Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 1 | /* -*- mode: c; c-basic-offset: 8; -*- |
| 2 | * vim: noexpandtab sw=8 ts=8 sts=0: |
| 3 | * |
| 4 | * reservations.c |
| 5 | * |
| 6 | * Allocation reservations implementation |
| 7 | * |
| 8 | * Some code borrowed from fs/ext3/balloc.c and is: |
| 9 | * |
| 10 | * Copyright (C) 1992, 1993, 1994, 1995 |
| 11 | * Remy Card (card@masi.ibp.fr) |
| 12 | * Laboratoire MASI - Institut Blaise Pascal |
| 13 | * Universite Pierre et Marie Curie (Paris VI) |
| 14 | * |
| 15 | * The rest is copyright (C) 2010 Novell. All rights reserved. |
| 16 | * |
| 17 | * This program is free software; you can redistribute it and/or |
| 18 | * modify it under the terms of the GNU General Public |
| 19 | * License version 2 as published by the Free Software Foundation. |
| 20 | * |
| 21 | * This program is distributed in the hope that it will be useful, |
| 22 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 23 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 24 | * General Public License for more details. |
| 25 | */ |
| 26 | |
| 27 | #include <linux/fs.h> |
| 28 | #include <linux/types.h> |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 29 | #include <linux/highmem.h> |
| 30 | #include <linux/bitops.h> |
| 31 | #include <linux/list.h> |
| 32 | |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 33 | #include <cluster/masklog.h> |
| 34 | |
| 35 | #include "ocfs2.h" |
Tao Ma | 2661836 | 2011-02-23 22:10:56 +0800 | [diff] [blame] | 36 | #include "ocfs2_trace.h" |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 37 | |
| 38 | #ifdef CONFIG_OCFS2_DEBUG_FS |
| 39 | #define OCFS2_CHECK_RESERVATIONS |
| 40 | #endif |
| 41 | |
| 42 | DEFINE_SPINLOCK(resv_lock); |
| 43 | |
| 44 | #define OCFS2_MIN_RESV_WINDOW_BITS 8 |
| 45 | #define OCFS2_MAX_RESV_WINDOW_BITS 1024 |
Mark Fasheh | 83f9231 | 2010-04-05 18:17:16 -0700 | [diff] [blame] | 46 | |
| 47 | int ocfs2_dir_resv_allowed(struct ocfs2_super *osb) |
| 48 | { |
| 49 | return (osb->osb_resv_level && osb->osb_dir_resv_level); |
| 50 | } |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 51 | |
| 52 | static unsigned int ocfs2_resv_window_bits(struct ocfs2_reservation_map *resmap, |
| 53 | struct ocfs2_alloc_reservation *resv) |
| 54 | { |
| 55 | struct ocfs2_super *osb = resmap->m_osb; |
| 56 | unsigned int bits; |
| 57 | |
Mark Fasheh | e3b4a97 | 2009-12-07 13:16:07 -0800 | [diff] [blame] | 58 | if (!(resv->r_flags & OCFS2_RESV_FLAG_DIR)) { |
| 59 | /* 8, 16, 32, 64, 128, 256, 512, 1024 */ |
| 60 | bits = 4 << osb->osb_resv_level; |
Mark Fasheh | b07f8f2 | 2010-04-05 18:17:15 -0700 | [diff] [blame] | 61 | } else { |
Mark Fasheh | 83f9231 | 2010-04-05 18:17:16 -0700 | [diff] [blame] | 62 | bits = 4 << osb->osb_dir_resv_level; |
Mark Fasheh | b07f8f2 | 2010-04-05 18:17:15 -0700 | [diff] [blame] | 63 | } |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 64 | return bits; |
| 65 | } |
| 66 | |
| 67 | static inline unsigned int ocfs2_resv_end(struct ocfs2_alloc_reservation *resv) |
| 68 | { |
| 69 | if (resv->r_len) |
| 70 | return resv->r_start + resv->r_len - 1; |
| 71 | return resv->r_start; |
| 72 | } |
| 73 | |
| 74 | static inline int ocfs2_resv_empty(struct ocfs2_alloc_reservation *resv) |
| 75 | { |
| 76 | return !!(resv->r_len == 0); |
| 77 | } |
| 78 | |
| 79 | static inline int ocfs2_resmap_disabled(struct ocfs2_reservation_map *resmap) |
| 80 | { |
| 81 | if (resmap->m_osb->osb_resv_level == 0) |
| 82 | return 1; |
| 83 | return 0; |
| 84 | } |
| 85 | |
| 86 | static void ocfs2_dump_resv(struct ocfs2_reservation_map *resmap) |
| 87 | { |
| 88 | struct ocfs2_super *osb = resmap->m_osb; |
| 89 | struct rb_node *node; |
| 90 | struct ocfs2_alloc_reservation *resv; |
| 91 | int i = 0; |
| 92 | |
| 93 | mlog(ML_NOTICE, "Dumping resmap for device %s. Bitmap length: %u\n", |
| 94 | osb->dev_str, resmap->m_bitmap_len); |
