1 /* 2 drbd_bitmap.c 3 4 This file is part of DRBD by Philipp Reisner and Lars Ellenberg. 5 6 Copyright (C) 2004-2008, LINBIT Information Technologies GmbH. 7 Copyright (C) 2004-2008, Philipp Reisner <philipp.reisner@linbit.com>. 8 Copyright (C) 2004-2008, Lars Ellenberg <lars.ellenberg@linbit.com>. 9 10 drbd is free software; you can redistribute it and/or modify 11 it under the terms of the GNU General Public License as published by 12 the Free Software Foundation; either version 2, or (at your option) 13 any later version. 14 15 drbd is distributed in the hope that it will be useful, 16 but WITHOUT ANY WARRANTY; without even the implied warranty of 17 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 18 GNU General Public License for more details. 19 20 You should have received a copy of the GNU General Public License 21 along with drbd; see the file COPYING. If not, write to 22 the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA. 23 */ 24 25 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 26 27 #include <linux/bitmap.h> 28 #include <linux/vmalloc.h> 29 #include <linux/string.h> 30 #include <linux/drbd.h> 31 #include <linux/slab.h> 32 #include <linux/highmem.h> 33 34 #include "drbd_int.h" 35 36 37 /* OPAQUE outside this file! 38 * interface defined in drbd_int.h 39 40 * convention: 41 * function name drbd_bm_... => used elsewhere, "public". 42 * function name bm_... => internal to implementation, "private". 43 */ 44 45 46 /* 47 * LIMITATIONS: 48 * We want to support >= peta byte of backend storage, while for now still using 49 * a granularity of one bit per 4KiB of storage. 50 * 1 << 50 bytes backend storage (1 PiB) 51 * 1 << (50 - 12) bits needed 52 * 38 --> we need u64 to index and count bits 53 * 1 << (38 - 3) bitmap bytes needed 54 * 35 --> we still need u64 to index and count bytes 55 * (that's 32 GiB of bitmap for 1 PiB storage) 56 * 1 << (35 - 2) 32bit longs needed 57 * 33 --> we'd even need u64 to index and count 32bit long words. 58 * 1 << (35 - 3) 64bit longs needed 59 * 32 --> we could get away with a 32bit unsigned int to index and count 60 * 64bit long words, but I rather stay with unsigned long for now. 61 * We probably should neither count nor point to bytes or long words 62 * directly, but either by bitnumber, or by page index and offset. 63 * 1 << (35 - 12) 64 * 22 --> we need that much 4KiB pages of bitmap. 65 * 1 << (22 + 3) --> on a 64bit arch, 66 * we need 32 MiB to store the array of page pointers. 67 * 68 * Because I'm lazy, and because the resulting patch was too large, too ugly 69 * and still incomplete, on 32bit we still "only" support 16 TiB (minus some), 70 * (1 << 32) bits * 4k storage. 71 * 72 73 * bitmap storage and IO: 74 * Bitmap is stored little endian on disk, and is kept little endian in 75 * core memory. Currently we still hold the full bitmap in core as long 76 * as we are "attached" to a local disk, which at 32 GiB for 1PiB storage 77 * seems excessive. 78 * 79 * We plan to reduce the amount of in-core bitmap pages by paging them in 80 * and out against their on-disk location as necessary, but need to make 81 * sure we don't cause too much meta data IO, and must not deadlock in 82 * tight memory situations. This needs some more work. 83 */ 84 85 /* 86 * NOTE 87 * Access to the *bm_pages is protected by bm_lock. 88 * It is safe to read the other members within the lock. 89 * 90 * drbd_bm_set_bits is called from bio_endio callbacks, 91 * We may be called with irq already disabled, 92 * so we need spin_lock_irqsave(). 93 * And we need the kmap_atomic. 94 */ 95 struct drbd_bitmap { 96 struct page **bm_pages; 97 spinlock_t bm_lock; 98 99 /* exclusively to be used by __al_write_transaction(), 100 * drbd_bm_mark_for_writeout() and 101 * and drbd_bm_write_hinted() -> bm_rw() called from there. 102 */ 103 unsigned int n_bitmap_hints; 104 unsigned int al_bitmap_hints[AL_UPDATES_PER_TRANSACTION]; 105 106 /* see LIMITATIONS: above */ 107 108 unsigned long bm_set; /* nr of set bits; THINK maybe atomic_t? */ 109 unsigned long bm_bits; 110 size_t bm_words; 111 size_t bm_number_of_pages; 112 sector_t bm_dev_capacity; 113 struct mutex bm_change; /* serializes resize operations */ 114 115 wait_queue_head_t bm_io_wait; /* used to serialize IO of single pages */ 116 117 enum bm_flag bm_flags; 118 119 /* debugging aid, in case we are still racy somewhere */ 120 char *bm_why; 121 struct task_struct *bm_task; 122 }; 123 124 #define bm_print_lock_info(m) __bm_print_lock_info(m, __func__) 125 static void __bm_print_lock_info(struct drbd_device *device, const char *func) 126 { 127 struct drbd_bitmap *b = device->bitmap; 128 if (!__ratelimit(&drbd_ratelimit_state)) 129 return; 130 drbd_err(device, "FIXME %s[%d] in %s, bitmap locked for '%s' by %s[%d]\n", 131 current->comm, task_pid_nr(current), 132 func, b->bm_why ?: "?", 133 b->bm_task->comm, task_pid_nr(b->bm_task)); 134 } 135 136 void drbd_bm_lock(struct drbd_device *device, char *why, enum bm_flag flags) 137 { 138 struct drbd_bitmap *b = device->bitmap; 139 int trylock_failed; 140 141 if (!b) { 142 drbd_err(device, "FIXME no bitmap in drbd_bm_lock!?\n"); 143 return; 144 } 145 146 trylock_failed = !mutex_trylock(&b->bm_change); 147 148 if (trylock_failed) { 149 drbd_warn(device, "%s[%d] going to '%s' but bitmap already locked for '%s' by %s[%d]\n", 150 current->comm, task_pid_nr(current), 151 why, b->bm_why ?: "?", 152 b->bm_task->comm, task_pid_nr(b->bm_task)); 153 mutex_lock(&b->bm_change); 154 } 155 if (BM_LOCKED_MASK & b->bm_flags) 156 drbd_err(device, "FIXME bitmap already locked in bm_lock\n"); 157 b->bm_flags |= flags & BM_LOCKED_MASK; 158 159 b->bm_why = why; 160 b->bm_task = current; 161 } 162 163 void drbd_bm_unlock(struct drbd_device *device) 164 { 165 struct drbd_bitmap *b = device->bitmap; 166 if (!b) { 167 drbd_err(device, "FIXME no bitmap in drbd_bm_unlock!?\n"); 168 return; 169 } 170 171 if (!(BM_LOCKED_MASK & device->bitmap->bm_flags)) 172 drbd_err(device, "FIXME bitmap not locked in bm_unlock\n"); 173 174 b->bm_flags &= ~BM_LOCKED_MASK; 175 b->bm_why = NULL; 176 b->bm_task = NULL; 177 mutex_unlock(&b->bm_change); 178 } 179 180 /* we store some "meta" info about our pages in page->private */ 181 /* at a granularity of 4k storage per bitmap bit: 182 * one peta byte storage: 1<<50 byte, 1<<38 * 4k storage blocks 183 * 1<<38 bits, 184 * 1<<23 4k bitmap pages. 185 * Use 24 bits as page index, covers 2 peta byte storage 186 * at a granularity of 4k per bit. 187 * Used to report the failed page idx on io error from the endio handlers. 