1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2021 Intel Corporation 4 */ 5 6 #include "xe_exec_queue.h" 7 8 #include <linux/nospec.h> 9 10 #include <drm/drm_device.h> 11 #include <drm/drm_file.h> 12 #include <uapi/drm/xe_drm.h> 13 14 #include "xe_device.h" 15 #include "xe_gt.h" 16 #include "xe_hw_engine_class_sysfs.h" 17 #include "xe_hw_engine_group.h" 18 #include "xe_hw_fence.h" 19 #include "xe_lrc.h" 20 #include "xe_macros.h" 21 #include "xe_migrate.h" 22 #include "xe_pm.h" 23 #include "xe_ring_ops_types.h" 24 #include "xe_trace.h" 25 #include "xe_vm.h" 26 27 enum xe_exec_queue_sched_prop { 28 XE_EXEC_QUEUE_JOB_TIMEOUT = 0, 29 XE_EXEC_QUEUE_TIMESLICE = 1, 30 XE_EXEC_QUEUE_PREEMPT_TIMEOUT = 2, 31 XE_EXEC_QUEUE_SCHED_PROP_MAX = 3, 32 }; 33 34 static int exec_queue_user_extensions(struct xe_device *xe, struct xe_exec_queue *q, 35 u64 extensions, int ext_number); 36 37 static void __xe_exec_queue_free(struct xe_exec_queue *q) 38 { 39 if (q->vm) 40 xe_vm_put(q->vm); 41 42 if (q->xef) 43 xe_file_put(q->xef); 44 45 kfree(q); 46 } 47 48 static struct xe_exec_queue *__xe_exec_queue_alloc(struct xe_device *xe, 49 struct xe_vm *vm, 50 u32 logical_mask, 51 u16 width, struct xe_hw_engine *hwe, 52 u32 flags, u64 extensions) 53 { 54 struct xe_exec_queue *q; 55 struct xe_gt *gt = hwe->gt; 56 int err; 57 58 /* only kernel queues can be permanent */ 59 XE_WARN_ON((flags & EXEC_QUEUE_FLAG_PERMANENT) && !(flags & EXEC_QUEUE_FLAG_KERNEL)); 60 61 q = kzalloc(struct_size(q, lrc, width), GFP_KERNEL); 62 if (!q) 63 return ERR_PTR(-ENOMEM); 64 65 kref_init(&q->refcount); 66 q->flags = flags; 67 q->hwe = hwe; 68 q->gt = gt; 69 q->class = hwe->class; 70 q->width = width; 71 q->logical_mask = logical_mask; 72 q->fence_irq = >->fence_irq[hwe->class]; 73 q->ring_ops = gt->ring_ops[hwe->class]; 74 q->ops = gt->exec_queue_ops; 75 INIT_LIST_HEAD(&q->lr.link); 76 INIT_LIST_HEAD(&q->multi_gt_link); 77 INIT_LIST_HEAD(&q->hw_engine_group_link); 78 79 q->sched_props.timeslice_us = hwe->eclass->sched_props.timeslice_us; 80 q->sched_props.preempt_timeout_us = 81 hwe->eclass->sched_props.preempt_timeout_us; 82 q->sched_props.job_timeout_ms = 83 hwe->eclass->sched_props.job_timeout_ms; 84 if (q->flags & EXEC_QUEUE_FLAG_KERNEL && 85 q->flags & EXEC_QUEUE_FLAG_HIGH_PRIORITY) 86 q->sched_props.priority = XE_EXEC_QUEUE_PRIORITY_KERNEL; 87 else 88 q->sched_props.priority = XE_EXEC_QUEUE_PRIORITY_NORMAL; 89 90 if (vm) 91 q->vm = xe_vm_get(vm); 92 93 if (extensions) { 94 /* 95 * may set q->usm, must come before xe_lrc_create(), 96 * may overwrite q->sched_props, must come before q->ops->init() 97 */ 98 err = exec_queue_user_extensions(xe, q, extensions, 0); 99 if (err) { 100 __xe_exec_queue_free(q); 101 return ERR_PTR(err); 102 } 103 } 104 105 return q; 106 } 107 108 static int __xe_exec_queue_init(struct xe_exec_queue *q) 109 { 110 struct xe_vm *vm = q->vm; 111 int i, err; 112 113 if (vm) { 114 err = xe_vm_lock(vm, true); 115 if (err) 116 return err; 117 } 118 119 for (i = 0; i < q->width; ++i) { 120 q->lrc[i] = xe_lrc_create(q->hwe, q->vm, SZ_16K); 121 if (IS_ERR(q->lrc[i])) { 122 err = PTR_ERR(q->lrc[i]); 123 goto err_unlock; 124 } 125 } 126 127 if (vm) 128 xe_vm_unlock(vm); 129 130 err = q->ops->init(q); 131 if (err) 132 goto err_lrc; 133 134 return 0; 135 136 err_unlock: 137 if (vm) 138 xe_vm_unlock(vm); 139 err_lrc: 140 for (i = i - 1; i >= 