| 95 | |
| 96 | node = rb_first(&resmap->m_reservations); |
| 97 | while (node) { |
| 98 | resv = rb_entry(node, struct ocfs2_alloc_reservation, r_node); |
| 99 | |
| 100 | mlog(ML_NOTICE, "start: %u\tend: %u\tlen: %u\tlast_start: %u" |
| 101 | "\tlast_len: %u\n", resv->r_start, |
| 102 | ocfs2_resv_end(resv), resv->r_len, resv->r_last_start, |
| 103 | resv->r_last_len); |
| 104 | |
| 105 | node = rb_next(node); |
| 106 | i++; |
| 107 | } |
| 108 | |
| 109 | mlog(ML_NOTICE, "%d reservations found. LRU follows\n", i); |
| 110 | |
| 111 | i = 0; |
| 112 | list_for_each_entry(resv, &resmap->m_lru, r_lru) { |
| 113 | mlog(ML_NOTICE, "LRU(%d) start: %u\tend: %u\tlen: %u\t" |
| 114 | "last_start: %u\tlast_len: %u\n", i, resv->r_start, |
| 115 | ocfs2_resv_end(resv), resv->r_len, resv->r_last_start, |
| 116 | resv->r_last_len); |
| 117 | |
| 118 | i++; |
| 119 | } |
| 120 | } |
| 121 | |
| 122 | #ifdef OCFS2_CHECK_RESERVATIONS |
| 123 | static int ocfs2_validate_resmap_bits(struct ocfs2_reservation_map *resmap, |
| 124 | int i, |
| 125 | struct ocfs2_alloc_reservation *resv) |
| 126 | { |
| 127 | char *disk_bitmap = resmap->m_disk_bitmap; |
| 128 | unsigned int start = resv->r_start; |
| 129 | unsigned int end = ocfs2_resv_end(resv); |
| 130 | |
| 131 | while (start <= end) { |
| 132 | if (ocfs2_test_bit(start, disk_bitmap)) { |
| 133 | mlog(ML_ERROR, |
| 134 | "reservation %d covers an allocated area " |
| 135 | "starting at bit %u!\n", i, start); |
| 136 | return 1; |
| 137 | } |
| 138 | |
| 139 | start++; |
| 140 | } |
| 141 | return 0; |
| 142 | } |
| 143 | |
| 144 | static void ocfs2_check_resmap(struct ocfs2_reservation_map *resmap) |
| 145 | { |
| 146 | unsigned int off = 0; |
| 147 | int i = 0; |
| 148 | struct rb_node *node; |
| 149 | struct ocfs2_alloc_reservation *resv; |
| 150 | |
| 151 | node = rb_first(&resmap->m_reservations); |
| 152 | while (node) { |
| 153 | resv = rb_entry(node, struct ocfs2_alloc_reservation, r_node); |
| 154 | |
| 155 | if (i > 0 && resv->r_start <= off) { |
| 156 | mlog(ML_ERROR, "reservation %d has bad start off!\n", |
| 157 | i); |
| 158 | goto bad; |
| 159 | } |
| 160 | |
| 161 | if (resv->r_len == 0) { |
| 162 | mlog(ML_ERROR, "reservation %d has no length!\n", |
| 163 | i); |
| 164 | goto bad; |
| 165 | } |
| 166 | |
| 167 | if (resv->r_start > ocfs2_resv_end(resv)) { |
| 168 | mlog(ML_ERROR, "reservation %d has invalid range!\n", |
| 169 | i); |
| 170 | goto bad; |
| 171 | } |
| 172 | |
| 173 | if (ocfs2_resv_end(resv) >= resmap->m_bitmap_len) { |
| 174 | mlog(ML_ERROR, "reservation %d extends past bitmap!\n", |
| 175 | i); |
| 176 | goto bad; |
| 177 | } |
| 178 | |
| 179 | if (ocfs2_validate_resmap_bits(resmap, i, resv)) |
| 180 | goto bad; |
| 181 | |
| 182 | off = ocfs2_resv_end(resv); |
| 183 | node = rb_next(node); |
| 184 | |
| 185 | i++; |
| 186 | } |
| 187 | return; |
| 188 | |
| 189 | bad: |
| 190 | ocfs2_dump_resv(resmap); |
| 191 | BUG(); |
| 192 | } |
| 193 | #else |
| 194 | static inline void ocfs2_check_resmap(struct ocfs2_reservation_map *resmap) |
| 195 | { |
| 196 | |
| 197 | } |
| 198 | #endif |
| 199 | |
| 200 | void ocfs2_resv_init_once(struct ocfs2_alloc_reservation *resv) |
| 201 | { |
| 202 | memset(resv, 0, sizeof(*resv)); |
| 203 | INIT_LIST_HEAD(&resv->r_lru); |
| 204 | } |
| 205 | |
| 206 | void ocfs2_resv_set_type(struct ocfs2_alloc_reservation *resv, |
| 207 | unsigned int flags) |
| 208 | { |
| 209 | BUG_ON(flags & ~OCFS2_RESV_TYPES); |
| 210 | |
| 211 | resv->r_flags |= flags; |
| 212 | } |
| 213 | |
| 214 | int ocfs2_resmap_init(struct ocfs2_super *osb, |
| 215 | struct ocfs2_reservation_map *resmap) |
| 216 | { |
| 217 | memset(resmap, 0, sizeof(*resmap)); |
| 218 | |
| 219 | resmap->m_osb = osb; |
| 220 | resmap->m_reservations = RB_ROOT; |
| 221 | /* m_bitmap_len is initialized to zero by the above memset. */ |
| 222 | INIT_LIST_HEAD(&resmap->m_lru); |
