1/* 2 * This file and its contents are supplied under the terms of the 3 * Common Development and Distribution License ("CDDL"), version 1.0. 4 * You may only use this file in accordance with the terms of version 5 * 1.0 of the CDDL. 6 * 7 * A full copy of the text of the CDDL should have accompanied this 8 * source. A copy of the CDDL is also available via the Internet at 9 * http://www.illumos.org/license/CDDL. 10 */ 11 12/* 13 * Copyright (c) 2014 by Chunwei Chen. All rights reserved. 14 * Copyright (c) 2016 by Delphix. All rights reserved. 15 */ 16 17/* 18 * ARC buffer data (ABD). 19 * 20 * ABDs are an abstract data structure for the ARC which can use two 21 * different ways of storing the underlying data: 22 * 23 * (a) Linear buffer. In this case, all the data in the ABD is stored in one 24 * contiguous buffer in memory (from a zio_[data_]buf_* kmem cache). 25 * 26 * +-------------------+ 27 * | ABD (linear) | 28 * | abd_flags = ... | 29 * | abd_size = ... | +--------------------------------+ 30 * | abd_buf ------------->| raw buffer of size abd_size | 31 * +-------------------+ +--------------------------------+ 32 * no abd_chunks 33 * 34 * (b) Scattered buffer. In this case, the data in the ABD is split into 35 * equal-sized chunks (from the abd_chunk_cache kmem_cache), with pointers 36 * to the chunks recorded in an array at the end of the ABD structure. 37 * 38 * +-------------------+ 39 * | ABD (scattered) | 40 * | abd_flags = ... | 41 * | abd_size = ... | 42 * | abd_offset = 0 | +-----------+ 43 * | abd_chunks[0] ----------------------------->| chunk 0 | 44 * | abd_chunks[1] ---------------------+ +-----------+ 45 * | ... | | +-----------+ 46 * | abd_chunks[N-1] ---------+ +------->| chunk 1 | 47 * +-------------------+ | +-----------+ 48 * | ... 49 * | +-----------+ 50 * +----------------->| chunk N-1 | 51 * +-----------+ 52 * 53 * Using a large proportion of scattered ABDs decreases ARC fragmentation since 54 * when we are at the limit of allocatable space, using equal-size chunks will 55 * allow us to quickly reclaim enough space for a new large allocation (assuming 56 * it is also scattered). 57 * 58 * In addition to directly allocating a linear or scattered ABD, it is also 59 * possible to create an ABD by requesting the "sub-ABD" starting at an offset 60 * within an existing ABD. In linear buffers this is simple (set abd_buf of 61 * the new ABD to the starting point within the original raw buffer), but 62 * scattered ABDs are a little more complex. The new ABD makes a copy of the 63 * relevant abd_chunks pointers (but not the underlying data). However, to 64 * provide arbitrary rather than only chunk-aligned starting offsets, it also 65 * tracks an abd_offset field which represents the starting point of the data 66 * within the first chunk in abd_chunks. For both linear and scattered ABDs, 67 * creating an offset ABD marks the original ABD as the offset's parent, and the 68 * original ABD's abd_children refcount is incremented. This data allows us to 69 * ensure the root ABD isn't deleted before its children. 70 * 71 * Most consumers should never need to know what type of ABD they're using -- 72 * the ABD public API ensures that it's possible to transparently switch from 73 * using a linear ABD to a scattered one when doing so would be beneficial. 