drivers/block/drbd/drbd_actlog.c

   1 /*
   2    drbd_actlog.c
   3
   4    This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
   5
   6    Copyright (C) 2003-2008, LINBIT Information Technologies GmbH.
   7    Copyright (C) 2003-2008, Philipp Reisner <philipp.reisner@linbit.com>.
   8    Copyright (C) 2003-2008, 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
  26 #include <linux/slab.h>
  27 #include <linux/drbd.h>
  28 #include "drbd_int.h"
  29 #include "drbd_wrappers.h"
  30
  31 /* We maintain a trivial check sum in our on disk activity log.
  32  * With that we can ensure correct operation even when the storage
  33  * device might do a partial (last) sector write while loosing power.
  34  */
  35 struct __packed al_transaction {
  36         u32       magic;
  37         u32       tr_number;
  38         struct __packed {
  39                 u32 pos;
  40                 u32 extent; } updates[1 + AL_EXTENTS_PT];
  41         u32       xor_sum;
  42 };
  43
  44 struct update_odbm_work {
  45         struct drbd_work w;
  46         unsigned int enr;
  47 };
  48
  49 struct update_al_work {
  50         struct drbd_work w;
  51         struct lc_element *al_ext;
  52         struct completion event;
  53         unsigned int enr;
  54         /* if old_enr != LC_FREE, write corresponding bitmap sector, too */
  55         unsigned int old_enr;
  56 };
  57
  58 struct drbd_atodb_wait {
  59         atomic_t           count;
  60         struct completion  io_done;
  61         struct drbd_conf   *mdev;
  62         int                error;
  63 };
  64
  65
  66 int w_al_write_transaction(struct drbd_conf *, struct drbd_work *, int);
  67
  68 static int _drbd_md_sync_page_io(struct drbd_conf *mdev,
  69                                  struct drbd_backing_dev *bdev,
  70                                  struct page *page, sector_t sector,
  71                                  int rw, int size)
  72 {
  73         struct bio *bio;
  74         struct drbd_md_io md_io;
  75         int ok;
  76
  77         md_io.mdev = mdev;
  78         init_completion(&md_io.event);
  79         md_io.error = 0;
  80
  81         if ((rw & WRITE) && !test_bit(MD_NO_BARRIER, &mdev->flags))
  82                 rw |= REQ_HARDBARRIER;
  83         rw |= REQ_UNPLUG | REQ_SYNC;
  84
  85  retry:
  86         bio = bio_alloc(GFP_NOIO, 1);
  87         bio->bi_bdev = bdev->md_bdev;
  88         bio->bi_sector = sector;
  89         ok = (bio_add_page(bio, page, size, 0) == size);
  90         if (!ok)
  91                 goto out;
  92         bio->bi_private = &md_io;
  93         bio->bi_end_io = drbd_md_io_complete;
  94         bio->bi_rw = rw;
  95
  96         if (FAULT_ACTIVE(mdev, (rw & WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD))
  97                 bio_endio(bio, -EIO);
  98         else
  99                 submit_bio(rw, bio);
 100         wait_for_completion(&md_io.event);
 101         ok = bio_flagged(bio, BIO_UPTODATE) && md_io.error == 0;
 102
 103         /* check for unsupported barrier op.
 104          * would rather check on EOPNOTSUPP, but that is not reliable.
 105          * don't try again for ANY return value != 0 */
 106         if (unlikely((bio->bi_rw & REQ_HARDBARRIER) && !ok)) {
 107                 /* Try again with no barrier */
 108                 dev_warn(DEV, "Barriers not supported on meta data device - disabling\n");
 109                 set_bit(MD_NO_BARRIER, &mdev->flags);
 110                 rw &= ~REQ_HARDBARRIER;
 111                 bio_put(bio);
 112                 goto retry;
 113         }
 114  out:
 115         bio_put(bio);
 116         return ok;
 117 }
 118
 119 int drbd_md_sync_page_io(struct drbd_conf *mdev, struct drbd_backing_dev *bdev,
 120                          sector_t sector, int rw)
 121 {
 122         int logical_block_size, mask, ok;
 123         int offset = 0;
 124         struct page *iop = mdev->md_io_page;
 125
 126         D_ASSERT(mutex_is_locked(&mdev->md_io_mutex));
 127
 128         BUG_ON(!bdev->md_bdev);
 129
 130         logical_block_size = bdev_logical_block_size(bdev->md_bdev);
 131         if (logical_block_size == 0)
 132                 logical_block_size = MD_SECTOR_SIZE;
 133
 134         /* in case logical_block_size != 512 [ s390 only? ] */
 135         if (logical_block_size != MD_SECTOR_SIZE) {
 136                 mask = (logical_block_size / MD_SECTOR_SIZE) - 1;
 137                 D_ASSERT(mask == 1 || mask == 3 || mask == 7);
 138                 D_ASSERT(logical_block_size == (mask+1) * MD_SECTOR_SIZE);
 139                 offset = sector & mask;
 140                 sector = sector & ~mask;
 141                 iop = mdev->md_io_tmpp;
 142
 143                 if (rw & WRITE) {
 144                         /* these are GFP_KERNEL pages, pre-allocated
 145                          * on device initialization */
 146                         void *p = page_address(mdev->md_io_page);
 147                         void *hp = page_address(mdev->md_io_tmpp);
 148
 149                         ok = _drbd_md_sync_page_io(mdev, bdev, iop, sector,
 150                                         READ, logical_block_size);
 151
 152                         if (unlikely(!ok)) {
 153                                 dev_err(DEV, "drbd_md_sync_page_io(,%llus,"
 154                                     "READ [logical_block_size!=512]) failed!\n",
 155                                     (unsigned long long)sector);
 156                                 return 0;
 157                         }
 158
 159                         memcpy(hp + offset*MD_SECTOR_SIZE, p, MD_SECTOR_SIZE);
 160                 }
 161         }
 162
 163         if (sector < drbd_md_first_sector(bdev) ||
 164             sector > drbd_md_last_sector(bdev))
 165                 dev_alert(DEV, "%s [%d]:%s(,%llus,%s) out of range md access!\n",
 166                      current->comm, current->pid, __func__,
 167                      (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ");
 168
 169         ok = _drbd_md_sync_page_io(mdev, bdev, iop, sector, rw, logical_block_size);
 170         if (unlikely(!ok)) {
 171                 dev_err(DEV, "drbd_md_sync_page_io(,%llus,%s) failed!\n",
 172                     (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ");
 173                 return 0;
 174         }
 175
 176         if (logical_block_size != MD_SECTOR_SIZE && !(rw & WRITE)) {
 177                 void *p = page_address(mdev->md_io_page);
 178                 void *hp = page_address(mdev->md_io_tmpp);
 179
 180                 memcpy(p, hp + offset*MD_SECTOR_SIZE, MD_SECTOR_SIZE);
 181         }
 182
 183         return ok;
 184 }
 185
 186 static struct lc_element *_al_get(struct drbd_conf *mdev, unsigned int enr)
 187 {
 188         struct lc_element *al_ext;
 189         struct lc_element *tmp;
 190         unsigned long     al_flags = 0;
 191
 192         spin_lock_irq(&mdev->al_lock);
 193         tmp = lc_find(mdev->resync, enr/AL_EXT_PER_BM_SECT);
 194         if (unlikely(tmp != NULL)) {
 195                 struct bm_extent  *bm_ext = lc_entry(tmp, struct bm_extent, lce);
 196                 if (test_bit(BME_NO_WRITES, &bm_ext->flags)) {
 197                         spin_unlock_irq(&mdev->al_lock);
 198                         return NULL;
 199                 }
 200         }
 201         al_ext   = lc_get(mdev->act_log, enr);
 202         al_flags = mdev->act_log->flags;
 203         spin_unlock_irq(&mdev->al_lock);
 204
 205         /*
 206         if (!al_ext) {
 207                 if (al_flags & LC_STARVING)
 208                         dev_warn(DEV, "Have to wait for LRU element (AL too small?)\n");
 209                 if (al_flags & LC_DIRTY)
 210                         dev_warn(DEV, "Ongoing AL update (AL device too slow?)\n");
 211         }
 212         */
 213
 214         return al_ext;
 215 }
 216
 217 void drbd_al_begin_io(struct drbd_conf *mdev, sector_t sector)
 218 {
 219         unsigned int enr = (sector >> (AL_EXTENT_SHIFT-9));
 220         struct lc_element *al_ext;
 221         struct update_al_work al_work;
 222
 223         D_ASSERT(atomic_read(&mdev->local_cnt) > 0);
 224
 225         wait_event(mdev->al_wait, (al_ext = _al_get(mdev, enr)));
 226
 227         if (al_ext->lc_number != enr) {
 228                 /* drbd_al_write_transaction(mdev,al_ext,enr);
 229                  * recurses into generic_make_request(), which
 230                  * disallows recursion, bios being serialized on the
 231                  * current->bio_tail list now.
