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/crc32c.h>
  28 #include <linux/drbd.h>
  29 #include <linux/drbd_limits.h>
  30 #include <linux/dynamic_debug.h>
  31 #include "drbd_int.h"
  32
  33
  34 enum al_transaction_types {
  35         AL_TR_UPDATE = 0,
  36         AL_TR_INITIALIZED = 0xffff
  37 };
  38 /* all fields on disc in big endian */
  39 struct __packed al_transaction_on_disk {
  40         /* don't we all like magic */
  41         __be32  magic;
  42
  43         /* to identify the most recent transaction block
  44          * in the on disk ring buffer */
  45         __be32  tr_number;
  46
  47         /* checksum on the full 4k block, with this field set to 0. */
  48         __be32  crc32c;
  49
  50         /* type of transaction, special transaction types like:
  51          * purge-all, set-all-idle, set-all-active, ... to-be-defined
  52          * see also enum al_transaction_types */
  53         __be16  transaction_type;
  54
  55         /* we currently allow only a few thousand extents,
  56          * so 16bit will be enough for the slot number. */
  57
  58         /* how many updates in this transaction */
  59         __be16  n_updates;
  60
  61         /* maximum slot number, "al-extents" in drbd.conf speak.
  62          * Having this in each transaction should make reconfiguration
  63          * of that parameter easier. */
  64         __be16  context_size;
  65
  66         /* slot number the context starts with */
  67         __be16  context_start_slot_nr;
  68
  69         /* Some reserved bytes.  Expected usage is a 64bit counter of
  70          * sectors-written since device creation, and other data generation tag
  71          * supporting usage */
  72         __be32  __reserved[4];
  73
  74         /* --- 36 byte used --- */
  75
  76         /* Reserve space for up to AL_UPDATES_PER_TRANSACTION changes
  77          * in one transaction, then use the remaining byte in the 4k block for
  78          * context information.  "Flexible" number of updates per transaction
  79          * does not help, as we have to account for the case when all update
  80          * slots are used anyways, so it would only complicate code without
  81          * additional benefit.
  82          */
  83         __be16  update_slot_nr[AL_UPDATES_PER_TRANSACTION];
  84
  85         /* but the extent number is 32bit, which at an extent size of 4 MiB
  86          * allows to cover device sizes of up to 2**54 Byte (16 PiB) */
  87         __be32  update_extent_nr[AL_UPDATES_PER_TRANSACTION];
  88
  89         /* --- 420 bytes used (36 + 64*6) --- */
  90
  91         /* 4096 - 420 = 3676 = 919 * 4 */
  92         __be32  context[AL_CONTEXT_PER_TRANSACTION];
  93 };
  94
  95 void *drbd_md_get_buffer(struct drbd_device *device, const char *intent)
  96 {
  97         int r;
  98
  99         wait_event(device->misc_wait,
 100                    (r = atomic_cmpxchg(&device->md_io.in_use, 0, 1)) == 0 ||
 101                    device->state.disk <= D_FAILED);
 102
 103         if (r)
 104                 return NULL;
 105
 106         device->md_io.current_use = intent;
 107         device->md_io.start_jif = jiffies;
 108         device->md_io.submit_jif = device->md_io.start_jif - 1;
 109         return page_address(device->md_io.page);
 110 }
 111
 112 void drbd_md_put_buffer(struct drbd_device *device)
 113 {
 114         if (atomic_dec_and_test(&device->md_io.in_use))
 115                 wake_up(&device->misc_wait);
 116 }
 117
 118 void wait_until_done_or_force_detached(struct drbd_device *device, struct drbd_backing_dev *bdev,
 119                                      unsigned int *done)
 120 {
 121         long dt;
 122
 123         rcu_read_lock();
 124         dt = rcu_dereference(bdev->disk_conf)->disk_timeout;
 125         rcu_read_unlock();
 126         dt = dt * HZ / 10;
 127         if (dt == 0)
 128                 dt = MAX_SCHEDULE_TIMEOUT;
 129
 130         dt = wait_event_timeout(device->misc_wait,
 131                         *done || test_bit(FORCE_DETACH, &device->flags), dt);
 132         if (dt == 0) {
 133                 drbd_err(device, "meta-data IO operation timed out\n");
 134                 drbd_chk_io_error(device, 1, DRBD_FORCE_DETACH);
 135         }
 136 }
 137
 138 static int _drbd_md_sync_page_io(struct drbd_device *device,
 139                                  struct drbd_backing_dev *bdev,
 140                                  struct page *page, sector_t sector,
 141                                  int rw, int size)
 142 {
 143         struct bio *bio;
 144         int err;
 145
 146         device->md_io.done = 0;
 147         device->md_io.error = -ENODEV;
 148
 149         if ((rw & WRITE) && !test_bit(MD_NO_FUA, &device->flags))
 150                 rw |= REQ_FUA | REQ_FLUSH;
 151         rw |= REQ_SYNC | REQ_NOIDLE;
 152
 153         bio = bio_alloc_drbd(GFP_NOIO);
 154         bio->bi_bdev = bdev->md_bdev;
 155         bio->bi_iter.bi_sector = sector;
 156         err = -EIO;
 157         if (bio_add_page(bio, page, size, 0) != size)
 158                 goto out;
 159         bio->bi_private = device;
 160         bio->bi_end_io = drbd_md_io_complete;
 161         bio->bi_rw = rw;
 162
 163         if (!(rw & WRITE) && device->state.disk == D_DISKLESS && device->ldev == NULL)
 164                 /* special case, drbd_md_read() during drbd_adm_attach(): no get_ldev */
 165                 ;
 166         else if (!get_ldev_if_state(device, D_ATTACHING)) {
 167                 /* Corresponding put_ldev in drbd_md_io_complete() */
