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_FUA, &mdev->flags))
  82                 rw |= REQ_FUA;
  83         rw |= REQ_SYNC;
  84
  85         bio = bio_alloc(GFP_NOIO, 1);
  86         bio->bi_bdev = bdev->md_bdev;
  87         bio->bi_sector = sector;
  88         ok = (bio_add_page(bio, page, size, 0) == size);
  89         if (!ok)
  90                 goto out;
  91         bio->bi_private = &md_io;
  92         bio->bi_end_io = drbd_md_io_complete;
  93         bio->bi_rw = rw;
  94
  95         if (FAULT_ACTIVE(mdev, (rw & WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD))
  96                 bio_endio(bio, -EIO);
  97         else
  98                 submit_bio(rw, bio);
  99         wait_for_completion(&md_io.event);
 100         ok = bio_flagged(bio, BIO_UPTODATE) && md_io.error == 0;
 101
 102  out:
 103         bio_put(bio);
 104         return ok;
 105 }
 106
 107 int drbd_md_sync_page_io(struct drbd_conf *mdev, struct drbd_backing_dev *bdev,
 108                          sector_t sector, int rw)
 109 {
 110         int logical_block_size, mask, ok;
 111         int offset = 0;
 112         struct page *iop = mdev->md_io_page;
 113
 114         D_ASSERT(mutex_is_locked(&mdev->md_io_mutex));
 115
 116         BUG_ON(!bdev->md_bdev);
 117
 118         logical_block_size = bdev_logical_block_size(bdev->md_bdev);
 119         if (logical_block_size == 0)
 120                 logical_block_size = MD_SECTOR_SIZE;
 121
 122         /* in case logical_block_size != 512 [ s390 only? ] */
 123         if (logical_block_size != MD_SECTOR_SIZE) {
 124                 mask = (logical_block_size / MD_SECTOR_SIZE) - 1;
 125                 D_ASSERT(mask == 1 || mask == 3 || mask == 7);
 126                 D_ASSERT(logical_block_size == (mask+1) * MD_SECTOR_SIZE);
 127                 offset = sector & mask;
 128                 sector = sector & ~mask;
 129                 iop = mdev->md_io_tmpp;
 130
 131                 if (rw & WRITE) {
 132                         /* these are GFP_KERNEL pages, pre-allocated
 133                          * on device initialization */
 134                         void *p = page_address(mdev->md_io_page);
 135                         void *hp = page_address(mdev->md_io_tmpp);
 136
 137                         ok = _drbd_md_sync_page_io(mdev, bdev, iop, sector,
 138                                         READ, logical_block_size);
 139
 140                         if (unlikely(!ok)) {
 141                                 dev_err(DEV, "drbd_md_sync_page_io(,%llus,"
 142                                     "READ [logical_block_size!=512]) failed!\n",
 143                                     (unsigned long long)sector);
 144                                 return 0;
 145                         }
 146
 147                         memcpy(hp + offset*MD_SECTOR_SIZE, p, MD_SECTOR_SIZE);
 148                 }
 149         }
 150
 151         if (sector < drbd_md_first_sector(bdev) ||
 152             sector > drbd_md_last_sector(bdev))
 153                 dev_alert(DEV, "%s [%d]:%s(,%llus,%s) out of range md access!\n",
 154                      current->comm, current->pid, __func__,
 155                      (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ");
 156
 157         ok = _drbd_md_sync_page_io(mdev, bdev, iop, sector, rw, logical_block_size);
 158         if (unlikely(!ok)) {
 159                 dev_err(DEV, "drbd_md_sync_page_io(,%llus,%s) failed!\n",
 160                     (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ");
 161                 return 0;
 162         }
 163
 164         if (logical_block_size != MD_SECTOR_SIZE && !(rw & WRITE)) {
 165                 void *p = page_address(mdev->md_io_page);
 166                 void *hp = page_address(mdev->md_io_tmpp);
 167
 168                 memcpy(p, hp + offset*MD_SECTOR_SIZE, MD_SECTOR_SIZE);
 169         }
 170
 171         return ok;
 172 }
 173
 174 static struct lc_element *_al_get(struct drbd_conf *mdev, unsigned int enr)
 175 {
 176         struct lc_element *al_ext;
 177         struct lc_element *tmp;
 178         unsigned long     al_flags = 0;
 179         int wake;
 180
 181         spin_lock_irq(&mdev->al_lock);
 182         tmp = lc_find(mdev->resync, enr/AL_EXT_PER_BM_SECT);
 183         if (unlikely(tmp != NULL)) {
 184                 struct bm_extent  *bm_ext = lc_entry(tmp, struct bm_extent, lce);
 185                 if (test_bit(BME_NO_WRITES, &bm_ext->flags)) {
 186                         wake = !test_and_set_bit(BME_PRIORITY, &bm_ext->flags);
 187                         spin_unlock_irq(&mdev->al_lock);
 188                         if (wake)
 189                                 wake_up(&mdev->al_wait);
 190                         return NULL;
 191                 }
 192         }
 193         al_ext   = lc_get(mdev->act_log, enr);
 194         al_flags = mdev->act_log->flags;
 195         spin_unlock_irq(&mdev->al_lock);
 196
 197         /*
 198         if (!al_ext) {
 199                 if (al_flags & LC_STARVING)
 200                         dev_warn(DEV, "Have to wait for LRU element (AL too small?)\n");
 201                 if (al_flags & LC_DIRTY)
 202                         dev_warn(DEV, "Ongoing AL update (AL device too slow?)\n");
 203         }
 204         */
 205
 206         return al_ext;
 207 }
 208
 209 void drbd_al_begin_io(struct drbd_conf *mdev, sector_t sector)
 210 {
 211         unsigned int enr = (sector >> (AL_EXTENT_SHIFT-9));
 212         struct lc_element *al_ext;
 213         struct update_al_work al_work;
 214
 215         D_ASSERT(atomic_read(&mdev->local_cnt) > 0);
 216
 217         wait_event(mdev->al_wait, (al_ext = _al_get(mdev, enr)));
 218
 219         if (al_ext->lc_number != enr) {
 220                 /* drbd_al_write_transaction(mdev,al_ext,enr);
 221                  * recurses into generic_make_request(), which
 222                  * disallows recursion, bios being serialized on the
 223                  * current->bio_tail list now.
