1 /* 2 * Copyright (C) 2012-2014 Canonical Ltd (Maarten Lankhorst) 3 * 4 * Based on bo.c which bears the following copyright notice, 5 * but is dual licensed: 6 * 7 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA 8 * All Rights Reserved. 9 * 10 * Permission is hereby granted, free of charge, to any person obtaining a 11 * copy of this software and associated documentation files (the 12 * "Software"), to deal in the Software without restriction, including 13 * without limitation the rights to use, copy, modify, merge, publish, 14 * distribute, sub license, and/or sell copies of the Software, and to 15 * permit persons to whom the Software is furnished to do so, subject to 16 * the following conditions: 17 * 18 * The above copyright notice and this permission notice (including the 19 * next paragraph) shall be included in all copies or substantial portions 20 * of the Software. 21 * 22 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 23 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 24 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 25 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, 26 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 27 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 28 * USE OR OTHER DEALINGS IN THE SOFTWARE. 29 * 30 **************************************************************************/ 31 /* 32 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com> 33 */ 34 35 #include <linux/dma-resv.h> 36 #include <linux/export.h> 37 #include <linux/mm.h> 38 #include <linux/sched/mm.h> 39 40 /** 41 * DOC: Reservation Object Overview 42 * 43 * The reservation object provides a mechanism to manage shared and 44 * exclusive fences associated with a buffer. A reservation object 45 * can have attached one exclusive fence (normally associated with 46 * write operations) or N shared fences (read operations). The RCU 47 * mechanism is used to protect read access to fences from locked 48 * write-side updates. 49 */ 50 51 DEFINE_WD_CLASS(reservation_ww_class); 52 EXPORT_SYMBOL(reservation_ww_class); 53 54 struct lock_class_key reservation_seqcount_class; 55 EXPORT_SYMBOL(reservation_seqcount_class); 56 57 const char reservation_seqcount_string[] = "reservation_seqcount"; 58 EXPORT_SYMBOL(reservation_seqcount_string); 59 60 /** 61 * dma_resv_list_alloc - allocate fence list 62 * @shared_max: number of fences we need space for 63 * 64 * Allocate a new dma_resv_list and make sure to correctly initialize 65 * shared_max. 66 */ 67 static struct dma_resv_list *dma_resv_list_alloc(unsigned int shared_max) 68 { 69 struct dma_resv_list *list; 70 71 list = kmalloc(offsetof(typeof(*list), shared[shared_max]), GFP_KERNEL); 72 if (!list) 73 return NULL; 74 75 list->shared_max = (ksize(list) - offsetof(typeof(*list), shared)) / 76 sizeof(*list->shared); 77 78 return list; 79 } 80 81 /** 82 * dma_resv_list_free - free fence list 83 * @list: list to free 84 * 85 * Free a dma_resv_list and make sure to drop all references. 86 */ 87 static void dma_resv_list_free(struct dma_resv_list *list) 88 { 89 unsigned int i; 90 91 if (!list) 92 return; 93 94 for (i = 0; i < list->shared_count; ++i) 95 dma_fence_put(rcu_dereference_protected(list->shared[i], true)); 96 97 kfree_rcu(list, rcu); 98 } 99 100 #if IS_ENABLED(CONFIG_LOCKDEP) 101 static int __init dma_resv_lockdep(void) 102 { 103 struct mm_struct *mm = mm_alloc(); 104 struct ww_acquire_ctx ctx; 105 struct dma_resv obj; 106 int ret; 107 108 if (!mm) 109 return -ENOMEM; 110 111 dma_resv_init(&obj); 112 113 mmap_read_lock(mm); 114 ww_acquire_init(&ctx, &reservation_ww_class); 115 ret = dma_resv_lock(&obj, &ctx); 