 168                 drbd_err(device, "ASSERT FAILED: get_ldev_if_state() == 1 in _drbd_md_sync_page_io()\n");
 169                 err = -ENODEV;
 170                 goto out;
 171         }
 172
 173         bio_get(bio); /* one bio_put() is in the completion handler */
 174         atomic_inc(&device->md_io.in_use); /* drbd_md_put_buffer() is in the completion handler */
 175         device->md_io.submit_jif = jiffies;
 176         if (drbd_insert_fault(device, (rw & WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD))
 177                 bio_endio(bio, -EIO);
 178         else
 179                 submit_bio(rw, bio);
 180         wait_until_done_or_force_detached(device, bdev, &device->md_io.done);
 181         if (bio_flagged(bio, BIO_UPTODATE))
 182                 err = device->md_io.error;
 183
 184  out:
 185         bio_put(bio);
 186         return err;
 187 }
 188
 189 int drbd_md_sync_page_io(struct drbd_device *device, struct drbd_backing_dev *bdev,
 190                          sector_t sector, int rw)
 191 {
 192         int err;
 193         struct page *iop = device->md_io.page;
 194
 195         D_ASSERT(device, atomic_read(&device->md_io.in_use) == 1);
 196
 197         BUG_ON(!bdev->md_bdev);
 198
 199         dynamic_drbd_dbg(device, "meta_data io: %s [%d]:%s(,%llus,%s) %pS\n",
 200              current->comm, current->pid, __func__,
 201              (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ",
 202              (void*)_RET_IP_ );
 203
 204         if (sector < drbd_md_first_sector(bdev) ||
 205             sector + 7 > drbd_md_last_sector(bdev))
 206                 drbd_alert(device, "%s [%d]:%s(,%llus,%s) out of range md access!\n",
 207                      current->comm, current->pid, __func__,
 208                      (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ");
 209
 210         /* we do all our meta data IO in aligned 4k blocks. */
 211         err = _drbd_md_sync_page_io(device, bdev, iop, sector, rw, 4096);
 212         if (err) {
 213                 drbd_err(device, "drbd_md_sync_page_io(,%llus,%s) failed with error %d\n",
 214                     (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ", err);
 215         }
 216         return err;
 217 }
 218
 219 static struct bm_extent *find_active_resync_extent(struct drbd_device *device, unsigned int enr)
 220 {
 221         struct lc_element *tmp;
 222         tmp = lc_find(device->resync, enr/AL_EXT_PER_BM_SECT);
 223         if (unlikely(tmp != NULL)) {
 224                 struct bm_extent  *bm_ext = lc_entry(tmp, struct bm_extent, lce);
 225                 if (test_bit(BME_NO_WRITES, &bm_ext->flags))
 226                         return bm_ext;
 227         }
 228         return NULL;
 229 }
 230
 231 static struct lc_element *_al_get(struct drbd_device *device, unsigned int enr, bool nonblock)
 232 {
 233         struct lc_element *al_ext;
 234         struct bm_extent *bm_ext;
 235         int wake;
 236
 237         spin_lock_irq(&device->al_lock);
 238         bm_ext = find_active_resync_extent(device, enr);
 239         if (bm_ext) {
 240                 wake = !test_and_set_bit(BME_PRIORITY, &bm_ext->flags);
 241                 spin_unlock_irq(&device->al_lock);
 242                 if (wake)
 243                         wake_up(&device->al_wait);
 244                 return NULL;
 245         }
 246         if (nonblock)
 247                 al_ext = lc_try_get(device->act_log, enr);
 248         else
 249                 al_ext = lc_get(device->act_log, enr);
 250         spin_unlock_irq(&device->al_lock);
 251         return al_ext;
 252 }
 253
 254 bool drbd_al_begin_io_fastpath(struct drbd_device *device, struct drbd_interval *i)
 255 {
 256         /* for bios crossing activity log extent boundaries,
 257          * we may need to activate two extents in one go */
 258         unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
 259         unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
 260
 261         D_ASSERT(device, (unsigned)(last - first) <= 1);
 262         D_ASSERT(device, atomic_read(&device->local_cnt) > 0);
 263
 264         /* FIXME figure out a fast path for bios crossing AL extent boundaries */
 265         if (first != last)
 266                 return false;
 267
 268         return _al_get(device, first, true);
 269 }
 270
 271 bool drbd_al_begin_io_prepare(struct drbd_device *device, struct drbd_interval *i)
 272 {
 273         /* for bios crossing activity log extent boundaries,
 274          * we may need to activate two extents in one go */
 275         unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
 276         unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
 277         unsigned enr;
 278         bool need_transaction = false;
 279
 280         D_ASSERT(device, first <= last);
 281         D_ASSERT(device, atomic_read(&device->local_cnt) > 0);
 282
 283         for (enr = first; enr <= last; enr++) {
 284                 struct lc_element *al_ext;
 285                 wait_event(device->al_wait,
 286                                 (al_ext = _al_get(device, enr, false)) != NULL);
 287                 if (al_ext->lc_number != enr)
 288                         need_transaction = true;
 289         }
 290         return need_transaction;
 291 }
 292
 293 static int al_write_transaction(struct drbd_device *device);
 294
 295 void drbd_al_begin_io_commit(struct drbd_device *device)
 296 {
 297         bool locked = false;
 298
 299         /* Serialize multiple transactions.
 300          * This uses test_and_set_bit, memory barrier is implicit.