74 * 75 * If you need to use the data within an ABD directly, if you know it's linear 76 * (because you allocated it) you can use abd_to_buf() to access the underlying 77 * raw buffer. Otherwise, you should use one of the abd_borrow_buf* functions 78 * which will allocate a raw buffer if necessary. Use the abd_return_buf* 79 * functions to return any raw buffers that are no longer necessary when you're 80 * done using them. 81 * 82 * There are a variety of ABD APIs that implement basic buffer operations: 83 * compare, copy, read, write, and fill with zeroes. If you need a custom 84 * function which progressively accesses the whole ABD, use the abd_iterate_* 85 * functions. 86 */ 87 88#include <sys/abd.h> 89#include <sys/param.h> 90#include <sys/zio.h> 91#include <sys/zfs_context.h> 92#include <sys/zfs_znode.h> 93 94typedef struct abd_stats { 95 kstat_named_t abdstat_struct_size; 96 kstat_named_t abdstat_scatter_cnt; 97 kstat_named_t abdstat_scatter_data_size; 98 kstat_named_t abdstat_scatter_chunk_waste; 99 kstat_named_t abdstat_linear_cnt; 100 kstat_named_t abdstat_linear_data_size; 101} abd_stats_t; 102 103static abd_stats_t abd_stats = { 104 /* Amount of memory occupied by all of the abd_t struct allocations */ 105 { "struct_size", KSTAT_DATA_UINT64 }, 106 /* 107 * The number of scatter ABDs which are currently allocated, excluding 108 * ABDs which don't own their data (for instance the ones which were 109 * allocated through abd_get_offset()). 110 */ 111 { "scatter_cnt", KSTAT_DATA_UINT64 }, 112 /* Amount of data stored in all scatter ABDs tracked by scatter_cnt */ 113 { "scatter_data_size", KSTAT_DATA_UINT64 }, 114 /* 115 * The amount of space wasted at the end of the last chunk across all 116 * scatter ABDs tracked by scatter_cnt. 117 */ 118 { "scatter_chunk_waste", KSTAT_DATA_UINT64 }, 119 /* 120 * The number of linear ABDs which are currently allocated, excluding 121 * ABDs which don't own their data (for instance the ones which were 122 * allocated through abd_get_offset() and abd_get_from_buf()). If an 123 * ABD takes ownership of its buf then it will become tracked. 124 */ 125 { "linear_cnt", KSTAT_DATA_UINT64 }, 126 /* Amount of data stored in all linear ABDs tracked by linear_cnt */ 127 { "linear_data_size", KSTAT_DATA_UINT64 }, 128}; 129 130#define ABDSTAT(stat) (abd_stats.stat.value.ui64) 131#define ABDSTAT_INCR(stat, val) \ 132 atomic_add_64(&abd_stats.stat.value.ui64, (val)) 133#define ABDSTAT_BUMP(stat) ABDSTAT_INCR(stat, 1) 134#define ABDSTAT_BUMPDOWN(stat) ABDSTAT_INCR(stat, -1) 135 136/* 137 * It is possible to make all future ABDs be linear by setting this to B_FALSE. 138 * Otherwise, ABDs are allocated scattered by default unless the caller uses 139 * abd_alloc_linear(). 140 */ 141boolean_t zfs_abd_scatter_enabled = B_TRUE; 142 143/* 144 * The size of the chunks ABD allocates. Because the sizes allocated from the 145 * kmem_cache can't change, this tunable can only be modified at boot. Changing 146 * it at runtime would cause ABD iteration to work incorrectly for ABDs which 147 * were allocated with the old size, so a safeguard has been put in place which 148 * will cause the machine to panic if you change it and try to access the data 149 * within a scattered ABD. 150 */ 151size_t zfs_abd_chunk_size = 4096; 152 153#if defined(__FreeBSD__) && defined(_KERNEL) 154SYSCTL_DECL(_vfs_zfs); 155 156SYSCTL_INT(_vfs_zfs, OID_AUTO, abd_scatter_enabled, CTLFLAG_RWTUN, 157 &zfs_abd_scatter_enabled, 0, "Enable scattered ARC data buffers"); 158SYSCTL_ULONG(_vfs_zfs, OID_AUTO, abd_chunk_size, CTLFLAG_RDTUN, 159 &zfs_abd_chunk_size, 0, "The size of the chunks ABD allocates"); 160#endif 161 162#ifdef _KERNEL 163extern vmem_t *zio_alloc_arena; 164#endif 165 166kmem_cache_t *abd_chunk_cache; 167static kstat_t *abd_ksp; 168 169extern inline boolean_t abd_is_linear(abd_t *abd); 170extern inline void abd_copy(abd_t *dabd, abd_t *sabd, size_t size); 171extern inline void abd_copy_from_buf(abd_t *abd, const void *buf, size_t size); 172extern inline void abd_copy_to_buf(void* buf, abd_t *abd, size_t size); 173extern inline int abd_cmp_buf(abd_t *abd, const void *buf, size_t size); 174extern inline void abd_zero(abd_t *abd, size_t size); 175 176static void * 177abd_alloc_chunk() 178{ 179 void *c = kmem_cache_alloc(abd_chunk_cache, KM_PUSHPAGE); 180 ASSERT3P(c, !