 224                  * we have to delegate updates to the activity log
 225                  * to the worker thread. */
 226                 init_completion(&al_work.event);
 227                 al_work.al_ext = al_ext;
 228                 al_work.enr = enr;
 229                 al_work.old_enr = al_ext->lc_number;
 230                 al_work.w.cb = w_al_write_transaction;
 231                 drbd_queue_work_front(&mdev->data.work, &al_work.w);
 232                 wait_for_completion(&al_work.event);
 233
 234                 mdev->al_writ_cnt++;
 235
 236                 spin_lock_irq(&mdev->al_lock);
 237                 lc_changed(mdev->act_log, al_ext);
 238                 spin_unlock_irq(&mdev->al_lock);
 239                 wake_up(&mdev->al_wait);
 240         }
 241 }
 242
 243 void drbd_al_complete_io(struct drbd_conf *mdev, sector_t sector)
 244 {
 245         unsigned int enr = (sector >> (AL_EXTENT_SHIFT-9));
 246         struct lc_element *extent;
 247         unsigned long flags;
 248
 249         spin_lock_irqsave(&mdev->al_lock, flags);
 250
 251         extent = lc_find(mdev->act_log, enr);
 252
 253         if (!extent) {
 254                 spin_unlock_irqrestore(&mdev->al_lock, flags);
 255                 dev_err(DEV, "al_complete_io() called on inactive extent %u\n", enr);
 256                 return;
 257         }
 258
 259         if (lc_put(mdev->act_log, extent) == 0)
 260                 wake_up(&mdev->al_wait);
 261
 262         spin_unlock_irqrestore(&mdev->al_lock, flags);
 263 }
 264
 265 int
 266 w_al_write_transaction(struct drbd_conf *mdev, struct drbd_work *w, int unused)
 267 {
 268         struct update_al_work *aw = container_of(w, struct update_al_work, w);
 269         struct lc_element *updated = aw->al_ext;
 270         const unsigned int new_enr = aw->enr;
 271         const unsigned int evicted = aw->old_enr;
 272         struct al_transaction *buffer;
 273         sector_t sector;
 274         int i, n, mx;
 275         unsigned int extent_nr;
 276         u32 xor_sum = 0;
 277
 278         if (!get_ldev(mdev)) {
 279                 dev_err(DEV,
 280                         "disk is %s, cannot start al transaction (-%d +%d)\n",
 281                         drbd_disk_str(mdev->state.disk), evicted, new_enr);
 282                 complete(&((struct update_al_work *)w)->event);
 283                 return 1;
 284         }
 285         /* do we have to do a bitmap write, first?
 286          * TODO reduce maximum latency:
 287          * submit both bios, then wait for both,
 288          * instead of doing two synchronous sector writes.
 289          * For now, we must not write the transaction,
 290          * if we cannot write out the bitmap of the evicted extent. */
 291         if (mdev->state.conn < C_CONNECTED && evicted != LC_FREE)
 292                 drbd_bm_write_sect(mdev, evicted/AL_EXT_PER_BM_SECT);
 293
 294         /* The bitmap write may have failed, causing a state change. */
 295         if (mdev->state.disk < D_INCONSISTENT) {
 296                 dev_err(DEV,
 297                         "disk is %s, cannot write al transaction (-%d +%d)\n",
 298                         drbd_disk_str(mdev->state.disk), evicted, new_enr);
 299                 complete(&((struct update_al_work *)w)->event);
 300                 put_ldev(mdev);
 301                 return 1;
 302         }
 303
 304         mutex_lock(&mdev->md_io_mutex); /* protects md_io_buffer, al_tr_cycle, ... */
 305         buffer = (struct al_transaction *)page_address(mdev->md_io_page);
 306
 307         buffer->magic = __constant_cpu_to_be32(DRBD_MAGIC);
 308         buffer->tr_number = cpu_to_be32(mdev->al_tr_number);
 309
 310         n = lc_index_of(mdev->act_log, updated);
 311
 312         buffer->updates[0].pos = cpu_to_be32(n);
 313         buffer->updates[0].extent = cpu_to_be32(new_enr);
 314
 315         xor_sum ^= new_enr;
 316
 317         mx = min_t(int, AL_EXTENTS_PT,
 318                    mdev->act_log->nr_elements - mdev->al_tr_cycle);
 319         for (i = 0; i < mx; i++) {
 320                 unsigned idx = mdev->al_tr_cycle + i;
 321                 extent_nr = lc_element_by_index(mdev->act_log, idx)->lc_number;
 322                 buffer->updates[i+1].pos = cpu_to_be32(idx);
 323                 buffer->updates[i+1].extent = cpu_to_be32(extent_nr);
 324                 xor_sum ^= extent_nr;
 325         }
 326         for (; i < AL_EXTENTS_PT; i++) {
 327                 buffer->updates[i+1].pos = __constant_cpu_to_be32(-1);
 328                 buffer->updates[i+1].extent = __constant_cpu_to_be32(LC_FREE);
 329                 xor_sum ^= LC_FREE;
 330         }
 331         mdev->al_tr_cycle += AL_EXTENTS_PT;
 332         if (mdev->al_tr_cycle >= mdev->act_log->nr_elements)
 333                 mdev->al_tr_cycle = 0;
 334
 335         buffer->xor_sum = cpu_to_be32(xor_sum);
 336
 337         sector =  mdev->ldev->md.md_offset
 338                 + mdev->ldev->md.al_offset + mdev->al_tr_pos;
 339
 340         if (!drbd_md_sync_page_io(mdev, mdev->ldev, sector, WRITE))
 341                 drbd_chk_io_error(mdev, 1, TRUE);
 342
 343         if (++mdev->al_tr_pos >
 344             div_ceil(mdev->act_log->nr_elements, AL_EXTENTS_PT))
 345                 mdev->al_tr_pos = 0;
 346
 347         D_ASSERT(mdev->al_tr_pos < MD_AL_MAX_SIZE);
 348         mdev->al_tr_number++;
 349
 350         mutex_unlock(&mdev->md_io_mutex);
 351
 352         complete(&((struct update_al_work *)w)->event);
 353         put_ldev(mdev);
 354
 355         return 1;
 356 }
 357
 358 /**
 359  * drbd_al_read_tr() - Read a single transaction from the on disk activity log
 360  * @mdev:       DRBD device.
 361  * @bdev:       Block device to read form.
 362  * @b:          pointer to an al_transaction.
 363  * @index:      On disk slot of the transaction to read.
 364  *
 365  * Returns -1 on IO error, 0 on checksum error and 1 upon success.
 366  */
 367 static int drbd_al_read_tr(struct drbd_conf *mdev,
 368                            struct drbd_backing_dev *bdev,
 369                            struct al_transaction *b,
 370                            int index)
 371 {
 372         sector_t sector;
 373         int rv, i;
 374         u32 xor_sum = 0;
 375
 376         sector = bdev->md.md_offset + bdev->md.al_offset + index;
 377
 378         /* Dont process error normally,
 379          * as this is done before disk is attached! */
 380         if (!drbd_md_sync_page_io(mdev, bdev, sector, READ))
 381                 return -1;
 382
 383         rv = (be32_to_cpu(b->magic) == DRBD_MAGIC);
 384
 385         for (i = 0; i < AL_EXTENTS_PT + 1; i++)
 386                 xor_sum ^= be32_to_cpu(b->updates[i].extent);
 387         rv &= (xor_sum == be32_to_cpu(b->xor_sum));
 388
 389         return rv;
 390 }
 391
 392 /**
 393  * drbd_al_read_log() - Restores the activity log from its on disk representation.
 394  * @mdev:       DRBD device.
 395  * @bdev:       Block device to read form.
 396  *
 397  * Returns 1 on success, returns 0 when reading the log failed due to IO errors.
 398  */
 399 int drbd_al_read_log(struct drbd_conf *mdev, struct drbd_backing_dev *bdev)
 400 {
 401         struct al_transaction *buffer;
 402         int i;
 403         int rv;
 404         int mx;
 405         int active_extents = 0;
 406         int transactions = 0;
 407         int found_valid = 0;
 408         int from = 0;
 409         int to = 0;
 410         u32 from_tnr = 0;
 411         u32 to_tnr = 0;
 412         u32 cnr;
 413
 414         mx = div_ceil(mdev->act_log->nr_elements, AL_EXTENTS_PT);
 415
 416         /* lock out all other meta data io for now,
 417          * and make sure the page is mapped.