 301          */
 302         wait_event(device->al_wait,
 303                         device->act_log->pending_changes == 0 ||
 304                         (locked = lc_try_lock_for_transaction(device->act_log)));
 305
 306         if (locked) {
 307                 /* Double check: it may have been committed by someone else,
 308                  * while we have been waiting for the lock. */
 309                 if (device->act_log->pending_changes) {
 310                         bool write_al_updates;
 311
 312                         rcu_read_lock();
 313                         write_al_updates = rcu_dereference(device->ldev->disk_conf)->al_updates;
 314                         rcu_read_unlock();
 315
 316                         if (write_al_updates)
 317                                 al_write_transaction(device);
 318                         spin_lock_irq(&device->al_lock);
 319                         /* FIXME
 320                         if (err)
 321                                 we need an "lc_cancel" here;
 322                         */
 323                         lc_committed(device->act_log);
 324                         spin_unlock_irq(&device->al_lock);
 325                 }
 326                 lc_unlock(device->act_log);
 327                 wake_up(&device->al_wait);
 328         }
 329 }
 330
 331 /*
 332  * @delegate:   delegate activity log I/O to the worker thread
 333  */
 334 void drbd_al_begin_io(struct drbd_device *device, struct drbd_interval *i)
 335 {
 336         if (drbd_al_begin_io_prepare(device, i))
 337                 drbd_al_begin_io_commit(device);
 338 }
 339
 340 int drbd_al_begin_io_nonblock(struct drbd_device *device, struct drbd_interval *i)
 341 {
 342         struct lru_cache *al = device->act_log;
 343         /* for bios crossing activity log extent boundaries,
 344          * we may need to activate two extents in one go */
 345         unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
 346         unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
 347         unsigned nr_al_extents;
 348         unsigned available_update_slots;
 349         unsigned enr;
 350
 351         D_ASSERT(device, first <= last);
 352
 353         nr_al_extents = 1 + last - first; /* worst case: all touched extends are cold. */
 354         available_update_slots = min(al->nr_elements - al->used,
 355                                 al->max_pending_changes - al->pending_changes);
 356
 357         /* We want all necessary updates for a given request within the same transaction
 358          * We could first check how many updates are *actually* needed,
 359          * and use that instead of the worst-case nr_al_extents */
 360         if (available_update_slots < nr_al_extents)
 361                 return -EWOULDBLOCK;
 362
 363         /* Is resync active in this area? */
 364         for (enr = first; enr <= last; enr++) {
 365                 struct lc_element *tmp;
 366                 tmp = lc_find(device->resync, enr/AL_EXT_PER_BM_SECT);
 367                 if (unlikely(tmp != NULL)) {
 368                         struct bm_extent  *bm_ext = lc_entry(tmp, struct bm_extent, lce);
 369                         if (test_bit(BME_NO_WRITES, &bm_ext->flags)) {
 370                                 if (!test_and_set_bit(BME_PRIORITY, &bm_ext->flags))
 371                                         return -EBUSY;
 372                                 return -EWOULDBLOCK;
 373                         }
 374                 }
 375         }
 376
 377         /* Checkout the refcounts.
 378          * Given that we checked for available elements and update slots above,
 379          * this has to be successful. */
 380         for (enr = first; enr <= last; enr++) {
 381                 struct lc_element *al_ext;
 382                 al_ext = lc_get_cumulative(device->act_log, enr);
 383                 if (!al_ext)
 384                         drbd_info(device, "LOGIC BUG for enr=%u\n", enr);
 385         }
 386         return 0;
 387 }
 388
 389 void drbd_al_complete_io(struct drbd_device *device, struct drbd_interval *i)
 390 {
 391         /* for bios crossing activity log extent boundaries,
 392          * we may need to activate two extents in one go */
 393         unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
 394         unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
 395         unsigned enr;
 396         struct lc_element *extent;
 397         unsigned long flags;
 398
 399         D_ASSERT(device, first <= last);
 400         spin_lock_irqsave(&device->al_lock, flags);
 401
 402         for (enr = first; enr <= last; enr++) {
 403                 extent = lc_find(device->act_log, enr);
 404                 if (!extent) {
 405                         drbd_err(device, "al_complete_io() called on inactive extent %u\n", enr);
 406                         continue;
 407                 }
 408                 lc_put(device->act_log, extent);
 409         }
 410         spin_unlock_irqrestore(&device->al_lock, flags);
 411         wake_up(&device->al_wait);
 412 }
 413
 414 #if (PAGE_SHIFT + 3) < (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT)
 415 /* Currently BM_BLOCK_SHIFT, BM_EXT_SHIFT and AL_EXTENT_SHIFT
 416  * are still coupled, or assume too much about their relation.
 417  * Code below will not work if this is violated.
 418  * Will be cleaned up with some followup patch.
 419  */
 420 # error FIXME
 421 #endif
 422
 423 static unsigned int al_extent_to_bm_page(unsigned int al_enr)
 424 {
 425         return al_enr >>
 426                 /* bit to page */
 427                 ((PAGE_SHIFT + 3) -
 428                 /* al extent number to bit */
 429                  (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT));
 430 }
 431
 432 static sector_t al_tr_number_to_on_disk_sector(struct drbd_device *device)
 433 {
 434         const unsigned int stripes = device->ldev->md.al_stripes;
 435         const unsigned int stripe_size_4kB = device->ldev->md.al_stripe_size_4k;
 436
 437         /* transaction number, modulo on-disk ring buffer wrap around */
 438         unsigned int t = device->al_tr_number % (device->ldev->md.al_size_4k);
 439
 440         /* ... to aligned 4k on disk block */
 441         t = ((t % stripes) * stripe_size_4kB) + t/stripes;
 442
 443         /* ... to 512 byte sector in activity log */
 444         t *= 8;
 445
 446         /* ... plus offset to the on disk position */
 447         return device->ldev->md.md_offset + device->ldev->md.al_offset + t;
 448 }
 449
 450 int al_write_transaction(struct drbd_device *device)
 451 {
 452         struct al_transaction_on_disk *buffer;
 453         struct lc_element *e;
 454         sector_t sector;
 455         int i, mx;
 456         unsigned extent_nr;
 457         unsigned crc = 0;
 458         int err = 0;
 459
 460         if (!get_ldev(device)) {