 232                  * we have to delegate updates to the activity log
 233                  * to the worker thread. */
 234                 init_completion(&al_work.event);
 235                 al_work.al_ext = al_ext;
 236                 al_work.enr = enr;
 237                 al_work.old_enr = al_ext->lc_number;
 238                 al_work.w.cb = w_al_write_transaction;
 239                 drbd_queue_work_front(&mdev->data.work, &al_work.w);
 240                 wait_for_completion(&al_work.event);
 241
 242                 mdev->al_writ_cnt++;
 243
 244                 spin_lock_irq(&mdev->al_lock);
 245                 lc_changed(mdev->act_log, al_ext);
 246                 spin_unlock_irq(&mdev->al_lock);
 247                 wake_up(&mdev->al_wait);
 248         }
 249 }
 250
 251 void drbd_al_complete_io(struct drbd_conf *mdev, sector_t sector)
 252 {
 253         unsigned int enr = (sector >> (AL_EXTENT_SHIFT-9));
 254         struct lc_element *extent;
 255         unsigned long flags;
 256
 257         spin_lock_irqsave(&mdev->al_lock, flags);
 258
 259         extent = lc_find(mdev->act_log, enr);
 260
 261         if (!extent) {
 262                 spin_unlock_irqrestore(&mdev->al_lock, flags);
 263                 dev_err(DEV, "al_complete_io() called on inactive extent %u\n", enr);
 264                 return;
 265         }
 266
 267         if (lc_put(mdev->act_log, extent) == 0)
 268                 wake_up(&mdev->al_wait);
 269
 270         spin_unlock_irqrestore(&mdev->al_lock, flags);
 271 }
 272
 273 int
 274 w_al_write_transaction(struct drbd_conf *mdev, struct drbd_work *w, int unused)
 275 {
 276         struct update_al_work *aw = container_of(w, struct update_al_work, w);
 277         struct lc_element *updated = aw->al_ext;
 278         const unsigned int new_enr = aw->enr;
 279         const unsigned int evicted = aw->old_enr;
 280         struct al_transaction *buffer;
 281         sector_t sector;
 282         int i, n, mx;
 283         unsigned int extent_nr;
 284         u32 xor_sum = 0;
 285
 286         if (!get_ldev(mdev)) {
 287                 dev_err(DEV,
 288                         "disk is %s, cannot start al transaction (-%d +%d)\n",
 289                         drbd_disk_str(mdev->state.disk), evicted, new_enr);
 290                 complete(&((struct update_al_work *)w)->event);
 291                 return 1;
 292         }
 293         /* do we have to do a bitmap write, first?
 294          * TODO reduce maximum latency:
 295          * submit both bios, then wait for both,
 296          * instead of doing two synchronous sector writes.
 297          * For now, we must not write the transaction,
 298          * if we cannot write out the bitmap of the evicted extent. */
 299         if (mdev->state.conn < C_CONNECTED && evicted != LC_FREE)
 300                 drbd_bm_write_sect(mdev, evicted/AL_EXT_PER_BM_SECT);
 301
 302         /* The bitmap write may have failed, causing a state change. */
 303         if (mdev->state.disk < D_INCONSISTENT) {
 304                 dev_err(DEV,
 305                         "disk is %s, cannot write al transaction (-%d +%d)\n",
 306                         drbd_disk_str(mdev->state.disk), evicted, new_enr);
 307                 complete(&((struct update_al_work *)w)->event);
 308                 put_ldev(mdev);
 309                 return 1;
 310         }
 311
 312         mutex_lock(&mdev->md_io_mutex); /* protects md_io_buffer, al_tr_cycle, ... */
 313         buffer = (struct al_transaction *)page_address(mdev->md_io_page);
 314
 315         buffer->magic = __constant_cpu_to_be32(DRBD_MAGIC);
 316         buffer->tr_number = cpu_to_be32(mdev->al_tr_number);
 317
 318         n = lc_index_of(mdev->act_log, updated);
 319
 320         buffer->updates[0].pos = cpu_to_be32(n);
 321         buffer->updates[0].extent = cpu_to_be32(new_enr);
 322
 323         xor_sum ^= new_enr;
 324
 325         mx = min_t(int, AL_EXTENTS_PT,
 326                    mdev->act_log->nr_elements - mdev->al_tr_cycle);
 327         for (i = 0; i < mx; i++) {
 328                 unsigned idx = mdev->al_tr_cycle + i;
 329                 extent_nr = lc_element_by_index(mdev->act_log, idx)->lc_number;
 330                 buffer->updates[i+1].pos = cpu_to_be32(idx);
 331                 buffer->updates[i+1].extent = cpu_to_be32(extent_nr);
 332                 xor_sum ^= extent_nr;
 333         }
 334         for (; i < AL_EXTENTS_PT; i++) {
 335                 buffer->updates[i+1].pos = __constant_cpu_to_be32(-1);
 336                 buffer->updates[i+1].extent = __constant_cpu_to_be32(LC_FREE);
 337                 xor_sum ^= LC_FREE;
 338         }
 339         mdev->al_tr_cycle += AL_EXTENTS_PT;
 340         if (mdev->al_tr_cycle >= mdev->act_log->nr_elements)
 341                 mdev->al_tr_cycle = 0;
 342
 343         buffer->xor_sum = cpu_to_be32(xor_sum);
 344
 345         sector =  mdev->ldev->md.md_offset
 346                 + mdev->ldev->md.al_offset + mdev->al_tr_pos;
 347
 348         if (!drbd_md_sync_page_io(mdev, mdev->ldev, sector, WRITE))
 349                 drbd_chk_io_error(mdev, 1, TRUE);
 350
 351         if (++mdev->al_tr_pos >
 352             div_ceil(mdev->act_log->nr_elements, AL_EXTENTS_PT))
 353                 mdev->al_tr_pos = 0;
 354
 355         D_ASSERT(mdev->al_tr_pos < MD_AL_MAX_SIZE);
 356         mdev->al_tr_number++;
 357
 358         mutex_unlock(&mdev->md_io_mutex);
 359
 360         complete(&((struct update_al_work *)w)->event);
 361         put_ldev(mdev);
 362
 363         return 1;
 364 }
 365
 366 /**
 367  * drbd_al_read_tr() - Read a single transaction from the on disk activity log
 368  * @mdev:       DRBD device.
