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 36 typedef 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 45 static 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 struct { 73 wmsum_t abdstat_struct_size; 74 wmsum_t abdstat_scatter_cnt; 75 wmsum_t abdstat_scatter_data_size; 76 wmsum_t abdstat_scatter_chunk_waste; 77 wmsum_t abdstat_linear_cnt; 78 wmsum_t abdstat_linear_data_size; 79 } abd_sums; 80 81 /* 82 * zfs_abd_scatter_min_size is the minimum allocation size to use scatter 83 * ABD's for. Smaller allocations will use linear ABD's which use 84 * zio_[data_]buf_alloc(). 85 * 86 * Scatter ABD's use at least one page each, so sub-page allocations waste 87 * some space when allocated as scatter (e.g. 2KB scatter allocation wastes 88 * half of each page). Using linear ABD's for small allocations means that 89 * they will be put on slabs which contain many allocations. 90 * 91 * Linear ABDs for multi-page allocations are easier to use, and in some cases 92 * it allows to avoid buffer copying. But allocation and especially free 93 * of multi-page linear ABDs are expensive operations due to KVA mapping and 94 * unmapping, and with time they cause KVA fragmentations. 95 */ 96 static size_t zfs_abd_scatter_min_size = PAGE_SIZE + 1; 97 98 #if defined(_KERNEL) 99 SYSCTL_DECL(_vfs_zfs); 100 101 SYSCTL_INT(_vfs_zfs, OID_AUTO, abd_scatter_enabled, CTLFLAG_RWTUN, 102 &zfs_abd_scatter_enabled, 0, "Enable scattered ARC data buffers"); 103 SYSCTL_ULONG(_vfs_zfs, OID_AUTO, abd_scatter_min_size, CTLFLAG_RWTUN, 104 &zfs_abd_scatter_min_size, 0, "Minimum size of scatter allocations."); 105 #endif 106 107 kmem_cache_t *abd_chunk_cache; 108 static kstat_t *abd_ksp; 109 110 /* 111 * We use a scattered SPA_MAXBLOCKSIZE sized ABD whose chunks are 112 * just a single zero'd page-sized buffer. This allows us to conserve 113 * memory by only using a single zero buffer for the scatter chunks. 114 */ 115 abd_t *abd_zero_scatter = NULL; 116 117 static uint_t 118 abd_chunkcnt_for_bytes(size_t size) 119 { 120 return ((size + PAGE_MASK) >> PAGE_SHIFT); 121 } 122 123 static inline uint_t 124 abd_scatter_chunkcnt(abd_t *abd) 125 { 126 ASSERT(!abd_is_linear(abd)); 127 return (abd_chunkcnt_for_bytes( 128 ABD_SCATTER(abd).abd_offset + abd->abd_size)); 129 } 130 131 boolean_t 132 abd_size_alloc_linear(size_t size) 133 { 134 return (!zfs_abd_scatter_enabled || size < zfs_abd_scatter_min_size); 135 } 136 137 void 138 abd_update_scatter_stats(abd_t *abd, abd_stats_op_t op) 139 { 140 uint_t n = abd_scatter_chunkcnt(abd); 141 ASSERT(op == ABDSTAT_INCR || op == ABDSTAT_DECR); 142 int waste = (n << PAGE_SHIFT) - abd->abd_size; 143 if (op == ABDSTAT_INCR) { 144 ABDSTAT_BUMP(abdstat_scatter_cnt); 145 ABDSTAT_INCR(abdstat_scatter_data_size, abd->abd_size); 146 ABDSTAT_INCR(abdstat_scatter_chunk_waste, waste); 147 arc_space_consume(waste, ARC_SPACE_ABD_CHUNK_WASTE); 148 } else { 149 ABDSTAT_BUMPDOWN(abdstat_scatter_cnt); 150 ABDSTAT_INCR(abdstat_scatter_data_size, -(int)abd->abd_size); 151 ABDSTAT_INCR(abdstat_scatter_chunk_waste, -waste); 152 arc_space_return(waste, ARC_SPACE_ABD_CHUNK_WASTE); 153 } 154 } 155 156 void 157 abd_update_linear_stats(abd_t *abd, abd_stats_op_t op) 158 { 159 ASSERT(op == ABDSTAT_INCR || op == ABDSTAT_DECR); 160 if (op == ABDSTAT_INCR) { 