0; --i) 141 xe_lrc_put(q->lrc[i]); 142 return err; 143 } 144 145 struct xe_exec_queue *xe_exec_queue_create(struct xe_device *xe, struct xe_vm *vm, 146 u32 logical_mask, u16 width, 147 struct xe_hw_engine *hwe, u32 flags, 148 u64 extensions) 149 { 150 struct xe_exec_queue *q; 151 int err; 152 153 q = __xe_exec_queue_alloc(xe, vm, logical_mask, width, hwe, flags, 154 extensions); 155 if (IS_ERR(q)) 156 return q; 157 158 err = __xe_exec_queue_init(q); 159 if (err) 160 goto err_post_alloc; 161 162 return q; 163 164 err_post_alloc: 165 __xe_exec_queue_free(q); 166 return ERR_PTR(err); 167 } 168 169 struct xe_exec_queue *xe_exec_queue_create_class(struct xe_device *xe, struct xe_gt *gt, 170 struct xe_vm *vm, 171 enum xe_engine_class class, 172 u32 flags, u64 extensions) 173 { 174 struct xe_hw_engine *hwe, *hwe0 = NULL; 175 enum xe_hw_engine_id id; 176 u32 logical_mask = 0; 177 178 for_each_hw_engine(hwe, gt, id) { 179 if (xe_hw_engine_is_reserved(hwe)) 180 continue; 181 182 if (hwe->class == class) { 183 logical_mask |= BIT(hwe->logical_instance); 184 if (!hwe0) 185 hwe0 = hwe; 186 } 187 } 188 189 if (!logical_mask) 190 return ERR_PTR(-ENODEV); 191 192 return xe_exec_queue_create(xe, vm, logical_mask, 1, hwe0, flags, extensions); 193 } 194 195 /** 196 * xe_exec_queue_create_bind() - Create bind exec queue. 197 * @xe: Xe device. 198 * @tile: tile which bind exec queue belongs to. 199 * @flags: exec queue creation flags 200 * @extensions: exec queue creation extensions 201 * 202 * Normalize bind exec queue creation. Bind exec queue is tied to migration VM 203 * for access to physical memory required for page table programming. On a 204 * faulting devices the reserved copy engine instance must be used to avoid 205 * deadlocking (user binds cannot get stuck behind faults as kernel binds which 206 * resolve faults depend on user binds). On non-faulting devices any copy engine 207 * can be used. 208 * 209 * Returns exec queue on success, ERR_PTR on failure 210 */ 211 struct xe_exec_queue *xe_exec_queue_create_bind(struct xe_device *xe, 212 struct xe_tile *tile, 213 u32 flags, u64 extensions) 214 { 215 struct xe_gt *gt = tile->primary_gt; 216 struct xe_exec_queue *q; 217 struct xe_vm *migrate_vm; 218 219 migrate_vm = xe_migrate_get_vm(tile->migrate); 220 if (xe->info.has_usm) { 221 struct xe_hw_engine *hwe = xe_gt_hw_engine(gt, 222 XE_ENGINE_CLASS_COPY, 223 gt->usm.reserved_bcs_instance, 224 false); 225 226 if (!hwe) { 227 xe_vm_put(migrate_vm); 228 return ERR_PTR(-EINVAL); 229 } 230 231 q = xe_exec_queue_create(xe, migrate_vm, 232 BIT(hwe->logical_instance), 1, hwe, 233 flags, extensions); 234 } else { 235 q = xe_exec_queue_create_class(xe, gt, migrate_vm, 236 XE_ENGINE_CLASS_COPY, flags, 237 extensions); 238 } 239 xe_vm_put(migrate_vm); 240 241 return q; 242 } 243 244 void xe_exec_queue_destroy(struct kref *ref) 245 { 246 struct xe_exec_queue *q = container_of(ref, struct xe_exec_queue, refcount); 247 struct xe_exec_queue *eq, *next; 248 249 xe_exec_queue_last_fence_put_unlocked(q); 250 if (!