 461                 drbd_err(device, "disk is %s, cannot start al transaction\n",
 462                         drbd_disk_str(device->state.disk));
 463                 return -EIO;
 464         }
 465
 466         /* The bitmap write may have failed, causing a state change. */
 467         if (device->state.disk < D_INCONSISTENT) {
 468                 drbd_err(device,
 469                         "disk is %s, cannot write al transaction\n",
 470                         drbd_disk_str(device->state.disk));
 471                 put_ldev(device);
 472                 return -EIO;
 473         }
 474
 475         /* protects md_io_buffer, al_tr_cycle, ... */
 476         buffer = drbd_md_get_buffer(device, __func__);
 477         if (!buffer) {
 478                 drbd_err(device, "disk failed while waiting for md_io buffer\n");
 479                 put_ldev(device);
 480                 return -ENODEV;
 481         }
 482
 483         memset(buffer, 0, sizeof(*buffer));
 484         buffer->magic = cpu_to_be32(DRBD_AL_MAGIC);
 485         buffer->tr_number = cpu_to_be32(device->al_tr_number);
 486
 487         i = 0;
 488
 489         /* Even though no one can start to change this list
 490          * once we set the LC_LOCKED -- from drbd_al_begin_io(),
 491          * lc_try_lock_for_transaction() --, someone may still
 492          * be in the process of changing it. */
 493         spin_lock_irq(&device->al_lock);
 494         list_for_each_entry(e, &device->act_log->to_be_changed, list) {
 495                 if (i == AL_UPDATES_PER_TRANSACTION) {
 496                         i++;
 497                         break;
 498                 }
 499                 buffer->update_slot_nr[i] = cpu_to_be16(e->lc_index);
 500                 buffer->update_extent_nr[i] = cpu_to_be32(e->lc_new_number);
 501                 if (e->lc_number != LC_FREE)
 502                         drbd_bm_mark_for_writeout(device,
 503                                         al_extent_to_bm_page(e->lc_number));
 504                 i++;
 505         }
 506         spin_unlock_irq(&device->al_lock);
 507         BUG_ON(i > AL_UPDATES_PER_TRANSACTION);
 508
 509         buffer->n_updates = cpu_to_be16(i);
 510         for ( ; i < AL_UPDATES_PER_TRANSACTION; i++) {
 511                 buffer->update_slot_nr[i] = cpu_to_be16(-1);
 512                 buffer->update_extent_nr[i] = cpu_to_be32(LC_FREE);
 513         }
 514
 515         buffer->context_size = cpu_to_be16(device->act_log->nr_elements);
 516         buffer->context_start_slot_nr = cpu_to_be16(device->al_tr_cycle);
 517
 518         mx = min_t(int, AL_CONTEXT_PER_TRANSACTION,
 519                    device->act_log->nr_elements - device->al_tr_cycle);
 520         for (i = 0; i < mx; i++) {
 521                 unsigned idx = device->al_tr_cycle + i;
 522                 extent_nr = lc_element_by_index(device->act_log, idx)->lc_number;
 523                 buffer->context[i] = cpu_to_be32(extent_nr);
 524         }
 525         for (; i < AL_CONTEXT_PER_TRANSACTION; i++)
 526                 buffer->context[i] = cpu_to_be32(LC_FREE);
 527
 528         device->al_tr_cycle += AL_CONTEXT_PER_TRANSACTION;
 529         if (device->al_tr_cycle >= device->act_log->nr_elements)
 530                 device->al_tr_cycle = 0;
 531
 532         sector = al_tr_number_to_on_disk_sector(device);
 533
 534         crc = crc32c(0, buffer, 4096);
 535         buffer->crc32c = cpu_to_be32(crc);
 536
 537         if (drbd_bm_write_hinted(device))
 538                 err = -EIO;
 539         else {
 540                 bool write_al_updates;
 541                 rcu_read_lock();
 542                 write_al_updates = rcu_dereference(device->ldev->disk_conf)->al_updates;
 543                 rcu_read_unlock();
 544                 if (write_al_updates) {
 545                         if (drbd_md_sync_page_io(device, device->ldev, sector, WRITE)) {
 546                                 err = -EIO;
 547                                 drbd_chk_io_error(device, 1, DRBD_META_IO_ERROR);
 548                         } else {
 549                                 device->al_tr_number++;
 550                                 device->al_writ_cnt++;
 551                         }
 552                 }
 553         }
 554
 555         drbd_md_put_buffer(device);
 556         put_ldev(device);
 557
 558         return err;
 559 }
 560
 561 static int _try_lc_del(struct drbd_device *device, struct lc_element *al_ext)
 562 {
 563         int rv;
 564
 565         spin_lock_irq(&device->al_lock);
 566         rv = (al_ext->refcnt == 0);
 567         if (likely(rv))
 568                 lc_del(device->act_log, al_ext);
 569         spin_unlock_irq(&device->al_lock);
 570
 571         return rv;
 572 }
 573
 574 /**
 575  * drbd_al_shrink() - Removes all active extents form the activity log
 576  * @device:     DRBD device.
 577  *
 578  * Removes all active extents form the activity log, waiting until
 579  * the reference count of each entry dropped to 0 first, of course.
 580  *
 581  * You need to lock device->act_log with lc_try_lock() / lc_unlock()
 582  */
 583 void drbd_al_shrink(struct drbd_device *device)
 584 {
 585         struct lc_element *al_ext;
 586         int i;
 587
 588         D_ASSERT(device, test_bit(__LC_LOCKED, &device->act_log->flags));
 589
 590         for (i = 0; i < device->act_log->nr_elements; i++) {
 591                 al_ext = lc_element_by_index(device->act_log, i);
 592                 if (al_ext->lc_number == LC_FREE)
 593                         continue;
 594                 wait_event(device->al_wait, _try_lc_del(device, al_ext));
 595         }
 596
 597         wake_up(&device->al_wait);
 598 }
 599
 600 int drbd_initialize_al(struct drbd_device *device, void *buffer)
 601 {
 602         struct al_transaction_on_disk *al = buffer;
 603         struct drbd_md *md = &device->ldev->md;
 604         sector_t al_base = md->md_offset + md->al_offset;
 605         int al_size_4k = md->al_stripes * md->al_stripe_size_4k;
 606         int i;
 607
 608         memset(al, 0, 4096);
 609         al->magic = cpu_to_be32(DRBD_AL_MAGIC);
 610         al->transaction_type = cpu_to_be16(AL_TR_INITIALIZED);
 611         al->crc32c = cpu_to_be32(crc32c(0, al, 4096));
 612
 613         for (i = 0; i < al_size_4k; i++) {
 614                 int err = drbd_md_sync_page_io(device, device->ldev, al_base + i * 8, WRITE);
 615                 if (err)
 616                         return err;
 617         }
 618         return 0;
 619 }
 620
 621 static const char *drbd_change_sync_fname[] = {
 622         [RECORD_RS_FAILED] = "drbd_rs_failed_io",
 623         [SET_IN_SYNC] = "drbd_set_in_sync",
 624         [SET_OUT_OF_SYNC] = "drbd_set_out_of_sync"
 625 };
 626
 627 /* ATTENTION. The AL's extents are 4MB each, while the extents in the
 628  * resync LRU-cache are 16MB each.