161 ABDSTAT_BUMP(abdstat_linear_cnt); 162 ABDSTAT_INCR(abdstat_linear_data_size, abd->abd_size); 163 } else { 164 ABDSTAT_BUMPDOWN(abdstat_linear_cnt); 165 ABDSTAT_INCR(abdstat_linear_data_size, -(int)abd->abd_size); 166 } 167 } 168 169 void 170 abd_verify_scatter(abd_t *abd) 171 { 172 uint_t i, n; 173 174 /* 175 * There is no scatter linear pages in FreeBSD so there is 176 * an error if the ABD has been marked as a linear page. 177 */ 178 ASSERT(!abd_is_linear_page(abd)); 179 ASSERT3U(ABD_SCATTER(abd).abd_offset, <, PAGE_SIZE); 180 n = abd_scatter_chunkcnt(abd); 181 for (i = 0; i < n; i++) { 182 ASSERT3P(ABD_SCATTER(abd).abd_chunks[i], !=, NULL); 183 } 184 } 185 186 void 187 abd_alloc_chunks(abd_t *abd, size_t size) 188 { 189 uint_t i, n; 190 191 n = abd_chunkcnt_for_bytes(size); 192 for (i = 0; i < n; i++) { 193 ABD_SCATTER(abd).abd_chunks[i] = 194 kmem_cache_alloc(abd_chunk_cache, KM_PUSHPAGE); 195 } 196 } 197 198 void 199 abd_free_chunks(abd_t *abd) 200 { 201 uint_t i, n; 202 203 n = abd_scatter_chunkcnt(abd); 204 for (i = 0; i < n; i++) { 205 kmem_cache_free(abd_chunk_cache, 206 ABD_SCATTER(abd).abd_chunks[i]); 207 } 208 } 209 210 abd_t * 211 abd_alloc_struct_impl(size_t size) 212 { 213 uint_t chunkcnt = abd_chunkcnt_for_bytes(size); 214 /* 215 * In the event we are allocating a gang ABD, the size passed in 216 * will be 0. We must make sure to set abd_size to the size of an 217 * ABD struct as opposed to an ABD scatter with 0 chunks. The gang 218 * ABD struct allocation accounts for an additional 24 bytes over 219 * a scatter ABD with 0 chunks. 220 */ 221 size_t abd_size = MAX(sizeof (abd_t), 222 offsetof(abd_t, abd_u.abd_scatter.abd_chunks[chunkcnt])); 223 abd_t *abd = kmem_alloc(abd_size, KM_PUSHPAGE); 224 ASSERT3P(abd, !=, NULL); 225 ABDSTAT_INCR(abdstat_struct_size, abd_size); 226 227 return (abd); 228 } 229 230 void 231 abd_free_struct_impl(abd_t *abd) 232 { 233 uint_t chunkcnt = abd_is_linear(abd) || abd_is_gang(abd) ? 0 : 234 abd_scatter_chunkcnt(abd); 235 ssize_t size = MAX(sizeof (abd_t), 236 offsetof(abd_t, abd_u.abd_scatter.abd_chunks[chunkcnt])); 237 kmem_free(abd, size); 238 ABDSTAT_INCR(abdstat_struct_size, -size); 239 } 240 241 /* 242 * Allocate scatter ABD of size SPA_MAXBLOCKSIZE, where 243 * each chunk in the scatterlist will be set to the same area. 244 */ 245 _Static_assert(ZERO_REGION_SIZE >= PAGE_SIZE, "zero_region too small"); 246 static void 247 abd_alloc_zero_scatter(void) 248 { 249 uint_t i, n; 250 251 n = abd_chunkcnt_for_bytes(SPA_MAXBLOCKSIZE); 252 abd_zero_scatter = abd_alloc_struct(SPA_MAXBLOCKSIZE); 253 abd_zero_scatter->abd_flags |= ABD_FLAG_OWNER | ABD_FLAG_ZEROS; 254 abd_zero_scatter->abd_size = SPA_MAXBLOCKSIZE; 255 256 ABD_SCATTER(abd_zero_scatter).abd_offset = 0; 257 258 for (i = 0; i < n; i++) { 259 ABD_SCATTER(abd_zero_scatter).abd_chunks[i] = 260 __DECONST(void *, zero_region); 261 } 262 263 ABDSTAT_BUMP(abdstat_scatter_cnt); 264 ABDSTAT_INCR(abdstat_scatter_data_size, PAGE_SIZE); 265 } 266 267 static void 268 abd_free_zero_scatter(void) 269 { 270 ABDSTAT_BUMPDOWN(abdstat_scatter_cnt); 271 ABDSTAT_INCR(abdstat_scatter_data_size, -(int)PAGE_SIZE); 272 273 abd_free_struct(abd_zero_scatter); 274 abd_zero_scatter = NULL; 275 } 276 277 static int 278 abd_kstats_update(kstat_t *ksp, int rw) 