| 223 | |
| 224 | return 0; |
| 225 | } |
| 226 | |
| 227 | static void ocfs2_resv_mark_lru(struct ocfs2_reservation_map *resmap, |
| 228 | struct ocfs2_alloc_reservation *resv) |
| 229 | { |
| 230 | assert_spin_locked(&resv_lock); |
| 231 | |
| 232 | if (!list_empty(&resv->r_lru)) |
| 233 | list_del_init(&resv->r_lru); |
| 234 | |
| 235 | list_add_tail(&resv->r_lru, &resmap->m_lru); |
| 236 | } |
| 237 | |
| 238 | static void __ocfs2_resv_trunc(struct ocfs2_alloc_reservation *resv) |
| 239 | { |
| 240 | resv->r_len = 0; |
| 241 | resv->r_start = 0; |
| 242 | } |
| 243 | |
| 244 | static void ocfs2_resv_remove(struct ocfs2_reservation_map *resmap, |
| 245 | struct ocfs2_alloc_reservation *resv) |
| 246 | { |
| 247 | if (resv->r_flags & OCFS2_RESV_FLAG_INUSE) { |
| 248 | list_del_init(&resv->r_lru); |
| 249 | rb_erase(&resv->r_node, &resmap->m_reservations); |
| 250 | resv->r_flags &= ~OCFS2_RESV_FLAG_INUSE; |
| 251 | } |
| 252 | } |
| 253 | |
| 254 | static void __ocfs2_resv_discard(struct ocfs2_reservation_map *resmap, |
| 255 | struct ocfs2_alloc_reservation *resv) |
| 256 | { |
| 257 | assert_spin_locked(&resv_lock); |
| 258 | |
| 259 | __ocfs2_resv_trunc(resv); |
| 260 | /* |
| 261 | * last_len and last_start no longer make sense if |
| 262 | * we're changing the range of our allocations. |
| 263 | */ |
| 264 | resv->r_last_len = resv->r_last_start = 0; |
| 265 | |
| 266 | ocfs2_resv_remove(resmap, resv); |
| 267 | } |
| 268 | |
| 269 | /* does nothing if 'resv' is null */ |
| 270 | void ocfs2_resv_discard(struct ocfs2_reservation_map *resmap, |
| 271 | struct ocfs2_alloc_reservation *resv) |
| 272 | { |
| 273 | if (resv) { |
| 274 | spin_lock(&resv_lock); |
| 275 | __ocfs2_resv_discard(resmap, resv); |
| 276 | spin_unlock(&resv_lock); |
| 277 | } |
| 278 | } |
| 279 | |
| 280 | static void ocfs2_resmap_clear_all_resv(struct ocfs2_reservation_map *resmap) |
| 281 | { |
| 282 | struct rb_node *node; |
| 283 | struct ocfs2_alloc_reservation *resv; |
| 284 | |
| 285 | assert_spin_locked(&resv_lock); |
| 286 | |
| 287 | while ((node = rb_last(&resmap->m_reservations)) != NULL) { |
| 288 | resv = rb_entry(node, struct ocfs2_alloc_reservation, r_node); |
| 289 | |
| 290 | __ocfs2_resv_discard(resmap, resv); |
| 291 | } |
| 292 | } |
| 293 | |
| 294 | void ocfs2_resmap_restart(struct ocfs2_reservation_map *resmap, |
| 295 | unsigned int clen, char *disk_bitmap) |
| 296 | { |
| 297 | if (ocfs2_resmap_disabled(resmap)) |
| 298 | return; |
| 299 | |
| 300 | spin_lock(&resv_lock); |
| 301 | |
| 302 | ocfs2_resmap_clear_all_resv(resmap); |
| 303 | resmap->m_bitmap_len = clen; |
| 304 | resmap->m_disk_bitmap = disk_bitmap; |
| 305 | |
| 306 | spin_unlock(&resv_lock); |
| 307 | } |
| 308 | |
| 309 | void ocfs2_resmap_uninit(struct ocfs2_reservation_map *resmap) |
| 310 | { |
| 311 | /* Does nothing for now. Keep this around for API symmetry */ |
| 312 | } |
| 313 | |
| 314 | static void ocfs2_resv_insert(struct ocfs2_reservation_map *resmap, |
| 315 | struct ocfs2_alloc_reservation *new) |
| 316 | { |
| 317 | struct rb_root *root = &resmap->m_reservations; |
| 318 | struct rb_node *parent = NULL; |
| 319 | struct rb_node **p = &root->rb_node; |
| 320 | struct ocfs2_alloc_reservation *tmp; |
| 321 | |
| 322 | assert_spin_locked(&resv_lock); |
| 323 | |
Tao Ma | 2661836 | 2011-02-23 22:10:56 +0800 | [diff] [blame] | 324 | trace_ocfs2_resv_insert(new->r_start, new->r_len); |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 325 | |
| 326 | while (*p) { |
| 327 | parent = *p; |
| 328 | |
| 329 | tmp = rb_entry(parent, struct ocfs2_alloc_reservation, r_node); |
| 330 | |
| 331 | if (new->r_start < tmp->r_start) { |
| 332 | p = &(*p)->rb_left; |
| 333 | |
| 334 | /* |
| 335 | * This is a good place to check for |
| 336 | * overlapping reservations. |
| 337 | */ |
| 338 | BUG_ON(ocfs2_resv_end(new) >= tmp->r_start); |