 418          */
 419         mutex_lock(&mdev->md_io_mutex);
 420         buffer = page_address(mdev->md_io_page);
 421
 422         /* Find the valid transaction in the log */
 423         for (i = 0; i <= mx; i++) {
 424                 rv = drbd_al_read_tr(mdev, bdev, buffer, i);
 425                 if (rv == 0)
 426                         continue;
 427                 if (rv == -1) {
 428                         mutex_unlock(&mdev->md_io_mutex);
 429                         return 0;
 430                 }
 431                 cnr = be32_to_cpu(buffer->tr_number);
 432
 433                 if (++found_valid == 1) {
 434                         from = i;
 435                         to = i;
 436                         from_tnr = cnr;
 437                         to_tnr = cnr;
 438                         continue;
 439                 }
 440                 if ((int)cnr - (int)from_tnr < 0) {
 441                         D_ASSERT(from_tnr - cnr + i - from == mx+1);
 442                         from = i;
 443                         from_tnr = cnr;
 444                 }
 445                 if ((int)cnr - (int)to_tnr > 0) {
 446                         D_ASSERT(cnr - to_tnr == i - to);
 447                         to = i;
 448                         to_tnr = cnr;
 449                 }
 450         }
 451
 452         if (!found_valid) {
 453                 dev_warn(DEV, "No usable activity log found.\n");
 454                 mutex_unlock(&mdev->md_io_mutex);
 455                 return 1;
 456         }
 457
 458         /* Read the valid transactions.
 459          * dev_info(DEV, "Reading from %d to %d.\n",from,to); */
 460         i = from;
 461         while (1) {
 462                 int j, pos;
 463                 unsigned int extent_nr;
 464                 unsigned int trn;
 465
 466                 rv = drbd_al_read_tr(mdev, bdev, buffer, i);
 467                 ERR_IF(rv == 0) goto cancel;
 468                 if (rv == -1) {
 469                         mutex_unlock(&mdev->md_io_mutex);
 470                         return 0;
 471                 }
 472
 473                 trn = be32_to_cpu(buffer->tr_number);
 474
 475                 spin_lock_irq(&mdev->al_lock);
 476
 477                 /* This loop runs backwards because in the cyclic
 478                    elements there might be an old version of the
 479                    updated element (in slot 0). So the element in slot 0
 480                    can overwrite old versions. */
 481                 for (j = AL_EXTENTS_PT; j >= 0; j--) {
 482                         pos = be32_to_cpu(buffer->updates[j].pos);
 483                         extent_nr = be32_to_cpu(buffer->updates[j].extent);
 484
 485                         if (extent_nr == LC_FREE)
 486                                 continue;
 487
 488                         lc_set(mdev->act_log, extent_nr, pos);
 489                         active_extents++;
 490                 }
 491                 spin_unlock_irq(&mdev->al_lock);
 492
 493                 transactions++;
 494
 495 cancel:
 496                 if (i == to)
 497                         break;
 498                 i++;
 499                 if (i > mx)
 500                         i = 0;
 501         }
 502
 503         mdev->al_tr_number = to_tnr+1;
 504         mdev->al_tr_pos = to;
 505         if (++mdev->al_tr_pos >
 506             div_ceil(mdev->act_log->nr_elements, AL_EXTENTS_PT))
 507                 mdev->al_tr_pos = 0;
 508
 509         /* ok, we are done with it */
 510         mutex_unlock(&mdev->md_io_mutex);
 511
 512         dev_info(DEV, "Found %d transactions (%d active extents) in activity log.\n",
 513              transactions, active_extents);
 514
 515         return 1;
 516 }
 517
 518 static void atodb_endio(struct bio *bio, int error)
 519 {
 520         struct drbd_atodb_wait *wc = bio->bi_private;
 521         struct drbd_conf *mdev = wc->mdev;
 522         struct page *page;
 523         int uptodate = bio_flagged(bio, BIO_UPTODATE);
 524
 525         /* strange behavior of some lower level drivers...
 526          * fail the request by clearing the uptodate flag,
 527          * but do not return any error?! */
 528         if (!error && !uptodate)
 529                 error = -EIO;
 530
 531         drbd_chk_io_error(mdev, error, TRUE);
 532         if (error && wc->error == 0)
 533                 wc->error = error;
 534
 535         if (atomic_dec_and_test(&wc->count))
 536                 complete(&wc->io_done);
 537
 538         page = bio->bi_io_vec[0].bv_page;
 539         put_page(page);
 540         bio_put(bio);
 541         mdev->bm_writ_cnt++;
 542         put_ldev(mdev);
 543 }
 544
 545 /* sector to word */
 546 #define S2W(s)  ((s)<<(BM_EXT_SHIFT-BM_BLOCK_SHIFT-LN2_BPL))
 547
 548 /* activity log to on disk bitmap -- prepare bio unless that sector
 549  * is already covered by previously prepared bios */
 550 static int atodb_prepare_unless_covered(struct drbd_conf *mdev,
 551                                         struct bio **bios,
 552                                         unsigned int enr,
 553                                         struct drbd_atodb_wait *wc) __must_hold(local)
 554 {
 555         struct bio *bio;
 556         struct page *page;
 557         sector_t on_disk_sector;
 558         unsigned int page_offset = PAGE_SIZE;
 559         int offset;
 560         int i = 0;
 561         int err = -ENOMEM;
 562
 563         /* We always write aligned, full 4k blocks,
 564          * so we can ignore the logical_block_size (for now) */
 565         enr &= ~7U;
 566         on_disk_sector = enr + mdev->ldev->md.md_offset
 567                              + mdev->ldev->md.bm_offset;
 568
 569         D_ASSERT(!(on_disk_sector & 7U));
 570
 571         /* Check if that enr is already covered by an already created bio.
 572          * Caution, bios[] is not NULL terminated,
 573          * but only initialized to all NULL.
 574          * For completely scattered activity log,
 575          * the last invocation iterates over all bios,
 576          * and finds the last NULL entry.
 577          */
 578         while ((bio = bios[i])) {
 579                 if (bio->bi_sector == on_disk_sector)
 580                         return 0;
 581                 i++;
 582         }
 583         /* bios[i] == NULL, the next not yet used slot */
 584
 585         /* GFP_KERNEL, we are not in the write-out path */
 586         bio = bio_alloc(GFP_KERNEL, 1);
 587         if (bio == NULL)
 588                 return -ENOMEM;
 589
 590         if (i > 0) {
 591                 const struct bio_vec *prev_bv = bios[i-1]->bi_io_vec;
 592                 page_offset = prev_bv->bv_offset + prev_bv->bv_len;
 593                 page = prev_bv->bv_page;
 594         }
 595         if (page_offset == PAGE_SIZE) {
 596                 page = alloc_page(__GFP_HIGHMEM);
 597                 if (page == NULL)
 598                         goto out_bio_put;
 599                 page_offset = 0;
 600         } else {
 601                 get_page(page);
 602         }
 603
 604         offset = S2W(enr);
 605         drbd_bm_get_lel(mdev, offset,
 606                         min_t(size_t, S2W(8), drbd_bm_words(mdev) - offset),
 607                         kmap(page) + page_offset);
 608         kunmap(page);
 609
 610         bio->bi_private = wc;
 611         bio->bi_end_io = atodb_endio;
 612         bio->bi_bdev = mdev->ldev->md_bdev;
 613         bio->bi_sector = on_disk_sector;
 614
 615         if (bio_add_page(bio, page, 4096, page_offset) != 4096)
 616                 goto out_put_page;
 617
 618         atomic_inc(&wc->count);
 619         /* we already know that we may do this...
