1/* 2 * SPDX-License-Identifier: MIT 3 * 4 * Copyright © 2016-2018 Intel Corporation 5 */ 6 7#include "i915_drv.h" 8 9#include "i915_active.h" 10#include "i915_syncmap.h" 11#include "intel_gt.h" 12#include "intel_ring.h" 13#include "intel_timeline.h" 14 15#define ptr_set_bit(ptr, bit) ((typeof(ptr))((unsigned long)(ptr) | BIT(bit))) 16#define ptr_test_bit(ptr, bit) ((unsigned long)(ptr) & BIT(bit)) 17 18#define CACHELINE_BITS 6 19#define CACHELINE_FREE CACHELINE_BITS 20 21struct intel_timeline_hwsp { 22 struct intel_gt *gt; 23 struct intel_gt_timelines *gt_timelines; 24 struct list_head free_link; 25 struct i915_vma *vma; 26 u64 free_bitmap; 27}; 28 29static struct i915_vma *__hwsp_alloc(struct intel_gt *gt) 30{ 31 struct drm_i915_private *i915 = gt->i915; 32 struct drm_i915_gem_object *obj; 33 struct i915_vma *vma; 34 35 obj = i915_gem_object_create_internal(i915, PAGE_SIZE); 36 if (IS_ERR(obj)) 37 return ERR_CAST(obj); 38 39 i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC); 40 41 vma = i915_vma_instance(obj, >->ggtt->vm, NULL); 42 if (IS_ERR(vma)) 43 i915_gem_object_put(obj); 44 45 return vma; 46} 47 48static struct i915_vma * 49hwsp_alloc(struct intel_timeline *timeline, unsigned int *cacheline) 50{ 51 struct intel_gt_timelines *gt = &timeline->gt->timelines; 52 struct intel_timeline_hwsp *hwsp; 53 54 BUILD_BUG_ON(BITS_PER_TYPE(u64) * CACHELINE_BYTES > PAGE_SIZE); 55 56 spin_lock_irq(>->hwsp_lock); 57 58 /* hwsp_free_list only contains HWSP that have available cachelines */ 59 hwsp = list_first_entry_or_null(>->hwsp_free_list, 60 typeof(*hwsp), free_link); 61 if (!hwsp) { 62 struct i915_vma *vma; 63 64 spin_unlock_irq(>->hwsp_lock); 65 66 hwsp = kmalloc(sizeof(*hwsp), GFP_KERNEL); 67 if (!hwsp) 68 return ERR_PTR(-ENOMEM); 69 70 vma = __hwsp_alloc(timeline->gt); 71 if (IS_ERR(vma)) { 72 kfree(hwsp); 73 return vma; 74 } 75 76 GT_TRACE(timeline->gt, "new HWSP allocated\n"); 77 78 vma->private = hwsp; 79 hwsp->gt = timeline->gt; 80 hwsp->vma = vma; 81 hwsp->free_bitmap = ~0ull; 82 hwsp->gt_timelines = gt; 83 84 spin_lock_irq(>->hwsp_lock); 85 list_add(&hwsp->free_link, >->hwsp_free_list); 86 } 87 88 GEM_BUG_ON(!hwsp->free_bitmap); 89 *cacheline = __ffs64(hwsp->free_bitmap); 90 hwsp->free_bitmap &= ~BIT_ULL(*cacheline); 91 if (!hwsp->free_bitmap) 92 list_del(&hwsp->free_link); 93 94 spin_unlock_irq(>->hwsp_lock); 95 96 GEM_BUG_ON(hwsp->vma->private != hwsp); 97 return hwsp->vma; 98} 99 100static void __idle_hwsp_free(struct intel_timeline_hwsp *hwsp, int cacheline) 101{ 102 struct intel_gt_timelines *gt = hwsp->gt_timelines; 103 unsigned long flags; 104 105 spin_lock_irqsave(>->hwsp_lock, flags); 106 107 /* As a cacheline becomes available, publish the HWSP on the freelist */ 108 if (!hwsp->free_bitmap) 109 list_add_tail(&hwsp->free_link, >->hwsp_free_list); 110 111 GEM_BUG_ON(cacheline >= BITS_PER_TYPE(hwsp->free_bitmap)); 112 hwsp->free_bitmap |= BIT_ULL(cacheline); 113 114 /* And if no one is left using it, give the page back to the system */ 115 if (hwsp->free_bitmap == ~0ull) { 116 i915_vma_put(hwsp->vma); 117 list_del(&hwsp->free_link); 118 kfree(hwsp); 119 } 120 121 spin_unlock_irqrestore(>->hwsp_lock, flags); 122} 123 124static void __rcu_cacheline_free(struct rcu_head *rcu) 125{ 126 struct intel_timeline_cacheline *cl = 127 container_of(rcu, typeof(*cl), rcu); 128 129 /* Must wait until after all *rq->hwsp are complete before removing */ 130 i915_gem_object_unpin_map(cl->hwsp->vma->obj); 131 __idle_hwsp_free(cl->hwsp, ptr_unmask_bits(cl->vaddr, CACHELINE_BITS)); 132 133 i915_active_fini(&cl->active); 134 kfree(cl); 135} 136 137static void __idle_cacheline_free(struct intel_timeline_cacheline *cl) 138{ 139 GEM_BUG_ON(!i915_active_is_idle(&cl->active)); 140 call_rcu(&cl->rcu, __rcu_cacheline_free); 141} 142 143__i915_active_call 144static void __cacheline_retire(struct i915_active *active) 145{ 146 struct intel_timeline_cacheline *cl = 147 container_of(active, typeof(*cl), active); 148 149 i915_vma_unpin(cl->hwsp->vma); 150 if (ptr_test_bit(cl->vaddr, CACHELINE_FREE)) 151 __idle_cacheline_free(cl); 152} 153 154static int __cacheline_active(struct i915_active *active) 155{ 156 struct intel_timeline_cacheline *cl = 157 container_of(active, typeof(*cl), active); 158 159 __i915_vma_pin(cl->hwsp->vma); 160 return 0; 161} 162 163static struct intel_timeline_cacheline * 164cacheline_alloc(struct intel_timeline_hwsp *hwsp, unsigned int cacheline) 165{ 166 struct intel_timeline_cacheline *cl; 167 void *vaddr; 168 169 GEM_BUG_ON(cacheline >= BIT(CACHELINE_BITS)); 170 171 cl = kmalloc(sizeof(*cl), GFP_KERNEL); 172 if (!cl) 173 return ERR_PTR(-ENOMEM); 174 175 vaddr = i915_gem_object_pin_map(hwsp->vma->obj, I915_MAP_WB); 176 if (IS_ERR(vaddr)) { 177 kfree(cl); 178 return ERR_CAST(vaddr); 179 } 180 181 cl->hwsp = hwsp; 182 cl->vaddr = page_pack_bits(vaddr, cacheline); 183 184 i915_active_init(&cl->active, __cacheline_active, __cacheline_retire); 185 186 return cl; 187} 188 189static void cacheline_acquire(struct intel_timeline_cacheline *cl, 190 u32 ggtt_offset) 191{ 192 if (!cl) 193 return; 194 195 cl->ggtt_offset = ggtt_offset; 196 i915_active_acquire(&cl->active); 197} 198 199static void cacheline_release(struct intel_timeline_cacheline *cl) 200{ 201 if (cl) 202 i915_active_release(&cl->active); 203} 204 205static void cacheline_free(struct intel_timeline_cacheline *cl) 206{ 207 if (!i915_active_acquire_if_busy(&cl->active)) { 208 __idle_cacheline_free(cl); 209 return; 210 } 211 212 GEM_BUG_ON(ptr_test_bit(cl->vaddr, CACHELINE_FREE)); 213 cl->vaddr = ptr_set_bit(cl->vaddr, CACHELINE_FREE); 214 215 i915_active_release(&cl->active); 216} 217 218static int intel_timeline_init(struct intel_timeline *timeline, 219 struct intel_gt *gt, 220 struct i915_vma *hwsp, 221 unsigned int offset) 222{ 223 void *vaddr; 224 225 kref_init(&timeline->kref); 226 atomic_set(&timeline->pin_count, 0); 227 228 timeline->gt = gt; 229 230 timeline->has_initial_breadcrumb = !hwsp; 231 timeline->hwsp_cacheline = NULL; 232 233 if (!hwsp) { 234 struct intel_timeline_cacheline *cl; 235 unsigned int cacheline; 236 237 hwsp = hwsp_alloc(timeline, &cacheline); 238 if (IS_ERR(hwsp)) 239 return PTR_ERR(hwsp); 240 241 cl = cacheline_alloc(hwsp->private, cacheline); 242 if (IS_ERR(cl)) { 243 __idle_hwsp_free(hwsp->private, cacheline); 244 return PTR_ERR(cl); 245 } 246 247 timeline->hwsp_cacheline = cl; 248 timeline->hwsp_offset = cacheline * CACHELINE_BYTES; 249 250 vaddr = page_mask_bits(cl->vaddr); 251 } else { 252 timeline->hwsp_offset = offset; 253 vaddr = i915_gem_object_pin_map(hwsp->obj, I915_MAP_WB); 254 if (IS_ERR(vaddr)) 255 return PTR_ERR(vaddr); 256 } 257 258 timeline->hwsp_seqno = 259 memset(vaddr + timeline->hwsp_offset, 0, CACHELINE_BYTES); 260 261 timeline->hwsp_ggtt = i915_vma_get(hwsp); 262 GEM_BUG_ON(timeline->hwsp_offset >= hwsp->size); 263 264 timeline->fence_context = dma_fence_context_alloc(1); 265 266 mutex_init(&timeline->mutex); 267 268 INIT_ACTIVE_FENCE(&timeline->last_request); 269 