| 339 | } else if (new->r_start > ocfs2_resv_end(tmp)) { |
| 340 | p = &(*p)->rb_right; |
| 341 | } else { |
| 342 | /* This should never happen! */ |
| 343 | mlog(ML_ERROR, "Duplicate reservation window!\n"); |
| 344 | BUG(); |
| 345 | } |
| 346 | } |
| 347 | |
| 348 | rb_link_node(&new->r_node, parent, p); |
| 349 | rb_insert_color(&new->r_node, root); |
| 350 | new->r_flags |= OCFS2_RESV_FLAG_INUSE; |
| 351 | |
| 352 | ocfs2_resv_mark_lru(resmap, new); |
| 353 | |
| 354 | ocfs2_check_resmap(resmap); |
| 355 | } |
| 356 | |
| 357 | /** |
| 358 | * ocfs2_find_resv_lhs() - find the window which contains goal |
| 359 | * @resmap: reservation map to search |
| 360 | * @goal: which bit to search for |
| 361 | * |
| 362 | * If a window containing that goal is not found, we return the window |
| 363 | * which comes before goal. Returns NULL on empty rbtree or no window |
| 364 | * before goal. |
| 365 | */ |
| 366 | static struct ocfs2_alloc_reservation * |
| 367 | ocfs2_find_resv_lhs(struct ocfs2_reservation_map *resmap, unsigned int goal) |
| 368 | { |
| 369 | struct ocfs2_alloc_reservation *resv = NULL; |
| 370 | struct ocfs2_alloc_reservation *prev_resv = NULL; |
| 371 | struct rb_node *node = resmap->m_reservations.rb_node; |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 372 | |
| 373 | assert_spin_locked(&resv_lock); |
| 374 | |
| 375 | if (!node) |
| 376 | return NULL; |
| 377 | |
| 378 | node = rb_first(&resmap->m_reservations); |
| 379 | while (node) { |
| 380 | resv = rb_entry(node, struct ocfs2_alloc_reservation, r_node); |
| 381 | |
| 382 | if (resv->r_start <= goal && ocfs2_resv_end(resv) >= goal) |
| 383 | break; |
| 384 | |
| 385 | /* Check if we overshot the reservation just before goal? */ |
| 386 | if (resv->r_start > goal) { |
| 387 | resv = prev_resv; |
| 388 | break; |
| 389 | } |
| 390 | |
| 391 | prev_resv = resv; |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 392 | node = rb_next(node); |
| 393 | } |
| 394 | |
| 395 | return resv; |
| 396 | } |
| 397 | |
| 398 | /* |
| 399 | * We are given a range within the bitmap, which corresponds to a gap |
| 400 | * inside the reservations tree (search_start, search_len). The range |
| 401 | * can be anything from the whole bitmap, to a gap between |
| 402 | * reservations. |
| 403 | * |
| 404 | * The start value of *rstart is insignificant. |
| 405 | * |
| 406 | * This function searches the bitmap range starting at search_start |
Tao Ma | b065556 | 2010-04-08 16:33:02 +0800 | [diff] [blame] | 407 | * with length search_len for a set of contiguous free bits. We try |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 408 | * to find up to 'wanted' bits, but can sometimes return less. |
| 409 | * |
| 410 | * Returns the length of allocation, 0 if no free bits are found. |
| 411 | * |
| 412 | * *cstart and *clen will also be populated with the result. |
| 413 | */ |
| 414 | static int ocfs2_resmap_find_free_bits(struct ocfs2_reservation_map *resmap, |
| 415 | unsigned int wanted, |
| 416 | unsigned int search_start, |
| 417 | unsigned int search_len, |
| 418 | unsigned int *rstart, |
| 419 | unsigned int *rlen) |
| 420 | { |
| 421 | void *bitmap = resmap->m_disk_bitmap; |
| 422 | unsigned int best_start, best_len = 0; |
| 423 | int offset, start, found; |
| 424 | |
Tao Ma | 2661836 | 2011-02-23 22:10:56 +0800 | [diff] [blame] | 425 | trace_ocfs2_resmap_find_free_bits_begin(search_start, search_len, |
| 426 | wanted, resmap->m_bitmap_len); |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 427 | |
| 428 | found = best_start = best_len = 0; |
| 429 | |
| 430 | start = search_start; |
| 431 | while ((offset = ocfs2_find_next_zero_bit(bitmap, resmap->m_bitmap_len, |
| 432 | start)) != -1) { |
| 433 | /* Search reached end of the region */ |
| 434 | if (offset >= (search_start + search_len)) |
| 435 | break; |
| 436 | |
| 437 | if (offset == start) { |