 620          * get_ldev_if_state(mdev,D_ATTACHING);
 621          * just get the extra reference, so that the local_cnt reflects
 622          * the number of pending IO requests DRBD at its backing device.
 623          */
 624         atomic_inc(&mdev->local_cnt);
 625
 626         bios[i] = bio;
 627
 628         return 0;
 629
 630 out_put_page:
 631         err = -EINVAL;
 632         put_page(page);
 633 out_bio_put:
 634         bio_put(bio);
 635         return err;
 636 }
 637
 638 /**
 639  * drbd_al_to_on_disk_bm() -  * Writes bitmap parts covered by active AL extents
 640  * @mdev:       DRBD device.
 641  *
 642  * Called when we detach (unconfigure) local storage,
 643  * or when we go from R_PRIMARY to R_SECONDARY role.
 644  */
 645 void drbd_al_to_on_disk_bm(struct drbd_conf *mdev)
 646 {
 647         int i, nr_elements;
 648         unsigned int enr;
 649         struct bio **bios;
 650         struct drbd_atodb_wait wc;
 651
 652         ERR_IF (!get_ldev_if_state(mdev, D_ATTACHING))
 653                 return; /* sorry, I don't have any act_log etc... */
 654
 655         wait_event(mdev->al_wait, lc_try_lock(mdev->act_log));
 656
 657         nr_elements = mdev->act_log->nr_elements;
 658
 659         /* GFP_KERNEL, we are not in anyone's write-out path */
 660         bios = kzalloc(sizeof(struct bio *) * nr_elements, GFP_KERNEL);
 661         if (!bios)
 662                 goto submit_one_by_one;
 663
 664         atomic_set(&wc.count, 0);
 665         init_completion(&wc.io_done);
 666         wc.mdev = mdev;
 667         wc.error = 0;
 668
 669         for (i = 0; i < nr_elements; i++) {
 670                 enr = lc_element_by_index(mdev->act_log, i)->lc_number;
 671                 if (enr == LC_FREE)
 672                         continue;
 673                 /* next statement also does atomic_inc wc.count and local_cnt */
 674                 if (atodb_prepare_unless_covered(mdev, bios,
 675                                                 enr/AL_EXT_PER_BM_SECT,
 676                                                 &wc))
 677                         goto free_bios_submit_one_by_one;
 678         }
 679
 680         /* unnecessary optimization? */
 681         lc_unlock(mdev->act_log);
 682         wake_up(&mdev->al_wait);
 683
 684         /* all prepared, submit them */
 685         for (i = 0; i < nr_elements; i++) {
 686                 if (bios[i] == NULL)
 687                         break;
 688                 if (FAULT_ACTIVE(mdev, DRBD_FAULT_MD_WR)) {
 689                         bios[i]->bi_rw = WRITE;
 690                         bio_endio(bios[i], -EIO);
 691                 } else {
 692                         submit_bio(WRITE, bios[i]);
 693                 }
 694         }
 695
 696         /* always (try to) flush bitmap to stable storage */
 697         drbd_md_flush(mdev);
 698
 699         /* In case we did not submit a single IO do not wait for
 700          * them to complete. ( Because we would wait forever here. )
 701          *
 702          * In case we had IOs and they are already complete, there
 703          * is not point in waiting anyways.
 704          * Therefore this if () ... */
 705         if (atomic_read(&wc.count))
 706                 wait_for_completion(&wc.io_done);
 707
 708         put_ldev(mdev);
 709
 710         kfree(bios);
 711         return;
 712
 713  free_bios_submit_one_by_one:
 714         /* free everything by calling the endio callback directly. */
 715         for (i = 0; i < nr_elements && bios[i]; i++)
 716                 bio_endio(bios[i], 0);
 717
 718         kfree(bios);
 719
 720  submit_one_by_one:
 721         dev_warn(DEV, "Using the slow drbd_al_to_on_disk_bm()\n");
 722
 723         for (i = 0; i < mdev->act_log->nr_elements; i++) {
 724                 enr = lc_element_by_index(mdev->act_log, i)->lc_number;
 725                 if (enr == LC_FREE)
 726                         continue;
 727                 /* Really slow: if we have al-extents 16..19 active,
 728                  * sector 4 will be written four times! Synchronous! */
 729                 drbd_bm_write_sect(mdev, enr/AL_EXT_PER_BM_SECT);
 730         }
 731
 732         lc_unlock(mdev->act_log);
 733         wake_up(&mdev->al_wait);
 734         put_ldev(mdev);
 735 }
 736
 737 /**
 738  * drbd_al_apply_to_bm() - Sets the bitmap to diry(1) where covered ba active AL extents
 739  * @mdev:       DRBD device.
 740  */
 741 void drbd_al_apply_to_bm(struct drbd_conf *mdev)
 742 {
 743         unsigned int enr;
 744         unsigned long add = 0;
 745         char ppb[10];
 746         int i, tmp;
 747
 748         wait_event(mdev->al_wait, lc_try_lock(mdev->act_log));
 749
 750         for (i = 0; i < mdev->act_log->nr_elements; i++) {
 751                 enr = lc_element_by_index(mdev->act_log, i)->lc_number;
 752                 if (enr == LC_FREE)
 753                         continue;
 754                 tmp = drbd_bm_ALe_set_all(mdev, enr);
 755                 dynamic_dev_dbg(DEV, "AL: set %d bits in extent %u\n", tmp, enr);
 756                 add += tmp;
 757         }
 758
 759         lc_unlock(mdev->act_log);
 760         wake_up(&mdev->al_wait);
 761
 762         dev_info(DEV, "Marked additional %s as out-of-sync based on AL.\n",
 763              ppsize(ppb, Bit2KB(add)));
 764 }
 765
 766 static int _try_lc_del(struct drbd_conf *mdev, struct lc_element *al_ext)
 767 {
 768         int rv;
 769
 770         spin_lock_irq(&mdev->al_lock);
 771         rv = (al_ext->refcnt == 0);
 772         if (likely(rv))
 773                 lc_del(mdev->act_log, al_ext);
 774         spin_unlock_irq(&mdev->al_lock);
 775
 776         return rv;
 777 }
 778
 779 /**
 780  * drbd_al_shrink() - Removes all active extents form the activity log
 781  * @mdev:       DRBD device.
 782  *
 783  * Removes all active extents form the activity log, waiting until
 784  * the reference count of each entry dropped to 0 first, of course.