188 */ 189 #define BM_PAGE_IDX_MASK ((1UL<<24)-1) 190 /* this page is currently read in, or written back */ 191 #define BM_PAGE_IO_LOCK 31 192 /* if there has been an IO error for this page */ 193 #define BM_PAGE_IO_ERROR 30 194 /* this is to be able to intelligently skip disk IO, 195 * set if bits have been set since last IO. */ 196 #define BM_PAGE_NEED_WRITEOUT 29 197 /* to mark for lazy writeout once syncer cleared all clearable bits, 198 * we if bits have been cleared since last IO. */ 199 #define BM_PAGE_LAZY_WRITEOUT 28 200 /* pages marked with this "HINT" will be considered for writeout 201 * on activity log transactions */ 202 #define BM_PAGE_HINT_WRITEOUT 27 203 204 /* store_page_idx uses non-atomic assignment. It is only used directly after 205 * allocating the page. All other bm_set_page_* and bm_clear_page_* need to 206 * use atomic bit manipulation, as set_out_of_sync (and therefore bitmap 207 * changes) may happen from various contexts, and wait_on_bit/wake_up_bit 208 * requires it all to be atomic as well. */ 209 static void bm_store_page_idx(struct page *page, unsigned long idx) 210 { 211 BUG_ON(0 != (idx & ~BM_PAGE_IDX_MASK)); 212 set_page_private(page, idx); 213 } 214 215 static unsigned long bm_page_to_idx(struct page *page) 216 { 217 return page_private(page) & BM_PAGE_IDX_MASK; 218 } 219 220 /* As is very unlikely that the same page is under IO from more than one 221 * context, we can get away with a bit per page and one wait queue per bitmap. 222 */ 223 static void bm_page_lock_io(struct drbd_device *device, int page_nr) 224 { 225 struct drbd_bitmap *b = device->bitmap; 226 void *addr = &page_private(b->bm_pages[page_nr]); 227 wait_event(b->bm_io_wait, !test_and_set_bit(BM_PAGE_IO_LOCK, addr)); 228 } 229 230 static void bm_page_unlock_io(struct drbd_device *device, int page_nr) 231 { 232 struct drbd_bitmap *b = device->bitmap; 233 void *addr = &page_private(b->bm_pages[page_nr]); 234 clear_bit_unlock(BM_PAGE_IO_LOCK, addr); 235 wake_up(&device->bitmap->bm_io_wait); 236 } 237 238 /* set _before_ submit_io, so it may be reset due to being changed 239 * while this page is in flight... will get submitted later again */ 240 static void bm_set_page_unchanged(struct page *page) 241 { 242 /* use cmpxchg? */ 243 clear_bit(BM_PAGE_NEED_WRITEOUT, &page_private(page)); 244 clear_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page)); 245 } 246 247 static void bm_set_page_need_writeout(struct page *page) 248 { 249 set_bit(BM_PAGE_NEED_WRITEOUT, &page_private(page)); 250 } 251 252 void drbd_bm_reset_al_hints(struct drbd_device *device) 253 { 254 device->bitmap->n_bitmap_hints = 0; 255 } 256 257 /** 258 * drbd_bm_mark_for_writeout() - mark a page with a "hint" to be considered for writeout 259 * @device: DRBD device. 260 * @page_nr: the bitmap page to mark with the "hint" flag 261 * 262 * From within an activity log transaction, we mark a few pages with these 263 * hints, then call drbd_bm_write_hinted(), which will only write out changed 264 * pages which are flagged with this mark. 265 */ 266 void drbd_bm_mark_for_writeout(struct drbd_device *device, int page_nr) 267 { 268 struct drbd_bitmap *b = device->bitmap; 269 struct page *page; 270 if (page_nr >= device->bitmap->bm_number_of_pages) { 271 drbd_warn(device, "BAD: page_nr: %u, number_of_pages: %u\n", 272 page_nr, (int)device->bitmap->bm_number_of_pages); 273 return; 274 } 275 page = device->bitmap->bm_pages[page_nr]; 276 BUG_ON(b->n_bitmap_hints >= ARRAY_SIZE(b->al_bitmap_hints)); 277 if (!test_and_set_bit(BM_PAGE_HINT_WRITEOUT, &page_private(page))) 278 b->al_bitmap_hints[b->n_bitmap_hints++] = page_nr; 279 } 280 281 static int bm_test_page_unchanged(struct page *page) 282 { 283 volatile const unsigned long *addr = &page_private(page); 284 return (*addr & ((1UL<<BM_PAGE_NEED_WRITEOUT)|(1UL<<BM_PAGE_LAZY_WRITEOUT))) == 0; 285 } 286 287 static void bm_set_page_io_err(struct page *page) 288 { 289 set_bit(BM_PAGE_IO_ERROR, &page_private(page)); 290 } 291 292 static void bm_clear_page_io_err(struct page *page) 293 { 294 clear_bit(BM_PAGE_IO_ERROR, &page_private(page)); 295 } 296 297 static void bm_set_page_lazy_writeout(struct page *page) 298 { 299 set_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page)); 300 } 301 302 static int bm_test_page_lazy_writeout(struct page *page) 303 { 304 return test_bit(BM_PAGE_LAZY_WRITEOUT, &page_private(page)); 305 } 306 307 /* on a 32bit box, this would allow for exactly (2<<38) bits. */ 308 static unsigned int bm_word_to_page_idx(struct drbd_bitmap *b, unsigned long long_nr) 309 { 310 /* page_nr = (word*sizeof(long)) >> PAGE_SHIFT; */ 311 unsigned int page_nr = long_nr >> (PAGE_SHIFT - LN2_BPL + 3); 312 BUG_ON(page_nr >= b->bm_number_of_pages); 313 return page_nr; 314 } 315 316 static unsigned int bm_bit_to_page_idx(struct drbd_bitmap *b, u64 bitnr) 317 { 318 /* page_nr = (bitnr/8) >> PAGE_SHIFT; */ 319 unsigned int page_nr = bitnr >> (PAGE_SHIFT + 3); 320 BUG_ON(page_nr >= b->bm_number_of_pages); 321 return page_nr; 322 } 323 324 static unsigned long *__bm_map_pidx(struct drbd_bitmap *b, unsigned int idx) 325 { 326 struct page *page = b->bm_pages[idx]; 327 return (unsigned long *) kmap_atomic(page); 328 } 329 330 static unsigned long *bm_map_pidx(struct drbd_bitmap *b, unsigned int idx) 331 { 332 return __bm_map_pidx(b, idx); 333 } 334 335 static void __bm_unmap(unsigned long *p_addr) 336 { 337 kunmap_atomic(p_addr); 338 }; 339 340 static void bm_unmap(unsigned long *p_addr) 341 { 342 return __bm_unmap(p_addr); 343 } 344 345 /* long word offset of _bitmap_ sector */ 346 #define S2W(s) ((s)<<(BM_EXT_SHIFT-BM_BLOCK_SHIFT-LN2_BPL)) 347 /* word offset from start of bitmap to word number _in_page_ 348 * modulo longs per page 349 #define MLPP(X) ((X) % (PAGE_SIZE/sizeof(long)) 350 hm, well, Philipp thinks gcc might not optimize the % into & (... - 1) 351 so do it explicitly: 352 */ 353 #define MLPP(X) ((X) & ((PAGE_SIZE/sizeof(long))-1)) 354 355 /* Long words per page */ 356 #define LWPP (PAGE_SIZE/sizeof(long)) 357 358 /* 359 * actually most functions herein should take a struct drbd_bitmap*, not a 360 * struct drbd_device*, but for the debug macros I like to have the device around 361 * to be able to report device specific. 362 */ 363 364 365 static void bm_free_pages(struct page **pages, unsigned long number) 366 { 367 unsigned long i; 368 if (!pages) 369 return; 370 371 for (i = 0; i < number; i++) { 372 if (!pages[i]) { 373 pr_alert("bm_free_pages tried to free a NULL pointer; i=%lu n=%lu\n", 374 i, number); 375 continue; 376 } 377 __free_page(pages[i]); 378 pages[i] = NULL; 379 } 380 } 381 382 static inline void bm_vk_free(void *ptr) 383 { 384 kvfree(ptr); 385 } 386 387 /* 388 * "have" and "want" are NUMBER OF PAGES. 