INIT_LIST_HEAD(&timeline->requests); 270 271 i915_syncmap_init(&timeline->sync); 272 273 return 0; 274} 275 276void intel_gt_init_timelines(struct intel_gt *gt) 277{ 278 struct intel_gt_timelines *timelines = >->timelines; 279 280 spin_lock_init(&timelines->lock); 281 INIT_LIST_HEAD(&timelines->active_list); 282 283 spin_lock_init(&timelines->hwsp_lock); 284 INIT_LIST_HEAD(&timelines->hwsp_free_list); 285} 286 287static void intel_timeline_fini(struct intel_timeline *timeline) 288{ 289 GEM_BUG_ON(atomic_read(&timeline->pin_count)); 290 GEM_BUG_ON(!list_empty(&timeline->requests)); 291 GEM_BUG_ON(timeline->retire); 292 293 if (timeline->hwsp_cacheline) 294 cacheline_free(timeline->hwsp_cacheline); 295 else 296 i915_gem_object_unpin_map(timeline->hwsp_ggtt->obj); 297 298 i915_vma_put(timeline->hwsp_ggtt); 299 300 /* 301 * A small race exists between intel_gt_retire_requests_timeout and 302 * intel_timeline_exit which could result in the syncmap not getting 303 * free'd. Rather than work to hard to seal this race, simply cleanup 304 * the syncmap on fini. 305 */ 306 i915_syncmap_free(&timeline->sync); 307} 308 309struct intel_timeline * 310__intel_timeline_create(struct intel_gt *gt, 311 struct i915_vma *global_hwsp, 312 unsigned int offset) 313{ 314 struct intel_timeline *timeline; 315 int err; 316 317 timeline = kzalloc(sizeof(*timeline), GFP_KERNEL); 318 if (!timeline) 319 return ERR_PTR(-ENOMEM); 320 321 err = intel_timeline_init(timeline, gt, global_hwsp, offset); 322 if (err) { 323 kfree(timeline); 324 return ERR_PTR(err); 325 } 326 327 return timeline; 328} 329 330void __intel_timeline_pin(struct intel_timeline *tl) 331{ 332 GEM_BUG_ON(!atomic_read(&tl->pin_count)); 333 atomic_inc(&tl->pin_count); 334} 335 336int intel_timeline_pin(struct intel_timeline *tl, struct i915_gem_ww_ctx *ww) 337{ 338 int err; 339 340 if (atomic_add_unless(&tl->pin_count, 1, 0)) 341 return 0; 342 343 err = i915_ggtt_pin(tl->hwsp_ggtt, ww, 0, PIN_HIGH); 344 if (err) 345 return err; 346 347 tl->hwsp_offset = 348 i915_ggtt_offset(tl->hwsp_ggtt) + 349 offset_in_page(tl->hwsp_offset); 350 GT_TRACE(tl->gt, "timeline:%llx using HWSP offset:%x\n", 351 tl->fence_context, tl->hwsp_offset); 352 353 cacheline_acquire(tl->hwsp_cacheline, tl->hwsp_offset); 354 if (atomic_fetch_inc(&tl->pin_count)) { 355 cacheline_release(tl->hwsp_cacheline); 356 __i915_vma_unpin(tl->hwsp_ggtt); 357 } 358 359 return 0; 360} 361 362void intel_timeline_reset_seqno(const struct intel_timeline *tl) 363{ 364 /* Must be pinned to be writable, and no requests in flight. */ 365 GEM_BUG_ON(!atomic_read(&tl->pin_count)); 366 WRITE_ONCE(*(u32 *)tl->hwsp_seqno, tl->seqno); 367} 368 369void intel_timeline_enter(struct intel_timeline *tl) 370{ 371 struct intel_gt_timelines *timelines = &tl->gt->timelines; 372 373 /* 374 * Pretend we are serialised by the timeline->mutex. 375 * 376 * While generally true, there are a few exceptions to the rule 377 * for the engine->kernel_context being used to manage power 378 * transitions. As the engine_park may be called from under any 379 * timeline, it uses the power mutex as a global serialisation 380 * lock to prevent any other request entering its timeline. 381 * 382 * The rule is generally tl->mutex, otherwise engine->wakeref.mutex. 383 * 384 * However, intel_gt_retire_request() does not know which engine 385 * it is retiring along and so cannot partake in the engine-pm 386 * barrier, and there we use the tl->active_count as a means to 387 * pin the timeline in the active_list while the locks are dropped. 