 369  * @bdev:       Block device to read form.
 370  * @b:          pointer to an al_transaction.
 371  * @index:      On disk slot of the transaction to read.
 372  *
 373  * Returns -1 on IO error, 0 on checksum error and 1 upon success.
 374  */
 375 static int drbd_al_read_tr(struct drbd_conf *mdev,
 376                            struct drbd_backing_dev *bdev,
 377                            struct al_transaction *b,
 378                            int index)
 379 {
 380         sector_t sector;
 381         int rv, i;
 382         u32 xor_sum = 0;
 383
 384         sector = bdev->md.md_offset + bdev->md.al_offset + index;
 385
 386         /* Dont process error normally,
 387          * as this is done before disk is attached! */
 388         if (!drbd_md_sync_page_io(mdev, bdev, sector, READ))
 389                 return -1;
 390
 391         rv = (be32_to_cpu(b->magic) == DRBD_MAGIC);
 392
 393         for (i = 0; i < AL_EXTENTS_PT + 1; i++)
 394                 xor_sum ^= be32_to_cpu(b->updates[i].extent);
 395         rv &= (xor_sum == be32_to_cpu(b->xor_sum));
 396
 397         return rv;
 398 }
 399
 400 /**
 401  * drbd_al_read_log() - Restores the activity log from its on disk representation.
 402  * @mdev:       DRBD device.
 403  * @bdev:       Block device to read form.
 404  *
 405  * Returns 1 on success, returns 0 when reading the log failed due to IO errors.
 406  */
 407 int drbd_al_read_log(struct drbd_conf *mdev, struct drbd_backing_dev *bdev)
 408 {
 409         struct al_transaction *buffer;
 410         int i;
 411         int rv;
 412         int mx;
 413         int active_extents = 0;
 414         int transactions = 0;
 415         int found_valid = 0;
 416         int from = 0;
 417         int to = 0;
 418         u32 from_tnr = 0;
 419         u32 to_tnr = 0;
 420         u32 cnr;
 421
 422         mx = div_ceil(mdev->act_log->nr_elements, AL_EXTENTS_PT);
 423
 424         /* lock out all other meta data io for now,
 425          * and make sure the page is mapped.
 426          */
 427         mutex_lock(&mdev->md_io_mutex);
 428         buffer = page_address(mdev->md_io_page);
 429
 430         /* Find the valid transaction in the log */
 431         for (i = 0; i <= mx; i++) {
 432                 rv = drbd_al_read_tr(mdev, bdev, buffer, i);
 433                 if (rv == 0)
 434                         continue;
 435                 if (rv == -1) {
 436                         mutex_unlock(&mdev->md_io_mutex);
 437                         return 0;
 438                 }
 439                 cnr = be32_to_cpu(buffer->tr_number);
 440
 441                 if (++found_valid == 1) {
 442                         from = i;
 443                         to = i;
 444                         from_tnr = cnr;
 445                         to_tnr = cnr;
 446                         continue;
 447                 }
 448                 if ((int)cnr - (int)from_tnr < 0) {
 449                         D_ASSERT(from_tnr - cnr + i - from == mx+1);
 450                         from = i;
 451                         from_tnr = cnr;
 452                 }
 453                 if ((int)cnr - (int)to_tnr > 0) {
 454                         D_ASSERT(cnr - to_tnr == i - to);
 455                         to = i;
 456                         to_tnr = cnr;
 457                 }
 458         }
 459
 460         if (!found_valid) {
 461                 dev_warn(DEV, "No usable activity log found.\n");
 462                 mutex_unlock(&mdev->md_io_mutex);
 463                 return 1;
 464         }
 465
 466         /* Read the valid transactions.
 467          * dev_info(DEV, "Reading from %d to %d.\n",from,to); */
 468         i = from;
 469         while (1) {
 470                 int j, pos;
 471                 unsigned int extent_nr;
 472                 unsigned int trn;
 473
 474                 rv = drbd_al_read_tr(mdev, bdev, buffer, i);
 475                 ERR_IF(rv == 0) goto cancel;
 476                 if (rv == -1) {
 477                         mutex_unlock(&mdev->md_io_mutex);
 478                         return 0;
 479                 }
 480
 481                 trn = be32_to_cpu(buffer->tr_number);
 482
 483                 spin_lock_irq(&mdev->al_lock);
 484
 485                 /* This loop runs backwards because in the cyclic
 486                    elements there might be an old version of the
 487                    updated element (in slot 0). So the element in slot 0
 488                    can overwrite old versions. */
 489                 for (j = AL_EXTENTS_PT; j >= 0; j--) {
 490                         pos = be32_to_cpu(buffer->updates[j].pos);
 491                         extent_nr = be32_to_cpu(buffer->updates[j].extent);
 492
 493                         if (extent_nr == LC_FREE)
 494                                 continue;
 495
 496                         lc_set(mdev->act_log, extent_nr, pos);
 497                         active_extents++;
 498                 }
 499                 spin_unlock_irq(&mdev->al_lock);
 500
 501                 transactions++;
 502
 503 cancel:
 504                 if (i == to)
 505                         break;
 506                 i++;
 507                 if (i > mx)
 508                         i = 0;
 509         }
 510
 511         mdev->al_tr_number = to_tnr+1;
 512         mdev->al_tr_pos = to;
 513         if (++mdev->al_tr_pos >
 514             div_ceil(mdev->act_log->nr_elements, AL_EXTENTS_PT))
 515                 mdev->al_tr_pos = 0;
 516
 517         /* ok, we are done with it */
 518         mutex_unlock(&mdev->md_io_mutex);
 519
 520         dev_info(DEV, "Found %d transactions (%d active extents) in activity log.\n",
 521              transactions, active_extents);
 522
 523         return 1;
 524 }
 525
 526 static void atodb_endio(struct bio *bio, int error)
 527 {
 528         struct drbd_atodb_wait *wc = bio->bi_private;
 529         struct drbd_conf *mdev = wc->mdev;
 530         struct page *page;
 531         int uptodate = bio_flagged(bio, BIO_UPTODATE);
 532
 533         /* strange behavior of some lower level drivers...
 534          * fail the request by clearing the uptodate flag,
 535          * but do not return any error?! */
 536         if (!error && !uptodate)
 537                 error = -EIO;
 538
 539         drbd_chk_io_error(mdev, error, TRUE);
 540         if (error && wc->error == 0)
 541                 wc->error = error;
 542
 543         if (atomic_dec_and_test(&wc->count))
 544                 complete(&wc->io_done);
 545
 546         page = bio->bi_io_vec[0].bv_page;
 547         put_page(page);
 548         bio_put(bio);
 549         mdev->bm_writ_cnt++;
 550         put_ldev(mdev);
 551 }
 552
 553 /* sector to word */
 554 #define S2W(s)  ((s)<<(BM_EXT_SHIFT-BM_BLOCK_SHIFT-LN2_BPL))
 555
 556 /* activity log to on disk bitmap -- prepare bio unless that sector
 557  * is already covered by previously prepared bios */
 558 static int atodb_prepare_unless_covered(struct drbd_conf *mdev,
 559                                         struct bio **bios,
 560                                         unsigned int enr,
 561                                         struct drbd_atodb_wait *wc) __must_hold(local)
 562 {
 563         struct bio *bio;
 564         struct page *page;
 565         sector_t on_disk_sector;
 566         unsigned int page_offset = PAGE_SIZE;
 567         int offset;
 568         int i = 0;
 569         int err = -ENOMEM;
 570
 571         /* We always write aligned, full 4k blocks,
 572          * so we can ignore the logical_block_size (for now) */
 573         enr &= ~7U;
 574         on_disk_sector = enr + mdev->ldev->md.md_offset
 575                              + mdev->ldev->md.bm_offset;
 576
 577         D_ASSERT(!(on_disk_sector & 7U));
 578
 579         /* Check if that enr is already covered by an already created bio.