| 438 | /* we found a zero */ |
| 439 | found++; |
| 440 | /* move start to the next bit to test */ |
| 441 | start++; |
| 442 | } else { |
| 443 | /* got a zero after some ones */ |
| 444 | found = 1; |
| 445 | start = offset + 1; |
| 446 | } |
| 447 | if (found > best_len) { |
| 448 | best_len = found; |
| 449 | best_start = start - found; |
| 450 | } |
| 451 | |
| 452 | if (found >= wanted) |
| 453 | break; |
| 454 | } |
| 455 | |
| 456 | if (best_len == 0) |
| 457 | return 0; |
| 458 | |
| 459 | if (best_len >= wanted) |
| 460 | best_len = wanted; |
| 461 | |
| 462 | *rlen = best_len; |
| 463 | *rstart = best_start; |
| 464 | |
Tao Ma | 2661836 | 2011-02-23 22:10:56 +0800 | [diff] [blame] | 465 | trace_ocfs2_resmap_find_free_bits_end(best_start, best_len); |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 466 | |
| 467 | return *rlen; |
| 468 | } |
| 469 | |
| 470 | static void __ocfs2_resv_find_window(struct ocfs2_reservation_map *resmap, |
| 471 | struct ocfs2_alloc_reservation *resv, |
| 472 | unsigned int goal, unsigned int wanted) |
| 473 | { |
| 474 | struct rb_root *root = &resmap->m_reservations; |
| 475 | unsigned int gap_start, gap_end, gap_len; |
| 476 | struct ocfs2_alloc_reservation *prev_resv, *next_resv; |
| 477 | struct rb_node *prev, *next; |
| 478 | unsigned int cstart, clen; |
| 479 | unsigned int best_start = 0, best_len = 0; |
| 480 | |
| 481 | /* |
| 482 | * Nasty cases to consider: |
| 483 | * |
| 484 | * - rbtree is empty |
| 485 | * - our window should be first in all reservations |
| 486 | * - our window should be last in all reservations |
| 487 | * - need to make sure we don't go past end of bitmap |
| 488 | */ |
Tao Ma | 2661836 | 2011-02-23 22:10:56 +0800 | [diff] [blame] | 489 | trace_ocfs2_resv_find_window_begin(resv->r_start, ocfs2_resv_end(resv), |
| 490 | goal, wanted, RB_EMPTY_ROOT(root)); |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 491 | |
| 492 | assert_spin_locked(&resv_lock); |
| 493 | |
| 494 | if (RB_EMPTY_ROOT(root)) { |
| 495 | /* |
| 496 | * Easiest case - empty tree. We can just take |
| 497 | * whatever window of free bits we want. |
| 498 | */ |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 499 | clen = ocfs2_resmap_find_free_bits(resmap, wanted, goal, |
| 500 | resmap->m_bitmap_len - goal, |
| 501 | &cstart, &clen); |
| 502 | |
| 503 | /* |
| 504 | * This should never happen - the local alloc window |
| 505 | * will always have free bits when we're called. |
| 506 | */ |
| 507 | BUG_ON(goal == 0 && clen == 0); |
| 508 | |
| 509 | if (clen == 0) |
| 510 | return; |
| 511 | |
| 512 | resv->r_start = cstart; |
| 513 | resv->r_len = clen; |
| 514 | |
| 515 | ocfs2_resv_insert(resmap, resv); |
| 516 | return; |
| 517 | } |
| 518 | |
| 519 | prev_resv = ocfs2_find_resv_lhs(resmap, goal); |
| 520 | |
| 521 | if (prev_resv == NULL) { |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 522 | /* |
| 523 | * A NULL here means that the search code couldn't |
| 524 | * find a window that starts before goal. |
| 525 | * |
| 526 | * However, we can take the first window after goal, |
| 527 | * which is also by definition, the leftmost window in |
| 528 | * the entire tree. If we can find free bits in the |
| 529 | * gap between goal and the LHS window, then the |
| 530 | * reservation can safely be placed there. |
| 531 | * |
| 532 | * Otherwise we fall back to a linear search, checking |
| 533 | * the gaps in between windows for a place to |
| 534 | * allocate. |
| 535 | */ |
| 536 | |
| 537 | next = rb_first(root); |
| 538 | next_resv = rb_entry(next, struct ocfs2_alloc_reservation, |
| 539 | r_node); |
| 540 | |
| 541 | /* |
| 542 | * The search should never return such a window. (see |
| 543 | * comment above |
| 544 | */ |
| 545 | if (next_resv->r_start <= goal) { |
| 546 | mlog(ML_ERROR, "goal: %u next_resv: start %u len %u\n", |
