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_drv.h> 12 #include <drm/drm_file.h> 13 #include <uapi/drm/xe_drm.h> 14 15 #include "xe_device.h" 16 #include "xe_gt.h" 17 #include "xe_hw_engine_class_sysfs.h" 18 #include "xe_hw_engine_group.h" 19 #include "xe_hw_fence.h" 20 #include "xe_irq.h" 21 #include "xe_lrc.h" 22 #include "xe_macros.h" 23 #include "xe_migrate.h" 24 #include "xe_pm.h" 25 #include "xe_ring_ops_types.h" 26 #include "xe_trace.h" 27 #include "xe_vm.h" 28 #include "xe_pxp.h" 29 30 enum xe_exec_queue_sched_prop { 31 XE_EXEC_QUEUE_JOB_TIMEOUT = 0, 32 XE_EXEC_QUEUE_TIMESLICE = 1, 33 XE_EXEC_QUEUE_PREEMPT_TIMEOUT = 2, 34 XE_EXEC_QUEUE_SCHED_PROP_MAX = 3, 35 }; 36 37 static int exec_queue_user_extensions(struct xe_device *xe, struct xe_exec_queue *q, 38 u64 extensions, int ext_number); 39 40 static void __xe_exec_queue_free(struct xe_exec_queue *q) 41 { 42 if (xe_exec_queue_uses_pxp(q)) 43 xe_pxp_exec_queue_remove(gt_to_xe(q->gt)->pxp, q); 44 if (q->vm) 45 xe_vm_put(q->vm); 46 47 if (q->xef) 48 xe_file_put(q->xef); 49 50 kfree(q); 51 } 52 53 static struct xe_exec_queue *__xe_exec_queue_alloc(struct xe_device *xe, 54 struct xe_vm *vm, 55 u32 logical_mask, 56 u16 width, struct xe_hw_engine *hwe, 57 u32 flags, u64 extensions) 58 { 59 struct xe_exec_queue *q; 60 struct xe_gt *gt = hwe->gt; 61 int err; 62 63 /* only kernel queues can be permanent */ 64 XE_WARN_ON((flags & EXEC_QUEUE_FLAG_PERMANENT) && !(flags & EXEC_QUEUE_FLAG_KERNEL)); 65 66 q = kzalloc(struct_size(q, lrc, width), GFP_KERNEL); 67 if (!q) 68 return ERR_PTR(-ENOMEM); 69 70 kref_init(&q->refcount); 71 q->flags = flags; 72 q->hwe = hwe; 73 q->gt = gt; 74 q->class = hwe->class; 75 q->width = width; 76 q->msix_vec = XE_IRQ_DEFAULT_MSIX; 77 q->logical_mask = logical_mask; 78 q->fence_irq = >->fence_irq[hwe->class]; 79 q->ring_ops = gt->ring_ops[hwe->class]; 80 q->ops = gt->exec_queue_ops; 81 INIT_LIST_HEAD(&q->lr.link); 82 INIT_LIST_HEAD(&q->multi_gt_link); 83 INIT_LIST_HEAD(&q->hw_engine_group_link); 84 INIT_LIST_HEAD(&q->pxp.link); 85 86 q->sched_props.timeslice_us = hwe->eclass->sched_props.timeslice_us; 87 q->sched_props.preempt_timeout_us = 88 hwe->eclass->sched_props.preempt_timeout_us; 89 q->sched_props.job_timeout_ms = 90 hwe->eclass->sched_props.job_timeout_ms; 91 if (q->flags & EXEC_QUEUE_FLAG_KERNEL && 92 q->flags & EXEC_QUEUE_FLAG_HIGH_PRIORITY) 93 q->sched_props.priority = XE_EXEC_QUEUE_PRIORITY_KERNEL; 94 else 95 q->sched_props.priority = XE_EXEC_QUEUE_PRIORITY_NORMAL; 96 97 if (vm) 98 q->vm = xe_vm_get(vm); 99 100 if (extensions) { 101 /* 102 * may set q->usm, must come before xe_lrc_create(), 103 * may overwrite q->sched_props, must come before q->ops->init() 104 */ 105 err = exec_queue_user_extensions(xe, q, extensions, 0); 106 if (err) { 107 __xe_exec_queue_free(q); 108 return ERR_PTR(err); 109 } 110 } 111 112 return q; 113 } 114 115 static int __xe_exec_queue_init(struct xe_exec_queue *q) 116 { 117 struct xe_vm *vm = q->vm; 118 int i, err; 119 u32 flags = 0; 120 121 /* 122 * PXP workloads executing on RCS or CCS must run in isolation (i.e. no 123 * other workload can use the EUs at the same time). On MTL this is done 124 * by setting the RUNALONE bit in the LRC, while starting on Xe2 there 125 * is a dedicated bit for it. 126 */ 127 if (xe_exec_queue_uses_pxp(q) && 128 (q->class == XE_ENGINE_CLASS_RENDER || q->class == XE_ENGINE_CLASS_COMPUTE)) { 129 if (GRAPHICS_VER(gt_to_xe(q->gt)) >= 20) 130 flags |= XE_LRC_CREATE_PXP; 131 else 132 flags |= XE_LRC_CREATE_RUNALONE; 133 } 134 135 if (vm) { 136 err = xe_vm_lock(vm, true); 137 if (err) 138 return err; 139 } 140 141 for (i = 0; i < q->width; ++i) { 142 