 580          * Caution, bios[] is not NULL terminated,
 581          * but only initialized to all NULL.
 582          * For completely scattered activity log,
 583          * the last invocation iterates over all bios,
 584          * and finds the last NULL entry.
 585          */
 586         while ((bio = bios[i])) {
 587                 if (bio->bi_sector == on_disk_sector)
 588                         return 0;
 589                 i++;
 590         }
 591         /* bios[i] == NULL, the next not yet used slot */
 592
 593         /* GFP_KERNEL, we are not in the write-out path */
 594         bio = bio_alloc(GFP_KERNEL, 1);
 595         if (bio == NULL)
 596                 return -ENOMEM;
 597
 598         if (i > 0) {
 599                 const struct bio_vec *prev_bv = bios[i-1]->bi_io_vec;
 600                 page_offset = prev_bv->bv_offset + prev_bv->bv_len;
 601                 page = prev_bv->bv_page;
 602         }
 603         if (page_offset == PAGE_SIZE) {
 604                 page = alloc_page(__GFP_HIGHMEM);
 605                 if (page == NULL)
 606                         goto out_bio_put;
 607                 page_offset = 0;
 608         } else {
 609                 get_page(page);
 610         }
 611
 612         offset = S2W(enr);
 613         drbd_bm_get_lel(mdev, offset,
 614                         min_t(size_t, S2W(8), drbd_bm_words(mdev) - offset),
 615                         kmap(page) + page_offset);
 616         kunmap(page);
 617
 618         bio->bi_private = wc;
 619         bio->bi_end_io = atodb_endio;
 620         bio->bi_bdev = mdev->ldev->md_bdev;
 621         bio->bi_sector = on_disk_sector;
 622
 623         if (bio_add_page(bio, page, 4096, page_offset) != 4096)
 624                 goto out_put_page;
 625
 626         atomic_inc(&wc->count);
 627         /* we already know that we may do this...
 628          * get_ldev_if_state(mdev,D_ATTACHING);
 629          * just get the extra reference, so that the local_cnt reflects
 630          * the number of pending IO requests DRBD at its backing device.
 631          */
 632         atomic_inc(&mdev->local_cnt);
 633
 634         bios[i] = bio;
 635
 636         return 0;
 637
 638 out_put_page:
 639         err = -EINVAL;
 640         put_page(page);
 641 out_bio_put:
 642         bio_put(bio);
 643         return err;
 644 }
 645
 646 /**
 647  * drbd_al_to_on_disk_bm() -  * Writes bitmap parts covered by active AL extents
 648  * @mdev:       DRBD device.
 649  *
 650  * Called when we detach (unconfigure) local storage,
 651  * or when we go from R_PRIMARY to R_SECONDARY role.
 652  */
 653 void drbd_al_to_on_disk_bm(struct drbd_conf *mdev)
 654 {
 655         int i, nr_elements;
 656         unsigned int enr;
 657         struct bio **bios;
 658         struct drbd_atodb_wait wc;
 659
 660         ERR_IF (!get_ldev_if_state(mdev, D_ATTACHING))
 661                 return; /* sorry, I don't have any act_log etc... */
 662
 663         wait_event(mdev->al_wait, lc_try_lock(mdev->act_log));
 664
 665         nr_elements = mdev->act_log->nr_elements;
 666
 667         /* GFP_KERNEL, we are not in anyone's write-out path */
 668         bios = kzalloc(sizeof(struct bio *) * nr_elements, GFP_KERNEL);
 669         if (!bios)
 670                 goto submit_one_by_one;
 671
 672         atomic_set(&wc.count, 0);
 673         init_completion(&wc.io_done);
 674         wc.mdev = mdev;
 675         wc.error = 0;
 676
 677         for (i = 0; i < nr_elements; i++) {
 678                 enr = lc_element_by_index(mdev->act_log, i)->lc_number;
 679                 if (enr == LC_FREE)
 680                         continue;
 681                 /* next statement also does atomic_inc wc.count and local_cnt */
 682                 if (atodb_prepare_unless_covered(mdev, bios,
 683                                                 enr/AL_EXT_PER_BM_SECT,
 684                                                 &wc))
 685                         goto free_bios_submit_one_by_one;
 686         }
 687
 688         /* unnecessary optimization? */
 689         lc_unlock(mdev->act_log);
 690         wake_up(&mdev->al_wait);
 691
 692         /* all prepared, submit them */
 693         for (i = 0; i < nr_elements; i++) {
 694                 if (bios[i] == NULL)
 695                         break;
 696                 if (FAULT_ACTIVE(mdev, DRBD_FAULT_MD_WR)) {
 697                         bios[i]->bi_rw = WRITE;
 698                         bio_endio(bios[i], -EIO);
 699                 } else {
 700                         submit_bio(WRITE, bios[i]);
 701                 }
 702         }
 703
 704         drbd_blk_run_queue(bdev_get_queue(mdev->ldev->md_bdev));
 705
 706         /* always (try to) flush bitmap to stable storage */
 707         drbd_md_flush(mdev);
 708
 709         /* In case we did not submit a single IO do not wait for
 710          * them to complete. ( Because we would wait forever here. )
 711          *
 712          * In case we had IOs and they are already complete, there
 713          * is not point in waiting anyways.
 714          * Therefore this if () ... */
 715         if (atomic_read(&wc.count))
 716                 wait_for_completion(&wc.io_done);
 717
 718         put_ldev(mdev);
 719
 720         kfree(bios);
 721         return;
 722
 723  free_bios_submit_one_by_one:
 724         /* free everything by calling the endio callback directly. */
 725         for (i = 0; i < nr_elements && bios[i]; i++)
 726                 bio_endio(bios[i], 0);
 727
 728         kfree(bios);
 729
 730  submit_one_by_one:
 731         dev_warn(DEV, "Using the slow drbd_al_to_on_disk_bm()\n");
 732
 733         for (i = 0; i < mdev->act_log->nr_elements; i++) {
 734                 enr = lc_element_by_index(mdev->act_log, i)->lc_number;
 735                 if (enr == LC_FREE)
 736                         continue;
 737                 /* Really slow: if we have al-extents 16..19 active,
 738                  * sector 4 will be written four times! Synchronous! */
 739                 drbd_bm_write_sect(mdev, enr/AL_EXT_PER_BM_SECT);
 740         }
 741
 742         lc_unlock(mdev->act_log);
 743         wake_up(&mdev->al_wait);
 744         put_ldev(mdev);
 745 }
 746
 747 /**
 748  * drbd_al_apply_to_bm() - Sets the bitmap to diry(1) where covered ba active AL extents
 749  * @mdev:       DRBD device.