116 if (ret == -EDEADLK) 117 dma_resv_lock_slow(&obj, &ctx); 118 fs_reclaim_acquire(GFP_KERNEL); 119 fs_reclaim_release(GFP_KERNEL); 120 ww_mutex_unlock(&obj.lock); 121 ww_acquire_fini(&ctx); 122 mmap_read_unlock(mm); 123 124 mmput(mm); 125 126 return 0; 127 } 128 subsys_initcall(dma_resv_lockdep); 129 #endif 130 131 /** 132 * dma_resv_init - initialize a reservation object 133 * @obj: the reservation object 134 */ 135 void dma_resv_init(struct dma_resv *obj) 136 { 137 ww_mutex_init(&obj->lock, &reservation_ww_class); 138 139 __seqcount_init(&obj->seq, reservation_seqcount_string, 140 &reservation_seqcount_class); 141 RCU_INIT_POINTER(obj->fence, NULL); 142 RCU_INIT_POINTER(obj->fence_excl, NULL); 143 } 144 EXPORT_SYMBOL(dma_resv_init); 145 146 /** 147 * dma_resv_fini - destroys a reservation object 148 * @obj: the reservation object 149 */ 150 void dma_resv_fini(struct dma_resv *obj) 151 { 152 struct dma_resv_list *fobj; 153 struct dma_fence *excl; 154 155 /* 156 * This object should be dead and all references must have 157 * been released to it, so no need to be protected with rcu. 158 */ 159 excl = rcu_dereference_protected(obj->fence_excl, 1); 160 if (excl) 161 dma_fence_put(excl); 162 163 fobj = rcu_dereference_protected(obj->fence, 1); 164 dma_resv_list_free(fobj); 165 ww_mutex_destroy(&obj->lock); 166 } 167 EXPORT_SYMBOL(dma_resv_fini); 168 169 /** 170 * dma_resv_reserve_shared - Reserve space to add shared fences to 171 * a dma_resv. 172 * @obj: reservation object 173 * @num_fences: number of fences we want to add 174 * 175 * Should be called before dma_resv_add_shared_fence(). Must 176 * be called with obj->lock held. 177 * 178 * RETURNS 179 * Zero for success, or -errno 180 */ 181 int dma_resv_reserve_shared(struct dma_resv *obj, unsigned int num_fences) 182 { 183 struct dma_resv_list *old, *new; 184 unsigned int i, j, k, max; 185 186 dma_resv_assert_held(obj); 187 188 old = dma_resv_get_list(obj); 189 190 if (old && old->shared_max) { 191 if ((old->shared_count + num_fences) <= old->shared_max) 192 return 0; 193 else 194 max = max(old->shared_count + num_fences, 195 old->shared_max * 2); 196 } else { 197 max = 4; 198 } 199 200 new = dma_resv_list_alloc(max); 201 if (!new) 202 return -ENOMEM; 203 204 /* 205 * no need to bump fence refcounts, rcu_read access 206 * requires the use of kref_get_unless_zero, and the 207 * references from the old struct are carried over to 208 * the new. 209 */ 210 for (i = 0, j = 0, k = max; i < (old ? old->shared_count : 0); ++i) { 211 struct dma_fence *fence; 212 213 fence = rcu_dereference_protected(old->shared[i], 214 dma_resv_held(obj)); 215 if (dma_fence_is_signaled(fence)) 216 RCU_INIT_POINTER(new->shared[--k], fence); 217 else 218 RCU_INIT_POINTER(new->shared[j++], fence); 219 } 220 new->shared_count = j; 221 222 /* 223 * We are not changing the effective set of fences here so can 224 * merely update the pointer to the new array; both existing 225 * readers and new readers will see exactly the same set of 226 * active (unsignaled) shared fences. Individual fences and the 227 * old array are protected by RCU and so will not vanish under 228 * the gaze of the rcu_read_lock() readers. 229 */ 230 rcu_assign_pointer(obj->fence, new); 231 232 if (!old) 233 return 0; 234 235 /* Drop the references to the signaled fences */ 236 for (i = k; i < max; ++i) { 237 struct dma_fence *fence; 238 239 fence = rcu_dereference_protected(new->shared[i], 240 dma_resv_held(obj)); 241 dma_fence_put(fence); 242 } 243 kfree_rcu(old, rcu); 244 245 return 0; 246 } 247 EXPORT_SYMBOL(dma_resv_reserve_shared); 248 249 /** 250 * dma_resv_add_shared_fence - Add a fence to a shared slot 251 * @obj: the reservation object 252 * @fence: the shared fence to add 253 * 254 * Add a fence to a shared slot, obj->lock must be held, and 255 * dma_resv_reserve_shared() has been called. 