q->lrc[i] = xe_lrc_create(q->hwe, q->vm, SZ_16K, q->msix_vec, flags); 143 if (IS_ERR(q->lrc[i])) { 144 err = PTR_ERR(q->lrc[i]); 145 goto err_unlock; 146 } 147 } 148 149 if (vm) 150 xe_vm_unlock(vm); 151 152 err = q->ops->init(q); 153 if (err) 154 goto err_lrc; 155 156 return 0; 157 158 err_unlock: 159 if (vm) 160 xe_vm_unlock(vm); 161 err_lrc: 162 for (i = i - 1; i >= 0; --i) 163 xe_lrc_put(q->lrc[i]); 164 return err; 165 } 166 167 struct xe_exec_queue *xe_exec_queue_create(struct xe_device *xe, struct xe_vm *vm, 168 u32 logical_mask, u16 width, 169 struct xe_hw_engine *hwe, u32 flags, 170 u64 extensions) 171 { 172 struct xe_exec_queue *q; 173 int err; 174 175 /* VMs for GSCCS queues (and only those) must have the XE_VM_FLAG_GSC flag */ 176 xe_assert(xe, !vm || (!!(vm->flags & XE_VM_FLAG_GSC) == !!(hwe->engine_id == XE_HW_ENGINE_GSCCS0))); 177 178 q = __xe_exec_queue_alloc(xe, vm, logical_mask, width, hwe, flags, 179 extensions); 180 if (IS_ERR(q)) 181 return q; 182 183 err = __xe_exec_queue_init(q); 184 if (err) 185 goto err_post_alloc; 186 187 /* 188 * We can only add the queue to the PXP list after the init is complete, 189 * because the PXP termination can call exec_queue_kill and that will 190 * go bad if the queue is only half-initialized. This means that we 191 * can't do it when we handle the PXP extension in __xe_exec_queue_alloc 192 * and we need to do it here instead. 193 */ 194 if (xe_exec_queue_uses_pxp(q)) { 195 err = xe_pxp_exec_queue_add(xe->pxp, q); 196 if (err) 197 goto err_post_alloc; 198 } 199 200 return q; 201 202 err_post_alloc: 203 __xe_exec_queue_free(q); 204 return ERR_PTR(err); 205 } 206 207 struct xe_exec_queue *xe_exec_queue_create_class(struct xe_device *xe, struct xe_gt *gt, 208 struct xe_vm *vm, 209 enum xe_engine_class class, 210 u32 flags, u64 extensions) 211 { 212 struct xe_hw_engine *hwe, *hwe0 = NULL; 213 enum xe_hw_engine_id id; 214 u32 logical_mask = 0; 215 216 for_each_hw_engine(hwe, gt, id) { 217 if (xe_hw_engine_is_reserved(hwe)) 218 continue; 219 220 if (hwe->class == class) { 221 logical_mask |= BIT(hwe->logical_instance); 222 if (!hwe0) 223 hwe0 = hwe; 224 } 225 } 226 227 if (!logical_mask) 228 return ERR_PTR(-ENODEV); 229 230 return xe_exec_queue_create(xe, vm, logical_mask, 1, hwe0, flags, extensions); 231 } 232 233 /** 234 * xe_exec_queue_create_bind() - Create bind exec queue. 235 * @xe: Xe device. 236 * @tile: tile which bind exec queue belongs to. 237 * @flags: exec queue creation flags 238 * @extensions: exec queue creation extensions 239 * 240 * Normalize bind exec queue creation. Bind exec queue is tied to migration VM 241 * for access to physical memory required for page table programming. On a 242 * faulting devices the reserved copy engine instance must be used to avoid 243 * deadlocking (user binds cannot get stuck behind faults as kernel binds which 244 * resolve faults depend on user binds). On non-faulting devices any copy engine 245 * can be used. 246 * 247 * Returns exec queue on success, ERR_PTR on failure 248 */ 249 struct xe_exec_queue *xe_exec_queue_create_bind(struct xe_device *xe, 250 struct xe_tile *tile, 251 u32 flags, u64 extensions) 252 { 253 struct xe_gt *gt = tile->primary_gt; 254 struct xe_exec_queue *q; 255 struct xe_vm *migrate_vm; 256 257 migrate_vm = xe_migrate_get_vm(tile->migrate); 258 if (xe->info.has_usm) { 259 struct xe_hw_engine *hwe = xe_gt_hw_engine(gt, 260 XE_ENGINE_CLASS_COPY, 261 gt->usm.reserved_bcs_instance, 262 false); 263 264 if (!hwe) { 265 xe_vm_put(migrate_vm); 266 return ERR_PTR(-EINVAL); 267 } 268 269 q = xe_exec_queue_create(xe, migrate_vm, 270 BIT(hwe->logical_instance), 1, hwe, 271 flags, extensions); 272 } else { 273 q = xe_exec_queue_create_class(xe, gt, migrate_vm, 274 XE_ENGINE_CLASS_COPY, flags, 275 extensions); 276 } 277 xe_vm_put(migrate_vm); 278 279 return q; 280 } 281 ALLOW_ERROR_INJECTION(xe_exec_queue_create_bind, ERRNO); 282 283 void xe_exec_queue_destroy(struct kref *ref) 284 { 285 struct xe_exec_queue *q = container_of(ref, struct xe_exec_queue, refcount); 286 struct xe_exec_queue *eq, *next; 287 288 if (xe_exec_queue_uses_pxp(q)) 289 xe_pxp_exec_queue_remove(gt_to_xe(q->gt)->pxp, q); 290 291 xe_exec_queue_last_fence_put_unlocked(q); 292 if (!