 750  */
 751 void drbd_al_apply_to_bm(struct drbd_conf *mdev)
 752 {
 753         unsigned int enr;
 754         unsigned long add = 0;
 755         char ppb[10];
 756         int i, tmp;
 757
 758         wait_event(mdev->al_wait, lc_try_lock(mdev->act_log));
 759
 760         for (i = 0; i < mdev->act_log->nr_elements; i++) {
 761                 enr = lc_element_by_index(mdev->act_log, i)->lc_number;
 762                 if (enr == LC_FREE)
 763                         continue;
 764                 tmp = drbd_bm_ALe_set_all(mdev, enr);
 765                 dynamic_dev_dbg(DEV, "AL: set %d bits in extent %u\n", tmp, enr);
 766                 add += tmp;
 767         }
 768
 769         lc_unlock(mdev->act_log);
 770         wake_up(&mdev->al_wait);
 771
 772         dev_info(DEV, "Marked additional %s as out-of-sync based on AL.\n",
 773              ppsize(ppb, Bit2KB(add)));
 774 }
 775
 776 static int _try_lc_del(struct drbd_conf *mdev, struct lc_element *al_ext)
 777 {
 778         int rv;
 779
 780         spin_lock_irq(&mdev->al_lock);
 781         rv = (al_ext->refcnt == 0);
 782         if (likely(rv))
 783                 lc_del(mdev->act_log, al_ext);
 784         spin_unlock_irq(&mdev->al_lock);
 785
 786         return rv;
 787 }
 788
 789 /**
 790  * drbd_al_shrink() - Removes all active extents form the activity log
 791  * @mdev:       DRBD device.
 792  *
 793  * Removes all active extents form the activity log, waiting until
 794  * the reference count of each entry dropped to 0 first, of course.
 795  *
 796  * You need to lock mdev->act_log with lc_try_lock() / lc_unlock()
 797  */
 798 void drbd_al_shrink(struct drbd_conf *mdev)
 799 {
 800         struct lc_element *al_ext;
 801         int i;
 802
 803         D_ASSERT(test_bit(__LC_DIRTY, &mdev->act_log->flags));
 804
 805         for (i = 0; i < mdev->act_log->nr_elements; i++) {
 806                 al_ext = lc_element_by_index(mdev->act_log, i);
 807                 if (al_ext->lc_number == LC_FREE)
 808                         continue;
 809                 wait_event(mdev->al_wait, _try_lc_del(mdev, al_ext));
 810         }
 811
 812         wake_up(&mdev->al_wait);
 813 }
 814
 815 static int w_update_odbm(struct drbd_conf *mdev, struct drbd_work *w, int unused)
 816 {
 817         struct update_odbm_work *udw = container_of(w, struct update_odbm_work, w);
 818
 819         if (!get_ldev(mdev)) {
 820                 if (__ratelimit(&drbd_ratelimit_state))
 821                         dev_warn(DEV, "Can not update on disk bitmap, local IO disabled.\n");
 822                 kfree(udw);
 823                 return 1;
 824         }
 825
 826         drbd_bm_write_sect(mdev, udw->enr);
 827         put_ldev(mdev);
 828
 829         kfree(udw);
 830
 831         if (drbd_bm_total_weight(mdev) <= mdev->rs_failed) {
 832                 switch (mdev->state.conn) {
 833                 case C_SYNC_SOURCE:  case C_SYNC_TARGET:
 834                 case C_PAUSED_SYNC_S: case C_PAUSED_SYNC_T:
 835                         drbd_resync_finished(mdev);
 836                 default:
 837                         /* nothing to do */
 838                         break;
 839                 }
 840         }
 841         drbd_bcast_sync_progress(mdev);
 842
 843         return 1;
 844 }
 845
 846
 847 /* ATTENTION. The AL's extents are 4MB each, while the extents in the
 848  * resync LRU-cache are 16MB each.
 849  * The caller of this function has to hold an get_ldev() reference.
 850  *
 851  * TODO will be obsoleted once we have a caching lru of the on disk bitmap
 852  */
 853 static void drbd_try_clear_on_disk_bm(struct drbd_conf *mdev, sector_t sector,
 854                                       int count, int success)
 855 {
 856         struct lc_element *e;
 857         struct update_odbm_work *udw;
 858
 859         unsigned int enr;
 860
 861         D_ASSERT(atomic_read(&mdev->local_cnt));
 862
 863         /* I simply assume that a sector/size pair never crosses
 864          * a 16 MB extent border. (Currently this is true...) */
 865         enr = BM_SECT_TO_EXT(sector);
 866
 867         e = lc_get(mdev->resync, enr);
 868         if (e) {
 869                 struct bm_extent *ext = lc_entry(e, struct bm_extent, lce);
 870                 if (ext->lce.lc_number == enr) {
 871                         if (success)
 872                                 ext->rs_left -= count;
 873                         else
 874                                 ext->rs_failed += count;
 875                         if (ext->rs_left < ext->rs_failed) {
 876                                 dev_err(DEV, "BAD! sector=%llus enr=%u rs_left=%d "
 877                                     "rs_failed=%d count=%d\n",
 878                                      (unsigned long long)sector,
 879                                      ext->lce.lc_number, ext->rs_left,
 880                                      ext->rs_failed, count);
 881                                 dump_stack();
 882
 883                                 lc_put(mdev->resync, &ext->lce);
 884                                 drbd_force_state(mdev, NS(conn, C_DISCONNECTING));
 885                                 return;
 886                         }
 887                 } else {
 888                         /* Normally this element should be in the cache,
 889                          * since drbd_rs_begin_io() pulled it already in.
 890                          *
 891                          * But maybe an application write finished, and we set
 892                          * something outside the resync lru_cache in sync.
 893                          */
 894                         int rs_left = drbd_bm_e_weight(mdev, enr);
 895                         if (ext->flags != 0) {
 896                                 dev_warn(DEV, "changing resync lce: %d[%u;%02lx]"
 897                                      " -> %d[%u;00]\n",
 898                                      ext->lce.lc_number, ext->rs_left,
 899                                      ext->flags, enr, rs_left);
 900                                 ext->flags = 0;
 901                         }
 902                         if (ext->rs_failed) {
 903                                 dev_warn(DEV, "Kicking resync_lru element enr=%u "
 904                                      "out with rs_failed=%d\n",
 905                                      ext->lce.lc_number, ext->rs_failed);
 906                                 set_bit(WRITE_BM_AFTER_RESYNC, &mdev->flags);
 907                         }
 908                         ext->rs_left = rs_left;
 909                         ext->rs_failed = success ? 0 : count;
 910                         lc_changed(mdev->resync, &ext->lce);
 911                 }
 912                 lc_put(mdev->resync, &ext->lce);
 913                 /* no race, we are within the al_lock! */
 914
 915                 if (ext->rs_left == ext->rs_failed) {
 916                         ext->rs_failed = 0;
 917
 918                         udw = kmalloc(sizeof(*udw), GFP_ATOMIC);
 919                         if (udw) {
 920                                 udw->enr = ext->lce.lc_number;
 921                                 udw->w.cb = w_update_odbm;
 922                                 drbd_queue_work_front(&mdev->data.work, &udw->w);
 923                         } else {
 924                                 dev_warn(DEV, "Could not kmalloc an udw\n");
 925                                 set_bit(WRITE_BM_AFTER_RESYNC, &mdev->flags);
 926                         }
 927                 }
 928         } else {
 929                 dev_err(DEV, "lc_get() failed! locked=%d/%d flags=%lu\n",
 930                     mdev->resync_locked,
 931                     mdev->resync->nr_elements,
 932                     mdev->resync->flags);
 933         }
 934 }
 935
 936 /* clear the bit corresponding to the piece of storage in question:
 937  * size byte of data starting from sector.  Only clear a bits of the affected
 938  * one ore more _aligned_ BM_BLOCK_SIZE blocks.
