma; 471 mqd->load_mqd = load_mqd_sdma; 472 mqd->update_mqd = update_mqd_sdma; 473 mqd->destroy_mqd = destroy_mqd_sdma; 474 mqd->is_occupied = is_occupied_sdma; 475 #if defined(CONFIG_DEBUG_FS) 476 mqd->debugfs_show_mqd = debugfs_show_mqd_sdma; 477 #endif 478 break; 479 default: 480 kfree(mqd); 481 return NULL; 482 } 483 484 return mqd; 485 } 486 487 struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type, 488 struct kfd_dev *dev) 489 { 490 struct mqd_manager *mqd; 491 492 mqd = mqd_manager_init_cik(type, dev); 493 if (!mqd) 494 return NULL; 495 if ((type == KFD_MQD_TYPE_CP) || (type == KFD_MQD_TYPE_COMPUTE)) 496 mqd->update_mqd = update_mqd_hawaii; 497 return mqd; 498 } 499