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