(q->flags & EXEC_QUEUE_FLAG_BIND_ENGINE_CHILD)) { 251 list_for_each_entry_safe(eq, next, &q->multi_gt_list, 252 multi_gt_link) 253 xe_exec_queue_put(eq); 254 } 255 256 q->ops->fini(q); 257 } 258 259 void xe_exec_queue_fini(struct xe_exec_queue *q) 260 { 261 int i; 262 263 /* 264 * Before releasing our ref to lrc and xef, accumulate our run ticks 265 */ 266 xe_exec_queue_update_run_ticks(q); 267 268 for (i = 0; i < q->width; ++i) 269 xe_lrc_put(q->lrc[i]); 270 271 __xe_exec_queue_free(q); 272 } 273 274 void xe_exec_queue_assign_name(struct xe_exec_queue *q, u32 instance) 275 { 276 switch (q->class) { 277 case XE_ENGINE_CLASS_RENDER: 278 snprintf(q->name, sizeof(q->name), "rcs%d", instance); 279 break; 280 case XE_ENGINE_CLASS_VIDEO_DECODE: 281 snprintf(q->name, sizeof(q->name), "vcs%d", instance); 282 break; 283 case XE_ENGINE_CLASS_VIDEO_ENHANCE: 284 snprintf(q->name, sizeof(q->name), "vecs%d", instance); 285 break; 286 case XE_ENGINE_CLASS_COPY: 287 snprintf(q->name, sizeof(q->name), "bcs%d", instance); 288 break; 289 case XE_ENGINE_CLASS_COMPUTE: 290 snprintf(q->name, sizeof(q->name), "ccs%d", instance); 291 break; 292 case XE_ENGINE_CLASS_OTHER: 293 snprintf(q->name, sizeof(q->name), "gsccs%d", instance); 294 break; 295 default: 296 XE_WARN_ON(q->class); 297 } 298 } 299 300 struct xe_exec_queue *xe_exec_queue_lookup(struct xe_file *xef, u32 id) 301 { 302 struct xe_exec_queue *q; 303 304 mutex_lock(&xef->exec_queue.lock); 305 q = xa_load(&xef->exec_queue.xa, id); 306 if (q) 307 xe_exec_queue_get(q); 308 mutex_unlock(&xef->exec_queue.lock); 309 310 return q; 311 } 312 313 enum xe_exec_queue_priority 314 xe_exec_queue_device_get_max_priority(struct xe_device *xe) 315 { 316 return capable(CAP_SYS_NICE) ? XE_EXEC_QUEUE_PRIORITY_HIGH : 317 XE_EXEC_QUEUE_PRIORITY_NORMAL; 318 } 319 320 static int exec_queue_set_priority(struct xe_device *xe, struct xe_exec_queue *q, 321 u64 value) 322 { 323 if (XE_IOCTL_DBG(xe, value > XE_EXEC_QUEUE_PRIORITY_HIGH)) 324 return -EINVAL; 325 326 if (XE_IOCTL_DBG(xe, value > xe_exec_queue_device_get_max_priority(xe))) 327 return -EPERM; 328 329 q->sched_props.priority = value; 330 return 0; 331 } 332 333 static bool xe_exec_queue_enforce_schedule_limit(void) 334 { 335 #if IS_ENABLED(CONFIG_DRM_XE_ENABLE_SCHEDTIMEOUT_LIMIT) 336 return true; 337 #else 338 return !capable(CAP_SYS_NICE); 339 #endif 340 } 341 342 static void 343 xe_exec_queue_get_prop_minmax(struct xe_hw_engine_class_intf *eclass, 344 enum xe_exec_queue_sched_prop prop, 345 u32 *min, u32 *max) 346 { 347 switch (prop) { 348 case XE_EXEC_QUEUE_JOB_TIMEOUT: 349 *min = eclass->sched_props.job_timeout_min; 350 *max = eclass->sched_props.job_timeout_max; 351 break; 352 case XE_EXEC_QUEUE_TIMESLICE: 353 *min = eclass->sched_props.timeslice_min; 354 *max = eclass->sched_props.timeslice_max; 355 break; 356 case XE_EXEC_QUEUE_PREEMPT_TIMEOUT: 357 *min = eclass->sched_props.preempt_timeout_min; 358 *max = eclass->sched_props.preempt_timeout_max; 359 break; 360 default: 361 break; 362 } 363 #if IS_ENABLED(CONFIG_DRM_XE_ENABLE_SCHEDTIMEOUT_LIMIT) 364 if (capable(CAP_SYS_NICE)) { 365 switch (prop) { 366 case XE_EXEC_QUEUE_JOB_TIMEOUT: 367 *min = XE_HW_ENGINE_JOB_TIMEOUT_MIN; 368 *max = XE_HW_ENGINE_JOB_TIMEOUT_MAX; 369 break; 370 case XE_EXEC_QUEUE_TIMESLICE: 371 *min = XE_HW_ENGINE_TIMESLICE_MIN; 372 *max = XE_HW_ENGINE_TIMESLICE_MAX; 373 break; 374 case XE_EXEC_QUEUE_PREEMPT_TIMEOUT: 375 *min = XE_HW_ENGINE_PREEMPT_TIMEOUT_MIN; 376 *max = XE_HW_ENGINE_PREEMPT_TIMEOUT_MAX; 377 break; 378 default: 379 break; 380 } 381 } 382 #endif 383 } 384 385 static int exec_queue_set_timeslice(struct xe_device *xe, struct xe_exec_queue *q, 386 u64 value) 387 { 388 u32 min = 0, max = 0; 