 629  * The caller of this function has to hold an get_ldev() reference.
 630  *
 631  * Adjusts the caching members ->rs_left (success) or ->rs_failed (!success),
 632  * potentially pulling in (and recounting the corresponding bits)
 633  * this resync extent into the resync extent lru cache.
 634  *
 635  * Returns whether all bits have been cleared for this resync extent,
 636  * precisely: (rs_left <= rs_failed)
 637  *
 638  * TODO will be obsoleted once we have a caching lru of the on disk bitmap
 639  */
 640 static bool update_rs_extent(struct drbd_device *device,
 641                 unsigned int enr, int count,
 642                 enum update_sync_bits_mode mode)
 643 {
 644         struct lc_element *e;
 645
 646         D_ASSERT(device, atomic_read(&device->local_cnt));
 647
 648         /* When setting out-of-sync bits,
 649          * we don't need it cached (lc_find).
 650          * But if it is present in the cache,
 651          * we should update the cached bit count.
 652          * Otherwise, that extent should be in the resync extent lru cache
 653          * already -- or we want to pull it in if necessary -- (lc_get),
 654          * then update and check rs_left and rs_failed. */
 655         if (mode == SET_OUT_OF_SYNC)
 656                 e = lc_find(device->resync, enr);
 657         else
 658                 e = lc_get(device->resync, enr);
 659         if (e) {
 660                 struct bm_extent *ext = lc_entry(e, struct bm_extent, lce);
 661                 if (ext->lce.lc_number == enr) {
 662                         if (mode == SET_IN_SYNC)
 663                                 ext->rs_left -= count;
 664                         else if (mode == SET_OUT_OF_SYNC)
 665                                 ext->rs_left += count;
 666                         else
 667                                 ext->rs_failed += count;
 668                         if (ext->rs_left < ext->rs_failed) {
 669                                 drbd_warn(device, "BAD! enr=%u rs_left=%d "
 670                                     "rs_failed=%d count=%d cstate=%s\n",
 671                                      ext->lce.lc_number, ext->rs_left,
 672                                      ext->rs_failed, count,
 673                                      drbd_conn_str(device->state.conn));
 674
 675                                 /* We don't expect to be able to clear more bits
 676                                  * than have been set when we originally counted
 677                                  * the set bits to cache that value in ext->rs_left.
 678                                  * Whatever the reason (disconnect during resync,
 679                                  * delayed local completion of an application write),
 680                                  * try to fix it up by recounting here. */
 681                                 ext->rs_left = drbd_bm_e_weight(device, enr);
 682                         }
 683                 } else {
 684                         /* Normally this element should be in the cache,
 685                          * since drbd_rs_begin_io() pulled it already in.
 686                          *
 687                          * But maybe an application write finished, and we set
 688                          * something outside the resync lru_cache in sync.
 689                          */
 690                         int rs_left = drbd_bm_e_weight(device, enr);
 691                         if (ext->flags != 0) {
 692                                 drbd_warn(device, "changing resync lce: %d[%u;%02lx]"
 693                                      " -> %d[%u;00]\n",
 694                                      ext->lce.lc_number, ext->rs_left,
 695                                      ext->flags, enr, rs_left);
 696                                 ext->flags = 0;
 697                         }
 698                         if (ext->rs_failed) {
 699                                 drbd_warn(device, "Kicking resync_lru element enr=%u "
 700                                      "out with rs_failed=%d\n",
 701                                      ext->lce.lc_number, ext->rs_failed);
 702                         }
 703                         ext->rs_left = rs_left;
 704                         ext->rs_failed = (mode == RECORD_RS_FAILED) ? count : 0;
 705                         /* we don't keep a persistent log of the resync lru,
 706                          * we can commit any change right away. */
 707                         lc_committed(device->resync);
 708                 }
 709                 if (mode != SET_OUT_OF_SYNC)
 710                         lc_put(device->resync, &ext->lce);
 711                 /* no race, we are within the al_lock! */
 712
 713                 if (ext->rs_left <= ext->rs_failed) {
 714                         ext->rs_failed = 0;
 715                         return true;
 716                 }
 717         } else if (mode != SET_OUT_OF_SYNC) {
 718                 /* be quiet if lc_find() did not find it. */
 719                 drbd_err(device, "lc_get() failed! locked=%d/%d flags=%lu\n",
 720                     device->resync_locked,
 721                     device->resync->nr_elements,
 722                     device->resync->flags);
 723         }
 724         return false;
 725 }
 726
 727 void drbd_advance_rs_marks(struct drbd_device *device, unsigned long still_to_go)
 728 {
 729         unsigned long now = jiffies;
 730         unsigned long last = device->rs_mark_time[device->rs_last_mark];
 731         int next = (device->rs_last_mark + 1) % DRBD_SYNC_MARKS;
 732         if (time_after_eq(now, last + DRBD_SYNC_MARK_STEP)) {
 733                 if (device->rs_mark_left[device->rs_last_mark] != still_to_go &&
 734                     device->state.conn != C_PAUSED_SYNC_T &&
 735                     device->state.conn != C_PAUSED_SYNC_S) {
 736                         device->rs_mark_time[next] = now;
 737                         device->rs_mark_left[next] = still_to_go;
 738                         device->rs_last_mark = next;
 739                 }
 740         }
 741 }
 742
 743 /* It is called lazy update, so don't do write-out too often. */
 744 static bool lazy_bitmap_update_due(struct drbd_device *device)
 745 {
 746         return time_after(jiffies, device->rs_last_bcast + 2*HZ);
 747 }
 748
 749 static void maybe_schedule_on_disk_bitmap_update(struct drbd_device *device, bool rs_done)
 750 {
 751         if (rs_done)
 752                 set_bit(RS_DONE, &device->flags);
 753                 /* and also set RS_PROGRESS below */
 754         else if (!lazy_bitmap_update_due(device))
 755                 return;
 756
 757         drbd_device_post_work(device, RS_PROGRESS);
 758 }
 759
 760 static int update_sync_bits(struct drbd_device *device,
 761                 unsigned long sbnr, unsigned long ebnr,
 762                 enum update_sync_bits_mode mode)
 763 {
 764         /*
 765          * We keep a count of set bits per resync-extent in the ->rs_left
 766          * caching member, so we need to loop and work within the resync extent
 767          * alignment. Typically this loop will execute exactly once.
