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 * See abd.c for a general overview of the arc buffered data (ABD). 19 * 20 * Using a large proportion of scattered ABDs decreases ARC fragmentation since 21 * when we are at the limit of allocatable space, using equal-size chunks will 22 * allow us to quickly reclaim enough space for a new large allocation (assuming 23 * it is also scattered). 24 * 25 * ABDs are allocated scattered by default unless the caller uses 26 * abd_alloc_linear() or zfs_abd_scatter_enabled is disabled. 27 */ 28 29#include <sys/abd_impl.h> 30#include <sys/param.h> 31#include <sys/types.h> 32#include <sys/zio.h> 33#include <sys/zfs_context.h> 34#include <sys/zfs_znode.h> 35 36typedef struct abd_stats { 37 kstat_named_t abdstat_struct_size; 38 kstat_named_t abdstat_scatter_cnt; 39 kstat_named_t abdstat_scatter_data_size; 40 kstat_named_t abdstat_scatter_chunk_waste; 41 kstat_named_t abdstat_linear_cnt; 42 kstat_named_t abdstat_linear_data_size; 43} abd_stats_t; 44 45static abd_stats_t abd_stats = { 46 /* Amount of memory occupied by all of the abd_t struct allocations */ 47 { "struct_size", KSTAT_DATA_UINT64 }, 48 /* 49 * The number of scatter ABDs which are currently allocated, excluding 50 * ABDs which don't own their data (for instance the ones which were 51 * allocated through abd_get_offset()). 52 */ 53 { "scatter_cnt", KSTAT_DATA_UINT64 }, 54 /* Amount of data stored in all scatter ABDs tracked by scatter_cnt */ 55 { "scatter_data_size", KSTAT_DATA_UINT64 }, 56 /* 57 * The amount of space wasted at the end of the last chunk across all 58 * scatter ABDs tracked by scatter_cnt. 59 */ 60 { "scatter_chunk_waste", KSTAT_DATA_UINT64 }, 61 /* 62 * The number of linear ABDs which are currently allocated, excluding 63 * ABDs which don't own their data (for instance the ones which were 64 * allocated through abd_get_offset() and abd_get_from_buf()). If an 65 * ABD takes ownership of its buf then it will become tracked. 66 */ 67 { "linear_cnt", KSTAT_DATA_UINT64 }, 68 /* Amount of data stored in all linear ABDs tracked by linear_cnt */ 69 { "linear_data_size", KSTAT_DATA_UINT64 }, 70}; 71 72/* 73 * The size of the chunks ABD allocates. Because the sizes allocated from the 74 * kmem_cache can't change, this tunable can only be modified at boot. Changing 75 * it at runtime would cause ABD iteration to work incorrectly for ABDs which 76 * were allocated with the old size, so a safeguard has been put in place which 77 * will cause the machine to panic if you change it and try to access the data 78 * within a scattered ABD. 79 */ 80size_t zfs_abd_chunk_size = 4096; 81 82#if defined(_KERNEL) 83SYSCTL_DECL(_vfs_zfs); 84 85SYSCTL_INT(_vfs_zfs, OID_AUTO, abd_scatter_enabled, CTLFLAG_RWTUN, 86 &zfs_abd_scatter_enabled, 0, "Enable scattered ARC data buffers"); 87SYSCTL_ULONG(_vfs_zfs, OID_AUTO, abd_chunk_size, CTLFLAG_RDTUN, 88 &zfs_abd_chunk_size, 0, "The size of the chunks ABD allocates"); 89#endif 90 91kmem_cache_t *abd_chunk_cache; 92static kstat_t *abd_ksp; 93 94/* 95 * We use a scattered SPA_MAXBLOCKSIZE sized ABD whose chunks are 96 * just a single zero'd sized zfs_abd_chunk_size buffer. This 97 * allows us to conserve memory by only using a single zero buffer 98 * for the scatter chunks. 