388 * Ergo, as that is outside of the engine-pm barrier, we need to 389 * use atomic to manipulate tl->active_count. 390 */ 391 lockdep_assert_held(&tl->mutex); 392 393 if (atomic_add_unless(&tl->active_count, 1, 0)) 394 return; 395 396 spin_lock(&timelines->lock); 397 if (!atomic_fetch_inc(&tl->active_count)) { 398 /* 399 * The HWSP is volatile, and may have been lost while inactive, 400 * e.g. across suspend/resume. Be paranoid, and ensure that 401 * the HWSP value matches our seqno so we don't proclaim 402 * the next request as already complete. 403 */ 404 intel_timeline_reset_seqno(tl); 405 list_add_tail(&tl->link, &timelines->active_list); 406 } 407 spin_unlock(&timelines->lock); 408} 409 410void intel_timeline_exit(struct intel_timeline *tl) 411{ 412 struct intel_gt_timelines *timelines = &tl->gt->timelines; 413 414 /* See intel_timeline_enter() */ 415 lockdep_assert_held(&tl->mutex); 416 417 GEM_BUG_ON(!atomic_read(&tl->active_count)); 418 if (atomic_add_unless(&tl->active_count, -1, 1)) 419 return; 420 421 spin_lock(&timelines->lock); 422 if (atomic_dec_and_test(&tl->active_count)) 423 list_del(&tl->link); 424 spin_unlock(&timelines->lock); 425 426 /* 427 * Since this timeline is idle, all bariers upon which we were waiting 428 * must also be complete and so we can discard the last used barriers 429 * without loss of information. 430 */ 431 i915_syncmap_free(&tl->sync); 432} 433 434static u32 timeline_advance(struct intel_timeline *tl) 435{ 436 GEM_BUG_ON(!atomic_read(&tl->pin_count)); 437 GEM_BUG_ON(tl->seqno & tl->has_initial_breadcrumb); 438 439 return tl->seqno += 1 + tl->has_initial_breadcrumb; 440} 441 442static void timeline_rollback(struct intel_timeline *tl) 443{ 444 tl->seqno -= 1 + tl->has_initial_breadcrumb; 445} 446 447static noinline int 448__intel_timeline_get_seqno(struct intel_timeline *tl, 449 struct i915_request *rq, 450 u32 *seqno) 451{ 452 struct intel_timeline_cacheline *cl; 453 unsigned int cacheline; 454 struct i915_vma *vma; 455 void *vaddr; 456 int err; 457 458 might_lock(&tl->gt->ggtt->vm.mutex); 459 GT_TRACE(tl->gt, "timeline:%llx wrapped\n", tl->fence_context); 460 461 /* 462 * If there is an outstanding GPU reference to this cacheline, 463 * such as it being sampled by a HW semaphore on another timeline, 464 * we cannot wraparound our seqno value (the HW semaphore does 465 * a strict greater-than-or-equals compare, not i915_seqno_passed). 466 * So if the cacheline is still busy, we must detach ourselves 467 * from it and leave it inflight alongside its users. 468 * 469 * However, if nobody is watching and we can guarantee that nobody 470 * will, we could simply reuse the same cacheline. 471 * 472 * if (i915_active_request_is_signaled(&tl->last_request) && 473 * i915_active_is_signaled(&tl->hwsp_cacheline->active)) 474 * return 0; 475 * 476 * That seems unlikely for a busy timeline that needed to wrap in 477 * the first place, so just replace the cacheline. 478 */ 479 480 vma = hwsp_alloc(tl, &cacheline); 481 if (IS_ERR(vma)) { 482 err = PTR_ERR(vma); 483 goto err_rollback; 484 } 485 486 err = i915_ggtt_pin(vma, NULL, 0, PIN_HIGH); 487 if (err) { 488 __idle_hwsp_free(vma->private, cacheline); 489 goto err_rollback; 490 } 491 492 cl = cacheline_alloc(vma->private, cacheline); 493 if (IS_ERR(cl)) { 494 err = PTR_ERR(cl); 495 __idle_hwsp_free(vma->private, cacheline); 496 goto err_unpin; 497 } 498 GEM_BUG_ON(cl->hwsp->vma != vma); 499 500 /* 501 * Attach the old cacheline to the current request, so that we only 502 * free it after the current request is retired, which ensures that 503 * all writes into the cacheline from previous requests are complete. 