 785  *
 786  * You need to lock mdev->act_log with lc_try_lock() / lc_unlock()
 787  */
 788 void drbd_al_shrink(struct drbd_conf *mdev)
 789 {
 790         struct lc_element *al_ext;
 791         int i;
 792
 793         D_ASSERT(test_bit(__LC_DIRTY, &mdev->act_log->flags));
 794
 795         for (i = 0; i < mdev->act_log->nr_elements; i++) {
 796                 al_ext = lc_element_by_index(mdev->act_log, i);
 797                 if (al_ext->lc_number == LC_FREE)
 798                         continue;
 799                 wait_event(mdev->al_wait, _try_lc_del(mdev, al_ext));
 800         }
 801
 802         wake_up(&mdev->al_wait);
 803 }
 804
 805 static int w_update_odbm(struct drbd_conf *mdev, struct drbd_work *w, int unused)
 806 {
 807         struct update_odbm_work *udw = container_of(w, struct update_odbm_work, w);
 808
 809         if (!get_ldev(mdev)) {
 810                 if (__ratelimit(&drbd_ratelimit_state))
 811                         dev_warn(DEV, "Can not update on disk bitmap, local IO disabled.\n");
 812                 kfree(udw);
 813                 return 1;
 814         }
 815
 816         drbd_bm_write_sect(mdev, udw->enr);
 817         put_ldev(mdev);
 818
 819         kfree(udw);
 820
 821         if (drbd_bm_total_weight(mdev) <= mdev->rs_failed) {
 822                 switch (mdev->state.conn) {
 823                 case C_SYNC_SOURCE:  case C_SYNC_TARGET:
 824                 case C_PAUSED_SYNC_S: case C_PAUSED_SYNC_T:
 825                         drbd_resync_finished(mdev);
 826                 default:
 827                         /* nothing to do */
 828                         break;
 829                 }
 830         }
 831         drbd_bcast_sync_progress(mdev);
 832
 833         return 1;
 834 }
 835
 836
 837 /* ATTENTION. The AL's extents are 4MB each, while the extents in the
 838  * resync LRU-cache are 16MB each.
 839  * The caller of this function has to hold an get_ldev() reference.
 840  *
 841  * TODO will be obsoleted once we have a caching lru of the on disk bitmap
 842  */
 843 static void drbd_try_clear_on_disk_bm(struct drbd_conf *mdev, sector_t sector,
 844                                       int count, int success)
 845 {
 846         struct lc_element *e;
 847         struct update_odbm_work *udw;
 848
 849         unsigned int enr;
 850
 851         D_ASSERT(atomic_read(&mdev->local_cnt));
 852
 853         /* I simply assume that a sector/size pair never crosses
 854          * a 16 MB extent border. (Currently this is true...) */
 855         enr = BM_SECT_TO_EXT(sector);
 856
 857         e = lc_get(mdev->resync, enr);
 858         if (e) {
 859                 struct bm_extent *ext = lc_entry(e, struct bm_extent, lce);
 860                 if (ext->lce.lc_number == enr) {
 861                         if (success)
 862                                 ext->rs_left -= count;
 863                         else
 864                                 ext->rs_failed += count;
 865                         if (ext->rs_left < ext->rs_failed) {
 866                                 dev_err(DEV, "BAD! sector=%llus enr=%u rs_left=%d "
 867                                     "rs_failed=%d count=%d\n",
 868                                      (unsigned long long)sector,
 869                                      ext->lce.lc_number, ext->rs_left,
 870                                      ext->rs_failed, count);
 871                                 dump_stack();
 872
 873                                 lc_put(mdev->resync, &ext->lce);
 874                                 drbd_force_state(mdev, NS(conn, C_DISCONNECTING));
 875                                 return;
 876                         }
 877                 } else {
 878                         /* Normally this element should be in the cache,
 879                          * since drbd_rs_begin_io() pulled it already in.
 880                          *
 881                          * But maybe an application write finished, and we set
 882                          * something outside the resync lru_cache in sync.
 883                          */
 884                         int rs_left = drbd_bm_e_weight(mdev, enr);
 885                         if (ext->flags != 0) {
 886                                 dev_warn(DEV, "changing resync lce: %d[%u;%02lx]"
 887                                      " -> %d[%u;00]\n",
 888                                      ext->lce.lc_number, ext->rs_left,
 889                                      ext->flags, enr, rs_left);
 890                                 ext->flags = 0;
 891                         }
 892                         if (ext->rs_failed) {
 893                                 dev_warn(DEV, "Kicking resync_lru element enr=%u "
 894                                      "out with rs_failed=%d\n",
 895                                      ext->lce.lc_number, ext->rs_failed);
 896                                 set_bit(WRITE_BM_AFTER_RESYNC, &mdev->flags);
 897                         }
 898                         ext->rs_left = rs_left;
 899                         ext->rs_failed = success ? 0 : count;
 900                         lc_changed(mdev->resync, &ext->lce);
 901                 }
 902                 lc_put(mdev->resync, &ext->lce);
 903                 /* no race, we are within the al_lock! */
 904
 905                 if (ext->rs_left == ext->rs_failed) {
 906                         ext->rs_failed = 0;
 907
 908                         udw = kmalloc(sizeof(*udw), GFP_ATOMIC);
 909                         if (udw) {
 910                                 udw->enr = ext->lce.lc_number;
 911                                 udw->w.cb = w_update_odbm;
 912                                 drbd_queue_work_front(&mdev->data.work, &udw->w);
 913                         } else {
 914                                 dev_warn(DEV, "Could not kmalloc an udw\n");
 915                                 set_bit(WRITE_BM_AFTER_RESYNC, &mdev->flags);
 916                         }
 917                 }
 918         } else {
 919                 dev_err(DEV, "lc_get() failed! locked=%d/%d flags=%lu\n",
 920                     mdev->resync_locked,
 921                     mdev->resync->nr_elements,
 922                     mdev->resync->flags);
 923         }
 924 }
 925
 926 void drbd_advance_rs_marks(struct drbd_conf *mdev, unsigned long still_to_go)
 927 {
 928         unsigned long now = jiffies;
 929         unsigned long last = mdev->rs_mark_time[mdev->rs_last_mark];
 930         int next = (mdev->rs_last_mark + 1) % DRBD_SYNC_MARKS;
 931         if (time_after_eq(now, last + DRBD_SYNC_MARK_STEP)) {
 932                 if (mdev->rs_mark_left[mdev->rs_last_mark] != still_to_go &&
 933                     mdev->state.conn != C_PAUSED_SYNC_T &&
 934                     mdev->state.conn != C_PAUSED_SYNC_S) {
 935                         mdev->rs_mark_time[next] = now;
 936                         mdev->rs_mark_left[next] = still_to_go;
 937                         mdev->rs_last_mark = next;
 938                 }
 939         }
 940 }
 941
 942 /* clear the bit corresponding to the piece of storage in question:
 943  * size byte of data starting from sector.  Only clear a bits of the affected
 944  * one ore more _aligned_ BM_BLOCK_SIZE blocks.
 945  *
 946  * called by worker on C_SYNC_TARGET and receiver on SyncSource.