389 */ 390 static struct page **bm_realloc_pages(struct drbd_bitmap *b, unsigned long want) 391 { 392 struct page **old_pages = b->bm_pages; 393 struct page **new_pages, *page; 394 unsigned int i, bytes; 395 unsigned long have = b->bm_number_of_pages; 396 397 BUG_ON(have == 0 && old_pages != NULL); 398 BUG_ON(have != 0 && old_pages == NULL); 399 400 if (have == want) 401 return old_pages; 402 403 /* Trying kmalloc first, falling back to vmalloc. 404 * GFP_NOIO, as this is called while drbd IO is "suspended", 405 * and during resize or attach on diskless Primary, 406 * we must not block on IO to ourselves. 407 * Context is receiver thread or dmsetup. */ 408 bytes = sizeof(struct page *)*want; 409 new_pages = kzalloc(bytes, GFP_NOIO | __GFP_NOWARN); 410 if (!new_pages) { 411 new_pages = __vmalloc(bytes, 412 GFP_NOIO | __GFP_HIGHMEM | __GFP_ZERO, 413 PAGE_KERNEL); 414 if (!new_pages) 415 return NULL; 416 } 417 418 if (want >= have) { 419 for (i = 0; i < have; i++) 420 new_pages[i] = old_pages[i]; 421 for (; i < want; i++) { 422 page = alloc_page(GFP_NOIO | __GFP_HIGHMEM); 423 if (!page) { 424 bm_free_pages(new_pages + have, i - have); 425 bm_vk_free(new_pages); 426 return NULL; 427 } 428 /* we want to know which page it is 429 * from the endio handlers */ 430 bm_store_page_idx(page, i); 431 new_pages[i] = page; 432 } 433 } else { 434 for (i = 0; i < want; i++) 435 new_pages[i] = old_pages[i]; 436 /* NOT HERE, we are outside the spinlock! 437 bm_free_pages(old_pages + want, have - want); 438 */ 439 } 440 441 return new_pages; 442 } 443 444 /* 445 * allocates the drbd_bitmap and stores it in device->bitmap. 446 */ 447 int drbd_bm_init(struct drbd_device *device) 448 { 449 struct drbd_bitmap *b = device->bitmap; 450 WARN_ON(b != NULL); 451 b = kzalloc(sizeof(struct drbd_bitmap), GFP_KERNEL); 452 if (!b) 453 return -ENOMEM; 454 spin_lock_init(&b->bm_lock); 455 mutex_init(&b->bm_change); 456 init_waitqueue_head(&b->bm_io_wait); 457 458 device->bitmap = b; 459 460 return 0; 461 } 462 463 sector_t drbd_bm_capacity(struct drbd_device *device) 464 { 465 if (!expect(device->bitmap)) 466 return 0; 467 return device->bitmap->bm_dev_capacity; 468 } 469 470 /* called on driver unload. TODO: call when a device is destroyed. 471 */ 472 void drbd_bm_cleanup(struct drbd_device *device) 473 { 474 if (!expect(device->bitmap)) 475 return; 476 bm_free_pages(device->bitmap->bm_pages, device->bitmap->bm_number_of_pages); 477 bm_vk_free(device->bitmap->bm_pages); 478 kfree(device->bitmap); 479 device->bitmap = NULL; 480 } 481 482 /* 483 * since (b->bm_bits % BITS_PER_LONG) != 0, 484 * this masks out the remaining bits. 485 * Returns the number of bits cleared. 486 */ 487 #ifndef BITS_PER_PAGE 488 #define BITS_PER_PAGE (1UL << (PAGE_SHIFT + 3)) 489 #define BITS_PER_PAGE_MASK (BITS_PER_PAGE - 1) 490 #else 491 # if BITS_PER_PAGE != (1UL << (PAGE_SHIFT + 3)) 492 # error "ambiguous BITS_PER_PAGE" 493 # endif 494 #endif 495 #define BITS_PER_LONG_MASK (BITS_PER_LONG - 1) 496 static int bm_clear_surplus(struct drbd_bitmap *b) 497 { 498 unsigned long mask; 499 unsigned long *p_addr, *bm; 500 int tmp; 501 int cleared = 0; 502 503 /* number of bits modulo bits per page */ 504 tmp = (b->bm_bits & BITS_PER_PAGE_MASK); 505 /* mask the used bits of the word containing the last bit */ 506 mask = (1UL << (tmp & BITS_PER_LONG_MASK)) -1; 507 /* bitmap is always stored little endian, 508 * on disk and in core memory alike */ 509 mask = cpu_to_lel(mask); 510 511 p_addr = bm_map_pidx(b, b->bm_number_of_pages - 1); 512 bm = p_addr + (tmp/BITS_PER_LONG); 513 if (mask) { 514 /* If mask != 0, we are not exactly aligned, so bm now points 515 * to the long containing the last bit. 516 * If mask == 0, bm already points to the word immediately 517 * after the last (long word aligned) bit. */ 518 cleared = hweight_long(*bm & ~mask); 519 *bm &= mask; 520 bm++; 521 } 522 523 if (BITS_PER_LONG == 32 && ((bm - p_addr) & 1) == 1) { 524 /* on a 32bit arch, we may need to zero out 525 * a padding long to align with a 64bit remote */ 526 cleared += hweight_long(*bm); 527 *bm = 0; 528 } 529 bm_unmap(p_addr); 530 return cleared; 531 } 532 533 static void bm_set_surplus(struct drbd_bitmap *b) 534 { 535 unsigned long mask; 536 unsigned long *p_addr, *bm; 537 int tmp; 538 539 /* number of bits modulo bits per page */ 540 tmp = (b->bm_bits & BITS_PER_PAGE_MASK); 541 /* mask the used bits of the word containing the last bit */ 542 mask = (1UL << (tmp & BITS_PER_LONG_MASK)) -1; 543 /* bitmap is always stored little endian, 544 * on disk and in core memory alike */ 545 mask = cpu_to_lel(mask); 546 547 p_addr = bm_map_pidx(b, b->bm_number_of_pages - 1); 548 bm = p_addr + (tmp/BITS_PER_LONG); 549 if (mask) { 550 /* If mask != 0, we are not exactly aligned, so bm now points 551 * to the long containing the last bit. 552 * If mask == 0, bm already points to the word immediately 553 * after the last (long word aligned) bit. */ 554 *bm |= ~mask; 555 bm++; 556 } 557 558 if (BITS_PER_LONG == 32 && ((bm - p_addr) & 1) == 1) { 559 /* on a 32bit arch, we may need to zero out 560 * a padding long to align with a 64bit remote */ 561 *bm = ~0UL; 562 } 563 bm_unmap(p_addr); 564 } 565 566 /* you better not modify the bitmap while this is running, 567 * or its results will be stale */ 568 static unsigned long bm_count_bits(struct drbd_bitmap *b) 569 { 570 unsigned long *p_addr; 571 unsigned long bits = 0; 572 unsigned long mask = (1UL << (b->bm_bits & BITS_PER_LONG_MASK)) -1; 573 int idx, last_word; 574 575 /* all but last page */ 576 for (idx = 0; idx < b->bm_number_of_pages - 1; idx++) { 577 p_addr = __bm_map_pidx(b, idx); 578 bits += bitmap_weight(p_addr, BITS_PER_PAGE); 579 __bm_unmap(p_addr); 580 cond_resched(); 581 } 582 /* last (or only) page */ 583 last_word = ((b->bm_bits - 1) & BITS_PER_PAGE_MASK) >> LN2_BPL; 584 p_addr = __bm_map_pidx(b, idx); 585 bits += bitmap_weight(p_addr, last_word * BITS_PER_LONG); 586 p_addr[last_word] &= cpu_to_lel(mask); 587 bits += hweight_long(p_addr[last_word]); 588 /* 32bit arch, may have an unused padding long */ 589 if (BITS_PER_LONG == 32 && (last_word & 1) == 0) 590 p_addr[last_word+1] = 0; 591 __bm_unmap(p_addr); 592 return bits; 593 } 594 595 /* offset and len in long words.