 939  *
 940  * called by worker on C_SYNC_TARGET and receiver on SyncSource.
 941  *
 942  */
 943 void __drbd_set_in_sync(struct drbd_conf *mdev, sector_t sector, int size,
 944                        const char *file, const unsigned int line)
 945 {
 946         /* Is called from worker and receiver context _only_ */
 947         unsigned long sbnr, ebnr, lbnr;
 948         unsigned long count = 0;
 949         sector_t esector, nr_sectors;
 950         int wake_up = 0;
 951         unsigned long flags;
 952
 953         if (size <= 0 || (size & 0x1ff) != 0 || size > DRBD_MAX_SEGMENT_SIZE) {
 954                 dev_err(DEV, "drbd_set_in_sync: sector=%llus size=%d nonsense!\n",
 955                                 (unsigned long long)sector, size);
 956                 return;
 957         }
 958         nr_sectors = drbd_get_capacity(mdev->this_bdev);
 959         esector = sector + (size >> 9) - 1;
 960
 961         ERR_IF(sector >= nr_sectors) return;
 962         ERR_IF(esector >= nr_sectors) esector = (nr_sectors-1);
 963
 964         lbnr = BM_SECT_TO_BIT(nr_sectors-1);
 965
 966         /* we clear it (in sync).
 967          * round up start sector, round down end sector.  we make sure we only
 968          * clear full, aligned, BM_BLOCK_SIZE (4K) blocks */
 969         if (unlikely(esector < BM_SECT_PER_BIT-1))
 970                 return;
 971         if (unlikely(esector == (nr_sectors-1)))
 972                 ebnr = lbnr;
 973         else
 974                 ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1));
 975         sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1);
 976
 977         if (sbnr > ebnr)
 978                 return;
 979
 980         /*
 981          * ok, (capacity & 7) != 0 sometimes, but who cares...
 982          * we count rs_{total,left} in bits, not sectors.
 983          */
 984         count = drbd_bm_clear_bits(mdev, sbnr, ebnr);
 985         if (count && get_ldev(mdev)) {
 986                 unsigned long now = jiffies;
 987                 unsigned long last = mdev->rs_mark_time[mdev->rs_last_mark];
 988                 int next = (mdev->rs_last_mark + 1) % DRBD_SYNC_MARKS;
 989                 if (time_after_eq(now, last + DRBD_SYNC_MARK_STEP)) {
 990                         unsigned long tw = drbd_bm_total_weight(mdev);
 991                         if (mdev->rs_mark_left[mdev->rs_last_mark] != tw &&
 992                             mdev->state.conn != C_PAUSED_SYNC_T &&
 993                             mdev->state.conn != C_PAUSED_SYNC_S) {
 994                                 mdev->rs_mark_time[next] = now;
 995                                 mdev->rs_mark_left[next] = tw;
 996                                 mdev->rs_last_mark = next;
 997                         }
 998                 }
 999                 spin_lock_irqsave(&mdev->al_lock, flags);
1000                 drbd_try_clear_on_disk_bm(mdev, sector, count, TRUE);
1001                 spin_unlock_irqrestore(&mdev->al_lock, flags);
1002
1003                 /* just wake_up unconditional now, various lc_chaged(),
1004                  * lc_put() in drbd_try_clear_on_disk_bm(). */
1005                 wake_up = 1;
1006                 put_ldev(mdev);
1007         }
1008         if (wake_up)
1009                 wake_up(&mdev->al_wait);
1010 }
1011
1012 /*
1013  * this is intended to set one request worth of data out of sync.
1014  * affects at least 1 bit,
1015  * and at most 1+DRBD_MAX_SEGMENT_SIZE/BM_BLOCK_SIZE bits.
1016  *
1017  * called by tl_clear and drbd_send_dblock (==drbd_make_request).
1018  * so this can be _any_ process.
1019  */
1020 void __drbd_set_out_of_sync(struct drbd_conf *mdev, sector_t sector, int size,
1021                             const char *file, const unsigned int line)
1022 {
1023         unsigned long sbnr, ebnr, lbnr, flags;
1024         sector_t esector, nr_sectors;
1025         unsigned int enr, count;
1026         struct lc_element *e;
1027
1028         if (size <= 0 || (size & 0x1ff) != 0 || size > DRBD_MAX_SEGMENT_SIZE) {
1029                 dev_err(DEV, "sector: %llus, size: %d\n",
1030                         (unsigned long long)sector, size);
1031                 return;
1032         }
1033
1034         if (!get_ldev(mdev))
1035                 return; /* no disk, no metadata, no bitmap to set bits in */
1036
1037         nr_sectors = drbd_get_capacity(mdev->this_bdev);
1038         esector = sector + (size >> 9) - 1;
1039
1040         ERR_IF(sector >= nr_sectors)
1041                 goto out;
1042         ERR_IF(esector >= nr_sectors)
1043                 esector = (nr_sectors-1);
1044
1045         lbnr = BM_SECT_TO_BIT(nr_sectors-1);
1046
1047         /* we set it out of sync,
1048          * we do not need to round anything here */
1049         sbnr = BM_SECT_TO_BIT(sector);
1050         ebnr = BM_SECT_TO_BIT(esector);
1051
1052         /* ok, (capacity & 7) != 0 sometimes, but who cares...
1053          * we count rs_{total,left} in bits, not sectors.  */
1054         spin_lock_irqsave(&mdev->al_lock, flags);
1055         count = drbd_bm_set_bits(mdev, sbnr, ebnr);
1056
1057         enr = BM_SECT_TO_EXT(sector);
1058         e = lc_find(mdev->resync, enr);
1059         if (e)
1060                 lc_entry(e, struct bm_extent, lce)->rs_left += count;
1061         spin_unlock_irqrestore(&mdev->al_lock, flags);
1062
1063 out:
1064         put_ldev(mdev);
1065 }
1066
1067 static
1068 struct bm_extent *_bme_get(struct drbd_conf *mdev, unsigned int enr)
1069 {
1070         struct lc_element *e;
1071         struct bm_extent *bm_ext;
1072         int wakeup = 0;
1073         unsigned long rs_flags;
1074
1075         spin_lock_irq(&mdev->al_lock);
1076         if (mdev->resync_locked > mdev->resync->nr_elements/2) {
1077                 spin_unlock_irq(&mdev->al_lock);
1078                 return NULL;
1079         }
1080         e = lc_get(mdev->resync, enr);
1081         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1082         if (bm_ext) {
1083                 if (bm_ext->lce.lc_number != enr) {
1084                         bm_ext->rs_left = drbd_bm_e_weight(mdev, enr);
1085                         bm_ext->rs_failed = 0;
1086                         lc_changed(mdev->resync, &bm_ext->lce);
1087                         wakeup = 1;
1088                 }
1089                 if (bm_ext->lce.refcnt == 1)
1090                         mdev->resync_locked++;
1091                 set_bit(BME_NO_WRITES, &bm_ext->flags);
1092         }
1093         rs_flags = mdev->resync->flags;
1094         spin_unlock_irq(&mdev->al_lock);
1095         if (wakeup)
1096                 wake_up(&mdev->al_wait);
1097
1098         if (!bm_ext) {
1099                 if (rs_flags & LC_STARVING)
1100                         dev_warn(DEV, "Have to wait for element"
1101                              " (resync LRU too small?)\n");
1102                 BUG_ON(rs_flags & LC_DIRTY);
1103         }
1104
1105         return bm_ext;
1106 }
1107
1108 static int _is_in_al(struct drbd_conf *mdev, unsigned int enr)
1109 {
1110         struct lc_element *al_ext;
1111         int rv = 0;
1112
1113         spin_lock_irq(&mdev->al_lock);
1114         if (unlikely(enr == mdev->act_log->new_number))
1115                 rv = 1;
1116         else {
1117                 al_ext = lc_find(mdev->act_log, enr);
1118                 if (al_ext) {
1119                         if (al_ext->refcnt)
1120                                 rv = 1;
1121                 }
1122         }
1123         spin_unlock_irq(&mdev->al_lock);
1124
1125         /*
1126         if (unlikely(rv)) {
1127                 dev_info(DEV, "Delaying sync read until app's write is done\n");
1128         }
1129         */
1130         return rv;
1131 }
1132
1133 /**
1134  * drbd_rs_begin_io() - Gets an extent in the resync LRU cache and sets it to BME_LOCKED
1135  * @mdev:       DRBD device.