(q->flags & EXEC_QUEUE_FLAG_BIND_ENGINE_CHILD)) { 293 list_for_each_entry_safe(eq, next, &q->multi_gt_list, 294 multi_gt_link) 295 xe_exec_queue_put(eq); 296 } 297 298 q->ops->fini(q); 299 } 300 301 void xe_exec_queue_fini(struct xe_exec_queue *q) 302 { 303 int i; 304 305 /* 306 * Before releasing our ref to lrc and xef, accumulate our run ticks 307 * and wakeup any waiters. 308 */ 309 xe_exec_queue_update_run_ticks(q); 310 if (q->xef && atomic_dec_and_test(&q->xef->exec_queue.pending_removal)) 311 wake_up_var(&q->xef->exec_queue.pending_removal); 312 313 for (i = 0; i < q->width; ++i) 314 xe_lrc_put(q->lrc[i]); 315 316 __xe_exec_queue_free(q); 317 } 318 319 void xe_exec_queue_assign_name(struct xe_exec_queue *q, u32 instance) 320 { 321 switch (q->class) { 322 case XE_ENGINE_CLASS_RENDER: 323 snprintf(q->name, sizeof(q->name), "rcs%d", instance); 324 break; 325 case XE_ENGINE_CLASS_VIDEO_DECODE: 326 snprintf(q->name, sizeof(q->name), "vcs%d", instance); 327 break; 328 case XE_ENGINE_CLASS_VIDEO_ENHANCE: 329 snprintf(q->name, sizeof(q->name), "vecs%d", instance); 330 break; 331 case XE_ENGINE_CLASS_COPY: 332 snprintf(q->name, sizeof(q->name), "bcs%d", instance); 333 break; 334 case XE_ENGINE_CLASS_COMPUTE: 335 snprintf(q->name, sizeof(q->name), "ccs%d", instance); 336 break; 337 case XE_ENGINE_CLASS_OTHER: 338 snprintf(q->name, sizeof(q->name), "gsccs%d", instance); 339 break; 340 default: 341 XE_WARN_ON(q->class); 342 } 343 } 344 345 struct xe_exec_queue *xe_exec_queue_lookup(struct xe_file *xef, u32 id) 346 { 347 struct xe_exec_queue *q; 348 349 mutex_lock(&xef->exec_queue.lock); 350 q = xa_load(&xef->exec_queue.xa, id); 351 if (q) 352 xe_exec_queue_get(q); 353 mutex_unlock(&xef->exec_queue.lock); 354 355 return q; 356 } 357 358 enum xe_exec_queue_priority 359 xe_exec_queue_device_get_max_priority(struct xe_device *xe) 360 { 361 return capable(CAP_SYS_NICE) ? XE_EXEC_QUEUE_PRIORITY_HIGH : 362 XE_EXEC_QUEUE_PRIORITY_NORMAL; 363 } 364 365 static int exec_queue_set_priority(struct xe_device *xe, struct xe_exec_queue *q, 366 u64 value) 367 { 368 if (XE_IOCTL_DBG(xe, value > XE_EXEC_QUEUE_PRIORITY_HIGH)) 369 return -EINVAL; 370 371 if (XE_IOCTL_DBG(xe, value > xe_exec_queue_device_get_max_priority(xe))) 372 return -EPERM; 373 374 q->sched_props.priority = value; 375 return 0; 376 } 377 378 static bool xe_exec_queue_enforce_schedule_limit(void) 379 { 380 #if IS_ENABLED(CONFIG_DRM_XE_ENABLE_SCHEDTIMEOUT_LIMIT) 381 return true; 382 #else 383 return !capable(CAP_SYS_NICE); 384 #endif 385 } 386 387 static void 388 xe_exec_queue_get_prop_minmax(struct xe_hw_engine_class_intf *eclass, 389 enum xe_exec_queue_sched_prop prop, 390 u32 *min, u32 *max) 391 { 392 switch (prop) { 393 case XE_EXEC_QUEUE_JOB_TIMEOUT: 394 *min = eclass->sched_props.job_timeout_min; 395 *max = eclass->sched_props.job_timeout_max; 396 break; 397 case XE_EXEC_QUEUE_TIMESLICE: 398 *min = eclass->sched_props.timeslice_min; 399 *max = eclass->sched_props.timeslice_max; 400 break; 401 case XE_EXEC_QUEUE_PREEMPT_TIMEOUT: 402 *min = eclass->sched_props.preempt_timeout_min; 403 *max = eclass->sched_props.preempt_timeout_max; 404 break; 405 