*/ 596 static void bm_memset(struct drbd_bitmap *b, size_t offset, int c, size_t len) 597 { 598 unsigned long *p_addr, *bm; 599 unsigned int idx; 600 size_t do_now, end; 601 602 end = offset + len; 603 604 if (end > b->bm_words) { 605 pr_alert("bm_memset end > bm_words\n"); 606 return; 607 } 608 609 while (offset < end) { 610 do_now = min_t(size_t, ALIGN(offset + 1, LWPP), end) - offset; 611 idx = bm_word_to_page_idx(b, offset); 612 p_addr = bm_map_pidx(b, idx); 613 bm = p_addr + MLPP(offset); 614 if (bm+do_now > p_addr + LWPP) { 615 pr_alert("BUG BUG BUG! p_addr:%p bm:%p do_now:%d\n", 616 p_addr, bm, (int)do_now); 617 } else 618 memset(bm, c, do_now * sizeof(long)); 619 bm_unmap(p_addr); 620 bm_set_page_need_writeout(b->bm_pages[idx]); 621 offset += do_now; 622 } 623 } 624 625 /* For the layout, see comment above drbd_md_set_sector_offsets(). */ 626 static u64 drbd_md_on_disk_bits(struct drbd_backing_dev *ldev) 627 { 628 u64 bitmap_sectors; 629 if (ldev->md.al_offset == 8) 630 bitmap_sectors = ldev->md.md_size_sect - ldev->md.bm_offset; 631 else 632 bitmap_sectors = ldev->md.al_offset - ldev->md.bm_offset; 633 return bitmap_sectors << (9 + 3); 634 } 635 636 /* 637 * make sure the bitmap has enough room for the attached storage, 638 * if necessary, resize. 639 * called whenever we may have changed the device size. 640 * returns -ENOMEM if we could not allocate enough memory, 0 on success. 641 * In case this is actually a resize, we copy the old bitmap into the new one. 642 * Otherwise, the bitmap is initialized to all bits set. 643 */ 644 int drbd_bm_resize(struct drbd_device *device, sector_t capacity, int set_new_bits) 645 { 646 struct drbd_bitmap *b = device->bitmap; 647 unsigned long bits, words, owords, obits; 648 unsigned long want, have, onpages; /* number of pages */ 649 struct page **npages, **opages = NULL; 650 int err = 0; 651 bool growing; 652 653 if (!expect(b)) 654 return -ENOMEM; 655 656 drbd_bm_lock(device, "resize", BM_LOCKED_MASK); 657 658 drbd_info(device, "drbd_bm_resize called with capacity == %llu\n", 659 (unsigned long long)capacity); 660 661 if (capacity == b->bm_dev_capacity) 662 goto out; 663 664 if (capacity == 0) { 665 spin_lock_irq(&b->bm_lock); 666 opages = b->bm_pages; 667 onpages = b->bm_number_of_pages; 668 owords = b->bm_words; 669 b->bm_pages = NULL; 670 b->bm_number_of_pages = 671 b->bm_set = 672 b->bm_bits = 673 b->bm_words = 674 b->bm_dev_capacity = 0; 675 spin_unlock_irq(&b->bm_lock); 676 bm_free_pages(opages, onpages); 677 bm_vk_free(opages); 678 goto out; 679 } 680 bits = BM_SECT_TO_BIT(ALIGN(capacity, BM_SECT_PER_BIT)); 681 682 /* if we would use 683 words = ALIGN(bits,BITS_PER_LONG) >> LN2_BPL; 684 a 32bit host could present the wrong number of words 685 to a 64bit host. 686 */ 687 words = ALIGN(bits, 64) >> LN2_BPL; 688 689 if (get_ldev(device)) { 690 u64 bits_on_disk = drbd_md_on_disk_bits(device->ldev); 691 put_ldev(device); 692 if (bits > bits_on_disk) { 693 drbd_info(device, "bits = %lu\n", bits); 694 drbd_info(device, "bits_on_disk = %llu\n", bits_on_disk); 695 err = -ENOSPC; 696 goto out; 697 } 698 } 699 700 want = ALIGN(words*sizeof(long), PAGE_SIZE) >> PAGE_SHIFT; 701 have = b->bm_number_of_pages; 702 if (want == have) { 703 D_ASSERT(device, b->bm_pages != NULL); 704 npages = b->bm_pages; 705 } else { 706 if (drbd_insert_fault(device, DRBD_FAULT_BM_ALLOC)) 707 npages = NULL; 708 else 709 npages = bm_realloc_pages(b, want); 710 } 711 712 if (!npages) { 713 err = -ENOMEM; 714 goto out; 715 } 716 717 spin_lock_irq(&b->bm_lock); 718 opages = b->bm_pages; 719 owords = b->bm_words; 720 obits = b->bm_bits; 721 722 growing = bits > obits; 723 if (opages && growing && set_new_bits) 724 bm_set_surplus(b); 725 726 b->bm_pages = npages; 727 b->bm_number_of_pages = want; 728 b->bm_bits = bits; 729 b->bm_words = words; 730 b->bm_dev_capacity = capacity; 731 732 if (growing) { 733 if (set_new_bits) { 734 bm_memset(b, owords, 0xff, words-owords); 735 b->bm_set += bits - obits; 736 } else 737 bm_memset(b, owords, 0x00, words-owords); 738 739 } 740 741 if (want < have) { 742 /* implicit: (opages != NULL) && (opages != npages) */ 743 bm_free_pages(opages + want, have - want); 744 } 745 746 (void)bm_clear_surplus(b); 747 748 spin_unlock_irq(&b->bm_lock); 749 if (opages != npages) 750 bm_vk_free(opages); 751 if (!growing) 752 b->bm_set = bm_count_bits(b); 753 drbd_info(device, "resync bitmap: bits=%lu words=%lu pages=%lu\n", bits, words, want); 754 755 out: 756 drbd_bm_unlock(device); 757 return err; 758 } 759 760 /* inherently racy: 761 * if not protected by other means, return value may be out of date when 762 * leaving this function... 763 * we still need to lock it, since it is important that this returns 764 * bm_set == 0 precisely. 765 * 766 * maybe bm_set should be atomic_t ? 767 */ 768 unsigned long _drbd_bm_total_weight(struct drbd_device *device) 769 { 770 struct drbd_bitmap *b = device->bitmap; 771 unsigned long s; 772 unsigned long flags; 773 774 if (!expect(b)) 775 return 0; 776 if (!expect(b->bm_pages)) 777 return 0; 778 779 spin_lock_irqsave(&b->bm_lock, flags); 780 s = b->bm_set; 781 spin_unlock_irqrestore(&b->bm_lock, flags); 782 783 return s; 784 } 785 786 unsigned long drbd_bm_total_weight(struct drbd_device *device) 787 { 788 unsigned long s; 789 /* if I don't have a disk, I don't know about out-of-sync status */ 790 if (!get_ldev_if_state(device, D_NEGOTIATING)) 791 return 0; 792 s = _drbd_bm_total_weight(device); 793 put_ldev(device); 794 return s; 795 } 796 797 size_t drbd_bm_words(struct drbd_device *device) 798 { 799 struct drbd_bitmap *b = device->bitmap; 800 if (!expect(b)) 801 return 0; 802 if (!expect(b->bm_pages)) 803 return 0; 804 805 return b->bm_words; 806 } 807 808 unsigned long drbd_bm_bits(struct drbd_device *device) 809 { 810 struct drbd_bitmap *b = device->bitmap; 811 if (!expect(b)) 812 return 0; 813 814 return b->bm_bits; 815 } 816 817 /* merge number words from buffer into the bitmap starting at offset. 818 * buffer[i] is expected to be little endian unsigned long. 819 * bitmap must be locked by drbd_bm_lock. 820 * currently only used from receive_bitmap. 821 */ 822 void drbd_bm_merge_lel(struct drbd_device *device, size_t offset, size_t number, 823 unsigned long *buffer) 824 { 825 struct drbd_bitmap *b = device->bitmap; 826 unsigned long *p_addr, *bm; 827 unsigned long word, bits; 828 unsigned int idx; 829 size_t end, do_now; 830 831 end = offset + number; 832 833 if (!expect(b)) 834 return; 835 if (!expect(b->bm_pages)) 836 return; 837 if (number == 0) 838 return; 839 WARN_ON(offset >= b->bm_words); 840 WARN_ON(end > b->bm_words); 841 842 spin_lock_irq(&b->bm_lock); 843 while (offset < end) { 844 do_now = min_t(size_t, ALIGN(offset+1, LWPP), end) - offset; 845 idx = bm_word_to_page_idx(b, offset); 846 p_addr = bm_map_pidx(b, idx); 847 bm = p_addr + MLPP(offset); 848 offset += do_now; 849 while (do_now--) { 850 bits = hweight_long(*bm); 851 word = *bm | *buffer++; 852 *bm++ = word; 853 b->bm_set += hweight_long(word) - bits; 854 } 855 bm_unmap(p_addr); 856 bm_set_page_need_writeout(b->bm_pages[idx]); 857 } 858 /* with 32bit <-> 64bit cross-platform connect 859 * this is only correct for current usage, 860 * where we _know_ that we are 64 bit aligned, 861 * and know that this function is used in this way, too... 862 */ 863 if (end == b->bm_words) 864 b->bm_set -= bm_clear_surplus(b); 865 spin_unlock_irq(&b->bm_lock); 866 } 867 868 /* copy number words from the bitmap starting at offset into the buffer. 