279 { 280 abd_stats_t *as = ksp->ks_data; 281 282 if (rw == KSTAT_WRITE) 283 return (EACCES); 284 as->abdstat_struct_size.value.ui64 = 285 wmsum_value(&abd_sums.abdstat_struct_size); 286 as->abdstat_scatter_cnt.value.ui64 = 287 wmsum_value(&abd_sums.abdstat_scatter_cnt); 288 as->abdstat_scatter_data_size.value.ui64 = 289 wmsum_value(&abd_sums.abdstat_scatter_data_size); 290 as->abdstat_scatter_chunk_waste.value.ui64 = 291 wmsum_value(&abd_sums.abdstat_scatter_chunk_waste); 292 as->abdstat_linear_cnt.value.ui64 = 293 wmsum_value(&abd_sums.abdstat_linear_cnt); 294 as->abdstat_linear_data_size.value.ui64 = 295 wmsum_value(&abd_sums.abdstat_linear_data_size); 296 return (0); 297 } 298 299 void 300 abd_init(void) 301 { 302 abd_chunk_cache = kmem_cache_create("abd_chunk", PAGE_SIZE, 0, 303 NULL, NULL, NULL, NULL, 0, KMC_NODEBUG); 304 305 wmsum_init(&abd_sums.abdstat_struct_size, 0); 306 wmsum_init(&abd_sums.abdstat_scatter_cnt, 0); 307 wmsum_init(&abd_sums.abdstat_scatter_data_size, 0); 308 wmsum_init(&abd_sums.abdstat_scatter_chunk_waste, 0); 309 wmsum_init(&abd_sums.abdstat_linear_cnt, 0); 310 wmsum_init(&abd_sums.abdstat_linear_data_size, 0); 311 312 abd_ksp = kstat_create("zfs", 0, "abdstats", "misc", KSTAT_TYPE_NAMED, 313 sizeof (abd_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 314 if (abd_ksp != NULL) { 315 abd_ksp->ks_data = &abd_stats; 316 abd_ksp->ks_update = abd_kstats_update; 317 kstat_install(abd_ksp); 318 } 319 320 abd_alloc_zero_scatter(); 321 } 322 323 void 324 abd_fini(void) 325 { 326 abd_free_zero_scatter(); 327 328 if (abd_ksp != NULL) { 329 kstat_delete(abd_ksp); 330 abd_ksp = NULL; 331 } 332 333 wmsum_fini(&abd_sums.abdstat_struct_size); 334 wmsum_fini(&abd_sums.abdstat_scatter_cnt); 335 wmsum_fini(&abd_sums.abdstat_scatter_data_size); 336 wmsum_fini(&abd_sums.abdstat_scatter_chunk_waste); 337 wmsum_fini(&abd_sums.abdstat_linear_cnt); 338 wmsum_fini(&abd_sums.abdstat_linear_data_size); 339 340 kmem_cache_destroy(abd_chunk_cache); 341 abd_chunk_cache = NULL; 342 } 343 344 void 345 abd_free_linear_page(abd_t *abd) 346 { 347 /* 348 * FreeBSD does not have scatter linear pages 349 * so there is an error. 350 */ 351 VERIFY(0); 352 } 353 354 /* 355 * If we're going to use this ABD for doing I/O using the block layer, the 356 * consumer of the ABD data doesn't care if it's scattered or not, and we don't 357 * plan to store this ABD in memory for a long period of time, we should 358 * allocate the ABD type that requires the least data copying to do the I/O. 359 * 360 * Currently this is linear ABDs, however if ldi_strategy() can ever issue I/Os 361 * using a scatter/gather list we should switch to that and replace this call 362 * with vanilla abd_alloc(). 363 */ 364 abd_t * 365 abd_alloc_for_io(size_t size, boolean_t is_metadata) 366 { 367 return (abd_alloc_linear(size, is_metadata)); 368 } 369 370 abd_t * 371 abd_get_offset_scatter(abd_t *abd, abd_t *sabd, size_t off, 372 size_t size) 373 { 374 abd_verify(sabd); 375 ASSERT3U(off, <=, sabd->abd_size); 376 377 size_t new_offset = ABD_SCATTER(sabd).abd_offset + off; 378 size_t chunkcnt = abd_chunkcnt_for_bytes( 379 (new_offset & PAGE_MASK) + size); 380 381 ASSERT3U(chunkcnt, <=, abd_scatter_chunkcnt(sabd)); 382 383 /* 384 * If an abd struct is provided, it is only the minimum size. If we 385 * need additional chunks, we need to allocate a new struct. 