256 */ 257 void dma_resv_add_shared_fence(struct dma_resv *obj, struct dma_fence *fence) 258 { 259 struct dma_resv_list *fobj; 260 struct dma_fence *old; 261 unsigned int i, count; 262 263 dma_fence_get(fence); 264 265 dma_resv_assert_held(obj); 266 267 fobj = dma_resv_get_list(obj); 268 count = fobj->shared_count; 269 270 preempt_disable(); 271 write_seqcount_begin(&obj->seq); 272 273 for (i = 0; i < count; ++i) { 274 275 old = rcu_dereference_protected(fobj->shared[i], 276 dma_resv_held(obj)); 277 if (old->context == fence->context || 278 dma_fence_is_signaled(old)) 279 goto replace; 280 } 281 282 BUG_ON(fobj->shared_count >= fobj->shared_max); 283 old = NULL; 284 count++; 285 286 replace: 287 RCU_INIT_POINTER(fobj->shared[i], fence); 288 /* pointer update must be visible before we extend the shared_count */ 289 smp_store_mb(fobj->shared_count, count); 290 291 write_seqcount_end(&obj->seq); 292 preempt_enable(); 293 dma_fence_put(old); 294 } 295 EXPORT_SYMBOL(dma_resv_add_shared_fence); 296 297 /** 298 * dma_resv_add_excl_fence - Add an exclusive fence. 299 * @obj: the reservation object 300 * @fence: the shared fence to add 301 * 302 * Add a fence to the exclusive slot. The obj->lock must be held. 303 */ 304 void dma_resv_add_excl_fence(struct dma_resv *obj, struct dma_fence *fence) 305 { 306 struct dma_fence *old_fence = dma_resv_get_excl(obj); 307 struct dma_resv_list *old; 308 u32 i = 0; 309 310 dma_resv_assert_held(obj); 311 312 old = dma_resv_get_list(obj); 313 if (old) 314 i = old->shared_count; 315 316 if (fence) 317 dma_fence_get(fence); 318 319 preempt_disable(); 320 write_seqcount_begin(&obj->seq); 321 /* write_seqcount_begin provides the necessary memory barrier */ 322 RCU_INIT_POINTER(obj->fence_excl, fence); 323 if (old) 324 old->shared_count = 0; 325 write_seqcount_end(&obj->seq); 326 preempt_enable(); 327 328 /* inplace update, no shared fences */ 329 while (i--) 330 dma_fence_put(rcu_dereference_protected(old->shared[i], 331 dma_resv_held(obj))); 332 333 dma_fence_put(old_fence); 334 } 335 EXPORT_SYMBOL(dma_resv_add_excl_fence); 336 337 /** 338 * dma_resv_copy_fences - Copy all fences from src to dst. 339 * @dst: the destination reservation object 340 * @src: the source reservation object 341 * 342 * Copy all fences from src to dst. dst-lock must be held. 343 */ 344 int dma_resv_copy_fences(struct dma_resv *dst, struct dma_resv *src) 345 { 346 struct dma_resv_list *src_list, *dst_list; 347 struct dma_fence *old, *new; 348 unsigned i; 349 350 dma_resv_assert_held(dst); 351 352 rcu_read_lock(); 353 src_list = rcu_dereference(src->fence); 354 355 retry: 356 if (src_list) { 357 unsigned shared_count = src_list->shared_count; 358 359 rcu_read_unlock(); 360 361 dst_list = dma_resv_list_alloc(shared_count); 362 if (!dst_list) 363 return -ENOMEM; 364 365 rcu_read_lock(); 366 src_list = rcu_dereference(src->fence); 367 if (!src_list || src_list->shared_count > shared_count) { 368 kfree(dst_list); 369 goto retry; 370 } 371 372 dst_list->shared_count = 0; 373 for (i = 0; i < src_list->shared_count; ++i) { 374 struct dma_fence *fence; 375 376 fence = rcu_dereference(src_list->shared[i]); 377 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, 378 &fence->flags)) 379 continue; 380 381 if (!dma_fence_get_rcu(fence)) { 382 dma_resv_list_free(dst_list); 383 src_list = rcu_dereference(src->fence); 384 goto retry; 385 } 386 387 if (dma_fence_is_signaled(fence)) { 388 dma_fence_put(fence); 389 continue; 390 } 391 392 rcu_assign_pointer(dst_list->shared[dst_list->shared_count++], fence); 393 } 394 } else { 395 dst_list = NULL; 396 } 397 398 new = dma_fence_get_rcu_safe(&src->fence_excl); 399 rcu_read_unlock(); 400 401 src_list = dma_resv_get_list(dst); 402 old = dma_resv_get_excl(dst); 403 404 preempt_disable(); 405 write_seqcount_begin(&dst->seq); 406 /* write_seqcount_begin provides the necessary memory barrier */ 407 RCU_INIT_POINTER(dst->fence_excl, new); 408 RCU_INIT_POINTER(dst->fence, dst_list); 409 write_seqcount_end(&dst->seq); 410 preempt_enable(); 411 412 dma_resv_list_free(src_list); 413 dma_fence_put(old); 414 415 return 0; 416 } 417 EXPORT_SYMBOL(dma_resv_copy_fences); 418 419 /** 420 * dma_resv_get_fences_rcu - Get an object's shared and exclusive 421 * fences without update side lock held 422 * @obj: the reservation object 423 * @pfence_excl: the returned exclusive fence (or NULL) 424 * @pshared_count: the number of shared fences returned 425 * @pshared: the array of shared fence ptrs returned (array is krealloc'd to 426 * the required size, and must be freed by caller) 427 * 428 * Retrieve all fences from the reservation object. If the pointer for the 429 * exclusive fence is not specified the fence is put into the array of the 430 * shared fences as well. Returns either zero or -ENOMEM. 431 */ 432 int dma_resv_get_fences_rcu(struct dma_resv *obj, 433 struct dma_fence **pfence_excl, 434 unsigned *pshared_count, 435 struct dma_fence ***pshared) 436 { 437 struct dma_fence **shared = NULL; 438 struct dma_fence *fence_excl; 439 unsigned int shared_count; 440 int ret = 1; 441 442 do { 443 struct dma_resv_list *fobj; 444 unsigned int i, seq; 445 size_t sz = 0; 446 447 shared_count = i = 0; 448 449 rcu_read_lock(); 450 seq = read_seqcount_begin(&obj->seq); 451 452 fence_excl = rcu_dereference(obj->fence_excl); 453 if (fence_excl && !dma_fence_get_rcu(fence_excl)) 454 goto unlock; 455 456 fobj = rcu_dereference(obj->fence); 457 if (fobj) 458 sz += sizeof(*shared) * fobj->shared_max; 459 460 if (!pfence_excl && fence_excl) 461 sz += sizeof(*shared); 462 463 if (sz) { 464 struct dma_fence **nshared; 465 466 nshared = krealloc(shared, sz, 467 GFP_NOWAIT | __GFP_NOWARN); 468 if (!nshared) { 469 rcu_read_unlock(); 470 471 dma_fence_put(fence_excl); 472 fence_excl = NULL; 473 474 nshared = krealloc(shared, sz, GFP_KERNEL); 475 if (nshared) { 476 shared = nshared; 477 continue; 478 } 479 480 ret = -ENOMEM; 481 break; 482 } 483 shared = nshared; 484 shared_count = fobj ? fobj->shared_count : 0; 485 for (i = 0; i < shared_count; ++i) { 486 shared[i] = rcu_dereference(fobj->shared[i]); 487 if (!dma_fence_get_rcu(shared[i])) 488 break; 489 } 490 } 491 492 if (i != shared_count || read_seqcount_retry(&obj->seq, seq)) { 493 while (i--) 494 dma_fence_put(shared[i]); 495 dma_fence_put(fence_excl); 496 goto unlock; 497 } 498 499 ret = 0; 500 unlock: 501 rcu_read_unlock(); 502 } while (ret); 503 504 if (pfence_excl) 505 *pfence_excl = fence_excl; 506 else if (fence_excl) 507 shared[shared_count++] = fence_excl; 508 509 if (!shared_count) { 510 kfree(shared); 511 shared = NULL; 512 } 513 514 *pshared_count = shared_count; 515 *pshared = shared; 516 return ret; 517 } 518 EXPORT_SYMBOL_GPL(dma_resv_get_fences_rcu); 519 520 /** 521 * dma_resv_wait_timeout_rcu - Wait on reservation's objects 522 * shared and/or exclusive fences. 523 * @obj: the reservation object 524 * @wait_all: if true, wait on all fences, else wait on just exclusive fence 525 * @intr: if true, do interruptible wait 526 * @timeout: timeout value in jiffies or zero to return immediately 527 * 528 * RETURNS 529 * Returns -ERESTARTSYS if interrupted, 0 if the wait timed out, or 530 * greater than zer on success. 