| 547 | goal, next_resv->r_start, next_resv->r_len); |
| 548 | ocfs2_dump_resv(resmap); |
| 549 | BUG(); |
| 550 | } |
| 551 | |
| 552 | clen = ocfs2_resmap_find_free_bits(resmap, wanted, goal, |
| 553 | next_resv->r_start - goal, |
| 554 | &cstart, &clen); |
| 555 | if (clen) { |
| 556 | best_len = clen; |
| 557 | best_start = cstart; |
| 558 | if (best_len == wanted) |
| 559 | goto out_insert; |
| 560 | } |
| 561 | |
| 562 | prev_resv = next_resv; |
| 563 | next_resv = NULL; |
| 564 | } |
| 565 | |
Tao Ma | 2661836 | 2011-02-23 22:10:56 +0800 | [diff] [blame] | 566 | trace_ocfs2_resv_find_window_prev(prev_resv->r_start, |
| 567 | ocfs2_resv_end(prev_resv)); |
| 568 | |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 569 | prev = &prev_resv->r_node; |
| 570 | |
| 571 | /* Now we do a linear search for a window, starting at 'prev_rsv' */ |
| 572 | while (1) { |
| 573 | next = rb_next(prev); |
| 574 | if (next) { |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 575 | next_resv = rb_entry(next, |
| 576 | struct ocfs2_alloc_reservation, |
| 577 | r_node); |
| 578 | |
| 579 | gap_start = ocfs2_resv_end(prev_resv) + 1; |
| 580 | gap_end = next_resv->r_start - 1; |
| 581 | gap_len = gap_end - gap_start + 1; |
| 582 | } else { |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 583 | /* |
| 584 | * We're at the rightmost edge of the |
| 585 | * tree. See if a reservation between this |
| 586 | * window and the end of the bitmap will work. |
| 587 | */ |
| 588 | gap_start = ocfs2_resv_end(prev_resv) + 1; |
| 589 | gap_len = resmap->m_bitmap_len - gap_start; |
| 590 | gap_end = resmap->m_bitmap_len - 1; |
| 591 | } |
| 592 | |
Tao Ma | 2661836 | 2011-02-23 22:10:56 +0800 | [diff] [blame] | 593 | trace_ocfs2_resv_find_window_next(next ? next_resv->r_start: -1, |
| 594 | next ? ocfs2_resv_end(next_resv) : -1); |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 595 | /* |
| 596 | * No need to check this gap if we have already found |
| 597 | * a larger region of free bits. |
| 598 | */ |
| 599 | if (gap_len <= best_len) |
| 600 | goto next_resv; |
| 601 | |
| 602 | clen = ocfs2_resmap_find_free_bits(resmap, wanted, gap_start, |
| 603 | gap_len, &cstart, &clen); |
| 604 | if (clen == wanted) { |
| 605 | best_len = clen; |
| 606 | best_start = cstart; |
| 607 | goto out_insert; |
| 608 | } else if (clen > best_len) { |
| 609 | best_len = clen; |
| 610 | best_start = cstart; |
| 611 | } |
| 612 | |
| 613 | next_resv: |
| 614 | if (!next) |
| 615 | break; |
| 616 | |
| 617 | prev = next; |
| 618 | prev_resv = rb_entry(prev, struct ocfs2_alloc_reservation, |
| 619 | r_node); |
| 620 | } |
| 621 | |
| 622 | out_insert: |
| 623 | if (best_len) { |
| 624 | resv->r_start = best_start; |
| 625 | resv->r_len = best_len; |
| 626 | ocfs2_resv_insert(resmap, resv); |
| 627 | } |
| 628 | } |
| 629 | |
| 630 | static void ocfs2_cannibalize_resv(struct ocfs2_reservation_map *resmap, |
| 631 | struct ocfs2_alloc_reservation *resv, |
| 632 | unsigned int wanted) |
| 633 | { |
| 634 | struct ocfs2_alloc_reservation *lru_resv; |
| 635 | int tmpwindow = !!(resv->r_flags & OCFS2_RESV_FLAG_TMP); |
| 636 | unsigned int min_bits; |
| 637 | |
| 638 | if (!tmpwindow) |
| 639 | min_bits = ocfs2_resv_window_bits(resmap, resv) >> 1; |
| 640 | else |
| 641 | min_bits = wanted; /* We at know the temp window will use all |
| 642 | * of these bits */ |
| 643 | |
| 644 | /* |
| 645 | * Take the first reservation off the LRU as our 'target'. We |
| 646 | * don't try to be smart about it. There might be a case for |
| 647 | * searching based on size but I don't have enough data to be |
| 648 | * sure. --Mark (3/16/2010) |
| 649 | */ |
| 650 | lru_resv = list_first_entry(&resmap->m_lru, |
| 651 | struct ocfs2_alloc_reservation, r_lru); |
| 652 | |
Tao Ma | 2661836 | 2011-02-23 22:10:56 +0800 | [diff] [blame] | 653 | trace_ocfs2_cannibalize_resv_begin(lru_resv->r_start, |