default: 406 break; 407 } 408 #if IS_ENABLED(CONFIG_DRM_XE_ENABLE_SCHEDTIMEOUT_LIMIT) 409 if (capable(CAP_SYS_NICE)) { 410 switch (prop) { 411 case XE_EXEC_QUEUE_JOB_TIMEOUT: 412 *min = XE_HW_ENGINE_JOB_TIMEOUT_MIN; 413 *max = XE_HW_ENGINE_JOB_TIMEOUT_MAX; 414 break; 415 case XE_EXEC_QUEUE_TIMESLICE: 416 *min = XE_HW_ENGINE_TIMESLICE_MIN; 417 *max = XE_HW_ENGINE_TIMESLICE_MAX; 418 break; 419 case XE_EXEC_QUEUE_PREEMPT_TIMEOUT: 420 *min = XE_HW_ENGINE_PREEMPT_TIMEOUT_MIN; 421 *max = XE_HW_ENGINE_PREEMPT_TIMEOUT_MAX; 422 break; 423 default: 424 break; 425 } 426 } 427 #endif 428 } 429 430 static int exec_queue_set_timeslice(struct xe_device *xe, struct xe_exec_queue *q, 431 u64 value) 432 { 433 u32 min = 0, max = 0; 434 435 xe_exec_queue_get_prop_minmax(q->hwe->eclass, 436 XE_EXEC_QUEUE_TIMESLICE, &min, &max); 437 438 if (xe_exec_queue_enforce_schedule_limit() && 439 !xe_hw_engine_timeout_in_range(value, min, max)) 440 return -EINVAL; 441 442 q->sched_props.timeslice_us = value; 443 return 0; 444 } 445 446 static int 447 exec_queue_set_pxp_type(struct xe_device *xe, struct xe_exec_queue *q, u64 value) 448 { 449 if (value == DRM_XE_PXP_TYPE_NONE) 450 return 0; 451 452 /* we only support HWDRM sessions right now */ 453 if (XE_IOCTL_DBG(xe, value != DRM_XE_PXP_TYPE_HWDRM)) 454 return -EINVAL; 455 456 if (!xe_pxp_is_enabled(xe->pxp)) 457 return -ENODEV; 458 459 return xe_pxp_exec_queue_set_type(xe->pxp, q, DRM_XE_PXP_TYPE_HWDRM); 460 } 461 462 typedef int (*xe_exec_queue_set_property_fn)(struct xe_device *xe, 463 struct xe_exec_queue *q, 464 u64 value); 465 466 static const xe_exec_queue_set_property_fn exec_queue_set_property_funcs[] = { 467 [DRM_XE_EXEC_QUEUE_SET_PROPERTY_PRIORITY] = exec_queue_set_priority, 468 [DRM_XE_EXEC_QUEUE_SET_PROPERTY_TIMESLICE] = exec_queue_set_timeslice, 469 [DRM_XE_EXEC_QUEUE_SET_PROPERTY_PXP_TYPE] = exec_queue_set_pxp_type, 470 }; 471 472 static int exec_queue_user_ext_set_property(struct xe_device *xe, 473 struct xe_exec_queue *q, 474 u64 extension) 475 { 476 u64 __user *address = u64_to_user_ptr(extension); 477 struct drm_xe_ext_set_property ext; 478 int err; 479 u32 idx; 480 481 err = __copy_from_user(&ext, address, sizeof(ext)); 482 if (XE_IOCTL_DBG(xe, err)) 483 return -EFAULT; 484 485 if (XE_IOCTL_DBG(xe, ext.property >= 486 ARRAY_SIZE(exec_queue_set_property_funcs)) || 487 XE_IOCTL_DBG(xe, ext.pad) || 488 XE_IOCTL_DBG(xe, ext.property != DRM_XE_EXEC_QUEUE_SET_PROPERTY_PRIORITY && 489 ext.property != DRM_XE_EXEC_QUEUE_SET_PROPERTY_TIMESLICE && 490 ext.property != DRM_XE_EXEC_QUEUE_SET_PROPERTY_PXP_TYPE)) 491 return -EINVAL; 492 493 idx = array_index_nospec(ext.property, ARRAY_SIZE(exec_queue_set_property_funcs)); 494 if (!exec_queue_set_property_funcs[idx]) 495 return -EINVAL; 496 497 return exec_queue_set_property_funcs[idx](xe, q, ext.value); 498 } 499 500 typedef int (*xe_exec_queue_user_extension_fn)(struct xe_device *xe, 501 struct xe_exec_queue *q, 502 u64 extension); 503 504 static const xe_exec_queue_user_extension_fn exec_queue_user_extension_funcs[] = { 505 [DRM_XE_EXEC_QUEUE_EXTENSION_SET_PROPERTY] = exec_queue_user_ext_set_property, 506 }; 507 508 #define MAX_USER_EXTENSIONS 16 509 static int exec_queue_user_extensions(struct xe_device *xe, struct xe_exec_queue *q, 510 u64 extensions, int ext_number) 511 { 512 u64 __user *address = u64_to_user_ptr(extensions); 513 struct drm_xe_user_extension ext; 514 int err; 515 u32 idx; 516 517 if (XE_IOCTL_DBG(xe, ext_number >= MAX_USER_EXTENSIONS)) 518 return -E2BIG; 519 520 err = __copy_from_user(&ext, address, sizeof(ext)); 521 if (XE_IOCTL_DBG(xe, err)) 522 return -EFAULT; 523 524 if (XE_IOCTL_DBG(xe, ext.pad) || 525 XE_IOCTL_DBG(xe, ext.name >= 526 ARRAY_SIZE(exec_queue_user_extension_funcs))) 527 return -EINVAL; 528 529 idx = array_index_nospec(ext.name, 530 ARRAY_SIZE(exec_queue_user_extension_funcs)); 531 err = exec_queue_user_extension_funcs[idx](xe, q, extensions); 532 if (XE_IOCTL_DBG(xe, err)) 533 return err; 534 535 if (ext.next_extension) 536 return exec_queue_user_extensions(xe, q, ext.next_extension, 537 ++ext_number); 538 539 return 0; 540 } 541 542 static u32 calc_validate_logical_mask(struct xe_device *xe, 543 struct drm_xe_engine_class_instance *eci, 544 u16 width, u16 num_placements) 545 { 546 int len = width * num_placements; 547 int i, j, n; 548 u16 class; 549 u16 gt_id; 550 u32 return_mask = 0, prev_mask; 551 552 if (XE_IOCTL_DBG(xe, !xe_device_uc_enabled(xe) && 553 len > 1)) 554 return 0; 555 556 for (i = 0; i < width; ++i) { 557 u32 current_mask = 0; 558 559 for (j = 0; j < num_placements; ++j) { 560 struct xe_hw_engine *hwe; 561 562 n = j * width + i; 563 564 hwe = xe_hw_engine_lookup(xe, eci[n]); 565 if (XE_IOCTL_DBG(xe, !hwe)) 566 return 0; 567 568 if (XE_IOCTL_DBG(xe, xe_hw_engine_is_reserved(hwe))) 569 return 0; 570 571 if (XE_IOCTL_DBG(xe, n && eci[n].gt_id != gt_id) || 572 XE_IOCTL_DBG(xe, n && eci[n].engine_class != class)) 573 return 0; 574 575 class = eci[n].engine_class; 576 gt_id = eci[n].gt_id; 577 578 if (width == 1 || !i) 579 return_mask |= BIT(eci[n].engine_instance); 580 current_mask |= BIT(eci[n].engine_instance); 581 } 582 583 /* Parallel submissions must be logically contiguous */ 584 if (i && XE_IOCTL_DBG(xe, current_mask != prev_mask << 1)) 585 return 0; 586 587 prev_mask = current_mask; 588 } 589 590 return return_mask; 591 } 592 593 int xe_exec_queue_create_ioctl(struct drm_device *dev, void *data, 594 struct drm_file *file) 595 { 596 struct xe_device *xe = to_xe_device(dev); 597 struct xe_file *xef = to_xe_file(file); 598 struct drm_xe_exec_queue_create *args = data; 599 struct drm_xe_engine_class_instance eci[XE_HW_ENGINE_MAX_INSTANCE]; 600 struct drm_xe_engine_class_instance __user *user_eci = 601 u64_to_user_ptr(args->instances); 602 struct xe_hw_engine *hwe; 603 struct xe_vm *vm; 604 struct xe_tile *tile; 605 struct xe_exec_queue *q = NULL; 606 u32 logical_mask; 607 u32 flags = 0; 608 u32 id; 609 u32 len; 610 int err; 611 612 if (XE_IOCTL_DBG(xe, args->flags & ~DRM_XE_EXEC_QUEUE_LOW_LATENCY_HINT) || 613 XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1])) 614 return -EINVAL; 615 616 len = args->width * args->num_placements; 617 if (XE_IOCTL_DBG(xe, !len || len > XE_HW_ENGINE_MAX_INSTANCE)) 618 return -EINVAL; 619 620 err = __copy_from_user(eci, user_eci, 621 sizeof(struct drm_xe_engine_class_instance) * 622 len); 623 if (XE_IOCTL_DBG(xe, err)) 624 return -EFAULT; 625 626 if (XE_IOCTL_DBG(xe, eci[0].gt_id >= xe->info.gt_count)) 627 return -EINVAL; 628 629 if (args->flags & DRM_XE_EXEC_QUEUE_LOW_LATENCY_HINT) 630 flags |= EXEC_QUEUE_FLAG_LOW_LATENCY; 631 632 if (eci[0].engine_class == DRM_XE_ENGINE_CLASS_VM_BIND) { 633 if (XE_IOCTL_DBG(xe, args->width != 1) || 634 XE_IOCTL_DBG(xe, args->num_placements != 1) || 635 XE_IOCTL_DBG(xe, eci[0].engine_instance != 0)) 636 return -EINVAL; 637 638 for_each_tile(tile, xe, id) { 639 struct xe_exec_queue *new; 640 641 flags |= EXEC_QUEUE_FLAG_VM; 642 if (id) 643 flags |= EXEC_QUEUE_FLAG_BIND_ENGINE_CHILD; 644 645 new = xe_exec_queue_create_bind(xe, tile, flags, 646 args->extensions); 647 if (IS_ERR(new)) { 648 err = PTR_ERR(new); 649 if (q) 650 goto put_exec_queue; 651 return err; 652 } 653 if (id == 0) 654 q = new; 655 else 656 list_add_tail(&new->multi_gt_list, 657 &q->multi_gt_link); 658 } 659 } else { 660 logical_mask = calc_validate_logical_mask(xe, eci, 