386 */ 387 if (abd != NULL && 388 offsetof(abd_t, abd_u.abd_scatter.abd_chunks[chunkcnt]) > 389 sizeof (abd_t)) { 390 abd = NULL; 391 } 392 393 if (abd == NULL) 394 abd = abd_alloc_struct(chunkcnt << PAGE_SHIFT); 395 396 /* 397 * Even if this buf is filesystem metadata, we only track that 398 * if we own the underlying data buffer, which is not true in 399 * this case. Therefore, we don't ever use ABD_FLAG_META here. 400 */ 401 402 ABD_SCATTER(abd).abd_offset = new_offset & PAGE_MASK; 403 404 /* Copy the scatterlist starting at the correct offset */ 405 (void) memcpy(&ABD_SCATTER(abd).abd_chunks, 406 &ABD_SCATTER(sabd).abd_chunks[new_offset >> PAGE_SHIFT], 407 chunkcnt * sizeof (void *)); 408 409 return (abd); 410 } 411 412 /* 413 * Initialize the abd_iter. 414 */ 415 void 416 abd_iter_init(struct abd_iter *aiter, abd_t *abd) 417 { 418 ASSERT(!abd_is_gang(abd)); 419 abd_verify(abd); 420 aiter->iter_abd = abd; 421 aiter->iter_pos = 0; 422 aiter->iter_mapaddr = NULL; 423 aiter->iter_mapsize = 0; 424 } 425 426 /* 427 * This is just a helper function to see if we have exhausted the 428 * abd_iter and reached the end. 429 */ 430 boolean_t 431 abd_iter_at_end(struct abd_iter *aiter) 432 { 433 return (aiter->iter_pos == aiter->iter_abd->abd_size); 434 } 435 436 /* 437 * Advance the iterator by a certain amount. Cannot be called when a chunk is 438 * in use. This can be safely called when the aiter has already exhausted, in 439 * which case this does nothing. 440 */ 441 void 442 abd_iter_advance(struct abd_iter *aiter, size_t amount) 443 { 444 ASSERT3P(aiter->iter_mapaddr, ==, NULL); 445 ASSERT0(aiter->iter_mapsize); 446 447 /* There's nothing left to advance to, so do nothing */ 448 if (abd_iter_at_end(aiter)) 449 return; 450 451 aiter->iter_pos += amount; 452 } 453 454 /* 455 * Map the current chunk into aiter. This can be safely called when the aiter 456 * has already exhausted, in which case this does nothing. 457 */ 458 void 459 abd_iter_map(struct abd_iter *aiter) 460 { 461 void *paddr; 462 463 ASSERT3P(aiter->iter_mapaddr, ==, NULL); 464 ASSERT0(aiter->iter_mapsize); 465 466 /* There's nothing left to iterate over, so do nothing */ 467 if (abd_iter_at_end(aiter)) 468 return; 469 470 abd_t *abd = aiter->iter_abd; 471 size_t offset = aiter->iter_pos; 472 if (abd_is_linear(abd)) { 473 aiter->iter_mapsize = abd->abd_size - offset; 474 paddr = ABD_LINEAR_BUF(abd); 475 } else { 476 offset += ABD_SCATTER(abd).abd_offset; 477 paddr = ABD_SCATTER(abd).abd_chunks[offset >> PAGE_SHIFT]; 478 offset &= PAGE_MASK; 479 aiter->iter_mapsize = MIN(PAGE_SIZE - offset, 480 abd->abd_size - aiter->iter_pos); 481 } 482 aiter->iter_mapaddr = (char *)paddr + offset; 483 } 484 485 /* 486 * Unmap the current chunk from aiter. This can be safely called when the aiter 487 * has already exhausted, in which case this does nothing. 488 */ 489 void 490 abd_iter_unmap(struct abd_iter *aiter) 491 { 492 if (!abd_iter_at_end(aiter)) { 493 ASSERT3P(aiter->iter_mapaddr, !=, NULL); 494 ASSERT3U(aiter->iter_mapsize, >, 0); 495 } 496 497 aiter->iter_mapaddr = NULL; 498 aiter->iter_mapsize = 0; 499 } 500 501 void 502 abd_cache_reap_now(void) 503 { 504 kmem_cache_reap_soon(abd_chunk_cache); 505 } 506