869 * buffer[i] will be little endian unsigned long. 870 */ 871 void drbd_bm_get_lel(struct drbd_device *device, size_t offset, size_t number, 872 unsigned long *buffer) 873 { 874 struct drbd_bitmap *b = device->bitmap; 875 unsigned long *p_addr, *bm; 876 size_t end, do_now; 877 878 end = offset + number; 879 880 if (!expect(b)) 881 return; 882 if (!expect(b->bm_pages)) 883 return; 884 885 spin_lock_irq(&b->bm_lock); 886 if ((offset >= b->bm_words) || 887 (end > b->bm_words) || 888 (number <= 0)) 889 drbd_err(device, "offset=%lu number=%lu bm_words=%lu\n", 890 (unsigned long) offset, 891 (unsigned long) number, 892 (unsigned long) b->bm_words); 893 else { 894 while (offset < end) { 895 do_now = min_t(size_t, ALIGN(offset+1, LWPP), end) - offset; 896 p_addr = bm_map_pidx(b, bm_word_to_page_idx(b, offset)); 897 bm = p_addr + MLPP(offset); 898 offset += do_now; 899 while (do_now--) 900 *buffer++ = *bm++; 901 bm_unmap(p_addr); 902 } 903 } 904 spin_unlock_irq(&b->bm_lock); 905 } 906 907 /* set all bits in the bitmap */ 908 void drbd_bm_set_all(struct drbd_device *device) 909 { 910 struct drbd_bitmap *b = device->bitmap; 911 if (!expect(b)) 912 return; 913 if (!expect(b->bm_pages)) 914 return; 915 916 spin_lock_irq(&b->bm_lock); 917 bm_memset(b, 0, 0xff, b->bm_words); 918 (void)bm_clear_surplus(b); 919 b->bm_set = b->bm_bits; 920 spin_unlock_irq(&b->bm_lock); 921 } 922 923 /* clear all bits in the bitmap */ 924 void drbd_bm_clear_all(struct drbd_device *device) 925 { 926 struct drbd_bitmap *b = device->bitmap; 927 if (!expect(b)) 928 return; 929 if (!expect(b->bm_pages)) 930 return; 931 932 spin_lock_irq(&b->bm_lock); 933 bm_memset(b, 0, 0, b->bm_words); 934 b->bm_set = 0; 935 spin_unlock_irq(&b->bm_lock); 936 } 937 938 static void drbd_bm_aio_ctx_destroy(struct kref *kref) 939 { 940 struct drbd_bm_aio_ctx *ctx = container_of(kref, struct drbd_bm_aio_ctx, kref); 941 unsigned long flags; 942 943 spin_lock_irqsave(&ctx->device->resource->req_lock, flags); 944 list_del(&ctx->list); 945 spin_unlock_irqrestore(&ctx->device->resource->req_lock, flags); 946 put_ldev(ctx->device); 947 kfree(ctx); 948 } 949 950 /* bv_page may be a copy, or may be the original */ 951 static void drbd_bm_endio(struct bio *bio) 952 { 953 struct drbd_bm_aio_ctx *ctx = bio->bi_private; 954 struct drbd_device *device = ctx->device; 955 struct drbd_bitmap *b = device->bitmap; 956 unsigned int idx = bm_page_to_idx(bio->bi_io_vec[0].bv_page); 957 958 if ((ctx->flags & BM_AIO_COPY_PAGES) == 0 && 959 !bm_test_page_unchanged(b->bm_pages[idx])) 960 drbd_warn(device, "bitmap page idx %u changed during IO!\n", idx); 961 962 if (bio->bi_error) { 963 /* ctx error will hold the completed-last non-zero error code, 964 * in case error codes differ. */ 965 ctx->error = bio->bi_error; 966 bm_set_page_io_err(b->bm_pages[idx]); 967 /* Not identical to on disk version of it. 968 * Is BM_PAGE_IO_ERROR enough? */ 969 if (__ratelimit(&drbd_ratelimit_state)) 970 drbd_err(device, "IO ERROR %d on bitmap page idx %u\n", 971 bio->bi_error, idx); 972 } else { 973 bm_clear_page_io_err(b->bm_pages[idx]); 974 dynamic_drbd_dbg(device, "bitmap page idx %u completed\n", idx); 975 } 976 977 bm_page_unlock_io(device, idx); 978 979 if (ctx->flags & BM_AIO_COPY_PAGES) 980 mempool_free(bio->bi_io_vec[0].bv_page, drbd_md_io_page_pool); 981 982 bio_put(bio); 983 984 if (atomic_dec_and_test(&ctx->in_flight)) { 985 ctx->done = 1; 986 wake_up(&device->misc_wait); 987 kref_put(&ctx->kref, &drbd_bm_aio_ctx_destroy); 988 } 989 } 990 991 static void bm_page_io_async(struct drbd_bm_aio_ctx *ctx, int page_nr) __must_hold(local) 992 { 993 struct bio *bio = bio_alloc_drbd(GFP_NOIO); 994 struct drbd_device *device = ctx->device; 995 struct drbd_bitmap *b = device->bitmap; 996 struct page *page; 997 unsigned int len; 998 unsigned int op = (ctx->flags & BM_AIO_READ) ? REQ_OP_READ : REQ_OP_WRITE; 999 1000 sector_t on_disk_sector = 1001 device->ldev->md.md_offset + device->ldev->md.bm_offset; 1002 on_disk_sector += ((sector_t)page_nr) << (PAGE_SHIFT-9); 1003 1004 /* this might happen with very small 1005 * flexible external meta data device, 1006 * or with PAGE_SIZE > 4k */ 1007 len = min_t(unsigned int, PAGE_SIZE, 1008 (drbd_md_last_sector(device->ldev) - on_disk_sector + 1)<<9); 1009 1010 /* serialize IO on this page */ 1011 bm_page_lock_io(device, page_nr); 1012 /* before memcpy and submit, 1013 * so it can be redirtied any time */ 1014 bm_set_page_unchanged(b->bm_pages[page_nr]); 1015 1016 if (ctx->flags & BM_AIO_COPY_PAGES) { 1017 page = mempool_alloc(drbd_md_io_page_pool, __GFP_HIGHMEM|__GFP_RECLAIM); 1018 copy_highpage(page, b->bm_pages[page_nr]); 1019 bm_store_page_idx(page, page_nr); 1020 } else 1021 page = b->bm_pages[page_nr]; 1022 bio->bi_bdev = device->ldev->md_bdev; 1023 bio->bi_iter.bi_sector = on_disk_sector; 1024 /* bio_add_page of a single page to an empty bio will always succeed, 1025 * according to api. Do we want to assert that? */ 1026 bio_add_page(bio, page, len, 0); 1027 bio->bi_private = ctx; 1028 bio->bi_end_io = drbd_bm_endio; 1029 bio_set_op_attrs(bio, op, 0); 1030 1031 if (drbd_insert_fault(device, (op == REQ_OP_WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD)) { 1032 bio_io_error(bio); 1033 } else { 1034 submit_bio(bio); 1035 /* this should not count as user activity and cause the 1036 * resync to throttle -- see drbd_rs_should_slow_down(). */ 1037 atomic_add(len >> 9, &device->rs_sect_ev); 1038 } 1039 } 1040 1041 /* 1042 * bm_rw: read/write the whole bitmap from/to its on disk location. 1043 */ 1044 static int bm_rw(struct drbd_device *device, const unsigned int flags, unsigned lazy_writeout_upper_idx) __must_hold(local) 1045 { 1046 struct drbd_bm_aio_ctx *ctx; 1047 struct drbd_bitmap *b = device->bitmap; 1048 unsigned int num_pages, i, count = 0; 1049 unsigned long now; 1050 char ppb[10]; 1051 int err = 0; 1052 1053 /* 1054 * We are protected against bitmap disappearing/resizing by holding an 1055 * ldev reference (caller must have called get_ldev()). 1056 * For read/write, we are protected against changes to the bitmap by 1057 * the bitmap lock (see drbd_bitmap_io). 1058 * For lazy writeout, we don't care for ongoing changes to the bitmap, 1059 * as we submit copies of pages anyways. 