661 args->width, 662 args->num_placements); 663 if (XE_IOCTL_DBG(xe, !logical_mask)) 664 return -EINVAL; 665 666 hwe = xe_hw_engine_lookup(xe, eci[0]); 667 if (XE_IOCTL_DBG(xe, !hwe)) 668 return -EINVAL; 669 670 vm = xe_vm_lookup(xef, args->vm_id); 671 if (XE_IOCTL_DBG(xe, !vm)) 672 return -ENOENT; 673 674 err = down_read_interruptible(&vm->lock); 675 if (err) { 676 xe_vm_put(vm); 677 return err; 678 } 679 680 if (XE_IOCTL_DBG(xe, xe_vm_is_closed_or_banned(vm))) { 681 up_read(&vm->lock); 682 xe_vm_put(vm); 683 return -ENOENT; 684 } 685 686 q = xe_exec_queue_create(xe, vm, logical_mask, 687 args->width, hwe, flags, 688 args->extensions); 689 up_read(&vm->lock); 690 xe_vm_put(vm); 691 if (IS_ERR(q)) 692 return PTR_ERR(q); 693 694 if (xe_vm_in_preempt_fence_mode(vm)) { 695 q->lr.context = dma_fence_context_alloc(1); 696 697 err = xe_vm_add_compute_exec_queue(vm, q); 698 if (XE_IOCTL_DBG(xe, err)) 699 goto put_exec_queue; 700 } 701 702 if (q->vm && q->hwe->hw_engine_group) { 703 err = xe_hw_engine_group_add_exec_queue(q->hwe->hw_engine_group, q); 704 if (err) 705 goto put_exec_queue; 706 } 707 } 708 709 q->xef = xe_file_get(xef); 710 711 /* user id alloc must always be last in ioctl to prevent UAF */ 712 err = xa_alloc(&xef->exec_queue.xa, &id, q, xa_limit_32b, GFP_KERNEL); 713 if (err) 714 goto kill_exec_queue; 715 716 args->exec_queue_id = id; 717 718 return 0; 719 720 kill_exec_queue: 721 xe_exec_queue_kill(q); 722 put_exec_queue: 723 xe_exec_queue_put(q); 724 return err; 725 } 726 727 int xe_exec_queue_get_property_ioctl(struct drm_device *dev, void *data, 728 struct drm_file *file) 729 { 730 struct xe_device *xe = to_xe_device(dev); 731 struct xe_file *xef = to_xe_file(file); 732 struct drm_xe_exec_queue_get_property *args = data; 733 struct xe_exec_queue *q; 734 int ret; 735 736 if (XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1])) 737 return -EINVAL; 738 739 q = xe_exec_queue_lookup(xef, args->exec_queue_id); 740 if (XE_IOCTL_DBG(xe, !q)) 741 return -ENOENT; 742 743 switch (args->property) { 744 case DRM_XE_EXEC_QUEUE_GET_PROPERTY_BAN: 745 args->value = q->ops->reset_status(q); 746 ret = 0; 747 break; 748 default: 749 ret = -EINVAL; 750 } 751 752 xe_exec_queue_put(q); 753 754 return ret; 755 } 756 757 /** 758 * xe_exec_queue_is_lr() - Whether an exec_queue is long-running 759 * @q: The exec_queue 760 * 761 * Return: True if the exec_queue is long-running, false otherwise. 762 */ 763 bool xe_exec_queue_is_lr(struct xe_exec_queue *q) 764 { 765 return q->vm && xe_vm_in_lr_mode(q->vm) && 766 !(q->flags & EXEC_QUEUE_FLAG_VM); 767 } 768 769 static s32 xe_exec_queue_num_job_inflight(struct xe_exec_queue *q) 770 { 771 return q->lrc[0]->fence_ctx.next_seqno - xe_lrc_seqno(q->lrc[0]) - 1; 772 } 773 774 /** 775 * xe_exec_queue_ring_full() - Whether an exec_queue's ring is full 776 * @q: The exec_queue 777 * 778 * Return: True if the exec_queue's ring is full, false otherwise. 779 */ 780 bool xe_exec_queue_ring_full(struct xe_exec_queue *q) 781 { 782 struct xe_lrc *lrc = q->lrc[0]; 783 s32 max_job = lrc->ring.size / MAX_JOB_SIZE_BYTES; 784 785 return xe_exec_queue_num_job_inflight(q) >= max_job; 786 } 787 788 /** 789 * xe_exec_queue_is_idle() - Whether an exec_queue is idle. 790 * @q: The exec_queue 791 * 792 * FIXME: Need to determine what to use as the short-lived 793 * timeline lock for the exec_queues, so that the return value 794 * of this function becomes more than just an advisory 795 * snapshot in time. The timeline lock must protect the 796 * seqno from racing submissions on the same exec_queue. 797 * Typically vm->resv, but user-created timeline locks use the migrate vm 798 * and never grabs the migrate vm->resv so we have a race there. 