389 390 xe_exec_queue_get_prop_minmax(q->hwe->eclass, 391 XE_EXEC_QUEUE_TIMESLICE, &min, &max); 392 393 if (xe_exec_queue_enforce_schedule_limit() && 394 !xe_hw_engine_timeout_in_range(value, min, max)) 395 return -EINVAL; 396 397 q->sched_props.timeslice_us = value; 398 return 0; 399 } 400 401 typedef int (*xe_exec_queue_set_property_fn)(struct xe_device *xe, 402 struct xe_exec_queue *q, 403 u64 value); 404 405 static const xe_exec_queue_set_property_fn exec_queue_set_property_funcs[] = { 406 [DRM_XE_EXEC_QUEUE_SET_PROPERTY_PRIORITY] = exec_queue_set_priority, 407 [DRM_XE_EXEC_QUEUE_SET_PROPERTY_TIMESLICE] = exec_queue_set_timeslice, 408 }; 409 410 static int exec_queue_user_ext_set_property(struct xe_device *xe, 411 struct xe_exec_queue *q, 412 u64 extension) 413 { 414 u64 __user *address = u64_to_user_ptr(extension); 415 struct drm_xe_ext_set_property ext; 416 int err; 417 u32 idx; 418 419 err = __copy_from_user(&ext, address, sizeof(ext)); 420 if (XE_IOCTL_DBG(xe, err)) 421 return -EFAULT; 422 423 if (XE_IOCTL_DBG(xe, ext.property >= 424 ARRAY_SIZE(exec_queue_set_property_funcs)) || 425 XE_IOCTL_DBG(xe, ext.pad) || 426 XE_IOCTL_DBG(xe, ext.property != DRM_XE_EXEC_QUEUE_SET_PROPERTY_PRIORITY && 427 ext.property != DRM_XE_EXEC_QUEUE_SET_PROPERTY_TIMESLICE)) 428 return -EINVAL; 429 430 idx = array_index_nospec(ext.property, ARRAY_SIZE(exec_queue_set_property_funcs)); 431 if (!exec_queue_set_property_funcs[idx]) 432 return -EINVAL; 433 434 return exec_queue_set_property_funcs[idx](xe, q, ext.value); 435 } 436 437 typedef int (*xe_exec_queue_user_extension_fn)(struct xe_device *xe, 438 struct xe_exec_queue *q, 439 u64 extension); 440 441 static const xe_exec_queue_user_extension_fn exec_queue_user_extension_funcs[] = { 442 [DRM_XE_EXEC_QUEUE_EXTENSION_SET_PROPERTY] = exec_queue_user_ext_set_property, 443 }; 444 445 #define MAX_USER_EXTENSIONS 16 446 static int exec_queue_user_extensions(struct xe_device *xe, struct xe_exec_queue *q, 447 u64 extensions, int ext_number) 448 { 449 u64 __user *address = u64_to_user_ptr(extensions); 450 struct drm_xe_user_extension ext; 451 int err; 452 u32 idx; 453 454 if (XE_IOCTL_DBG(xe, ext_number >= MAX_USER_EXTENSIONS)) 455 return -E2BIG; 456 457 err = __copy_from_user(&ext, address, sizeof(ext)); 458 if (XE_IOCTL_DBG(xe, err)) 459 return -EFAULT; 460 461 if (XE_IOCTL_DBG(xe, ext.pad) || 462 XE_IOCTL_DBG(xe, ext.name >= 463 ARRAY_SIZE(exec_queue_user_extension_funcs))) 464 return -EINVAL; 465 466 idx = array_index_nospec(ext.name, 467 ARRAY_SIZE(exec_queue_user_extension_funcs)); 468 err = exec_queue_user_extension_funcs[idx](xe, q, extensions); 469 if (XE_IOCTL_DBG(xe, err)) 470 return err; 471 472 if (ext.next_extension) 473 return exec_queue_user_extensions(xe, q, ext.next_extension, 474 ++ext_number); 475 476 return 0; 477 } 478 479 static u32 calc_validate_logical_mask(struct xe_device *xe, struct xe_gt *gt, 480 struct drm_xe_engine_class_instance *eci, 481 u16 width, u16 num_placements) 482 { 483 int len = width * num_placements; 484 int i, j, n; 485 u16 class; 486 u16 gt_id; 487 u32 return_mask = 0, prev_mask; 488 489 if (XE_IOCTL_DBG(xe, !xe_device_uc_enabled(xe) && 490 len > 1)) 491 return 0; 492 493 for (i = 0; i < width; ++i) { 494 u32 current_mask = 0; 495 496 for (j = 0; j < num_placements; ++j) { 497 struct xe_hw_engine *hwe; 498 499 n = j * width + i; 500 501 hwe = xe_hw_engine_lookup(xe, eci[n]); 502 if (XE_IOCTL_DBG(xe, !hwe)) 503 return 0; 504 505 if (XE_IOCTL_DBG(xe, xe_hw_engine_is_reserved(hwe))) 506 return 0; 507 508 if (XE_IOCTL_DBG(xe, n && eci[n].gt_id != gt_id) || 509 XE_IOCTL_DBG(xe, n && eci[n].engine_class != class)) 510 return 0; 511 512 class = eci[n].engine_class; 513 gt_id = eci[n].gt_id; 514 515 if (width == 1 || !i) 516 return_mask |= BIT(eci[n].engine_instance); 517 current_mask |= BIT(eci[n].engine_instance); 518 } 519 520 /* Parallel submissions must be logically contiguous */ 521 if (i && XE_IOCTL_DBG(xe, current_mask != prev_mask << 1)) 522 return 0; 523 524 prev_mask = current_mask; 525 } 526 527 return return_mask; 528 } 529 530 int xe_exec_queue_create_ioctl(struct drm_device *dev, void *data, 531 struct drm_file *file) 532 { 533 struct xe_device *xe = to_xe_device(dev); 534 struct xe_file *xef = to_xe_file(file); 535 struct drm_xe_exec_queue_create *args = data; 536 struct drm_xe_engine_class_instance eci[XE_HW_ENGINE_MAX_INSTANCE]; 537 struct drm_xe_engine_class_instance __user *user_eci = 538 u64_to_user_ptr(args->instances); 539 struct xe_hw_engine *hwe; 540 struct xe_vm *vm; 541 struct xe_gt *gt; 542 struct xe_tile *tile; 543 struct xe_exec_queue *q = NULL; 544 u32 logical_mask; 545 u32 id; 546 u32 len; 547 int err; 548 549 if (XE_IOCTL_DBG(xe, args->flags) || 550 XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1])) 551 return -EINVAL; 552 553 len = args->width * args->num_placements; 554 if (XE_IOCTL_DBG(xe, !len || len > XE_HW_ENGINE_MAX_INSTANCE)) 555 return -EINVAL; 556 557 err = __copy_from_user(eci, user_eci, 558 sizeof(struct drm_xe_engine_class_instance) * 559 len); 560 if (XE_IOCTL_DBG(xe, err)) 561 return -EFAULT; 562 563 if (XE_IOCTL_DBG(xe, eci[0].gt_id >= xe->info.gt_count)) 564 return -EINVAL; 565 566 if (eci[0].engine_class == DRM_XE_ENGINE_CLASS_VM_BIND) { 567 if (XE_IOCTL_DBG(xe, args->width != 1) || 568 XE_IOCTL_DBG(xe, args->num_placements != 1) || 569 XE_IOCTL_DBG(xe, eci[0].engine_instance != 0)) 570 return -EINVAL; 571 572 for_each_tile(tile, xe, id) { 573 struct xe_exec_queue *new; 574 u32 flags = EXEC_QUEUE_FLAG_VM; 575 576 if (id) 577 flags |= EXEC_QUEUE_FLAG_BIND_ENGINE_CHILD; 578 579 new = xe_exec_queue_create_bind(xe, tile, flags, 580 args->extensions); 581 if (IS_ERR(new)) { 582 err = PTR_ERR(new); 583 if (q) 584 goto put_exec_queue; 585 return err; 586 } 587 if (id == 0) 588 q = new; 589 else 590 list_add_tail(&new->multi_gt_list, 591 &q->multi_gt_link); 592 } 593 } else { 594 gt = xe_device_get_gt(xe, eci[0].gt_id); 595 logical_mask = calc_validate_logical_mask(xe, gt, eci, 596 args->width, 597 args->num_placements); 598 if (XE_IOCTL_DBG(xe, !logical_mask)) 599 return -EINVAL; 600 601 hwe = xe_hw_engine_lookup(xe, eci[0]); 602 if (XE_IOCTL_DBG(xe, !hwe)) 603 return -EINVAL; 604 605 vm = xe_vm_lookup(xef, args->vm_id); 606 if (XE_IOCTL_DBG(xe, !vm)) 607 return -ENOENT; 608 609 err = down_read_interruptible(&vm->lock); 610 if (err) { 611 xe_vm_put(vm); 612 return err; 613 } 614 615 if (XE_IOCTL_DBG(xe, xe_vm_is_closed_or_banned(vm))) { 616 up_read(&vm->lock); 617 xe_vm_put(vm); 618 return -ENOENT; 619 } 620 621 q = xe_exec_queue_create(xe, vm, logical_mask, 622 args->width, hwe, 0, 623 args->extensions); 624 up_read(&vm->lock); 625 xe_vm_put(vm); 626 if (IS_ERR(q)) 627 return PTR_ERR(q); 628 629 if (xe_vm_in_preempt_fence_mode(vm)) { 630 q->lr.context = dma_fence_context_alloc(1); 631 632 err = xe_vm_add_compute_exec_queue(vm, q); 633 if (XE_IOCTL_DBG(xe, err)) 634 goto put_exec_queue; 635 } 636 637 if (q->vm && q->hwe->hw_engine_group) { 638 err = xe_hw_engine_group_add_exec_queue(q->hwe->hw_engine_group, q); 639 if (err) 640 goto put_exec_queue; 641 } 642 } 643 644 q->xef = xe_file_get(xef); 645 646 /* user id alloc must always be last in ioctl to prevent UAF */ 647 err = xa_alloc(&xef->exec_queue.xa, &id, q, xa_limit_32b, GFP_KERNEL); 648 if (err) 649 goto kill_exec_queue; 650 651 args->exec_queue_id = id; 652 653 return 0; 654 655 kill_exec_queue: 656 xe_exec_queue_kill(q); 657 put_exec_queue: 658 xe_exec_queue_put(q); 659 return err; 660 } 661 662 int xe_exec_queue_get_property_ioctl(struct drm_device *dev, void *data, 663 struct drm_file *file) 664 { 665 struct xe_device *xe = to_xe_device(dev); 666 struct xe_file *xef = to_xe_file(file); 667 struct drm_xe_exec_queue_get_property *args = data; 668 struct xe_exec_queue *q; 669 int ret; 670 671 if (XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1])) 672 return -EINVAL; 673 674 q = xe_exec_queue_lookup(xef, args->exec_queue_id); 675 if (XE_IOCTL_DBG(xe, !q)) 676 return -ENOENT; 677 678 switch (args->property) { 679 case DRM_XE_EXEC_QUEUE_GET_PROPERTY_BAN: 680 args->value = q->ops->reset_status(q); 681 ret = 0; 682 break; 683 default: 684 ret = -EINVAL; 685 } 686 687 xe_exec_queue_put(q); 688 689 return ret; 690 } 691 692 /** 693 * xe_exec_queue_is_lr() - Whether an exec_queue is long-running 694 * @q: The exec_queue 695 * 696 * Return: True if the exec_queue is long-running, false otherwise. 697 */ 698 bool xe_exec_queue_is_lr(struct xe_exec_queue *q) 699 { 700 return q->vm && xe_vm_in_lr_mode(q->vm) && 701 !(q->flags & EXEC_QUEUE_FLAG_VM); 702 } 703 704 static s32 xe_exec_queue_num_job_inflight(struct xe_exec_queue *q) 705 { 706 return q->lrc[0]->fence_ctx.next_seqno - xe_lrc_seqno(q->lrc[0]) - 1; 707 } 708 709 /** 710 * xe_exec_queue_ring_full() - Whether an exec_queue's ring is full 711 * @q: The exec_queue 712 * 713 * Return: True if the exec_queue's ring is full, false otherwise. 714 */ 715 bool xe_exec_queue_ring_full(struct xe_exec_queue *q) 716 { 717 struct xe_lrc *lrc = q->lrc[0]; 718 s32 max_job = lrc->ring.size / MAX_JOB_SIZE_BYTES; 719 720 return xe_exec_queue_num_job_inflight(q) >= max_job; 721 } 722 723 /** 724 * xe_exec_queue_is_idle() - Whether an exec_queue is idle. 725 * @q: The exec_queue 726 * 727 * FIXME: Need to determine what to use as the short-lived 728 * timeline lock for the exec_queues, so that the return value 729 * of this function becomes more than just an advisory 730 * snapshot in time. The timeline lock must protect the 731 * seqno from racing submissions on the same exec_queue. 732 * Typically vm->resv, but user-created timeline locks use the migrate vm 733 * and never grabs the migrate vm->resv so we have a race there. 734 * 735 * Return: True if the exec_queue is idle, false otherwise. 736 */ 737 bool xe_exec_queue_is_idle(struct xe_exec_queue *q) 738 { 739 if (xe_exec_queue_is_parallel(q)) { 740 int i; 741 742 for (i = 0; i < q->width; ++i) { 743 if (xe_lrc_seqno(q->lrc[i]) != 744 q->lrc[i]->fence_ctx.next_seqno - 1) 745 return false; 746 } 747 748 return true; 749 } 750 751 return xe_lrc_seqno(q->lrc[0]) == 752 q->lrc[0]->fence_ctx.next_seqno - 1; 753 } 754 755 /** 756 * xe_exec_queue_update_run_ticks() - Update run time in ticks for this exec queue 757 * from hw 758 * @q: The exec queue 759 * 760 * Update the timestamp saved by HW for this exec queue and save run ticks 761 * calculated by using the delta from last update. 