=, NULL); 181 return (c); 182} 183 184static void 185abd_free_chunk(void *c) 186{ 187 kmem_cache_free(abd_chunk_cache, c); 188} 189 190void 191abd_init(void) 192{ 193#ifdef illumos 194 vmem_t *data_alloc_arena = NULL; 195 196#ifdef _KERNEL 197 data_alloc_arena = zio_alloc_arena; 198#endif 199 200 /* 201 * Since ABD chunks do not appear in crash dumps, we pass KMC_NOTOUCH 202 * so that no allocator metadata is stored with the buffers. 203 */ 204 abd_chunk_cache = kmem_cache_create("abd_chunk", zfs_abd_chunk_size, 0, 205 NULL, NULL, NULL, NULL, data_alloc_arena, KMC_NOTOUCH); 206#else 207 abd_chunk_cache = kmem_cache_create("abd_chunk", zfs_abd_chunk_size, 0, 208 NULL, NULL, NULL, NULL, 0, KMC_NOTOUCH | KMC_NODEBUG); 209#endif 210 abd_ksp = kstat_create("zfs", 0, "abdstats", "misc", KSTAT_TYPE_NAMED, 211 sizeof (abd_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 212 if (abd_ksp != NULL) { 213 abd_ksp->ks_data = &abd_stats; 214 kstat_install(abd_ksp); 215 } 216} 217 218void 219abd_fini(void) 220{ 221 if (abd_ksp != NULL) { 222 kstat_delete(abd_ksp); 223 abd_ksp = NULL; 224 } 225 226 kmem_cache_destroy(abd_chunk_cache); 227 abd_chunk_cache = NULL; 228} 229 230static inline size_t 231abd_chunkcnt_for_bytes(size_t size) 232{ 233 return (P2ROUNDUP(size, zfs_abd_chunk_size) / zfs_abd_chunk_size); 234} 235 236static inline size_t 237abd_scatter_chunkcnt(abd_t *abd) 238{ 239 ASSERT(!abd_is_linear(abd)); 240 return (abd_chunkcnt_for_bytes( 241 abd->abd_u.abd_scatter.abd_offset + abd->abd_size)); 242} 243 244static inline void 245abd_verify(abd_t *abd) 246{ 247 ASSERT3U(abd->abd_size, >, 0); 248 ASSERT3U(abd->abd_size, <=, SPA_MAXBLOCKSIZE); 249 ASSERT3U(abd->abd_flags, ==, abd->abd_flags & (ABD_FLAG_LINEAR | 250 ABD_FLAG_OWNER | ABD_FLAG_META)); 251 IMPLY(abd->abd_parent != NULL, !(abd->abd_flags & ABD_FLAG_OWNER)); 252 IMPLY(abd->abd_flags & ABD_FLAG_META, abd->abd_flags & ABD_FLAG_OWNER); 253 if (abd_is_linear(abd)) { 254 ASSERT3P(abd->abd_u.abd_linear.abd_buf, !=, NULL); 255 } else { 256 ASSERT3U(abd->abd_u.abd_scatter.abd_offset, <, 257 zfs_abd_chunk_size); 258 size_t n = abd_scatter_chunkcnt(abd); 259 for (int i = 0; i < n; i++) { 260 ASSERT3P( 261 abd->abd_u.abd_scatter.abd_chunks[i], !=, NULL); 262 } 263 } 264} 265 266static inline abd_t * 267abd_alloc_struct(size_t chunkcnt) 268{ 269 size_t size = offsetof(abd_t, abd_u.abd_scatter.abd_chunks[chunkcnt]); 270 abd_t *abd = kmem_alloc(size, KM_PUSHPAGE); 271 ASSERT3P(abd, !=, NULL); 272 ABDSTAT_INCR(abdstat_struct_size, size); 273 274 return (abd); 275} 276 277static inline void 278abd_free_struct(abd_t *abd) 279{ 280 size_t chunkcnt = abd_is_linear(abd) ? 