1060 */ 1061 1062 ctx = kmalloc(sizeof(struct drbd_bm_aio_ctx), GFP_NOIO); 1063 if (!ctx) 1064 return -ENOMEM; 1065 1066 *ctx = (struct drbd_bm_aio_ctx) { 1067 .device = device, 1068 .start_jif = jiffies, 1069 .in_flight = ATOMIC_INIT(1), 1070 .done = 0, 1071 .flags = flags, 1072 .error = 0, 1073 .kref = { ATOMIC_INIT(2) }, 1074 }; 1075 1076 if (!get_ldev_if_state(device, D_ATTACHING)) { /* put is in drbd_bm_aio_ctx_destroy() */ 1077 drbd_err(device, "ASSERT FAILED: get_ldev_if_state() == 1 in bm_rw()\n"); 1078 kfree(ctx); 1079 return -ENODEV; 1080 } 1081 /* Here D_ATTACHING is sufficient since drbd_bm_read() is called only from 1082 drbd_adm_attach(), after device->ldev was assigned. */ 1083 1084 if (0 == (ctx->flags & ~BM_AIO_READ)) 1085 WARN_ON(!(BM_LOCKED_MASK & b->bm_flags)); 1086 1087 spin_lock_irq(&device->resource->req_lock); 1088 list_add_tail(&ctx->list, &device->pending_bitmap_io); 1089 spin_unlock_irq(&device->resource->req_lock); 1090 1091 num_pages = b->bm_number_of_pages; 1092 1093 now = jiffies; 1094 1095 /* let the layers below us try to merge these bios... */ 1096 1097 if (flags & BM_AIO_READ) { 1098 for (i = 0; i < num_pages; i++) { 1099 atomic_inc(&ctx->in_flight); 1100 bm_page_io_async(ctx, i); 1101 ++count; 1102 cond_resched(); 1103 } 1104 } else if (flags & BM_AIO_WRITE_HINTED) { 1105 /* ASSERT: BM_AIO_WRITE_ALL_PAGES is not set. */ 1106 unsigned int hint; 1107 for (hint = 0; hint < b->n_bitmap_hints; hint++) { 1108 i = b->al_bitmap_hints[hint]; 1109 if (i >= num_pages) /* == -1U: no hint here. */ 1110 continue; 1111 /* Several AL-extents may point to the same page. */ 1112 if (!test_and_clear_bit(BM_PAGE_HINT_WRITEOUT, 1113 &page_private(b->bm_pages[i]))) 1114 continue; 1115 /* Has it even changed? */ 1116 if (bm_test_page_unchanged(b->bm_pages[i])) 1117 continue; 1118 atomic_inc(&ctx->in_flight); 1119 bm_page_io_async(ctx, i); 1120 ++count; 1121 } 1122 } else { 1123 for (i = 0; i < num_pages; i++) { 1124 /* ignore completely unchanged pages */ 1125 if (lazy_writeout_upper_idx && i == lazy_writeout_upper_idx) 1126 break; 1127 if (!(flags & BM_AIO_WRITE_ALL_PAGES) && 1128 bm_test_page_unchanged(b->bm_pages[i])) { 1129 dynamic_drbd_dbg(device, "skipped bm write for idx %u\n", i); 1130 continue; 1131 } 1132 /* during lazy writeout, 1133 * ignore those pages not marked for lazy writeout. */ 1134 if (lazy_writeout_upper_idx && 1135 !bm_test_page_lazy_writeout(b->bm_pages[i])) { 1136 dynamic_drbd_dbg(device, "skipped bm lazy write for idx %u\n", i); 1137 continue; 1138 } 1139 atomic_inc(&ctx->in_flight); 1140 bm_page_io_async(ctx, i); 1141 ++count; 1142 cond_resched(); 1143 } 1144 } 1145 1146 /* 1147 * We initialize ctx->in_flight to one to make sure drbd_bm_endio 1148 * will not set ctx->done early, and decrement / test it here. If there 1149 * are still some bios in flight, we need to wait for them here. 1150 * If all IO is done already (or nothing had been submitted), there is 1151 * no need to wait. Still, we need to put the kref associated with the 1152 * "in_flight reached zero, all done" event. 1153 */ 1154 if (!atomic_dec_and_test(&ctx->in_flight)) 1155 wait_until_done_or_force_detached(device, device->ldev, &ctx->done); 1156 else 1157 kref_put(&ctx->kref, &drbd_bm_aio_ctx_destroy); 1158 1159 /* summary for global bitmap IO */ 1160 if (flags == 0) { 1161 unsigned int ms = jiffies_to_msecs(jiffies - now); 1162 if (ms > 5) { 1163 drbd_info(device, "bitmap %s of %u pages took %u ms\n", 1164 (flags & BM_AIO_READ) ? "READ" : "WRITE", 1165 count, ms); 1166 } 1167 } 1168 1169 if (ctx->error) { 1170 drbd_alert(device, "we had at least one MD IO ERROR during bitmap IO\n"); 1171 drbd_chk_io_error(device, 1, DRBD_META_IO_ERROR); 1172 err = -EIO; /* ctx->error ? */ 1173 } 1174 1175 if (atomic_read(&ctx->in_flight)) 1176 err = -EIO; /* Disk timeout/force-detach during IO... */ 1177 1178 now = jiffies; 1179 if (flags & BM_AIO_READ) { 1180 b->bm_set = bm_count_bits(b); 1181 drbd_info(device, "recounting of set bits took additional %lu jiffies\n", 1182 jiffies - now); 1183 } 1184 now = b->bm_set; 1185 1186 if ((flags & ~BM_AIO_READ) == 0) 1187 drbd_info(device, "%s (%lu bits) marked out-of-sync by on disk bit-map.\n", 1188 ppsize(ppb, now << (BM_BLOCK_SHIFT-10)), now); 1189 1190 kref_put(&ctx->kref, &drbd_bm_aio_ctx_destroy); 1191 return err; 1192 } 1193 1194 /** 1195 * drbd_bm_read() - Read the whole bitmap from its on disk location. 1196 * @device: DRBD device. 1197 */ 1198 int drbd_bm_read(struct drbd_device *device) __must_hold(local) 1199 { 1200 return bm_rw(device, BM_AIO_READ, 0); 1201 } 1202 1203 /** 1204 * drbd_bm_write() - Write the whole bitmap to its on disk location. 1205 * @device: DRBD device. 1206 * 1207 * Will only write pages that have changed since last IO. 1208 */ 1209 int drbd_bm_write(struct drbd_device *device) __must_hold(local) 1210 { 1211 return bm_rw(device, 0, 0); 1212 } 1213 1214 /** 1215 * drbd_bm_write_all() - Write the whole bitmap to its on disk location. 1216 * @device: DRBD device. 1217 * 1218 * Will write all pages. 1219 */ 1220 int drbd_bm_write_all(struct drbd_device *device) __must_hold(local) 1221 { 1222 return bm_rw(device, BM_AIO_WRITE_ALL_PAGES, 0); 1223 } 1224 1225 /** 1226 * drbd_bm_write_lazy() - Write bitmap pages 0 to @upper_idx-1, if they have changed. 1227 * @device: DRBD device. 1228 * @upper_idx: 0: write all changed pages; +ve: page index to stop scanning for changed pages 1229 */ 1230 int drbd_bm_write_lazy(struct drbd_device *device, unsigned upper_idx) __must_hold(local) 1231 { 1232 return bm_rw(device, BM_AIO_COPY_PAGES, upper_idx); 1233 } 1234 1235 /** 1236 * drbd_bm_write_copy_pages() - Write the whole bitmap to its on disk location. 1237 * @device: DRBD device. 1238 * 1239 * Will only write pages that have changed since last IO. 1240 * In contrast to drbd_bm_write(), this will copy the bitmap pages 1241 * to temporary writeout pages. It is intended to trigger a full write-out 1242 * while still allowing the bitmap to change, for example if a resync or online 1243 * verify is aborted due to a failed peer disk, while local IO continues, or 1244 * pending resync acks are still being processed. 1245 */ 1246 int drbd_bm_write_copy_pages(struct drbd_device *device) __must_hold(local) 1247 { 1248 return bm_rw(device, BM_AIO_COPY_PAGES, 0); 1249 } 1250 1251 /** 1252 * drbd_bm_write_hinted() - Write bitmap pages with "hint" marks, if they have changed. 1253 * @device: DRBD device. 1254 */ 1255 int drbd_bm_write_hinted(struct drbd_device *device) __must_hold(local) 1256 { 1257 return bm_rw(device, BM_AIO_WRITE_HINTED | BM_AIO_COPY_PAGES, 0); 1258 } 1259 1260 /* NOTE 1261 * find_first_bit returns int, we return unsigned long. 1262 * For this to work on 32bit arch with bitnumbers > (1<<32), 1263 * we'd need to return u64, and get a whole lot of other places 1264 * fixed where we still use unsigned long. 1265 * 1266 * this returns a bit number, NOT a sector! 