762 */ 763 void xe_exec_queue_update_run_ticks(struct xe_exec_queue *q) 764 { 765 struct xe_file *xef; 766 struct xe_lrc *lrc; 767 u32 old_ts, new_ts; 768 769 /* 770 * Jobs that are run during driver load may use an exec_queue, but are 771 * not associated with a user xe file, so avoid accumulating busyness 772 * for kernel specific work. 773 */ 774 if (!q->vm || !q->vm->xef) 775 return; 776 777 xef = q->vm->xef; 778 779 /* 780 * Only sample the first LRC. For parallel submission, all of them are 781 * scheduled together and we compensate that below by multiplying by 782 * width - this may introduce errors if that premise is not true and 783 * they don't exit 100% aligned. On the other hand, looping through 784 * the LRCs and reading them in different time could also introduce 785 * errors. 786 */ 787 lrc = q->lrc[0]; 788 new_ts = xe_lrc_update_timestamp(lrc, &old_ts); 789 xef->run_ticks[q->class] += (new_ts - old_ts) * q->width; 790 } 791 792 /** 793 * xe_exec_queue_kill - permanently stop all execution from an exec queue 794 * @q: The exec queue 795 * 796 * This function permanently stops all activity on an exec queue. If the queue 797 * is actively executing on the HW, it will be kicked off the engine; any 798 * pending jobs are discarded and all future submissions are rejected. 799 * This function is safe to call multiple times. 800 */ 801 void xe_exec_queue_kill(struct xe_exec_queue *q) 802 { 803 struct xe_exec_queue *eq = q, *next; 804 805 list_for_each_entry_safe(eq, next, &eq->multi_gt_list, 806 multi_gt_link) { 807 q->ops->kill(eq); 808 xe_vm_remove_compute_exec_queue(q->vm, eq); 809 } 810 811 q->ops->kill(q); 812 xe_vm_remove_compute_exec_queue(q->vm, q); 813 } 814 815 int xe_exec_queue_destroy_ioctl(struct drm_device *dev, void *data, 816 struct drm_file *file) 817 { 818 struct xe_device *xe = to_xe_device(dev); 819 struct xe_file *xef = to_xe_file(file); 820 struct drm_xe_exec_queue_destroy *args = data; 821 struct xe_exec_queue *q; 822 823 if (XE_IOCTL_DBG(xe, args->pad) || 824 XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1])) 825 return -EINVAL; 826 827 mutex_lock(&xef->exec_queue.lock); 828 q = xa_erase(&xef->exec_queue.xa, args->exec_queue_id); 829 mutex_unlock(&xef->exec_queue.lock); 830 if (XE_IOCTL_DBG(xe, !q)) 831 return -ENOENT; 832 833 if (q->vm && q->hwe->hw_engine_group) 834 xe_hw_engine_group_del_exec_queue(q->hwe->hw_engine_group, q); 835 836 xe_exec_queue_kill(q); 837 838 trace_xe_exec_queue_close(q); 839 xe_exec_queue_put(q); 840 841 return 0; 842 } 843 844 static void xe_exec_queue_last_fence_lockdep_assert(struct xe_exec_queue *q, 845 struct xe_vm *vm) 846 { 847 if (q->flags & EXEC_QUEUE_FLAG_VM) { 848 lockdep_assert_held(&vm->lock); 849 } else { 850 xe_vm_assert_held(vm); 851 lockdep_assert_held(&q->hwe->hw_engine_group->mode_sem); 852 } 853 } 854 855 /** 856 * xe_exec_queue_last_fence_put() - Drop ref to last fence 857 * @q: The exec queue 858 * @vm: The VM the engine does a bind or exec for 859 */ 860 void xe_exec_queue_last_fence_put(struct xe_exec_queue *q, struct xe_vm *vm) 861 { 862 xe_exec_queue_last_fence_lockdep_assert(q, vm); 863 864 xe_exec_queue_last_fence_put_unlocked(q); 865 } 866 867 /** 868 * xe_exec_queue_last_fence_put_unlocked() - Drop ref to last fence unlocked 869 * @q: The exec queue 870 * 871 * Only safe to be called from xe_exec_queue_destroy(). 