 768          */
 769         unsigned long flags;
 770         unsigned long count = 0;
 771         unsigned int cleared = 0;
 772         while (sbnr <= ebnr) {
 773                 /* set temporary boundary bit number to last bit number within
 774                  * the resync extent of the current start bit number,
 775                  * but cap at provided end bit number */
 776                 unsigned long tbnr = min(ebnr, sbnr | BM_BLOCKS_PER_BM_EXT_MASK);
 777                 unsigned long c;
 778
 779                 if (mode == RECORD_RS_FAILED)
 780                         /* Only called from drbd_rs_failed_io(), bits
 781                          * supposedly still set.  Recount, maybe some
 782                          * of the bits have been successfully cleared
 783                          * by application IO meanwhile.
 784                          */
 785                         c = drbd_bm_count_bits(device, sbnr, tbnr);
 786                 else if (mode == SET_IN_SYNC)
 787                         c = drbd_bm_clear_bits(device, sbnr, tbnr);
 788                 else /* if (mode == SET_OUT_OF_SYNC) */
 789                         c = drbd_bm_set_bits(device, sbnr, tbnr);
 790
 791                 if (c) {
 792                         spin_lock_irqsave(&device->al_lock, flags);
 793                         cleared += update_rs_extent(device, BM_BIT_TO_EXT(sbnr), c, mode);
 794                         spin_unlock_irqrestore(&device->al_lock, flags);
 795                         count += c;
 796                 }
 797                 sbnr = tbnr + 1;
 798         }
 799         if (count) {
 800                 if (mode == SET_IN_SYNC) {
 801                         unsigned long still_to_go = drbd_bm_total_weight(device);
 802                         bool rs_is_done = (still_to_go <= device->rs_failed);
 803                         drbd_advance_rs_marks(device, still_to_go);
 804                         if (cleared || rs_is_done)
 805                                 maybe_schedule_on_disk_bitmap_update(device, rs_is_done);
 806                 } else if (mode == RECORD_RS_FAILED)
 807                         device->rs_failed += count;
 808                 wake_up(&device->al_wait);
 809         }
 810         return count;
 811 }
 812
 813 /* clear the bit corresponding to the piece of storage in question:
 814  * size byte of data starting from sector.  Only clear a bits of the affected
 815  * one ore more _aligned_ BM_BLOCK_SIZE blocks.
 816  *
 817  * called by worker on C_SYNC_TARGET and receiver on SyncSource.
 818  *
 819  */
 820 int __drbd_change_sync(struct drbd_device *device, sector_t sector, int size,
 821                 enum update_sync_bits_mode mode,
 822                 const char *file, const unsigned int line)
 823 {
 824         /* Is called from worker and receiver context _only_ */
 825         unsigned long sbnr, ebnr, lbnr;
 826         unsigned long count = 0;
 827         sector_t esector, nr_sectors;
 828
 829         /* This would be an empty REQ_FLUSH, be silent. */
 830         if ((mode == SET_OUT_OF_SYNC) && size == 0)
 831                 return 0;
 832
 833         if (size <= 0 || !IS_ALIGNED(size, 512) || size > DRBD_MAX_DISCARD_SIZE) {
 834                 drbd_err(device, "%s: sector=%llus size=%d nonsense!\n",
 835                                 drbd_change_sync_fname[mode],
 836                                 (unsigned long long)sector, size);
 837                 return 0;
 838         }
 839
 840         if (!get_ldev(device))
 841                 return 0; /* no disk, no metadata, no bitmap to manipulate bits in */
 842
 843         nr_sectors = drbd_get_capacity(device->this_bdev);
 844         esector = sector + (size >> 9) - 1;
 845
 846         if (!expect(sector < nr_sectors))
 847                 goto out;
 848         if (!expect(esector < nr_sectors))
 849                 esector = nr_sectors - 1;
 850
 851         lbnr = BM_SECT_TO_BIT(nr_sectors-1);
 852
 853         if (mode == SET_IN_SYNC) {
 854                 /* Round up start sector, round down end sector.  We make sure
 855                  * we only clear full, aligned, BM_BLOCK_SIZE blocks. */
 856                 if (unlikely(esector < BM_SECT_PER_BIT-1))
 857                         goto out;
 858                 if (unlikely(esector == (nr_sectors-1)))
 859                         ebnr = lbnr;
 860                 else
 861                         ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1));
 862                 sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1);
 863         } else {
 864                 /* We set it out of sync, or record resync failure.