 947  *
 948  */
 949 void __drbd_set_in_sync(struct drbd_conf *mdev, sector_t sector, int size,
 950                        const char *file, const unsigned int line)
 951 {
 952         /* Is called from worker and receiver context _only_ */
 953         unsigned long sbnr, ebnr, lbnr;
 954         unsigned long count = 0;
 955         sector_t esector, nr_sectors;
 956         int wake_up = 0;
 957         unsigned long flags;
 958
 959         if (size <= 0 || (size & 0x1ff) != 0 || size > DRBD_MAX_BIO_SIZE) {
 960                 dev_err(DEV, "drbd_set_in_sync: sector=%llus size=%d nonsense!\n",
 961                                 (unsigned long long)sector, size);
 962                 return;
 963         }
 964         nr_sectors = drbd_get_capacity(mdev->this_bdev);
 965         esector = sector + (size >> 9) - 1;
 966
 967         ERR_IF(sector >= nr_sectors) return;
 968         ERR_IF(esector >= nr_sectors) esector = (nr_sectors-1);
 969
 970         lbnr = BM_SECT_TO_BIT(nr_sectors-1);
 971
 972         /* we clear it (in sync).
 973          * round up start sector, round down end sector.  we make sure we only
 974          * clear full, aligned, BM_BLOCK_SIZE (4K) blocks */
 975         if (unlikely(esector < BM_SECT_PER_BIT-1))
 976                 return;
 977         if (unlikely(esector == (nr_sectors-1)))
 978                 ebnr = lbnr;
 979         else
 980                 ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1));
 981         sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1);
 982
 983         if (sbnr > ebnr)
 984                 return;
 985
 986         /*
 987          * ok, (capacity & 7) != 0 sometimes, but who cares...
 988          * we count rs_{total,left} in bits, not sectors.
 989          */
 990         count = drbd_bm_clear_bits(mdev, sbnr, ebnr);
 991         if (count && get_ldev(mdev)) {
 992                 drbd_advance_rs_marks(mdev, drbd_bm_total_weight(mdev));
 993                 spin_lock_irqsave(&mdev->al_lock, flags);
 994                 drbd_try_clear_on_disk_bm(mdev, sector, count, TRUE);
 995                 spin_unlock_irqrestore(&mdev->al_lock, flags);
 996
 997                 /* just wake_up unconditional now, various lc_chaged(),
 998                  * lc_put() in drbd_try_clear_on_disk_bm(). */
 999                 wake_up = 1;
1000                 put_ldev(mdev);
1001         }
1002         if (wake_up)
1003                 wake_up(&mdev->al_wait);
1004 }
1005
1006 /*
1007  * this is intended to set one request worth of data out of sync.
1008  * affects at least 1 bit,
1009  * and at most 1+DRBD_MAX_BIO_SIZE/BM_BLOCK_SIZE bits.
1010  *
1011  * called by tl_clear and drbd_send_dblock (==drbd_make_request).
1012  * so this can be _any_ process.
1013  */
1014 int __drbd_set_out_of_sync(struct drbd_conf *mdev, sector_t sector, int size,
1015                             const char *file, const unsigned int line)
1016 {
1017         unsigned long sbnr, ebnr, lbnr, flags;
1018         sector_t esector, nr_sectors;
1019         unsigned int enr, count = 0;
1020         struct lc_element *e;
1021
1022         if (size <= 0 || (size & 0x1ff) != 0 || size > DRBD_MAX_BIO_SIZE) {
1023                 dev_err(DEV, "sector: %llus, size: %d\n",
1024                         (unsigned long long)sector, size);
1025                 return 0;
1026         }
1027
1028         if (!get_ldev(mdev))
1029                 return 0; /* no disk, no metadata, no bitmap to set bits in */
1030
1031         nr_sectors = drbd_get_capacity(mdev->this_bdev);
1032         esector = sector + (size >> 9) - 1;
1033
1034         ERR_IF(sector >= nr_sectors)
1035                 goto out;
1036         ERR_IF(esector >= nr_sectors)
1037                 esector = (nr_sectors-1);
1038
1039         lbnr = BM_SECT_TO_BIT(nr_sectors-1);
1040
1041         /* we set it out of sync,
1042          * we do not need to round anything here */
1043         sbnr = BM_SECT_TO_BIT(sector);
1044         ebnr = BM_SECT_TO_BIT(esector);
1045
1046         /* ok, (capacity & 7) != 0 sometimes, but who cares...
1047          * we count rs_{total,left} in bits, not sectors.  */
1048         spin_lock_irqsave(&mdev->al_lock, flags);
1049         count = drbd_bm_set_bits(mdev, sbnr, ebnr);
1050
1051         enr = BM_SECT_TO_EXT(sector);
1052         e = lc_find(mdev->resync, enr);
1053         if (e)
1054                 lc_entry(e, struct bm_extent, lce)->rs_left += count;
1055         spin_unlock_irqrestore(&mdev->al_lock, flags);
1056
1057 out:
1058         put_ldev(mdev);
1059
1060         return count;
1061 }
1062
1063 static
1064 struct bm_extent *_bme_get(struct drbd_conf *mdev, unsigned int enr)
1065 {
1066         struct lc_element *e;
1067         struct bm_extent *bm_ext;
1068         int wakeup = 0;
1069         unsigned long rs_flags;
1070
1071         spin_lock_irq(&mdev->al_lock);
1072         if (mdev->resync_locked > mdev->resync->nr_elements/2) {
1073                 spin_unlock_irq(&mdev->al_lock);
1074                 return NULL;
1075         }
1076         e = lc_get(mdev->resync, enr);
1077         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1078         if (bm_ext) {
1079                 if (bm_ext->lce.lc_number != enr) {
1080                         bm_ext->rs_left = drbd_bm_e_weight(mdev, enr);
1081                         bm_ext->rs_failed = 0;
1082                         lc_changed(mdev->resync, &bm_ext->lce);
1083                         wakeup = 1;
1084                 }
1085                 if (bm_ext->lce.refcnt == 1)
1086                         mdev->resync_locked++;
1087                 set_bit(BME_NO_WRITES, &bm_ext->flags);
1088         }
1089         rs_flags = mdev->resync->flags;
1090         spin_unlock_irq(&mdev->al_lock);
1091         if (wakeup)
1092                 wake_up(&mdev->al_wait);
1093
1094         if (!bm_ext) {
1095                 if (rs_flags & LC_STARVING)
1096                         dev_warn(DEV, "Have to wait for element"
1097                              " (resync LRU too small?)\n");
1098                 BUG_ON(rs_flags & LC_DIRTY);
1099         }
1100
1101         return bm_ext;
1102 }
1103
1104 static int _is_in_al(struct drbd_conf *mdev, unsigned int enr)
1105 {
1106         struct lc_element *al_ext;
1107         int rv = 0;
1108
1109         spin_lock_irq(&mdev->al_lock);
1110         if (unlikely(enr == mdev->act_log->new_number))
1111                 rv = 1;
1112         else {
1113                 al_ext = lc_find(mdev->act_log, enr);
1114                 if (al_ext) {
1115                         if (al_ext->refcnt)
1116                                 rv = 1;
1117                 }
1118         }
1119         spin_unlock_irq(&mdev->al_lock);
1120
1121         /*
1122         if (unlikely(rv)) {
1123                 dev_info(DEV, "Delaying sync read until app's write is done\n");
1124         }
1125         */
1126         return rv;
1127 }
1128
1129 /**
1130  * drbd_rs_begin_io() - Gets an extent in the resync LRU cache and sets it to BME_LOCKED
1131  * @mdev:       DRBD device.