504 */ 505 err = i915_active_ref(&tl->hwsp_cacheline->active, 506 tl->fence_context, 507 &rq->fence); 508 if (err) 509 goto err_cacheline; 510 511 cacheline_release(tl->hwsp_cacheline); /* ownership now xfered to rq */ 512 cacheline_free(tl->hwsp_cacheline); 513 514 i915_vma_unpin(tl->hwsp_ggtt); /* binding kept alive by old cacheline */ 515 i915_vma_put(tl->hwsp_ggtt); 516 517 tl->hwsp_ggtt = i915_vma_get(vma); 518 519 vaddr = page_mask_bits(cl->vaddr); 520 tl->hwsp_offset = cacheline * CACHELINE_BYTES; 521 tl->hwsp_seqno = 522 memset(vaddr + tl->hwsp_offset, 0, CACHELINE_BYTES); 523 524 tl->hwsp_offset += i915_ggtt_offset(vma); 525 GT_TRACE(tl->gt, "timeline:%llx using HWSP offset:%x\n", 526 tl->fence_context, tl->hwsp_offset); 527 528 cacheline_acquire(cl, tl->hwsp_offset); 529 tl->hwsp_cacheline = cl; 530 531 *seqno = timeline_advance(tl); 532 GEM_BUG_ON(i915_seqno_passed(*tl->hwsp_seqno, *seqno)); 533 return 0; 534 535err_cacheline: 536 cacheline_free(cl); 537err_unpin: 538 i915_vma_unpin(vma); 539err_rollback: 540 timeline_rollback(tl); 541 return err; 542} 543 544int intel_timeline_get_seqno(struct intel_timeline *tl, 545 struct i915_request *rq, 546 u32 *seqno) 547{ 548 *seqno = timeline_advance(tl); 549 550 /* Replace the HWSP on wraparound for HW semaphores */ 551 if (unlikely(!*seqno && tl->hwsp_cacheline)) 552 return __intel_timeline_get_seqno(tl, rq, seqno); 553 554 return 0; 555} 556 557static int cacheline_ref(struct intel_timeline_cacheline *cl, 558 struct i915_request *rq) 559{ 560 return i915_active_add_request(&cl->active, rq); 561} 562 563int intel_timeline_read_hwsp(struct i915_request *from, 564 struct i915_request *to, 565 u32 *hwsp) 566{ 567 struct intel_timeline_cacheline *cl; 568 int err; 569 570 GEM_BUG_ON(!rcu_access_pointer(from->hwsp_cacheline)); 571 572 rcu_read_lock(); 573 cl = rcu_dereference(from->hwsp_cacheline); 574 if (i915_request_completed(from)) /* confirm cacheline is valid */ 575 goto unlock; 576 if (unlikely(!i915_active_acquire_if_busy(&cl->active))) 577 goto unlock; /* seqno wrapped and completed! */ 578 if (unlikely(i915_request_completed(from))) 579 goto release; 580 rcu_read_unlock(); 581 582 err = cacheline_ref(cl, to); 583 if (err) 584 goto out; 585 586 *hwsp = cl->ggtt_offset; 587out: 588 i915_active_release(&cl->active); 589 return err; 590 591release: 592 i915_active_release(&cl->active); 593unlock: 594 rcu_read_unlock(); 595 return 1; 596} 597 598void intel_timeline_unpin(struct intel_timeline *tl) 599{ 600 GEM_BUG_ON(!atomic_read(&tl->pin_count)); 601 if (!atomic_dec_and_test(&tl->pin_count)) 602 return; 603 604 cacheline_release(tl->hwsp_cacheline); 605 606 __i915_vma_unpin(tl->hwsp_ggtt); 607} 608 609void __intel_timeline_free(struct kref *kref) 610{ 611 struct intel_timeline *timeline = 612 container_of(kref, typeof(*timeline), kref); 613 614 intel_timeline_fini(timeline); 615 kfree_rcu(timeline, rcu); 616} 617 618void intel_gt_fini_timelines(struct intel_gt *gt) 619{ 620 struct intel_gt_timelines *timelines = >->timelines; 621 622 GEM_BUG_ON(!list_empty(&timelines->active_list)); 623 GEM_BUG_ON(!list_empty(&timelines->hwsp_free_list)); 624} 625 626#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 627#include "gt/selftests/mock_timeline.c" 628#include "gt/selftest_timeline.c" 629#endif 630