1267 */ 1268 static unsigned long __bm_find_next(struct drbd_device *device, unsigned long bm_fo, 1269 const int find_zero_bit) 1270 { 1271 struct drbd_bitmap *b = device->bitmap; 1272 unsigned long *p_addr; 1273 unsigned long bit_offset; 1274 unsigned i; 1275 1276 1277 if (bm_fo > b->bm_bits) { 1278 drbd_err(device, "bm_fo=%lu bm_bits=%lu\n", bm_fo, b->bm_bits); 1279 bm_fo = DRBD_END_OF_BITMAP; 1280 } else { 1281 while (bm_fo < b->bm_bits) { 1282 /* bit offset of the first bit in the page */ 1283 bit_offset = bm_fo & ~BITS_PER_PAGE_MASK; 1284 p_addr = __bm_map_pidx(b, bm_bit_to_page_idx(b, bm_fo)); 1285 1286 if (find_zero_bit) 1287 i = find_next_zero_bit_le(p_addr, 1288 PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK); 1289 else 1290 i = find_next_bit_le(p_addr, 1291 PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK); 1292 1293 __bm_unmap(p_addr); 1294 if (i < PAGE_SIZE*8) { 1295 bm_fo = bit_offset + i; 1296 if (bm_fo >= b->bm_bits) 1297 break; 1298 goto found; 1299 } 1300 bm_fo = bit_offset + PAGE_SIZE*8; 1301 } 1302 bm_fo = DRBD_END_OF_BITMAP; 1303 } 1304 found: 1305 return bm_fo; 1306 } 1307 1308 static unsigned long bm_find_next(struct drbd_device *device, 1309 unsigned long bm_fo, const int find_zero_bit) 1310 { 1311 struct drbd_bitmap *b = device->bitmap; 1312 unsigned long i = DRBD_END_OF_BITMAP; 1313 1314 if (!expect(b)) 1315 return i; 1316 if (!expect(b->bm_pages)) 1317 return i; 1318 1319 spin_lock_irq(&b->bm_lock); 1320 if (BM_DONT_TEST & b->bm_flags) 1321 bm_print_lock_info(device); 1322 1323 i = __bm_find_next(device, bm_fo, find_zero_bit); 1324 1325 spin_unlock_irq(&b->bm_lock); 1326 return i; 1327 } 1328 1329 unsigned long drbd_bm_find_next(struct drbd_device *device, unsigned long bm_fo) 1330 { 1331 return bm_find_next(device, bm_fo, 0); 1332 } 1333 1334 #if 0 1335 /* not yet needed for anything. */ 1336 unsigned long drbd_bm_find_next_zero(struct drbd_device *device, unsigned long bm_fo) 1337 { 1338 return bm_find_next(device, bm_fo, 1); 1339 } 1340 #endif 1341 1342 /* does not spin_lock_irqsave. 1343 * you must take drbd_bm_lock() first */ 1344 unsigned long _drbd_bm_find_next(struct drbd_device *device, unsigned long bm_fo) 1345 { 1346 /* WARN_ON(!(BM_DONT_SET & device->b->bm_flags)); */ 1347 return __bm_find_next(device, bm_fo, 0); 1348 } 1349 1350 unsigned long _drbd_bm_find_next_zero(struct drbd_device *device, unsigned long bm_fo) 1351 { 1352 /* WARN_ON(!(BM_DONT_SET & device->b->bm_flags)); */ 1353 return __bm_find_next(device, bm_fo, 1); 1354 } 1355 1356 /* returns number of bits actually changed. 1357 * for val != 0, we change 0 -> 1, return code positive 1358 * for val == 0, we change 1 -> 0, return code negative 1359 * wants bitnr, not sector. 1360 * expected to be called for only a few bits (e - s about BITS_PER_LONG). 1361 * Must hold bitmap lock already. */ 1362 static int __bm_change_bits_to(struct drbd_device *device, const unsigned long s, 1363 unsigned long e, int val) 1364 { 1365 struct drbd_bitmap *b = device->bitmap; 1366 unsigned long *p_addr = NULL; 1367 unsigned long bitnr; 1368 unsigned int last_page_nr = -1U; 1369 int c = 0; 1370 int changed_total = 0; 1371 1372 if (e >= b->bm_bits) { 1373 drbd_err(device, "ASSERT FAILED: bit_s=%lu bit_e=%lu bm_bits=%lu\n", 1374 s, e, b->bm_bits); 1375 e = b->bm_bits ? b->bm_bits -1 : 0; 1376 } 1377 for (bitnr = s; bitnr <= e; bitnr++) { 1378 unsigned int page_nr = bm_bit_to_page_idx(b, bitnr); 1379 if (page_nr != last_page_nr) { 1380 if (p_addr) 1381 __bm_unmap(p_addr); 1382 if (c < 0) 1383 bm_set_page_lazy_writeout(b->bm_pages[last_page_nr]); 1384 else if (c > 0) 1385 bm_set_page_need_writeout(b->bm_pages[last_page_nr]); 1386 changed_total += c; 1387 c = 0; 1388 p_addr = __bm_map_pidx(b, page_nr); 1389 last_page_nr = page_nr; 1390 } 1391 if (val) 1392 c += (0 == __test_and_set_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr)); 1393 else 1394 c -= (0 != __test_and_clear_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr)); 1395 } 1396 if (p_addr) 1397 __bm_unmap(p_addr); 1398 if (c < 0) 1399 bm_set_page_lazy_writeout(b->bm_pages[last_page_nr]); 1400 else if (c > 0) 1401 bm_set_page_need_writeout(b->bm_pages[last_page_nr]); 1402 changed_total += c; 1403 b->bm_set += changed_total; 1404 return changed_total; 1405 } 1406 1407 /* returns number of bits actually changed. 1408 * for val != 0, we change 0 -> 1, return code positive 1409 * for val == 0, we change 1 -> 0, return code negative 1410 * wants bitnr, not sector */ 1411 static int bm_change_bits_to(struct drbd_device *device, const unsigned long s, 1412 const unsigned long e, int val) 1413 { 1414 unsigned long flags; 1415 struct drbd_bitmap *b = device->bitmap; 1416 int c = 0; 1417 1418 if (!expect(b)) 1419 return 1; 1420 if (!expect(b->bm_pages)) 1421 return 0; 1422 1423 spin_lock_irqsave(&b->bm_lock, flags); 1424 if ((val ? BM_DONT_SET : BM_DONT_CLEAR) & b->bm_flags) 1425 bm_print_lock_info(device); 1426 1427 c = __bm_change_bits_to(device, s, e, val); 1428 1429 spin_unlock_irqrestore(&b->bm_lock, flags); 1430 return c; 1431 } 1432 1433 /* returns number of bits changed 0 -> 1 */ 1434 int drbd_bm_set_bits(struct drbd_device *device, const unsigned long s, const unsigned long e) 1435 { 1436 return bm_change_bits_to(device, s, e, 1); 1437 } 1438 1439 /* returns number of bits changed 1 -> 0 */ 1440 int drbd_bm_clear_bits(struct drbd_device *device, const unsigned long s, const unsigned long e) 1441 { 1442 return -bm_change_bits_to(device, s, e, 0); 1443 } 1444 1445 /* sets all bits in full words, 1446 * from first_word up to, but not including, last_word */ 1447 static inline void bm_set_full_words_within_one_page(struct drbd_bitmap *b, 1448 int page_nr, int first_word, int last_word) 1449 { 1450 int i; 1451 int bits; 1452 int changed = 0; 1453 unsigned long *paddr = kmap_atomic(b->bm_pages[page_nr]); 1454 1455 /* I think it is more cache line friendly to hweight_long then set to ~0UL, 1456 * than to first bitmap_weight() all words, then bitmap_fill() all words */ 1457 for (i = first_word; i < last_word; i++) { 1458 bits = hweight_long(paddr[i]); 1459 paddr[i] = ~0UL; 1460 changed += BITS_PER_LONG - bits; 1461 } 1462 kunmap_atomic(paddr); 1463 if (changed) { 1464 /* We only need lazy writeout, the information is still in the 1465 * remote bitmap as well, and is reconstructed during the next 1466 * bitmap exchange, if lost locally due to a crash. */ 1467 bm_set_page_lazy_writeout(b->bm_pages[page_nr]); 1468 b->bm_set += changed; 1469 } 1470 } 1471 1472 /* Same thing as drbd_bm_set_bits, 1473 * but more efficient for a large bit range. 1474 * You must first drbd_bm_lock(). 