 865                  * Should not round anything here. */
 866                 sbnr = BM_SECT_TO_BIT(sector);
 867                 ebnr = BM_SECT_TO_BIT(esector);
 868         }
 869
 870         count = update_sync_bits(device, sbnr, ebnr, mode);
 871 out:
 872         put_ldev(device);
 873         return count;
 874 }
 875
 876 static
 877 struct bm_extent *_bme_get(struct drbd_device *device, unsigned int enr)
 878 {
 879         struct lc_element *e;
 880         struct bm_extent *bm_ext;
 881         int wakeup = 0;
 882         unsigned long rs_flags;
 883
 884         spin_lock_irq(&device->al_lock);
 885         if (device->resync_locked > device->resync->nr_elements/2) {
 886                 spin_unlock_irq(&device->al_lock);
 887                 return NULL;
 888         }
 889         e = lc_get(device->resync, enr);
 890         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
 891         if (bm_ext) {
 892                 if (bm_ext->lce.lc_number != enr) {
 893                         bm_ext->rs_left = drbd_bm_e_weight(device, enr);
 894                         bm_ext->rs_failed = 0;
 895                         lc_committed(device->resync);
 896                         wakeup = 1;
 897                 }
 898                 if (bm_ext->lce.refcnt == 1)
 899                         device->resync_locked++;
 900                 set_bit(BME_NO_WRITES, &bm_ext->flags);
 901         }
 902         rs_flags = device->resync->flags;
 903         spin_unlock_irq(&device->al_lock);
 904         if (wakeup)
 905                 wake_up(&device->al_wait);
 906
 907         if (!bm_ext) {
 908                 if (rs_flags & LC_STARVING)
 909                         drbd_warn(device, "Have to wait for element"
 910                              " (resync LRU too small?)\n");
 911                 BUG_ON(rs_flags & LC_LOCKED);
 912         }
 913
 914         return bm_ext;
 915 }
 916
 917 static int _is_in_al(struct drbd_device *device, unsigned int enr)
 918 {
 919         int rv;
 920
 921         spin_lock_irq(&device->al_lock);
 922         rv = lc_is_used(device->act_log, enr);
 923         spin_unlock_irq(&device->al_lock);
 924
 925         return rv;
 926 }
 927
 928 /**
 929  * drbd_rs_begin_io() - Gets an extent in the resync LRU cache and sets it to BME_LOCKED
 930  * @device:     DRBD device.
 931  * @sector:     The sector number.
 932  *
 933  * This functions sleeps on al_wait. Returns 0 on success, -EINTR if interrupted.
 934  */
 935 int drbd_rs_begin_io(struct drbd_device *device, sector_t sector)
 936 {
 937         unsigned int enr = BM_SECT_TO_EXT(sector);
 938         struct bm_extent *bm_ext;
 939         int i, sig;
 940         bool sa;
 941
 942 retry:
 943         sig = wait_event_interruptible(device->al_wait,
 944                         (bm_ext = _bme_get(device, enr)));
 945         if (sig)
 946                 return -EINTR;
 947
 948         if (test_bit(BME_LOCKED, &bm_ext->flags))
 949                 return 0;
 950
 951         /* step aside only while we are above c-min-rate; unless disabled. */
 952         sa = drbd_rs_c_min_rate_throttle(device);
 953
 954         for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
 955                 sig = wait_event_interruptible(device->al_wait,
 956                                                !_is_in_al(device, enr * AL_EXT_PER_BM_SECT + i) ||
 957                                                (sa && test_bit(BME_PRIORITY, &bm_ext->flags)));
 958
 959                 if (sig || (sa && test_bit(BME_PRIORITY, &bm_ext->flags))) {
 960                         spin_lock_irq(&device->al_lock);
 961                         if (lc_put(device->resync, &bm_ext->lce) == 0) {
 962                                 bm_ext->flags = 0; /* clears BME_NO_WRITES and eventually BME_PRIORITY */
 963                                 device->resync_locked--;
 964                                 wake_up(&device->al_wait);
 965                         }
 966                         spin_unlock_irq(&device->al_lock);
 967                         if (sig)
 968                                 return -EINTR;
 969                         if (schedule_timeout_interruptible(HZ/10))
 970                                 return -EINTR;
 971                         goto retry;
 972                 }
 973         }
 974         set_bit(BME_LOCKED, &bm_ext->flags);
 975         return 0;
 976 }
 977
 978 /**
 979  * drbd_try_rs_begin_io() - Gets an extent in the resync LRU cache, does not sleep
 980  * @device:     DRBD device.
 981  * @sector:     The sector number.
 982  *
 983  * Gets an extent in the resync LRU cache, sets it to BME_NO_WRITES, then
 984  * tries to set it to BME_LOCKED. Returns 0 upon success, and -EAGAIN
 985  * if there is still application IO going on in this area.
 986  */
 987 int drbd_try_rs_begin_io(struct drbd_device *device, sector_t sector)
 988 {
 989         unsigned int enr = BM_SECT_TO_EXT(sector);
 990         const unsigned int al_enr = enr*AL_EXT_PER_BM_SECT;
 991         struct lc_element *e;
 992         struct bm_extent *bm_ext;
 993         int i;
 994
 995         spin_lock_irq(&device->al_lock);
 996         if (device->resync_wenr != LC_FREE && device->resync_wenr != enr) {
 997                 /* in case you have very heavy scattered io, it may
 998                  * stall the syncer undefined if we give up the ref count
 999                  * when we try again and requeue.
1000                  *
1001                  * if we don't give up the refcount, but the next time
1002                  * we are scheduled this extent has been "synced" by new
1003                  * application writes, we'd miss the lc_put on the
1004                  * extent we keep the refcount on.
1005                  * so we remembered which extent we had to try again, and
1006                  * if the next requested one is something else, we do
1007                  * the lc_put here...
1008                  * we also have to wake_up
1009                  */
1010                 e = lc_find(device->resync, device->resync_wenr);
1011                 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1012                 if (bm_ext) {
1013                         D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1014                         D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1015                         clear_bit(BME_NO_WRITES, &bm_ext->flags);
1016                         device->resync_wenr = LC_FREE;
1017                         if (lc_put(device->resync, &bm_ext->lce) == 0)
1018                                 device->resync_locked--;
1019                         wake_up(&device->al_wait);
1020                 } else {
1021                         drbd_alert(device, "LOGIC BUG\n");
1022                 }
1023         }
1024         /* TRY. */
1025         e = lc_try_get(device->resync, enr);
1026         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1027         if (bm_ext) {
1028                 if (test_bit(BME_LOCKED, &bm_ext->flags))
1029                         goto proceed;
1030                 if (!test_and_set_bit(BME_NO_WRITES, &bm_ext->flags)) {
1031                         device->resync_locked++;
1032                 } else {
1033                         /* we did set the BME_NO_WRITES,
1034                          * but then could not set BME_LOCKED,
1035                          * so we tried again.