0 : abd_scatter_chunkcnt(abd); 281 int size = offsetof(abd_t, abd_u.abd_scatter.abd_chunks[chunkcnt]); 282 kmem_free(abd, size); 283 ABDSTAT_INCR(abdstat_struct_size, -size); 284} 285 286/* 287 * Allocate an ABD, along with its own underlying data buffers. Use this if you 288 * don't care whether the ABD is linear or not. 289 */ 290abd_t * 291abd_alloc(size_t size, boolean_t is_metadata) 292{ 293 if (!zfs_abd_scatter_enabled || size <= zfs_abd_chunk_size) 294 return (abd_alloc_linear(size, is_metadata)); 295 296 VERIFY3U(size, <=, SPA_MAXBLOCKSIZE); 297 298 size_t n = abd_chunkcnt_for_bytes(size); 299 abd_t *abd = abd_alloc_struct(n); 300 301 abd->abd_flags = ABD_FLAG_OWNER; 302 if (is_metadata) { 303 abd->abd_flags |= ABD_FLAG_META; 304 } 305 abd->abd_size = size; 306 abd->abd_parent = NULL; 307 zfs_refcount_create(&abd->abd_children); 308 309 abd->abd_u.abd_scatter.abd_offset = 0; 310 abd->abd_u.abd_scatter.abd_chunk_size = zfs_abd_chunk_size; 311 312 for (int i = 0; i < n; i++) { 313 void *c = abd_alloc_chunk(); 314 ASSERT3P(c, !=, NULL); 315 abd->abd_u.abd_scatter.abd_chunks[i] = c; 316 } 317 318 ABDSTAT_BUMP(abdstat_scatter_cnt); 319 ABDSTAT_INCR(abdstat_scatter_data_size, size); 320 ABDSTAT_INCR(abdstat_scatter_chunk_waste, 321 n * zfs_abd_chunk_size - size); 322 323 return (abd); 324} 325 326static void 327abd_free_scatter(abd_t *abd) 328{ 329 size_t n = abd_scatter_chunkcnt(abd); 330 for (int i = 0; i < n; i++) { 331 abd_free_chunk(abd->abd_u.abd_scatter.abd_chunks[i]); 332 } 333 334 zfs_refcount_destroy(&abd->abd_children); 335 ABDSTAT_BUMPDOWN(abdstat_scatter_cnt); 336 ABDSTAT_INCR(abdstat_scatter_data_size, -(int)abd->abd_size); 337 ABDSTAT_INCR(abdstat_scatter_chunk_waste, 338 abd->abd_size - n * zfs_abd_chunk_size); 339 340 abd_free_struct(abd); 341} 342 343/* 344 * Allocate an ABD that must be linear, along with its own underlying data 345 * buffer. Only use this when it would be very annoying to write your ABD 346 * consumer with a scattered ABD. 347 */ 348abd_t * 349abd_alloc_linear(size_t size, boolean_t is_metadata) 350{ 351 abd_t *abd = abd_alloc_struct(0); 352 353 VERIFY3U(size, <=, SPA_MAXBLOCKSIZE); 354 355 abd->abd_flags = ABD_FLAG_LINEAR | ABD_FLAG_OWNER; 356 if (is_metadata) { 357 abd->abd_flags |= ABD_FLAG_META; 358 } 359 abd->abd_size = size; 360 abd->abd_parent = NULL; 361 zfs_refcount_create(&abd->abd_children); 362 363 if (is_metadata) { 364 abd->abd_u.abd_linear.abd_buf = zio_buf_alloc(size); 365 } else { 366 abd->abd_u.abd_linear.abd_buf = zio_data_buf_alloc(size); 367 } 368 369 ABDSTAT_BUMP(abdstat_linear_cnt); 370 ABDSTAT_INCR(abdstat_linear_data_size, size); 371 372 return (abd); 373} 374 375static void 376abd_free_linear(abd_t *abd) 377{ 378 if (abd->abd_flags & ABD_FLAG_META) { 379 zio_buf_free(abd->abd_u.abd_linear.abd_buf, abd->abd_size); 380 } else { 381 zio_data_buf_free(abd->abd_u.abd_linear.abd_buf, abd->abd_size); 382 } 383 384 zfs_refcount_destroy(&abd->abd_children); 385 ABDSTAT_BUMPDOWN(abdstat_linear_cnt); 386 ABDSTAT_INCR(abdstat_linear_data_size, -(int)abd->abd_size); 387 388 abd_free_struct(abd); 389} 390 391/* 392 * Free an ABD. Only use this on ABDs allocated with abd_alloc() or 393 * abd_alloc_linear(). 394 */ 395void 396abd_free(abd_t *abd) 397{ 398 abd_verify(abd); 399 ASSERT3P(abd->abd_parent, ==, NULL); 400 ASSERT(abd->abd_flags & ABD_FLAG_OWNER); 401 if (abd_is_linear(abd)) 402 abd_free_linear(abd); 403 else 404 abd_free_scatter(abd); 405} 406 407/* 408 * Allocate an ABD of the same format (same metadata flag, same scatterize 409 * setting) as another ABD. 410 */ 411abd_t * 412abd_alloc_sametype(abd_t *sabd, size_t size) 413{ 414 boolean_t is_metadata = (sabd->abd_flags & ABD_FLAG_META) != 0; 415 if (abd_is_linear(sabd)) { 416 return (abd_alloc_linear(size, is_metadata)); 417 } else { 418 return (abd_alloc(size, is_metadata)); 419 } 420} 421 422/* 423 * If we're going to use this ABD for doing I/O using the block layer, the 424 * consumer of the ABD data doesn't care if it's scattered or not, and we don't 425 * plan to store this ABD in memory for a long period of time, we should 426 * allocate the ABD type that requires the least data copying to do the I/O. 427 * 428 * Currently this is linear ABDs, however if ldi_strategy() can ever issue I/Os 429 * using a scatter/gather list we should switch to that and replace this call 430 * with vanilla abd_alloc(). 