1136  * @sector:     The sector number.
1137  *
1138  * This functions sleeps on al_wait. Returns 0 on success, -EINTR if interrupted.
1139  */
1140 int drbd_rs_begin_io(struct drbd_conf *mdev, sector_t sector)
1141 {
1142         unsigned int enr = BM_SECT_TO_EXT(sector);
1143         struct bm_extent *bm_ext;
1144         int i, sig;
1145
1146         sig = wait_event_interruptible(mdev->al_wait,
1147                         (bm_ext = _bme_get(mdev, enr)));
1148         if (sig)
1149                 return -EINTR;
1150
1151         if (test_bit(BME_LOCKED, &bm_ext->flags))
1152                 return 0;
1153
1154         for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
1155                 sig = wait_event_interruptible(mdev->al_wait,
1156                                 !_is_in_al(mdev, enr * AL_EXT_PER_BM_SECT + i));
1157                 if (sig) {
1158                         spin_lock_irq(&mdev->al_lock);
1159                         if (lc_put(mdev->resync, &bm_ext->lce) == 0) {
1160                                 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1161                                 mdev->resync_locked--;
1162                                 wake_up(&mdev->al_wait);
1163                         }
1164                         spin_unlock_irq(&mdev->al_lock);
1165                         return -EINTR;
1166                 }
1167         }
1168         set_bit(BME_LOCKED, &bm_ext->flags);
1169         return 0;
1170 }
1171
1172 /**
1173  * drbd_try_rs_begin_io() - Gets an extent in the resync LRU cache, does not sleep
1174  * @mdev:       DRBD device.
1175  * @sector:     The sector number.
1176  *
1177  * Gets an extent in the resync LRU cache, sets it to BME_NO_WRITES, then
1178  * tries to set it to BME_LOCKED. Returns 0 upon success, and -EAGAIN
1179  * if there is still application IO going on in this area.
1180  */
1181 int drbd_try_rs_begin_io(struct drbd_conf *mdev, sector_t sector)
1182 {
1183         unsigned int enr = BM_SECT_TO_EXT(sector);
1184         const unsigned int al_enr = enr*AL_EXT_PER_BM_SECT;
1185         struct lc_element *e;
1186         struct bm_extent *bm_ext;
1187         int i;
1188
1189         spin_lock_irq(&mdev->al_lock);
1190         if (mdev->resync_wenr != LC_FREE && mdev->resync_wenr != enr) {
1191                 /* in case you have very heavy scattered io, it may
1192                  * stall the syncer undefined if we give up the ref count
1193                  * when we try again and requeue.
1194                  *
1195                  * if we don't give up the refcount, but the next time
1196                  * we are scheduled this extent has been "synced" by new
1197                  * application writes, we'd miss the lc_put on the
1198                  * extent we keep the refcount on.
1199                  * so we remembered which extent we had to try again, and
1200                  * if the next requested one is something else, we do
1201                  * the lc_put here...
1202                  * we also have to wake_up
1203                  */
1204                 e = lc_find(mdev->resync, mdev->resync_wenr);
1205                 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1206                 if (bm_ext) {
1207                         D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags));
1208                         D_ASSERT(test_bit(BME_NO_WRITES, &bm_ext->flags));
1209                         clear_bit(BME_NO_WRITES, &bm_ext->flags);
1210                         mdev->resync_wenr = LC_FREE;
1211                         if (lc_put(mdev->resync, &bm_ext->lce) == 0)
1212                                 mdev->resync_locked--;
1213                         wake_up(&mdev->al_wait);
1214                 } else {
1215                         dev_alert(DEV, "LOGIC BUG\n");
1216                 }
1217         }
1218         /* TRY. */
1219         e = lc_try_get(mdev->resync, enr);
1220         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1221         if (bm_ext) {
1222                 if (test_bit(BME_LOCKED, &bm_ext->flags))
1223                         goto proceed;
1224                 if (!test_and_set_bit(BME_NO_WRITES, &bm_ext->flags)) {
1225                         mdev->resync_locked++;
1226                 } else {
1227                         /* we did set the BME_NO_WRITES,
1228                          * but then could not set BME_LOCKED,
1229                          * so we tried again.
1230                          * drop the extra reference. */
1231                         bm_ext->lce.refcnt--;
1232                         D_ASSERT(bm_ext->lce.refcnt > 0);
1233                 }
1234                 goto check_al;
1235         } else {
1236                 /* do we rather want to try later? */
1237                 if (mdev->resync_locked > mdev->resync->nr_elements-3)
1238                         goto try_again;
1239                 /* Do or do not. There is no try. -- Yoda */
1240                 e = lc_get(mdev->resync, enr);
1241                 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1242                 if (!bm_ext) {
1243                         const unsigned long rs_flags = mdev->resync->flags;
1244                         if (rs_flags & LC_STARVING)
1245                                 dev_warn(DEV, "Have to wait for element"
1246                                      " (resync LRU too small?)\n");
1247                         BUG_ON(rs_flags & LC_DIRTY);
1248                         goto try_again;
1249                 }
1250                 if (bm_ext->lce.lc_number != enr) {
1251                         bm_ext->rs_left = drbd_bm_e_weight(mdev, enr);
1252                         bm_ext->rs_failed = 0;
1253                         lc_changed(mdev->resync, &bm_ext->lce);
1254                         wake_up(&mdev->al_wait);
1255                         D_ASSERT(test_bit(BME_LOCKED, &bm_ext->flags) == 0);
1256                 }
1257                 set_bit(BME_NO_WRITES, &bm_ext->flags);
1258                 D_ASSERT(bm_ext->lce.refcnt == 1);
1259                 mdev->resync_locked++;
1260                 goto check_al;
1261         }
1262 check_al:
1263         for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
1264                 if (unlikely(al_enr+i == mdev->act_log->new_number))
1265                         goto try_again;
1266                 if (lc_is_used(mdev->act_log, al_enr+i))
1267                         goto try_again;
1268         }
1269         set_bit(BME_LOCKED, &bm_ext->flags);
1270 proceed:
1271         mdev->resync_wenr = LC_FREE;
1272         spin_unlock_irq(&mdev->al_lock);
1273         return 0;
1274
1275 try_again:
1276         if (bm_ext)
1277                 mdev->resync_wenr = enr;
1278         spin_unlock_irq(&mdev->al_lock);
1279         return -EAGAIN;
1280 }
1281
1282 void drbd_rs_complete_io(struct drbd_conf *mdev, sector_t sector)
1283 {
1284         unsigned int enr = BM_SECT_TO_EXT(sector);
1285         struct lc_element *e;
1286         struct bm_extent *bm_ext;
1287         unsigned long flags;
1288
1289         spin_lock_irqsave(&mdev->al_lock, flags);
1290         e = lc_find(mdev->resync, enr);
1291         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1292         if (!bm_ext) {
1293                 spin_unlock_irqrestore(&mdev->al_lock, flags);
1294                 if (__ratelimit(&drbd_ratelimit_state))
1295                         dev_err(DEV, "drbd_rs_complete_io() called, but extent not found\n");
1296                 return;
1297         }
1298
1299         if (bm_ext->lce.refcnt == 0) {
1300                 spin_unlock_irqrestore(&mdev->al_lock, flags);
1301                 dev_err(DEV, "drbd_rs_complete_io(,%llu [=%u]) called, "
1302                     "but refcnt is 0!?\n",
1303                     (unsigned long long)sector, enr);
1304                 return;
1305         }
1306
1307         if (lc_put(mdev->resync, &bm_ext->lce) == 0) {
1308                 clear_bit(BME_LOCKED, &bm_ext->flags);
1309                 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1310                 mdev->resync_locked--;
1311                 wake_up(&mdev->al_wait);
1312         }
1313
1314         spin_unlock_irqrestore(&mdev->al_lock, flags);
1315 }
1316
1317 /**
1318  * drbd_rs_cancel_all() - Removes all extents from the resync LRU (even BME_LOCKED)
1319  * @mdev:       DRBD device.