1036                          * drop the extra reference. */
1037                         bm_ext->lce.refcnt--;
1038                         D_ASSERT(device, bm_ext->lce.refcnt > 0);
1039                 }
1040                 goto check_al;
1041         } else {
1042                 /* do we rather want to try later? */
1043                 if (device->resync_locked > device->resync->nr_elements-3)
1044                         goto try_again;
1045                 /* Do or do not. There is no try. -- Yoda */
1046                 e = lc_get(device->resync, enr);
1047                 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1048                 if (!bm_ext) {
1049                         const unsigned long rs_flags = device->resync->flags;
1050                         if (rs_flags & LC_STARVING)
1051                                 drbd_warn(device, "Have to wait for element"
1052                                      " (resync LRU too small?)\n");
1053                         BUG_ON(rs_flags & LC_LOCKED);
1054                         goto try_again;
1055                 }
1056                 if (bm_ext->lce.lc_number != enr) {
1057                         bm_ext->rs_left = drbd_bm_e_weight(device, enr);
1058                         bm_ext->rs_failed = 0;
1059                         lc_committed(device->resync);
1060                         wake_up(&device->al_wait);
1061                         D_ASSERT(device, test_bit(BME_LOCKED, &bm_ext->flags) == 0);
1062                 }
1063                 set_bit(BME_NO_WRITES, &bm_ext->flags);
1064                 D_ASSERT(device, bm_ext->lce.refcnt == 1);
1065                 device->resync_locked++;
1066                 goto check_al;
1067         }
1068 check_al:
1069         for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
1070                 if (lc_is_used(device->act_log, al_enr+i))
1071                         goto try_again;
1072         }
1073         set_bit(BME_LOCKED, &bm_ext->flags);
1074 proceed:
1075         device->resync_wenr = LC_FREE;
1076         spin_unlock_irq(&device->al_lock);
1077         return 0;
1078
1079 try_again:
1080         if (bm_ext)
1081                 device->resync_wenr = enr;
1082         spin_unlock_irq(&device->al_lock);
1083         return -EAGAIN;
1084 }
1085
1086 void drbd_rs_complete_io(struct drbd_device *device, sector_t sector)
1087 {
1088         unsigned int enr = BM_SECT_TO_EXT(sector);
1089         struct lc_element *e;
1090         struct bm_extent *bm_ext;
1091         unsigned long flags;
1092
1093         spin_lock_irqsave(&device->al_lock, flags);
1094         e = lc_find(device->resync, enr);
1095         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1096         if (!bm_ext) {
1097                 spin_unlock_irqrestore(&device->al_lock, flags);
1098                 if (__ratelimit(&drbd_ratelimit_state))
1099                         drbd_err(device, "drbd_rs_complete_io() called, but extent not found\n");
1100                 return;
1101         }
1102
1103         if (bm_ext->lce.refcnt == 0) {
1104                 spin_unlock_irqrestore(&device->al_lock, flags);
1105                 drbd_err(device, "drbd_rs_complete_io(,%llu [=%u]) called, "
1106                     "but refcnt is 0!?\n",
1107                     (unsigned long long)sector, enr);
1108                 return;
1109         }
1110
1111         if (lc_put(device->resync, &bm_ext->lce) == 0) {
1112                 bm_ext->flags = 0; /* clear BME_LOCKED, BME_NO_WRITES and BME_PRIORITY */
1113                 device->resync_locked--;
1114                 wake_up(&device->al_wait);
1115         }
1116
1117         spin_unlock_irqrestore(&device->al_lock, flags);
1118 }
1119
1120 /**
1121  * drbd_rs_cancel_all() - Removes all extents from the resync LRU (even BME_LOCKED)
1122  * @device:     DRBD device.
1123  */
1124 void drbd_rs_cancel_all(struct drbd_device *device)
1125 {
1126         spin_lock_irq(&device->al_lock);
1127
1128         if (get_ldev_if_state(device, D_FAILED)) { /* Makes sure ->resync is there. */
1129                 lc_reset(device->resync);
1130                 put_ldev(device);
1131         }
1132         device->resync_locked = 0;
1133         device->resync_wenr = LC_FREE;
1134         spin_unlock_irq(&device->al_lock);
1135         wake_up(&device->al_wait);
1136 }
1137
1138 /**
1139  * drbd_rs_del_all() - Gracefully remove all extents from the resync LRU
1140  * @device:     DRBD device.
1141  *
1142  * Returns 0 upon success, -EAGAIN if at least one reference count was
1143  * not zero.
1144  */
1145 int drbd_rs_del_all(struct drbd_device *device)
1146 {
1147         struct lc_element *e;
1148         struct bm_extent *bm_ext;
1149         int i;
1150
1151         spin_lock_irq(&device->al_lock);
1152
1153         if (get_ldev_if_state(device, D_FAILED)) {
1154                 /* ok, ->resync is there. */
1155                 for (i = 0; i < device->resync->nr_elements; i++) {
1156                         e = lc_element_by_index(device->resync, i);
1157                         bm_ext = lc_entry(e, struct bm_extent, lce);
1158                         if (bm_ext->lce.lc_number == LC_FREE)
1159                                 continue;
1160                         if (bm_ext->lce.lc_number == device->resync_wenr) {
1161                                 drbd_info(device, "dropping %u in drbd_rs_del_all, apparently"
1162                                      " got 'synced' by application io\n",
1163                                      device->resync_wenr);
1164                                 D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1165                                 D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1166                                 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1167                                 device->resync_wenr = LC_FREE;
1168                                 lc_put(device->resync, &bm_ext->lce);
1169                         }
1170                         if (bm_ext->lce.refcnt != 0) {
1171                                 drbd_info(device, "Retrying drbd_rs_del_all() later. "
1172                                      "refcnt=%d\n", bm_ext->lce.refcnt);
1173                                 put_ldev(device);
1174                                 spin_unlock_irq(&device->al_lock);
1175                                 return -EAGAIN;
1176                         }
1177                         D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1178                         D_ASSERT(device, !test_bit(BME_NO_WRITES, &bm_ext->flags));
1179                         lc_del(device->resync, &bm_ext->lce);
1180                 }
1181                 D_ASSERT(device, device->resync->used == 0);
1182                 put_ldev(device);
1183         }
1184         spin_unlock_irq(&device->al_lock);
1185         wake_up(&device->al_wait);
1186
1187         return 0;
1188 }