531 */ 532 long dma_resv_wait_timeout_rcu(struct dma_resv *obj, 533 bool wait_all, bool intr, 534 unsigned long timeout) 535 { 536 struct dma_fence *fence; 537 unsigned seq, shared_count; 538 long ret = timeout ? timeout : 1; 539 int i; 540 541 retry: 542 shared_count = 0; 543 seq = read_seqcount_begin(&obj->seq); 544 rcu_read_lock(); 545 i = -1; 546 547 fence = rcu_dereference(obj->fence_excl); 548 if (fence && !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) { 549 if (!dma_fence_get_rcu(fence)) 550 goto unlock_retry; 551 552 if (dma_fence_is_signaled(fence)) { 553 dma_fence_put(fence); 554 fence = NULL; 555 } 556 557 } else { 558 fence = NULL; 559 } 560 561 if (wait_all) { 562 struct dma_resv_list *fobj = rcu_dereference(obj->fence); 563 564 if (fobj) 565 shared_count = fobj->shared_count; 566 567 for (i = 0; !fence && i < shared_count; ++i) { 568 struct dma_fence *lfence = rcu_dereference(fobj->shared[i]); 569 570 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, 571 &lfence->flags)) 572 continue; 573 574 if (!dma_fence_get_rcu(lfence)) 575 goto unlock_retry; 576 577 if (dma_fence_is_signaled(lfence)) { 578 dma_fence_put(lfence); 579 continue; 580 } 581 582 fence = lfence; 583 break; 584 } 585 } 586 587 rcu_read_unlock(); 588 if (fence) { 589 if (read_seqcount_retry(&obj->seq, seq)) { 590 dma_fence_put(fence); 591 goto retry; 592 } 593 594 ret = dma_fence_wait_timeout(fence, intr, ret); 595 dma_fence_put(fence); 596 if (ret > 0 && wait_all && (i + 1 < shared_count)) 597 goto retry; 598 } 599 return ret; 600 601 unlock_retry: 602 rcu_read_unlock(); 603 goto retry; 604 } 605 EXPORT_SYMBOL_GPL(dma_resv_wait_timeout_rcu); 606 607 608 static inline int dma_resv_test_signaled_single(struct dma_fence *passed_fence) 609 { 610 struct dma_fence *fence, *lfence = passed_fence; 611 int ret = 1; 612 613 if (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &lfence->flags)) { 614 fence = dma_fence_get_rcu(lfence); 615 if (!fence) 616 return -1; 617 618 ret = !!dma_fence_is_signaled(fence); 619 dma_fence_put(fence); 620 } 621 return ret; 622 } 623 624 /** 625 * dma_resv_test_signaled_rcu - Test if a reservation object's 626 * fences have been signaled. 627 * @obj: the reservation object 628 * @test_all: if true, test all fences, otherwise only test the exclusive 629 * fence 630 * 631 * RETURNS 632 * true if all fences signaled, else false 633 */ 634 bool dma_resv_test_signaled_rcu(struct dma_resv *obj, bool test_all) 635 { 636 unsigned seq, shared_count; 637 int ret; 638 639 rcu_read_lock(); 640 retry: 641 ret = true; 642 shared_count = 0; 643 seq = read_seqcount_begin(&obj->seq); 644 645 if (test_all) { 646 unsigned i; 647 648 struct dma_resv_list *fobj = rcu_dereference(obj->fence); 649 650 if (fobj) 651 shared_count = fobj->shared_count; 652 653 for (i = 0; i < shared_count; ++i) { 654 struct dma_fence *fence = rcu_dereference(fobj->shared[i]); 655 656 ret = dma_resv_test_signaled_single(fence); 657 if (ret < 0) 658 goto retry; 659 else if (!ret) 660 break; 661 } 662 663 if (read_seqcount_retry(&obj->seq, seq)) 664 goto retry; 665 } 666 667 if (!shared_count) { 668 struct dma_fence *fence_excl = rcu_dereference(obj->fence_excl); 669 670 if (fence_excl) { 671 ret = dma_resv_test_signaled_single(fence_excl); 672 if (ret < 0) 673 goto retry; 674 675 if (read_seqcount_retry(&obj->seq, seq)) 676 goto retry; 677 } 678 } 679 680 rcu_read_unlock(); 681 return ret; 682 } 683 EXPORT_SYMBOL_GPL(dma_resv_test_signaled_rcu); 684