99 */ 100abd_t *abd_zero_scatter = NULL; 101static char *abd_zero_buf = NULL; 102 103static void 104abd_free_chunk(void *c) 105{ 106 kmem_cache_free(abd_chunk_cache, c); 107} 108 109static uint_t 110abd_chunkcnt_for_bytes(size_t size) 111{ 112 return (P2ROUNDUP(size, zfs_abd_chunk_size) / zfs_abd_chunk_size); 113} 114 115static inline uint_t 116abd_scatter_chunkcnt(abd_t *abd) 117{ 118 ASSERT(!abd_is_linear(abd)); 119 return (abd_chunkcnt_for_bytes( 120 ABD_SCATTER(abd).abd_offset + abd->abd_size)); 121} 122 123boolean_t 124abd_size_alloc_linear(size_t size) 125{ 126 return (size <= zfs_abd_chunk_size ? B_TRUE : B_FALSE); 127} 128 129void 130abd_update_scatter_stats(abd_t *abd, abd_stats_op_t op) 131{ 132 uint_t n = abd_scatter_chunkcnt(abd); 133 ASSERT(op == ABDSTAT_INCR || op == ABDSTAT_DECR); 134 int waste = n * zfs_abd_chunk_size - abd->abd_size; 135 if (op == ABDSTAT_INCR) { 136 ABDSTAT_BUMP(abdstat_scatter_cnt); 137 ABDSTAT_INCR(abdstat_scatter_data_size, abd->abd_size); 138 ABDSTAT_INCR(abdstat_scatter_chunk_waste, waste); 139 arc_space_consume(waste, ARC_SPACE_ABD_CHUNK_WASTE); 140 } else { 141 ABDSTAT_BUMPDOWN(abdstat_scatter_cnt); 142 ABDSTAT_INCR(abdstat_scatter_data_size, -(int)abd->abd_size); 143 ABDSTAT_INCR(abdstat_scatter_chunk_waste, -waste); 144 arc_space_return(waste, ARC_SPACE_ABD_CHUNK_WASTE); 145 } 146} 147 148void 149abd_update_linear_stats(abd_t *abd, abd_stats_op_t op) 150{ 151 ASSERT(op == ABDSTAT_INCR || op == ABDSTAT_DECR); 152 if (op == ABDSTAT_INCR) { 153 ABDSTAT_BUMP(abdstat_linear_cnt); 154 ABDSTAT_INCR(abdstat_linear_data_size, abd->abd_size); 155 } else { 156 ABDSTAT_BUMPDOWN(abdstat_linear_cnt); 157 ABDSTAT_INCR(abdstat_linear_data_size, -(int)abd->abd_size); 158 } 159} 160 161void 162abd_verify_scatter(abd_t *abd) 163{ 164 uint_t i, n; 165 166 /* 167 * There is no scatter linear pages in FreeBSD so there is an 168 * if an error if the ABD has been marked as a linear page. 169 */ 170 ASSERT(!abd_is_linear_page(abd)); 171 ASSERT3U(ABD_SCATTER(abd).abd_offset, <, 172 zfs_abd_chunk_size); 173 n = abd_scatter_chunkcnt(abd); 174 for (i = 0; i < n; i++) { 175 ASSERT3P(ABD_SCATTER(abd).abd_chunks[i], !=, NULL); 176 } 177} 178 179void 180abd_alloc_chunks(abd_t *abd, size_t size) 181{ 182 uint_t i, n; 183 184 n = abd_chunkcnt_for_bytes(size); 185 for (i = 0; i < n; i++) { 186 void *c = kmem_cache_alloc(abd_chunk_cache, KM_PUSHPAGE); 187 ASSERT3P(c, !=, NULL); 188 ABD_SCATTER(abd).abd_chunks[i] = c; 189 } 190 ABD_SCATTER(abd).abd_chunk_size = zfs_abd_chunk_size; 191} 192 193void 194abd_free_chunks(abd_t *abd) 195{ 196 uint_t i, n; 197 198 n = abd_scatter_chunkcnt(abd); 199 for (i = 0; i < n; i++) { 200 abd_free_chunk(ABD_SCATTER(abd).abd_chunks[i]); 201 } 202} 203 204abd_t * 205abd_alloc_struct_impl(size_t size) 206{ 207 uint_t chunkcnt = abd_chunkcnt_for_bytes(size); 208 /* 209 * In the event we are allocating a gang ABD, the size passed in 210 * will be 0. We must make sure to set abd_size to the size of an 211 * ABD struct as opposed to an ABD scatter with 0 chunks. The gang 212 * ABD struct allocation accounts for an additional 24 bytes over 213 * a scatter ABD with 0 chunks. 214 */ 215 size_t abd_size = MAX(sizeof (abd_t), 216 offsetof(abd_t, abd_u.abd_scatter.abd_chunks[chunkcnt])); 217 abd_t *abd = kmem_alloc(abd_size, KM_PUSHPAGE); 218 ASSERT3P(abd, !=, NULL); 219 ABDSTAT_INCR(abdstat_struct_size, abd_size); 220 221 return (abd); 222} 223 224void 225abd_free_struct_impl(abd_t *abd) 226{ 227 uint_t chunkcnt = abd_is_linear(abd) || abd_is_gang(abd) ? 