431 */ 432abd_t * 433abd_alloc_for_io(size_t size, boolean_t is_metadata) 434{ 435 return (abd_alloc_linear(size, is_metadata)); 436} 437 438/* 439 * Allocate a new ABD to point to offset off of sabd. It shares the underlying 440 * buffer data with sabd. Use abd_put() to free. sabd must not be freed while 441 * any derived ABDs exist. 442 */ 443abd_t * 444abd_get_offset(abd_t *sabd, size_t off) 445{ 446 abd_t *abd; 447 448 abd_verify(sabd); 449 ASSERT3U(off, <=, sabd->abd_size); 450 451 if (abd_is_linear(sabd)) { 452 abd = abd_alloc_struct(0); 453 454 /* 455 * Even if this buf is filesystem metadata, we only track that 456 * if we own the underlying data buffer, which is not true in 457 * this case. Therefore, we don't ever use ABD_FLAG_META here. 458 */ 459 abd->abd_flags = ABD_FLAG_LINEAR; 460 461 abd->abd_u.abd_linear.abd_buf = 462 (char *)sabd->abd_u.abd_linear.abd_buf + off; 463 } else { 464 size_t new_offset = sabd->abd_u.abd_scatter.abd_offset + off; 465 size_t chunkcnt = abd_scatter_chunkcnt(sabd) - 466 (new_offset / zfs_abd_chunk_size); 467 468 abd = abd_alloc_struct(chunkcnt); 469 470 /* 471 * Even if this buf is filesystem metadata, we only track that 472 * if we own the underlying data buffer, which is not true in 473 * this case. Therefore, we don't ever use ABD_FLAG_META here. 474 */ 475 abd->abd_flags = 0; 476 477 abd->abd_u.abd_scatter.abd_offset = 478 new_offset % zfs_abd_chunk_size; 479 abd->abd_u.abd_scatter.abd_chunk_size = zfs_abd_chunk_size; 480 481 /* Copy the scatterlist starting at the correct offset */ 482 (void) memcpy(&abd->abd_u.abd_scatter.abd_chunks, 483 &sabd->abd_u.abd_scatter.abd_chunks[new_offset / 484 zfs_abd_chunk_size], 485 chunkcnt * sizeof (void *)); 486 } 487 488 abd->abd_size = sabd->abd_size - off; 489 abd->abd_parent = sabd; 490 zfs_refcount_create(&abd->abd_children); 491 (void) zfs_refcount_add_many(&sabd->abd_children, abd->abd_size, abd); 492 493 return (abd); 494} 495 496/* 497 * Allocate a linear ABD structure for buf. You must free this with abd_put() 498 * since the resulting ABD doesn't own its own buffer. 499 */ 500abd_t * 501abd_get_from_buf(void *buf, size_t size) 502{ 503 abd_t *abd = abd_alloc_struct(0); 504 505 VERIFY3U(size, <=, SPA_MAXBLOCKSIZE); 506 507 /* 508 * Even if this buf is filesystem metadata, we only track that if we 509 * own the underlying data buffer, which is not true in this case. 510 * Therefore, we don't ever use ABD_FLAG_META here. 511 */ 512 abd->abd_flags = ABD_FLAG_LINEAR; 513 abd->abd_size = size; 514 abd->abd_parent = NULL; 515 zfs_refcount_create(&abd->abd_children); 516 517 abd->abd_u.abd_linear.abd_buf = buf; 518 519 return (abd); 520} 521 522/* 523 * Free an ABD allocated from abd_get_offset() or abd_get_from_buf(). Will not 524 * free the underlying scatterlist or buffer. 525 */ 526void 527abd_put(abd_t *abd) 528{ 529 abd_verify(abd); 530 ASSERT(!(abd->abd_flags & ABD_FLAG_OWNER)); 531 532 if (abd->abd_parent != NULL) { 533 (void) zfs_refcount_remove_many(&abd->abd_parent->abd_children, 534 abd->abd_size, abd); 535 } 536 537 zfs_refcount_destroy(&abd->abd_children); 538 abd_free_struct(abd); 539} 540 541/* 542 * Get the raw buffer associated with a linear ABD. 543 */ 544void * 545abd_to_buf(abd_t *abd) 546{ 547 ASSERT(abd_is_linear(abd)); 548 abd_verify(abd); 549 return (abd->abd_u.abd_linear.abd_buf); 550} 551 552/* 553 * Borrow a raw buffer from an ABD without copying the contents of the ABD 554 * into the buffer. If the ABD is scattered, this will allocate a raw buffer 555 * whose contents are undefined. To copy over the existing data in the ABD, use 556 * abd_borrow_buf_copy() instead. 