799 * 800 * Return: True if the exec_queue is idle, false otherwise. 801 */ 802 bool xe_exec_queue_is_idle(struct xe_exec_queue *q) 803 { 804 if (xe_exec_queue_is_parallel(q)) { 805 int i; 806 807 for (i = 0; i < q->width; ++i) { 808 if (xe_lrc_seqno(q->lrc[i]) != 809 q->lrc[i]->fence_ctx.next_seqno - 1) 810 return false; 811 } 812 813 return true; 814 } 815 816 return xe_lrc_seqno(q->lrc[0]) == 817 q->lrc[0]->fence_ctx.next_seqno - 1; 818 } 819 820 /** 821 * xe_exec_queue_update_run_ticks() - Update run time in ticks for this exec queue 822 * from hw 823 * @q: The exec queue 824 * 825 * Update the timestamp saved by HW for this exec queue and save run ticks 826 * calculated by using the delta from last update. 827 */ 828 void xe_exec_queue_update_run_ticks(struct xe_exec_queue *q) 829 { 830 struct xe_device *xe = gt_to_xe(q->gt); 831 struct xe_lrc *lrc; 832 u32 old_ts, new_ts; 833 int idx; 834 835 /* 836 * Jobs that are executed by kernel doesn't have a corresponding xe_file 837 * and thus are not accounted. 838 */ 839 if (!q->xef) 840 return; 841 842 /* Synchronize with unbind while holding the xe file open */ 843 if (!drm_dev_enter(&xe->drm, &idx)) 844 return; 845 /* 846 * Only sample the first LRC. For parallel submission, all of them are 847 * scheduled together and we compensate that below by multiplying by 848 * width - this may introduce errors if that premise is not true and 849 * they don't exit 100% aligned. On the other hand, looping through 850 * the LRCs and reading them in different time could also introduce 851 * errors. 852 */ 853 lrc = q->lrc[0]; 854 new_ts = xe_lrc_update_timestamp(lrc, &old_ts); 855 q->xef->run_ticks[q->class] += (new_ts - old_ts) * q->width; 856 857 drm_dev_exit(idx); 858 } 859 860 /** 861 * xe_exec_queue_kill - permanently stop all execution from an exec queue 862 * @q: The exec queue 863 * 864 * This function permanently stops all activity on an exec queue. If the queue 865 * is actively executing on the HW, it will be kicked off the engine; any 866 * pending jobs are discarded and all future submissions are rejected. 867 * This function is safe to call multiple times. 868 */ 869 void xe_exec_queue_kill(struct xe_exec_queue *q) 870 { 871 struct xe_exec_queue *eq = q, *next; 872 873 list_for_each_entry_safe(eq, next, &eq->multi_gt_list, 874 multi_gt_link) { 875 q->ops->kill(eq); 876 xe_vm_remove_compute_exec_queue(q->vm, eq); 877 } 878 879 q->ops->kill(q); 880 xe_vm_remove_compute_exec_queue(q->vm, q); 881 } 882 883 int xe_exec_queue_destroy_ioctl(struct drm_device *dev, void *data, 884 struct drm_file *file) 885 { 886 struct xe_device *xe = to_xe_device(dev); 887 struct xe_file *xef = to_xe_file(file); 888 struct drm_xe_exec_queue_destroy *args = data; 889 struct xe_exec_queue *q; 890 891 if (XE_IOCTL_DBG(xe, args->pad) || 892 XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1])) 893 return -EINVAL; 894 895 mutex_lock(&xef->exec_queue.lock); 896 q = xa_erase(&xef->exec_queue.xa, args->exec_queue_id); 897 if (q) 898 atomic_inc(&xef->exec_queue.pending_removal); 899 mutex_unlock(&xef->exec_queue.lock); 900 901 if (XE_IOCTL_DBG(xe, !q)) 902 return -ENOENT; 903 904 if (q->vm && q->hwe->hw_engine_group) 905 xe_hw_engine_group_del_exec_queue(q->hwe->hw_engine_group, q); 906 907 xe_exec_queue_kill(q); 908 909 trace_xe_exec_queue_close(q); 910 xe_exec_queue_put(q); 911 912 return 0; 913 } 914 915 static void xe_exec_queue_last_fence_lockdep_assert(struct xe_exec_queue *q, 916 struct xe_vm *vm) 917 { 918 if (q->flags & EXEC_QUEUE_FLAG_VM) { 919 lockdep_assert_held(&vm->lock); 920 } else { 921 xe_vm_assert_held(vm); 922 