1320  */
1321 void drbd_rs_cancel_all(struct drbd_conf *mdev)
1322 {
1323         spin_lock_irq(&mdev->al_lock);
1324
1325         if (get_ldev_if_state(mdev, D_FAILED)) { /* Makes sure ->resync is there. */
1326                 lc_reset(mdev->resync);
1327                 put_ldev(mdev);
1328         }
1329         mdev->resync_locked = 0;
1330         mdev->resync_wenr = LC_FREE;
1331         spin_unlock_irq(&mdev->al_lock);
1332         wake_up(&mdev->al_wait);
1333 }
1334
1335 /**
1336  * drbd_rs_del_all() - Gracefully remove all extents from the resync LRU
1337  * @mdev:       DRBD device.
1338  *
1339  * Returns 0 upon success, -EAGAIN if at least one reference count was
1340  * not zero.
1341  */
1342 int drbd_rs_del_all(struct drbd_conf *mdev)
1343 {
1344         struct lc_element *e;
1345         struct bm_extent *bm_ext;
1346         int i;
1347
1348         spin_lock_irq(&mdev->al_lock);
1349
1350         if (get_ldev_if_state(mdev, D_FAILED)) {
1351                 /* ok, ->resync is there. */
1352                 for (i = 0; i < mdev->resync->nr_elements; i++) {
1353                         e = lc_element_by_index(mdev->resync, i);
1354                         bm_ext = lc_entry(e, struct bm_extent, lce);
1355                         if (bm_ext->lce.lc_number == LC_FREE)
1356                                 continue;
1357                         if (bm_ext->lce.lc_number == mdev->resync_wenr) {
1358                                 dev_info(DEV, "dropping %u in drbd_rs_del_all, apparently"
1359                                      " got 'synced' by application io\n",
1360                                      mdev->resync_wenr);
1361                                 D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags));
1362                                 D_ASSERT(test_bit(BME_NO_WRITES, &bm_ext->flags));
1363                                 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1364                                 mdev->resync_wenr = LC_FREE;
1365                                 lc_put(mdev->resync, &bm_ext->lce);
1366                         }
1367                         if (bm_ext->lce.refcnt != 0) {
1368                                 dev_info(DEV, "Retrying drbd_rs_del_all() later. "
1369                                      "refcnt=%d\n", bm_ext->lce.refcnt);
1370                                 put_ldev(mdev);
1371                                 spin_unlock_irq(&mdev->al_lock);
1372                                 return -EAGAIN;
1373                         }
1374                         D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags));
1375                         D_ASSERT(!test_bit(BME_NO_WRITES, &bm_ext->flags));
1376                         lc_del(mdev->resync, &bm_ext->lce);
1377                 }
1378                 D_ASSERT(mdev->resync->used == 0);
1379                 put_ldev(mdev);
1380         }
1381         spin_unlock_irq(&mdev->al_lock);
1382
1383         return 0;
1384 }
1385
1386 /**
1387  * drbd_rs_failed_io() - Record information on a failure to resync the specified blocks
1388  * @mdev:       DRBD device.
1389  * @sector:     The sector number.
1390  * @size:       Size of failed IO operation, in byte.
1391  */
1392 void drbd_rs_failed_io(struct drbd_conf *mdev, sector_t sector, int size)
1393 {
1394         /* Is called from worker and receiver context _only_ */
1395         unsigned long sbnr, ebnr, lbnr;
1396         unsigned long count;
1397         sector_t esector, nr_sectors;
1398         int wake_up = 0;
1399
1400         if (size <= 0 || (size & 0x1ff) != 0 || size > DRBD_MAX_SEGMENT_SIZE) {
1401                 dev_err(DEV, "drbd_rs_failed_io: sector=%llus size=%d nonsense!\n",
1402                                 (unsigned long long)sector, size);
1403                 return;
1404         }
1405         nr_sectors = drbd_get_capacity(mdev->this_bdev);
1406         esector = sector + (size >> 9) - 1;
1407
1408         ERR_IF(sector >= nr_sectors) return;
1409         ERR_IF(esector >= nr_sectors) esector = (nr_sectors-1);
1410
1411         lbnr = BM_SECT_TO_BIT(nr_sectors-1);
1412
1413         /*
1414          * round up start sector, round down end sector.  we make sure we only
1415          * handle full, aligned, BM_BLOCK_SIZE (4K) blocks */
1416         if (unlikely(esector < BM_SECT_PER_BIT-1))
1417                 return;
1418         if (unlikely(esector == (nr_sectors-1)))
1419                 ebnr = lbnr;
1420         else
1421                 ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1));
1422         sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1);
1423
1424         if (sbnr > ebnr)
1425                 return;
1426
1427         /*
1428          * ok, (capacity & 7) != 0 sometimes, but who cares...
1429          * we count rs_{total,left} in bits, not sectors.
1430          */
1431         spin_lock_irq(&mdev->al_lock);
1432         count = drbd_bm_count_bits(mdev, sbnr, ebnr);
1433         if (count) {
1434                 mdev->rs_failed += count;
1435
1436                 if (get_ldev(mdev)) {
1437                         drbd_try_clear_on_disk_bm(mdev, sector, count, FALSE);
1438                         put_ldev(mdev);
1439                 }
1440
1441                 /* just wake_up unconditional now, various lc_chaged(),
1442                  * lc_put() in drbd_try_clear_on_disk_bm(). */
1443                 wake_up = 1;
1444         }
1445         spin_unlock_irq(&mdev->al_lock);
1446         if (wake_up)
1447                 wake_up(&mdev->al_wait);
1448 }