557 */ 558void * 559abd_borrow_buf(abd_t *abd, size_t n) 560{ 561 void *buf; 562 abd_verify(abd); 563 ASSERT3U(abd->abd_size, >=, n); 564 if (abd_is_linear(abd)) { 565 buf = abd_to_buf(abd); 566 } else { 567 buf = zio_buf_alloc(n); 568 } 569 (void) zfs_refcount_add_many(&abd->abd_children, n, buf); 570 571 return (buf); 572} 573 574void * 575abd_borrow_buf_copy(abd_t *abd, size_t n) 576{ 577 void *buf = abd_borrow_buf(abd, n); 578 if (!abd_is_linear(abd)) { 579 abd_copy_to_buf(buf, abd, n); 580 } 581 return (buf); 582} 583 584/* 585 * Return a borrowed raw buffer to an ABD. If the ABD is scattered, this will 586 * not change the contents of the ABD and will ASSERT that you didn't modify 587 * the buffer since it was borrowed. If you want any changes you made to buf to 588 * be copied back to abd, use abd_return_buf_copy() instead. 589 */ 590void 591abd_return_buf(abd_t *abd, void *buf, size_t n) 592{ 593 abd_verify(abd); 594 ASSERT3U(abd->abd_size, >=, n); 595 if (abd_is_linear(abd)) { 596 ASSERT3P(buf, ==, abd_to_buf(abd)); 597 } else { 598 ASSERT0(abd_cmp_buf(abd, buf, n)); 599 zio_buf_free(buf, n); 600 } 601 (void) zfs_refcount_remove_many(&abd->abd_children, n, buf); 602} 603 604void 605abd_return_buf_copy(abd_t *abd, void *buf, size_t n) 606{ 607 if (!abd_is_linear(abd)) { 608 abd_copy_from_buf(abd, buf, n); 609 } 610 abd_return_buf(abd, buf, n); 611} 612 613/* 614 * Give this ABD ownership of the buffer that it's storing. Can only be used on 615 * linear ABDs which were allocated via abd_get_from_buf(), or ones allocated 616 * with abd_alloc_linear() which subsequently released ownership of their buf 617 * with abd_release_ownership_of_buf(). 618 */ 619void 620abd_take_ownership_of_buf(abd_t *abd, boolean_t is_metadata) 621{ 622 ASSERT(abd_is_linear(abd)); 623 ASSERT(!(abd->abd_flags & ABD_FLAG_OWNER)); 624 abd_verify(abd); 625 626 abd->abd_flags |= ABD_FLAG_OWNER; 627 if (is_metadata) { 628 abd->abd_flags |= ABD_FLAG_META; 629 } 630 631 ABDSTAT_BUMP(abdstat_linear_cnt); 632 ABDSTAT_INCR(abdstat_linear_data_size, abd->abd_size); 633} 634 635void 636abd_release_ownership_of_buf(abd_t *abd) 637{ 638 ASSERT(abd_is_linear(abd)); 639 ASSERT(abd->abd_flags & ABD_FLAG_OWNER); 640 abd_verify(abd); 641 642 abd->abd_flags &= ~ABD_FLAG_OWNER; 643 /* Disable this flag since we no longer own the data buffer */ 644 abd->abd_flags &= ~ABD_FLAG_META; 645 646 ABDSTAT_BUMPDOWN(abdstat_linear_cnt); 647 ABDSTAT_INCR(abdstat_linear_data_size, -(int)abd->abd_size); 648} 649 650struct abd_iter { 651 abd_t *iter_abd; /* ABD being iterated through */ 652 size_t iter_pos; /* position (relative to abd_offset) */ 653 void *iter_mapaddr; /* addr corresponding to iter_pos */ 654 size_t iter_mapsize; /* length of data valid at mapaddr */ 655}; 656 657static inline size_t 658abd_iter_scatter_chunk_offset(struct abd_iter *aiter) 659{ 660 ASSERT(!abd_is_linear(aiter->iter_abd)); 661 return ((aiter->iter_abd->abd_u.abd_scatter.abd_offset + 662 aiter->iter_pos) % zfs_abd_chunk_size); 663} 664 665static inline size_t 666abd_iter_scatter_chunk_index(struct abd_iter *aiter) 667{ 668 ASSERT(!abd_is_linear(aiter->iter_abd)); 669 return ((aiter->iter_abd->abd_u.abd_scatter.abd_offset + 670 aiter->iter_pos) / zfs_abd_chunk_size); 671} 672 673/* 674 * Initialize the abd_iter. 675 */ 676static void 677abd_iter_init(struct abd_iter *aiter, abd_t *abd) 678{ 679 abd_verify(abd); 680 aiter->iter_abd = abd; 681 aiter->iter_pos = 0; 682 aiter->iter_mapaddr = NULL; 683 aiter->iter_mapsize = 0; 684} 685 686/* 687 * Advance the iterator by a certain amount. Cannot be called when a chunk is 688 * in use. This can be safely called when the aiter has already exhausted, in 689 * which case this does nothing. 