| 654 | lru_resv->r_len, |
| 655 | ocfs2_resv_end(lru_resv)); |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 656 | |
| 657 | /* |
| 658 | * Cannibalize (some or all) of the target reservation and |
| 659 | * feed it to the current window. |
| 660 | */ |
| 661 | if (lru_resv->r_len <= min_bits) { |
| 662 | /* |
| 663 | * Discard completely if size is less than or equal to a |
| 664 | * reasonable threshold - 50% of window bits for non temporary |
| 665 | * windows. |
| 666 | */ |
| 667 | resv->r_start = lru_resv->r_start; |
| 668 | resv->r_len = lru_resv->r_len; |
| 669 | |
| 670 | __ocfs2_resv_discard(resmap, lru_resv); |
| 671 | } else { |
| 672 | unsigned int shrink; |
| 673 | if (tmpwindow) |
| 674 | shrink = min_bits; |
| 675 | else |
| 676 | shrink = lru_resv->r_len / 2; |
| 677 | |
| 678 | lru_resv->r_len -= shrink; |
| 679 | |
| 680 | resv->r_start = ocfs2_resv_end(lru_resv) + 1; |
| 681 | resv->r_len = shrink; |
| 682 | } |
| 683 | |
Tao Ma | 2661836 | 2011-02-23 22:10:56 +0800 | [diff] [blame] | 684 | trace_ocfs2_cannibalize_resv_end(resv->r_start, ocfs2_resv_end(resv), |
| 685 | resv->r_len, resv->r_last_start, |
| 686 | resv->r_last_len); |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 687 | |
| 688 | ocfs2_resv_insert(resmap, resv); |
| 689 | } |
| 690 | |
| 691 | static void ocfs2_resv_find_window(struct ocfs2_reservation_map *resmap, |
| 692 | struct ocfs2_alloc_reservation *resv, |
| 693 | unsigned int wanted) |
| 694 | { |
| 695 | unsigned int goal = 0; |
| 696 | |
| 697 | BUG_ON(!ocfs2_resv_empty(resv)); |
| 698 | |
| 699 | /* |
| 700 | * Begin by trying to get a window as close to the previous |
| 701 | * one as possible. Using the most recent allocation as a |
| 702 | * start goal makes sense. |
| 703 | */ |
| 704 | if (resv->r_last_len) { |
| 705 | goal = resv->r_last_start + resv->r_last_len; |
| 706 | if (goal >= resmap->m_bitmap_len) |
| 707 | goal = 0; |
| 708 | } |
| 709 | |
| 710 | __ocfs2_resv_find_window(resmap, resv, goal, wanted); |
| 711 | |
| 712 | /* Search from last alloc didn't work, try once more from beginning. */ |
| 713 | if (ocfs2_resv_empty(resv) && goal != 0) |
| 714 | __ocfs2_resv_find_window(resmap, resv, 0, wanted); |
| 715 | |
| 716 | if (ocfs2_resv_empty(resv)) { |
| 717 | /* |
| 718 | * Still empty? Pull oldest one off the LRU, remove it from |
| 719 | * tree, put this one in it's place. |
| 720 | */ |
| 721 | ocfs2_cannibalize_resv(resmap, resv, wanted); |
| 722 | } |
| 723 | |
| 724 | BUG_ON(ocfs2_resv_empty(resv)); |
| 725 | } |
| 726 | |
| 727 | int ocfs2_resmap_resv_bits(struct ocfs2_reservation_map *resmap, |
| 728 | struct ocfs2_alloc_reservation *resv, |
| 729 | int *cstart, int *clen) |
| 730 | { |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 731 | if (resv == NULL || ocfs2_resmap_disabled(resmap)) |
| 732 | return -ENOSPC; |
| 733 | |
| 734 | spin_lock(&resv_lock); |
| 735 | |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 736 | if (ocfs2_resv_empty(resv)) { |
Tao Ma | 4a452de | 2010-09-19 13:42:28 +0800 | [diff] [blame] | 737 | /* |
| 738 | * We don't want to over-allocate for temporary |
| 739 | * windows. Otherwise, we run the risk of fragmenting the |
| 740 | * allocation space. |
| 741 | */ |
| 742 | unsigned int wanted = ocfs2_resv_window_bits(resmap, resv); |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 743 | |
Tao Ma | 4a452de | 2010-09-19 13:42:28 +0800 | [diff] [blame] | 744 | if ((resv->r_flags & OCFS2_RESV_FLAG_TMP) || wanted < *clen) |
| 745 | wanted = *clen; |
| 746 | |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 747 | /* |
| 748 | * Try to get a window here. If it works, we must fall |
| 749 | * through and test the bitmap . This avoids some |
| 750 | * ping-ponging of windows due to non-reserved space |
| 751 | * being allocation before we initialize a window for |
| 752 | * that inode. |