0 : 228 abd_scatter_chunkcnt(abd); 229 ssize_t size = MAX(sizeof (abd_t), 230 offsetof(abd_t, abd_u.abd_scatter.abd_chunks[chunkcnt])); 231 kmem_free(abd, size); 232 ABDSTAT_INCR(abdstat_struct_size, -size); 233} 234 235/* 236 * Allocate scatter ABD of size SPA_MAXBLOCKSIZE, where 237 * each chunk in the scatterlist will be set to abd_zero_buf. 238 */ 239static void 240abd_alloc_zero_scatter(void) 241{ 242 uint_t i, n; 243 244 n = abd_chunkcnt_for_bytes(SPA_MAXBLOCKSIZE); 245 abd_zero_buf = kmem_zalloc(zfs_abd_chunk_size, KM_SLEEP); 246 abd_zero_scatter = abd_alloc_struct(SPA_MAXBLOCKSIZE); 247 248 abd_zero_scatter->abd_flags |= ABD_FLAG_OWNER | ABD_FLAG_ZEROS; 249 abd_zero_scatter->abd_size = SPA_MAXBLOCKSIZE; 250 251 ABD_SCATTER(abd_zero_scatter).abd_offset = 0; 252 ABD_SCATTER(abd_zero_scatter).abd_chunk_size = 253 zfs_abd_chunk_size; 254 255 for (i = 0; i < n; i++) { 256 ABD_SCATTER(abd_zero_scatter).abd_chunks[i] = 257 abd_zero_buf; 258 } 259 260 ABDSTAT_BUMP(abdstat_scatter_cnt); 261 ABDSTAT_INCR(abdstat_scatter_data_size, zfs_abd_chunk_size); 262} 263 264static void 265abd_free_zero_scatter(void) 266{ 267 ABDSTAT_BUMPDOWN(abdstat_scatter_cnt); 268 ABDSTAT_INCR(abdstat_scatter_data_size, -(int)zfs_abd_chunk_size); 269 270 abd_free_struct(abd_zero_scatter); 271 abd_zero_scatter = NULL; 272 kmem_free(abd_zero_buf, zfs_abd_chunk_size); 273} 274 275void 276abd_init(void) 277{ 278 abd_chunk_cache = kmem_cache_create("abd_chunk", zfs_abd_chunk_size, 0, 279 NULL, NULL, NULL, NULL, 0, KMC_NODEBUG); 280 281 abd_ksp = kstat_create("zfs", 0, "abdstats", "misc", KSTAT_TYPE_NAMED, 282 sizeof (abd_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 283 if (abd_ksp != NULL) { 284 abd_ksp->ks_data = &abd_stats; 285 kstat_install(abd_ksp); 286 } 287 288 abd_alloc_zero_scatter(); 289} 290 291void 292abd_fini(void) 293{ 294 abd_free_zero_scatter(); 295 296 if (abd_ksp != NULL) { 297 kstat_delete(abd_ksp); 298 abd_ksp = NULL; 299 } 300 301 kmem_cache_destroy(abd_chunk_cache); 302 abd_chunk_cache = NULL; 303} 304 305void 306abd_free_linear_page(abd_t *abd) 307{ 308 /* 309 * FreeBSD does not have scatter linear pages 310 * so there is an error. 311 */ 312 VERIFY(0); 313} 314 315/* 316 * If we're going to use this ABD for doing I/O using the block layer, the 317 * consumer of the ABD data doesn't care if it's scattered or not, and we don't 318 * plan to store this ABD in memory for a long period of time, we should 319 * allocate the ABD type that requires the least data copying to do the I/O. 320 * 321 * Currently this is linear ABDs, however if ldi_strategy() can ever issue I/Os 322 * using a scatter/gather list we should switch to that and replace this call 323 * with vanilla abd_alloc(). 324 */ 325abd_t * 326abd_alloc_for_io(size_t size, boolean_t is_metadata) 327{ 328 return (abd_alloc_linear(size, is_metadata)); 329} 330 331abd_t * 332abd_get_offset_scatter(abd_t *abd, abd_t *sabd, size_t off) 333{ 334 abd_verify(sabd); 335 ASSERT3U(off, <=, sabd->abd_size); 336 337 size_t new_offset = ABD_SCATTER(sabd).abd_offset + off; 338 uint_t chunkcnt = abd_scatter_chunkcnt(sabd) - 339 (new_offset / zfs_abd_chunk_size); 340 341 /* 342 * If an abd struct is provided, it is only the minimum size. If we 343 * need additional chunks, we need to allocate a new struct. 