690 */ 691static void 692abd_iter_advance(struct abd_iter *aiter, size_t amount) 693{ 694 ASSERT3P(aiter->iter_mapaddr, ==, NULL); 695 ASSERT0(aiter->iter_mapsize); 696 697 /* There's nothing left to advance to, so do nothing */ 698 if (aiter->iter_pos == aiter->iter_abd->abd_size) 699 return; 700 701 aiter->iter_pos += amount; 702} 703 704/* 705 * Map the current chunk into aiter. This can be safely called when the aiter 706 * has already exhausted, in which case this does nothing. 707 */ 708static void 709abd_iter_map(struct abd_iter *aiter) 710{ 711 void *paddr; 712 size_t offset = 0; 713 714 ASSERT3P(aiter->iter_mapaddr, ==, NULL); 715 ASSERT0(aiter->iter_mapsize); 716 717 /* Panic if someone has changed zfs_abd_chunk_size */ 718 IMPLY(!abd_is_linear(aiter->iter_abd), zfs_abd_chunk_size == 719 aiter->iter_abd->abd_u.abd_scatter.abd_chunk_size); 720 721 /* There's nothing left to iterate over, so do nothing */ 722 if (aiter->iter_pos == aiter->iter_abd->abd_size) 723 return; 724 725 if (abd_is_linear(aiter->iter_abd)) { 726 offset = aiter->iter_pos; 727 aiter->iter_mapsize = aiter->iter_abd->abd_size - offset; 728 paddr = aiter->iter_abd->abd_u.abd_linear.abd_buf; 729 } else { 730 size_t index = abd_iter_scatter_chunk_index(aiter); 731 offset = abd_iter_scatter_chunk_offset(aiter); 732 aiter->iter_mapsize = zfs_abd_chunk_size - offset; 733 paddr = aiter->iter_abd->abd_u.abd_scatter.abd_chunks[index]; 734 } 735 aiter->iter_mapaddr = (char *)paddr + offset; 736} 737 738/* 739 * Unmap the current chunk from aiter. This can be safely called when the aiter 740 * has already exhausted, in which case this does nothing. 741 */ 742static void 743abd_iter_unmap(struct abd_iter *aiter) 744{ 745 /* There's nothing left to unmap, so do nothing */ 746 if (aiter->iter_pos == aiter->iter_abd->abd_size) 747 return; 748 749 ASSERT3P(aiter->iter_mapaddr, !=, NULL); 750 ASSERT3U(aiter->iter_mapsize, >, 0); 751 752 aiter->iter_mapaddr = NULL; 753 aiter->iter_mapsize = 0; 754} 755 756int 757abd_iterate_func(abd_t *abd, size_t off, size_t size, 758 abd_iter_func_t *func, void *private) 759{ 760 int ret = 0; 761 struct abd_iter aiter; 762 763 abd_verify(abd); 764 ASSERT3U(off + size, <=, abd->abd_size); 765 766 abd_iter_init(&aiter, abd); 767 abd_iter_advance(&aiter, off); 768 769 while (size > 0) { 770 abd_iter_map(&aiter); 771 772 size_t len = MIN(aiter.iter_mapsize, size); 773 ASSERT3U(len, >, 0); 774 775 ret = func(aiter.iter_mapaddr, len, private); 776 777 abd_iter_unmap(&aiter); 778 779 if (ret != 0) 780 break; 781 782 size -= len; 783 abd_iter_advance(&aiter, len); 784 } 785 786 return (ret); 787} 788 789struct buf_arg { 790 void *arg_buf; 791}; 792 793static int 794abd_copy_to_buf_off_cb(void *buf, size_t size, void *private) 795{ 796 struct buf_arg *ba_ptr = private; 797 798 (void) memcpy(ba_ptr->arg_buf, buf, size); 799 ba_ptr->arg_buf = (char *)ba_ptr->arg_buf + size; 800 801 return (0); 802} 803 804/* 805 * Copy abd to buf. (off is the offset in abd.) 806 */ 807void 808abd_copy_to_buf_off(void *buf, abd_t *abd, size_t off, size_t size) 809{ 810 struct buf_arg ba_ptr = { buf }; 811 812 (void) abd_iterate_func(abd, off, size, abd_copy_to_buf_off_cb, 813 &ba_ptr); 814} 815 816static int 817abd_cmp_buf_off_cb(void *buf, size_t size, void *private) 818{ 819 int ret; 820 struct buf_arg *ba_ptr = private; 821 822 ret = memcmp(buf, ba_ptr->arg_buf, size); 823 ba_ptr->arg_buf = (char *)ba_ptr->arg_buf + size; 824 825 return (ret); 826} 827 828/* 829 * Compare the contents of abd to buf. (off is the offset in abd.) 830 */ 831int 832abd_cmp_buf_off(abd_t *abd, const void *buf, size_t off, size_t size) 833{ 834 struct buf_arg ba_ptr = { (void *) buf }; 835 836 return (abd_iterate_func(abd, off, size, abd_cmp_buf_off_cb, &ba_ptr)); 837} 838 839static int 840abd_copy_from_buf_off_cb(void *buf, size_t size, void *private) 841{ 842 struct buf_arg *ba_ptr = private; 843 844 (void) memcpy(buf, ba_ptr->arg_buf, size); 845 ba_ptr->arg_buf = (char *)ba_ptr->arg_buf + size; 846 847 return (0); 848} 849 850/* 851 * Copy from buf to abd. (off is the offset in abd.) 852 */ 853void 854abd_copy_from_buf_off(abd_t *abd, const void *buf, size_t off, size_t size) 855{ 856 struct buf_arg ba_ptr = { (void *) buf }; 857 858 (void) abd_iterate_func(abd, off, size, abd_copy_from_buf_off_cb, 859 &ba_ptr); 860} 861 862/*ARGSUSED*/ 863static int 864abd_zero_off_cb(void *buf, size_t size, void *private) 865{ 866 (void) memset(buf, 0, size); 867 return (0); 868} 869 870/* 871 * Zero out the abd from a particular offset to the end. 872 */ 873void 874abd_zero_off(abd_t *abd, size_t off, size_t size) 875{ 876 (void) abd_iterate_func(abd, off, size, abd_zero_off_cb, NULL); 877} 878 879/* 880 * Iterate over two ABDs and call func incrementally on the two ABDs' data in 881 * equal-sized chunks (passed to func as raw buffers). func could be called many 882 * times during this iteration. 883 */ 884int 885abd_iterate_func2(abd_t *dabd, abd_t *sabd, size_t doff, size_t soff, 886 size_t size, abd_iter_func2_t *func, void *private) 887{ 888 int ret = 0; 889 struct abd_iter daiter, saiter; 890 891 abd_verify(dabd); 892 abd_verify(sabd); 893 894 ASSERT3U(doff + size, <=, dabd->abd_size); 895 ASSERT3U(soff + size, <=, sabd->abd_size); 896 897 abd_iter_init(&daiter, dabd); 898 abd_iter_init(&saiter, sabd); 899 abd_iter_advance(&daiter, doff); 900 abd_iter_advance(&saiter, soff); 901 902 while (size > 0) { 903 abd_iter_map(&daiter); 904 abd_iter_map(&saiter); 905 906 size_t dlen = MIN(daiter.iter_mapsize, size); 907 size_t slen = MIN(saiter.iter_mapsize, size); 908 size_t len = MIN(dlen, slen); 909 ASSERT(dlen > 0 || slen > 0); 910 911 ret = func(daiter.iter_mapaddr, saiter.iter_mapaddr, len, 912 private); 913 914 abd_iter_unmap(&saiter); 915 abd_iter_unmap(&daiter); 916 917 if (ret != 0) 918 break; 919 920 size -= len; 921 abd_iter_advance(&daiter, len); 922 abd_iter_advance(&saiter, len); 923 } 924 925 return (ret); 926} 927 928/*ARGSUSED*/ 929static int 930abd_copy_off_cb(void *dbuf, void *sbuf, size_t size, void *private) 931{ 932 (void) memcpy(dbuf, sbuf, size); 933 return (0); 934} 935 936/* 937 * Copy from sabd to dabd starting from soff and doff. 938 */ 939void 940abd_copy_off(abd_t *dabd, abd_t *sabd, size_t doff, size_t soff, size_t size) 941{ 942 (void) abd_iterate_func2(dabd, sabd, doff, soff, size, 943 abd_copy_off_cb, NULL); 944} 945 946/*ARGSUSED*/ 947static int 948abd_cmp_cb(void *bufa, void *bufb, size_t size, void *private) 949{ 950 return (memcmp(bufa, bufb, size)); 951} 952 953/* 954 * Compares the first size bytes of two ABDs. 955 */ 956int 957abd_cmp(abd_t *dabd, abd_t *sabd, size_t size) 958{ 959 return (abd_iterate_func2(dabd, sabd, 0, 0, size, abd_cmp_cb, NULL)); 960} 961