| 753 | */ |
| 754 | ocfs2_resv_find_window(resmap, resv, wanted); |
Tao Ma | 2661836 | 2011-02-23 22:10:56 +0800 | [diff] [blame] | 755 | trace_ocfs2_resmap_resv_bits(resv->r_start, resv->r_len); |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 756 | } |
| 757 | |
| 758 | BUG_ON(ocfs2_resv_empty(resv)); |
| 759 | |
| 760 | *cstart = resv->r_start; |
| 761 | *clen = resv->r_len; |
| 762 | |
| 763 | spin_unlock(&resv_lock); |
| 764 | return 0; |
| 765 | } |
| 766 | |
| 767 | static void |
| 768 | ocfs2_adjust_resv_from_alloc(struct ocfs2_reservation_map *resmap, |
| 769 | struct ocfs2_alloc_reservation *resv, |
| 770 | unsigned int start, unsigned int end) |
| 771 | { |
Tao Ma | b065556 | 2010-04-08 16:33:02 +0800 | [diff] [blame] | 772 | unsigned int rhs = 0; |
| 773 | unsigned int old_end = ocfs2_resv_end(resv); |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 774 | |
Tao Ma | b065556 | 2010-04-08 16:33:02 +0800 | [diff] [blame] | 775 | BUG_ON(start != resv->r_start || old_end < end); |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 776 | |
| 777 | /* |
| 778 | * Completely used? We can remove it then. |
| 779 | */ |
Tao Ma | b065556 | 2010-04-08 16:33:02 +0800 | [diff] [blame] | 780 | if (old_end == end) { |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 781 | __ocfs2_resv_discard(resmap, resv); |
| 782 | return; |
| 783 | } |
| 784 | |
Tao Ma | b065556 | 2010-04-08 16:33:02 +0800 | [diff] [blame] | 785 | rhs = old_end - end; |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 786 | |
| 787 | /* |
Tao Ma | b065556 | 2010-04-08 16:33:02 +0800 | [diff] [blame] | 788 | * This should have been trapped above. |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 789 | */ |
Tao Ma | b065556 | 2010-04-08 16:33:02 +0800 | [diff] [blame] | 790 | BUG_ON(rhs == 0); |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 791 | |
Tao Ma | b065556 | 2010-04-08 16:33:02 +0800 | [diff] [blame] | 792 | resv->r_start = end + 1; |
| 793 | resv->r_len = old_end - resv->r_start + 1; |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 794 | } |
| 795 | |
| 796 | void ocfs2_resmap_claimed_bits(struct ocfs2_reservation_map *resmap, |
| 797 | struct ocfs2_alloc_reservation *resv, |
| 798 | u32 cstart, u32 clen) |
| 799 | { |
| 800 | unsigned int cend = cstart + clen - 1; |
| 801 | |
| 802 | if (resmap == NULL || ocfs2_resmap_disabled(resmap)) |
| 803 | return; |
| 804 | |
| 805 | if (resv == NULL) |
| 806 | return; |
| 807 | |
Tao Ma | b065556 | 2010-04-08 16:33:02 +0800 | [diff] [blame] | 808 | BUG_ON(cstart != resv->r_start); |
| 809 | |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 810 | spin_lock(&resv_lock); |
| 811 | |
Tao Ma | 2661836 | 2011-02-23 22:10:56 +0800 | [diff] [blame] | 812 | trace_ocfs2_resmap_claimed_bits_begin(cstart, cend, clen, resv->r_start, |
| 813 | ocfs2_resv_end(resv), resv->r_len, |
| 814 | resv->r_last_start, |
| 815 | resv->r_last_len); |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 816 | |
| 817 | BUG_ON(cstart < resv->r_start); |
| 818 | BUG_ON(cstart > ocfs2_resv_end(resv)); |
| 819 | BUG_ON(cend > ocfs2_resv_end(resv)); |
| 820 | |
| 821 | ocfs2_adjust_resv_from_alloc(resmap, resv, cstart, cend); |
| 822 | resv->r_last_start = cstart; |
| 823 | resv->r_last_len = clen; |
| 824 | |
| 825 | /* |
| 826 | * May have been discarded above from |
| 827 | * ocfs2_adjust_resv_from_alloc(). |
| 828 | */ |
| 829 | if (!ocfs2_resv_empty(resv)) |
| 830 | ocfs2_resv_mark_lru(resmap, resv); |
| 831 | |
Tao Ma | 2661836 | 2011-02-23 22:10:56 +0800 | [diff] [blame] | 832 | trace_ocfs2_resmap_claimed_bits_end(resv->r_start, ocfs2_resv_end(resv), |
| 833 | resv->r_len, resv->r_last_start, |
| 834 | resv->r_last_len); |
Mark Fasheh | d02f00c | 2009-12-07 13:10:48 -0800 | [diff] [blame] | 835 | |
| 836 | ocfs2_check_resmap(resmap); |
| 837 | |
| 838 | spin_unlock(&resv_lock); |
| 839 | } |