344 */ 345 if (abd != NULL && 346 offsetof(abd_t, abd_u.abd_scatter.abd_chunks[chunkcnt]) > 347 sizeof (abd_t)) { 348 abd = NULL; 349 } 350 351 if (abd == NULL) 352 abd = abd_alloc_struct(chunkcnt * zfs_abd_chunk_size); 353 354 /* 355 * Even if this buf is filesystem metadata, we only track that 356 * if we own the underlying data buffer, which is not true in 357 * this case. Therefore, we don't ever use ABD_FLAG_META here. 358 */ 359 360 ABD_SCATTER(abd).abd_offset = new_offset % zfs_abd_chunk_size; 361 ABD_SCATTER(abd).abd_chunk_size = zfs_abd_chunk_size; 362 363 /* Copy the scatterlist starting at the correct offset */ 364 (void) memcpy(&ABD_SCATTER(abd).abd_chunks, 365 &ABD_SCATTER(sabd).abd_chunks[new_offset / 366 zfs_abd_chunk_size], 367 chunkcnt * sizeof (void *)); 368 369 return (abd); 370} 371 372static inline size_t 373abd_iter_scatter_chunk_offset(struct abd_iter *aiter) 374{ 375 ASSERT(!abd_is_linear(aiter->iter_abd)); 376 return ((ABD_SCATTER(aiter->iter_abd).abd_offset + 377 aiter->iter_pos) % zfs_abd_chunk_size); 378} 379 380static inline size_t 381abd_iter_scatter_chunk_index(struct abd_iter *aiter) 382{ 383 ASSERT(!abd_is_linear(aiter->iter_abd)); 384 return ((ABD_SCATTER(aiter->iter_abd).abd_offset + 385 aiter->iter_pos) / zfs_abd_chunk_size); 386} 387 388/* 389 * Initialize the abd_iter. 390 */ 391void 392abd_iter_init(struct abd_iter *aiter, abd_t *abd) 393{ 394 ASSERT(!abd_is_gang(abd)); 395 abd_verify(abd); 396 aiter->iter_abd = abd; 397 aiter->iter_pos = 0; 398 aiter->iter_mapaddr = NULL; 399 aiter->iter_mapsize = 0; 400} 401 402/* 403 * This is just a helper function to see if we have exhausted the 404 * abd_iter and reached the end. 405 */ 406boolean_t 407abd_iter_at_end(struct abd_iter *aiter) 408{ 409 return (aiter->iter_pos == aiter->iter_abd->abd_size); 410} 411 412/* 413 * Advance the iterator by a certain amount. Cannot be called when a chunk is 414 * in use. This can be safely called when the aiter has already exhausted, in 415 * which case this does nothing. 416 */ 417void 418abd_iter_advance(struct abd_iter *aiter, size_t amount) 419{ 420 ASSERT3P(aiter->iter_mapaddr, ==, NULL); 421 ASSERT0(aiter->iter_mapsize); 422 423 /* There's nothing left to advance to, so do nothing */ 424 if (abd_iter_at_end(aiter)) 425 return; 426 427 aiter->iter_pos += amount; 428} 429 430/* 431 * Map the current chunk into aiter. This can be safely called when the aiter 432 * has already exhausted, in which case this does nothing. 433 */ 434void 435abd_iter_map(struct abd_iter *aiter) 436{ 437 void *paddr; 438 size_t offset = 0; 439 440 ASSERT3P(aiter->iter_mapaddr, ==, NULL); 441 ASSERT0(aiter->iter_mapsize); 442 443 /* Panic if someone has changed zfs_abd_chunk_size */ 444 IMPLY(!abd_is_linear(aiter->iter_abd), zfs_abd_chunk_size == 445 ABD_SCATTER(aiter->iter_abd).abd_chunk_size); 446 447 /* There's nothing left to iterate over, so do nothing */ 448 if (abd_iter_at_end(aiter)) 449 return; 450 451 if (abd_is_linear(aiter->iter_abd)) { 452 offset = aiter->iter_pos; 453 aiter->iter_mapsize = aiter->iter_abd->abd_size - offset; 454 paddr = ABD_LINEAR_BUF(aiter->iter_abd); 455 } else { 456 size_t index = abd_iter_scatter_chunk_index(aiter); 457 offset = abd_iter_scatter_chunk_offset(aiter); 458 aiter->iter_mapsize = MIN(zfs_abd_chunk_size - offset, 459 aiter->iter_abd->abd_size - aiter->iter_pos); 460 paddr = ABD_SCATTER(aiter->iter_abd).abd_chunks[index]; 461 } 462 aiter->iter_mapaddr = (char *)paddr + offset; 463} 464 465/* 466 * Unmap the current chunk from aiter. This can be safely called when the aiter 467 * has already exhausted, in which case this does nothing. 468 */ 469void 470abd_iter_unmap(struct abd_iter *aiter) 471{ 472 /* There's nothing left to unmap, so do nothing */ 473 if (abd_iter_at_end(aiter)) 474 return; 475 476 ASSERT3P(aiter->iter_mapaddr, !=, NULL); 477 ASSERT3U(aiter->iter_mapsize, >, 0); 478 479 aiter->iter_mapaddr = NULL; 480 aiter->iter_mapsize = 0; 481} 482 483void 484abd_cache_reap_now(void) 485{ 486 kmem_cache_reap_soon(abd_chunk_cache); 487} 488