lockdep_assert_held(&q->hwe->hw_engine_group->mode_sem); 923 } 924 } 925 926 /** 927 * xe_exec_queue_last_fence_put() - Drop ref to last fence 928 * @q: The exec queue 929 * @vm: The VM the engine does a bind or exec for 930 */ 931 void xe_exec_queue_last_fence_put(struct xe_exec_queue *q, struct xe_vm *vm) 932 { 933 xe_exec_queue_last_fence_lockdep_assert(q, vm); 934 935 xe_exec_queue_last_fence_put_unlocked(q); 936 } 937 938 /** 939 * xe_exec_queue_last_fence_put_unlocked() - Drop ref to last fence unlocked 940 * @q: The exec queue 941 * 942 * Only safe to be called from xe_exec_queue_destroy(). 943 */ 944 void xe_exec_queue_last_fence_put_unlocked(struct xe_exec_queue *q) 945 { 946 if (q->last_fence) { 947 dma_fence_put(q->last_fence); 948 q->last_fence = NULL; 949 } 950 } 951 952 /** 953 * xe_exec_queue_last_fence_get() - Get last fence 954 * @q: The exec queue 955 * @vm: The VM the engine does a bind or exec for 956 * 957 * Get last fence, takes a ref 958 * 959 * Returns: last fence if not signaled, dma fence stub if signaled 960 */ 961 struct dma_fence *xe_exec_queue_last_fence_get(struct xe_exec_queue *q, 962 struct xe_vm *vm) 963 { 964 struct dma_fence *fence; 965 966 xe_exec_queue_last_fence_lockdep_assert(q, vm); 967 968 if (q->last_fence && 969 test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &q->last_fence->flags)) 970 xe_exec_queue_last_fence_put(q, vm); 971 972 fence = q->last_fence ? q->last_fence : dma_fence_get_stub(); 973 dma_fence_get(fence); 974 return fence; 975 } 976 977 /** 978 * xe_exec_queue_last_fence_get_for_resume() - Get last fence 979 * @q: The exec queue 980 * @vm: The VM the engine does a bind or exec for 981 * 982 * Get last fence, takes a ref. Only safe to be called in the context of 983 * resuming the hw engine group's long-running exec queue, when the group 984 * semaphore is held. 985 * 986 * Returns: last fence if not signaled, dma fence stub if signaled 987 */ 988 struct dma_fence *xe_exec_queue_last_fence_get_for_resume(struct xe_exec_queue *q, 989 struct xe_vm *vm) 990 { 991 struct dma_fence *fence; 992 993 lockdep_assert_held_write(&q->hwe->hw_engine_group->mode_sem); 994 995 if (q->last_fence && 996 test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &q->last_fence->flags)) 997 xe_exec_queue_last_fence_put_unlocked(q); 998 999 fence = q->last_fence ? q->last_fence : dma_fence_get_stub(); 1000 dma_fence_get(fence); 1001 return fence; 1002 } 1003 1004 /** 1005 * xe_exec_queue_last_fence_set() - Set last fence 1006 * @q: The exec queue 1007 * @vm: The VM the engine does a bind or exec for 1008 * @fence: The fence 1009 * 1010 * Set the last fence for the engine. Increases reference count for fence, when 1011 * closing engine xe_exec_queue_last_fence_put should be called. 1012 */ 1013 void xe_exec_queue_last_fence_set(struct xe_exec_queue *q, struct xe_vm *vm, 1014 struct dma_fence *fence) 1015 { 1016 xe_exec_queue_last_fence_lockdep_assert(q, vm); 1017 1018 xe_exec_queue_last_fence_put(q, vm); 1019 q->last_fence = dma_fence_get(fence); 1020 } 1021 1022 /** 1023 * xe_exec_queue_last_fence_test_dep - Test last fence dependency of queue 1024 * @q: The exec queue 1025 * @vm: The VM the engine does a bind or exec for 1026 * 1027 * Returns: 1028 * -ETIME if there exists an unsignalled last fence dependency, zero otherwise. 1029 */ 1030 int xe_exec_queue_last_fence_test_dep(struct xe_exec_queue *q, struct xe_vm *vm) 1031 { 1032 struct dma_fence *fence; 1033 int err = 0; 1034 1035 fence = xe_exec_queue_last_fence_get(q, vm); 1036 if (fence) { 1037 err = test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags) ? 1038 0 : -ETIME; 1039 dma_fence_put(fence); 1040 } 1041 1042 return err; 1043 } 1044