1132  * @sector:     The sector number.
1133  *
1134  * This functions sleeps on al_wait. Returns 0 on success, -EINTR if interrupted.
1135  */
1136 int drbd_rs_begin_io(struct drbd_conf *mdev, sector_t sector)
1137 {
1138         unsigned int enr = BM_SECT_TO_EXT(sector);
1139         struct bm_extent *bm_ext;
1140         int i, sig;
1141         int sa = 200; /* Step aside 200 times, then grab the extent and let app-IO wait.
1142                          200 times -> 20 seconds. */
1143
1144 retry:
1145         sig = wait_event_interruptible(mdev->al_wait,
1146                         (bm_ext = _bme_get(mdev, enr)));
1147         if (sig)
1148                 return -EINTR;
1149
1150         if (test_bit(BME_LOCKED, &bm_ext->flags))
1151                 return 0;
1152
1153         for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
1154                 sig = wait_event_interruptible(mdev->al_wait,
1155                                                !_is_in_al(mdev, enr * AL_EXT_PER_BM_SECT + i) ||
1156                                                test_bit(BME_PRIORITY, &bm_ext->flags));
1157
1158                 if (sig || (test_bit(BME_PRIORITY, &bm_ext->flags) && sa)) {
1159                         spin_lock_irq(&mdev->al_lock);
1160                         if (lc_put(mdev->resync, &bm_ext->lce) == 0) {
1161                                 bm_ext->flags = 0; /* clears BME_NO_WRITES and eventually BME_PRIORITY */
1162                                 mdev->resync_locked--;
1163                                 wake_up(&mdev->al_wait);
1164                         }
1165                         spin_unlock_irq(&mdev->al_lock);
1166                         if (sig)
1167                                 return -EINTR;
1168                         if (schedule_timeout_interruptible(HZ/10))
1169                                 return -EINTR;
1170                         if (sa && --sa == 0)
1171                                 dev_warn(DEV,"drbd_rs_begin_io() stepped aside for 20sec."
1172                                          "Resync stalled?\n");
1173                         goto retry;
1174                 }
1175         }
1176         set_bit(BME_LOCKED, &bm_ext->flags);
1177         return 0;
1178 }
1179
1180 /**
1181  * drbd_try_rs_begin_io() - Gets an extent in the resync LRU cache, does not sleep
1182  * @mdev:       DRBD device.
1183  * @sector:     The sector number.
1184  *
1185  * Gets an extent in the resync LRU cache, sets it to BME_NO_WRITES, then
1186  * tries to set it to BME_LOCKED. Returns 0 upon success, and -EAGAIN
1187  * if there is still application IO going on in this area.
1188  */
1189 int drbd_try_rs_begin_io(struct drbd_conf *mdev, sector_t sector)
1190 {
1191         unsigned int enr = BM_SECT_TO_EXT(sector);
1192         const unsigned int al_enr = enr*AL_EXT_PER_BM_SECT;
1193         struct lc_element *e;
1194         struct bm_extent *bm_ext;
1195         int i;
1196
1197         spin_lock_irq(&mdev->al_lock);
1198         if (mdev->resync_wenr != LC_FREE && mdev->resync_wenr != enr) {
1199                 /* in case you have very heavy scattered io, it may
1200                  * stall the syncer undefined if we give up the ref count
1201                  * when we try again and requeue.
1202                  *
1203                  * if we don't give up the refcount, but the next time
1204                  * we are scheduled this extent has been "synced" by new
1205                  * application writes, we'd miss the lc_put on the
1206                  * extent we keep the refcount on.
1207                  * so we remembered which extent we had to try again, and
1208                  * if the next requested one is something else, we do
1209                  * the lc_put here...
1210                  * we also have to wake_up
1211                  */
1212                 e = lc_find(mdev->resync, mdev->resync_wenr);
1213                 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1214                 if (bm_ext) {
1215                         D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags));
1216                         D_ASSERT(test_bit(BME_NO_WRITES, &bm_ext->flags));
1217                         clear_bit(BME_NO_WRITES, &bm_ext->flags);
1218                         mdev->resync_wenr = LC_FREE;
1219                         if (lc_put(mdev->resync, &bm_ext->lce) == 0)
1220                                 mdev->resync_locked--;
1221                         wake_up(&mdev->al_wait);
1222                 } else {
1223                         dev_alert(DEV, "LOGIC BUG\n");
1224                 }
1225         }
1226         /* TRY. */
1227         e = lc_try_get(mdev->resync, enr);
1228         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1229         if (bm_ext) {
1230                 if (test_bit(BME_LOCKED, &bm_ext->flags))
1231                         goto proceed;
1232                 if (!test_and_set_bit(BME_NO_WRITES, &bm_ext->flags)) {
1233                         mdev->resync_locked++;
1234                 } else {
1235                         /* we did set the BME_NO_WRITES,
1236                          * but then could not set BME_LOCKED,
1237                          * so we tried again.
1238                          * drop the extra reference. */
1239                         bm_ext->lce.refcnt--;
1240                         D_ASSERT(bm_ext->lce.refcnt > 0);
1241                 }
1242                 goto check_al;
1243         } else {
1244                 /* do we rather want to try later? */
1245                 if (mdev->resync_locked > mdev->resync->nr_elements-3)
1246                         goto try_again;
1247                 /* Do or do not. There is no try. -- Yoda */
1248                 e = lc_get(mdev->resync, enr);
1249                 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1250                 if (!bm_ext) {
1251                         const unsigned long rs_flags = mdev->resync->flags;
1252                         if (rs_flags & LC_STARVING)
1253                                 dev_warn(DEV, "Have to wait for element"
1254                                      " (resync LRU too small?)\n");
1255                         BUG_ON(rs_flags & LC_DIRTY);
1256                         goto try_again;
1257                 }
1258                 if (bm_ext->lce.lc_number != enr) {
1259                         bm_ext->rs_left = drbd_bm_e_weight(mdev, enr);
1260                         bm_ext->rs_failed = 0;
1261                         lc_changed(mdev->resync, &bm_ext->lce);
1262                         wake_up(&mdev->al_wait);
1263                         D_ASSERT(test_bit(BME_LOCKED, &bm_ext->flags) == 0);
1264                 }
1265                 set_bit(BME_NO_WRITES, &bm_ext->flags);
1266                 D_ASSERT(bm_ext->lce.refcnt == 1);
1267                 mdev->resync_locked++;
1268                 goto check_al;
1269         }
1270 check_al:
1271         for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
1272                 if (unlikely(al_enr+i == mdev->act_log->new_number))
1273                         goto try_again;
1274                 if (lc_is_used(mdev->act_log, al_enr+i))
1275                         goto try_again;
1276         }
1277         set_bit(BME_LOCKED, &bm_ext->flags);
1278 proceed:
1279         mdev->resync_wenr = LC_FREE;
1280         spin_unlock_irq(&mdev->al_lock);
1281         return 0;
1282
1283 try_again:
1284         if (bm_ext)
1285                 mdev->resync_wenr = enr;
1286         spin_unlock_irq(&mdev->al_lock);
1287         return -EAGAIN;
1288 }
1289
1290 void drbd_rs_complete_io(struct drbd_conf *mdev, sector_t sector)
1291 {
1292         unsigned int enr = BM_SECT_TO_EXT(sector);
1293         struct lc_element *e;
1294         struct bm_extent *bm_ext;
1295         unsigned long flags;
1296
1297         spin_lock_irqsave(&mdev->al_lock, flags);
1298         e = lc_find(mdev->resync, enr);
1299         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1300         if (!bm_ext) {
1301                 spin_unlock_irqrestore(&mdev->al_lock, flags);
1302                 if (__ratelimit(&drbd_ratelimit_state))
1303                         dev_err(DEV, "drbd_rs_complete_io() called, but extent not found\n");
1304                 return;
1305         }
1306
1307         if (bm_ext->lce.refcnt == 0) {
1308                 spin_unlock_irqrestore(&mdev->al_lock, flags);
1309                 dev_err(DEV, "drbd_rs_complete_io(,%llu [=%u]) called, "
1310                     "but refcnt is 0!?\n",
1311                     (unsigned long long)sector, enr);
1312                 return;
1313         }
1314
1315         if (lc_put(mdev->resync, &bm_ext->lce) == 0) {
1316                 bm_ext->flags = 0; /* clear BME_LOCKED, BME_NO_WRITES and BME_PRIORITY */
1317                 mdev->resync_locked--;
1318                 wake_up(&mdev->al_wait);
1319         }
1320
1321         spin_unlock_irqrestore(&mdev->al_lock, flags);
1322 }
1323
1324 /**
1325  * drbd_rs_cancel_all() - Removes all extents from the resync LRU (even BME_LOCKED)
1326  * @mdev:       DRBD device.