872 */ 873 void xe_exec_queue_last_fence_put_unlocked(struct xe_exec_queue *q) 874 { 875 if (q->last_fence) { 876 dma_fence_put(q->last_fence); 877 q->last_fence = NULL; 878 } 879 } 880 881 /** 882 * xe_exec_queue_last_fence_get() - Get last fence 883 * @q: The exec queue 884 * @vm: The VM the engine does a bind or exec for 885 * 886 * Get last fence, takes a ref 887 * 888 * Returns: last fence if not signaled, dma fence stub if signaled 889 */ 890 struct dma_fence *xe_exec_queue_last_fence_get(struct xe_exec_queue *q, 891 struct xe_vm *vm) 892 { 893 struct dma_fence *fence; 894 895 xe_exec_queue_last_fence_lockdep_assert(q, vm); 896 897 if (q->last_fence && 898 test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &q->last_fence->flags)) 899 xe_exec_queue_last_fence_put(q, vm); 900 901 fence = q->last_fence ? q->last_fence : dma_fence_get_stub(); 902 dma_fence_get(fence); 903 return fence; 904 } 905 906 /** 907 * xe_exec_queue_last_fence_get_for_resume() - Get last fence 908 * @q: The exec queue 909 * @vm: The VM the engine does a bind or exec for 910 * 911 * Get last fence, takes a ref. Only safe to be called in the context of 912 * resuming the hw engine group's long-running exec queue, when the group 913 * semaphore is held. 914 * 915 * Returns: last fence if not signaled, dma fence stub if signaled 916 */ 917 struct dma_fence *xe_exec_queue_last_fence_get_for_resume(struct xe_exec_queue *q, 918 struct xe_vm *vm) 919 { 920 struct dma_fence *fence; 921 922 lockdep_assert_held_write(&q->hwe->hw_engine_group->mode_sem); 923 924 if (q->last_fence && 925 test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &q->last_fence->flags)) 926 xe_exec_queue_last_fence_put_unlocked(q); 927 928 fence = q->last_fence ? q->last_fence : dma_fence_get_stub(); 929 dma_fence_get(fence); 930 return fence; 931 } 932 933 /** 934 * xe_exec_queue_last_fence_set() - Set last fence 935 * @q: The exec queue 936 * @vm: The VM the engine does a bind or exec for 937 * @fence: The fence 938 * 939 * Set the last fence for the engine. Increases reference count for fence, when 940 * closing engine xe_exec_queue_last_fence_put should be called. 941 */ 942 void xe_exec_queue_last_fence_set(struct xe_exec_queue *q, struct xe_vm *vm, 943 struct dma_fence *fence) 944 { 945 xe_exec_queue_last_fence_lockdep_assert(q, vm); 946 947 xe_exec_queue_last_fence_put(q, vm); 948 q->last_fence = dma_fence_get(fence); 949 } 950 951 /** 952 * xe_exec_queue_last_fence_test_dep - Test last fence dependency of queue 953 * @q: The exec queue 954 * @vm: The VM the engine does a bind or exec for 955 * 956 * Returns: 957 * -ETIME if there exists an unsignalled last fence dependency, zero otherwise. 958 */ 959 int xe_exec_queue_last_fence_test_dep(struct xe_exec_queue *q, struct xe_vm *vm) 960 { 961 struct dma_fence *fence; 962 int err = 0; 963 964 fence = xe_exec_queue_last_fence_get(q, vm); 965 if (fence) { 966 err = test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags) ? 967 0 : -ETIME; 968 dma_fence_put(fence); 969 } 970 971 return err; 972 } 973