1475 * Can be called to set the whole bitmap in one go. 1476 * Sets bits from s to e _inclusive_. */ 1477 void _drbd_bm_set_bits(struct drbd_device *device, const unsigned long s, const unsigned long e) 1478 { 1479 /* First set_bit from the first bit (s) 1480 * up to the next long boundary (sl), 1481 * then assign full words up to the last long boundary (el), 1482 * then set_bit up to and including the last bit (e). 1483 * 1484 * Do not use memset, because we must account for changes, 1485 * so we need to loop over the words with hweight() anyways. 1486 */ 1487 struct drbd_bitmap *b = device->bitmap; 1488 unsigned long sl = ALIGN(s,BITS_PER_LONG); 1489 unsigned long el = (e+1) & ~((unsigned long)BITS_PER_LONG-1); 1490 int first_page; 1491 int last_page; 1492 int page_nr; 1493 int first_word; 1494 int last_word; 1495 1496 if (e - s <= 3*BITS_PER_LONG) { 1497 /* don't bother; el and sl may even be wrong. */ 1498 spin_lock_irq(&b->bm_lock); 1499 __bm_change_bits_to(device, s, e, 1); 1500 spin_unlock_irq(&b->bm_lock); 1501 return; 1502 } 1503 1504 /* difference is large enough that we can trust sl and el */ 1505 1506 spin_lock_irq(&b->bm_lock); 1507 1508 /* bits filling the current long */ 1509 if (sl) 1510 __bm_change_bits_to(device, s, sl-1, 1); 1511 1512 first_page = sl >> (3 + PAGE_SHIFT); 1513 last_page = el >> (3 + PAGE_SHIFT); 1514 1515 /* MLPP: modulo longs per page */ 1516 /* LWPP: long words per page */ 1517 first_word = MLPP(sl >> LN2_BPL); 1518 last_word = LWPP; 1519 1520 /* first and full pages, unless first page == last page */ 1521 for (page_nr = first_page; page_nr < last_page; page_nr++) { 1522 bm_set_full_words_within_one_page(device->bitmap, page_nr, first_word, last_word); 1523 spin_unlock_irq(&b->bm_lock); 1524 cond_resched(); 1525 first_word = 0; 1526 spin_lock_irq(&b->bm_lock); 1527 } 1528 /* last page (respectively only page, for first page == last page) */ 1529 last_word = MLPP(el >> LN2_BPL); 1530 1531 /* consider bitmap->bm_bits = 32768, bitmap->bm_number_of_pages = 1. (or multiples). 1532 * ==> e = 32767, el = 32768, last_page = 2, 1533 * and now last_word = 0. 1534 * We do not want to touch last_page in this case, 1535 * as we did not allocate it, it is not present in bitmap->bm_pages. 1536 */ 1537 if (last_word) 1538 bm_set_full_words_within_one_page(device->bitmap, last_page, first_word, last_word); 1539 1540 /* possibly trailing bits. 1541 * example: (e & 63) == 63, el will be e+1. 1542 * if that even was the very last bit, 1543 * it would trigger an assert in __bm_change_bits_to() 1544 */ 1545 if (el <= e) 1546 __bm_change_bits_to(device, el, e, 1); 1547 spin_unlock_irq(&b->bm_lock); 1548 } 1549 1550 /* returns bit state 1551 * wants bitnr, NOT sector. 1552 * inherently racy... area needs to be locked by means of {al,rs}_lru 1553 * 1 ... bit set 1554 * 0 ... bit not set 1555 * -1 ... first out of bounds access, stop testing for bits! 1556 */ 1557 int drbd_bm_test_bit(struct drbd_device *device, const unsigned long bitnr) 1558 { 1559 unsigned long flags; 1560 struct drbd_bitmap *b = device->bitmap; 1561 unsigned long *p_addr; 1562 int i; 1563 1564 if (!expect(b)) 1565 return 0; 1566 if (!expect(b->bm_pages)) 1567 return 0; 1568 1569 spin_lock_irqsave(&b->bm_lock, flags); 1570 if (BM_DONT_TEST & b->bm_flags) 1571 bm_print_lock_info(device); 1572 if (bitnr < b->bm_bits) { 1573 p_addr = bm_map_pidx(b, bm_bit_to_page_idx(b, bitnr)); 1574 i = test_bit_le(bitnr & BITS_PER_PAGE_MASK, p_addr) ? 1 : 0; 1575 bm_unmap(p_addr); 1576 } else if (bitnr == b->bm_bits) { 1577 i = -1; 1578 } else { /* (bitnr > b->bm_bits) */ 1579 drbd_err(device, "bitnr=%lu > bm_bits=%lu\n", bitnr, b->bm_bits); 1580 i = 0; 1581 } 1582 1583 spin_unlock_irqrestore(&b->bm_lock, flags); 1584 return i; 1585 } 1586 1587 /* returns number of bits set in the range [s, e] */ 1588 int drbd_bm_count_bits(struct drbd_device *device, const unsigned long s, const unsigned long e) 1589 { 1590 unsigned long flags; 1591 struct drbd_bitmap *b = device->bitmap; 1592 unsigned long *p_addr = NULL; 1593 unsigned long bitnr; 1594 unsigned int page_nr = -1U; 1595 int c = 0; 1596 1597 /* If this is called without a bitmap, that is a bug. But just to be 1598 * robust in case we screwed up elsewhere, in that case pretend there 1599 * was one dirty bit in the requested area, so we won't try to do a 1600 * local read there (no bitmap probably implies no disk) */ 1601 if (!expect(b)) 1602 return 1; 1603 if (!expect(b->bm_pages)) 1604 return 1; 1605 1606 spin_lock_irqsave(&b->bm_lock, flags); 1607 if (BM_DONT_TEST & b->bm_flags) 1608 bm_print_lock_info(device); 1609 for (bitnr = s; bitnr <= e; bitnr++) { 1610 unsigned int idx = bm_bit_to_page_idx(b, bitnr); 1611 if (page_nr != idx) { 1612 page_nr = idx; 1613 if (p_addr) 1614 bm_unmap(p_addr); 1615 p_addr = bm_map_pidx(b, idx); 1616 } 1617 if (expect(bitnr < b->bm_bits)) 1618 c += (0 != test_bit_le(bitnr - (page_nr << (PAGE_SHIFT+3)), p_addr)); 1619 else 1620 drbd_err(device, "bitnr=%lu bm_bits=%lu\n", bitnr, b->bm_bits); 1621 } 1622 if (p_addr) 1623 bm_unmap(p_addr); 1624 spin_unlock_irqrestore(&b->bm_lock, flags); 1625 return c; 1626 } 1627 1628 1629 /* inherently racy... 1630 * return value may be already out-of-date when this function returns. 1631 * but the general usage is that this is only use during a cstate when bits are 1632 * only cleared, not set, and typically only care for the case when the return 1633 * value is zero, or we already "locked" this "bitmap extent" by other means. 1634 * 1635 * enr is bm-extent number, since we chose to name one sector (512 bytes) 1636 * worth of the bitmap a "bitmap extent". 1637 * 1638 * TODO 1639 * I think since we use it like a reference count, we should use the real 1640 * reference count of some bitmap extent element from some lru instead... 1641 * 1642 */ 1643 int drbd_bm_e_weight(struct drbd_device *device, unsigned long enr) 1644 { 1645 struct drbd_bitmap *b = device->bitmap; 1646 int count, s, e; 1647 unsigned long flags; 1648 unsigned long *p_addr, *bm; 1649 1650 if (!expect(b)) 1651 return 0; 1652 if (!expect(b->bm_pages)) 1653 return 0; 1654 1655 spin_lock_irqsave(&b->bm_lock, flags); 1656 if (BM_DONT_TEST & b->bm_flags) 1657 bm_print_lock_info(device); 1658 1659 s = S2W(enr); 1660 e = min((size_t)S2W(enr+1), b->bm_words); 1661 count = 0; 1662 if (s < b->bm_words) { 1663 int n = e-s; 1664 p_addr = bm_map_pidx(b, bm_word_to_page_idx(b, s)); 1665 bm = p_addr + MLPP(s); 1666 count += bitmap_weight(bm, n * BITS_PER_LONG); 1667 bm_unmap(p_addr); 1668 } else { 1669 drbd_err(device, "start offset (%d) too large in drbd_bm_e_weight\n", s); 1670 } 1671 spin_unlock_irqrestore(&b->bm_lock, flags); 1672 return count; 1673 } 1674