1327  */
1328 void drbd_rs_cancel_all(struct drbd_conf *mdev)
1329 {
1330         spin_lock_irq(&mdev->al_lock);
1331
1332         if (get_ldev_if_state(mdev, D_FAILED)) { /* Makes sure ->resync is there. */
1333                 lc_reset(mdev->resync);
1334                 put_ldev(mdev);
1335         }
1336         mdev->resync_locked = 0;
1337         mdev->resync_wenr = LC_FREE;
1338         spin_unlock_irq(&mdev->al_lock);
1339         wake_up(&mdev->al_wait);
1340 }
1341
1342 /**
1343  * drbd_rs_del_all() - Gracefully remove all extents from the resync LRU
1344  * @mdev:       DRBD device.
1345  *
1346  * Returns 0 upon success, -EAGAIN if at least one reference count was
1347  * not zero.
1348  */
1349 int drbd_rs_del_all(struct drbd_conf *mdev)
1350 {
1351         struct lc_element *e;
1352         struct bm_extent *bm_ext;
1353         int i;
1354
1355         spin_lock_irq(&mdev->al_lock);
1356
1357         if (get_ldev_if_state(mdev, D_FAILED)) {
1358                 /* ok, ->resync is there. */
1359                 for (i = 0; i < mdev->resync->nr_elements; i++) {
1360                         e = lc_element_by_index(mdev->resync, i);
1361                         bm_ext = lc_entry(e, struct bm_extent, lce);
1362                         if (bm_ext->lce.lc_number == LC_FREE)
1363                                 continue;
1364                         if (bm_ext->lce.lc_number == mdev->resync_wenr) {
1365                                 dev_info(DEV, "dropping %u in drbd_rs_del_all, apparently"
1366                                      " got 'synced' by application io\n",
1367                                      mdev->resync_wenr);
1368                                 D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags));
1369                                 D_ASSERT(test_bit(BME_NO_WRITES, &bm_ext->flags));
1370                                 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1371                                 mdev->resync_wenr = LC_FREE;
1372                                 lc_put(mdev->resync, &bm_ext->lce);
1373                         }
1374                         if (bm_ext->lce.refcnt != 0) {
1375                                 dev_info(DEV, "Retrying drbd_rs_del_all() later. "
1376                                      "refcnt=%d\n", bm_ext->lce.refcnt);
1377                                 put_ldev(mdev);
1378                                 spin_unlock_irq(&mdev->al_lock);
1379                                 return -EAGAIN;
1380                         }
1381                         D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags));
1382                         D_ASSERT(!test_bit(BME_NO_WRITES, &bm_ext->flags));
1383                         lc_del(mdev->resync, &bm_ext->lce);
1384                 }
1385                 D_ASSERT(mdev->resync->used == 0);
1386                 put_ldev(mdev);
1387         }
1388         spin_unlock_irq(&mdev->al_lock);
1389
1390         return 0;
1391 }
1392
1393 /**
1394  * drbd_rs_failed_io() - Record information on a failure to resync the specified blocks
1395  * @mdev:       DRBD device.
1396  * @sector:     The sector number.
1397  * @size:       Size of failed IO operation, in byte.
1398  */
1399 void drbd_rs_failed_io(struct drbd_conf *mdev, sector_t sector, int size)
1400 {
1401         /* Is called from worker and receiver context _only_ */
1402         unsigned long sbnr, ebnr, lbnr;
1403         unsigned long count;
1404         sector_t esector, nr_sectors;
1405         int wake_up = 0;
1406
1407         if (size <= 0 || (size & 0x1ff) != 0 || size > DRBD_MAX_BIO_SIZE) {
1408                 dev_err(DEV, "drbd_rs_failed_io: sector=%llus size=%d nonsense!\n",
1409                                 (unsigned long long)sector, size);
1410                 return;
1411         }
1412         nr_sectors = drbd_get_capacity(mdev->this_bdev);
1413         esector = sector + (size >> 9) - 1;
1414
1415         ERR_IF(sector >= nr_sectors) return;
1416         ERR_IF(esector >= nr_sectors) esector = (nr_sectors-1);
1417
1418         lbnr = BM_SECT_TO_BIT(nr_sectors-1);
1419
1420         /*
1421          * round up start sector, round down end sector.  we make sure we only
1422          * handle full, aligned, BM_BLOCK_SIZE (4K) blocks */
1423         if (unlikely(esector < BM_SECT_PER_BIT-1))
1424                 return;
1425         if (unlikely(esector == (nr_sectors-1)))
1426                 ebnr = lbnr;
1427         else
1428                 ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1));
1429         sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1);
1430
1431         if (sbnr > ebnr)
1432                 return;
1433
1434         /*
1435          * ok, (capacity & 7) != 0 sometimes, but who cares...
1436          * we count rs_{total,left} in bits, not sectors.
1437          */
1438         spin_lock_irq(&mdev->al_lock);
1439         count = drbd_bm_count_bits(mdev, sbnr, ebnr);
1440         if (count) {
1441                 mdev->rs_failed += count;
1442
1443                 if (get_ldev(mdev)) {
1444                         drbd_try_clear_on_disk_bm(mdev, sector, count, FALSE);
1445                         put_ldev(mdev);
1446                 }
1447
1448                 /* just wake_up unconditional now, various lc_chaged(),
1449                  * lc_put() in drbd_try_clear_on_disk_bm(). */
1450                 wake_up = 1;
1451         }
1452         spin_unlock_irq(&mdev->al_lock);
1453         if (wake_up)
1454                 wake_up(&mdev->al_wait);
1455 }