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 #include "drbd_wrappers.h"
  33
  34
  35 enum al_transaction_types {
  36         AL_TR_UPDATE = 0,
  37         AL_TR_INITIALIZED = 0xffff
  38 };
  39 /* all fields on disc in big endian */
  40 struct __packed al_transaction_on_disk {
  41         /* don't we all like magic */
  42         __be32  magic;
  43
  44         /* to identify the most recent transaction block
  45          * in the on disk ring buffer */
  46         __be32  tr_number;
  47
  48         /* checksum on the full 4k block, with this field set to 0. */
  49         __be32  crc32c;
  50
  51         /* type of transaction, special transaction types like:
  52          * purge-all, set-all-idle, set-all-active, ... to-be-defined
  53          * see also enum al_transaction_types */
  54         __be16  transaction_type;
  55
  56         /* we currently allow only a few thousand extents,
  57          * so 16bit will be enough for the slot number. */
  58
  59         /* how many updates in this transaction */
  60         __be16  n_updates;
  61
  62         /* maximum slot number, "al-extents" in drbd.conf speak.
  63          * Having this in each transaction should make reconfiguration
  64          * of that parameter easier. */
  65         __be16  context_size;
  66
  67         /* slot number the context starts with */
  68         __be16  context_start_slot_nr;
  69
  70         /* Some reserved bytes.  Expected usage is a 64bit counter of
  71          * sectors-written since device creation, and other data generation tag
  72          * supporting usage */
  73         __be32  __reserved[4];
  74
  75         /* --- 36 byte used --- */
  76
  77         /* Reserve space for up to AL_UPDATES_PER_TRANSACTION changes
  78          * in one transaction, then use the remaining byte in the 4k block for
  79          * context information.  "Flexible" number of updates per transaction
  80          * does not help, as we have to account for the case when all update
  81          * slots are used anyways, so it would only complicate code without
  82          * additional benefit.
  83          */
  84         __be16  update_slot_nr[AL_UPDATES_PER_TRANSACTION];
  85
  86         /* but the extent number is 32bit, which at an extent size of 4 MiB
  87          * allows to cover device sizes of up to 2**54 Byte (16 PiB) */
  88         __be32  update_extent_nr[AL_UPDATES_PER_TRANSACTION];
  89
  90         /* --- 420 bytes used (36 + 64*6) --- */
  91
  92         /* 4096 - 420 = 3676 = 919 * 4 */
  93         __be32  context[AL_CONTEXT_PER_TRANSACTION];
  94 };
  95
  96 struct update_odbm_work {
  97         struct drbd_work w;
  98         unsigned int enr;
  99 };
 100
 101 struct update_al_work {
 102         struct drbd_work w;
 103         struct completion event;
 104         int err;
 105 };
 106
 107
 108 void *drbd_md_get_buffer(struct drbd_conf *mdev)
 109 {
 110         int r;
 111
 112         wait_event(mdev->misc_wait,
 113                    (r = atomic_cmpxchg(&mdev->md_io_in_use, 0, 1)) == 0 ||
 114                    mdev->state.disk <= D_FAILED);
 115
 116         return r ? NULL : page_address(mdev->md_io_page);
 117 }
 118
 119 void drbd_md_put_buffer(struct drbd_conf *mdev)
 120 {
 121         if (atomic_dec_and_test(&mdev->md_io_in_use))
 122                 wake_up(&mdev->misc_wait);
 123 }
 124
 125 void wait_until_done_or_force_detached(struct drbd_conf *mdev, struct drbd_backing_dev *bdev,
 126                                      unsigned int *done)
 127 {
 128         long dt;
 129
 130         rcu_read_lock();
 131         dt = rcu_dereference(bdev->disk_conf)->disk_timeout;
 132         rcu_read_unlock();
 133         dt = dt * HZ / 10;
 134         if (dt == 0)
 135                 dt = MAX_SCHEDULE_TIMEOUT;
 136
 137         dt = wait_event_timeout(mdev->misc_wait,
 138                         *done || test_bit(FORCE_DETACH, &mdev->flags), dt);
 139         if (dt == 0) {
 140                 dev_err(DEV, "meta-data IO operation timed out\n");
 141                 drbd_chk_io_error(mdev, 1, DRBD_FORCE_DETACH);
 142         }
 143 }
 144
 145 static int _drbd_md_sync_page_io(struct drbd_conf *mdev,
 146                                  struct drbd_backing_dev *bdev,
 147                                  struct page *page, sector_t sector,
 148                                  int rw, int size)
 149 {
 150         struct bio *bio;
 151         int err;
 152
 153         mdev->md_io.done = 0;
 154         mdev->md_io.error = -ENODEV;
 155
 156         if ((rw & WRITE) && !test_bit(MD_NO_FUA, &mdev->flags))
 157                 rw |= REQ_FUA | REQ_FLUSH;
 158         rw |= REQ_SYNC;
 159
 160         bio = bio_alloc_drbd(GFP_NOIO);
 161         bio->bi_bdev = bdev->md_bdev;
 162         bio->bi_sector = sector;
 163         err = -EIO;
 164         if (bio_add_page(bio, page, size, 0) != size)
 165                 goto out;
 166         bio->bi_private = &mdev->md_io;
 167         bio->bi_end_io = drbd_md_io_complete;
 168         bio->bi_rw = rw;
 169
 170         if (!(rw & WRITE) && mdev->state.disk == D_DISKLESS && mdev->ldev == NULL)
 171                 /* special case, drbd_md_read() during drbd_adm_attach(): no get_ldev */
 172                 ;
 173         else if (!get_ldev_if_state(mdev, D_ATTACHING)) {
 174                 /* Corresponding put_ldev in drbd_md_io_complete() */
 175                 dev_err(DEV, "ASSERT FAILED: get_ldev_if_state() == 1 in _drbd_md_sync_page_io()\n");
 176                 err = -ENODEV;
 177                 goto out;
 178         }
 179
 180         bio_get(bio); /* one bio_put() is in the completion handler */
 181         atomic_inc(&mdev->md_io_in_use); /* drbd_md_put_buffer() is in the completion handler */
 182         if (drbd_insert_fault(mdev, (rw & WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD))
 183                 bio_endio(bio, -EIO);
 184         else
 185                 submit_bio(rw, bio);
 186         wait_until_done_or_force_detached(mdev, bdev, &mdev->md_io.done);
 187         if (bio_flagged(bio, BIO_UPTODATE))
 188                 err = mdev->md_io.error;
 189
 190  out:
 191         bio_put(bio);
 192         return err;
 193 }
 194
 195 int drbd_md_sync_page_io(struct drbd_conf *mdev, struct drbd_backing_dev *bdev,
 196                          sector_t sector, int rw)
 197 {
 198         int err;
 199         struct page *iop = mdev->md_io_page;
 200
 201         D_ASSERT(atomic_read(&mdev->md_io_in_use) == 1);
 202
 203         BUG_ON(!bdev->md_bdev);
 204
 205         dev_dbg(DEV, "meta_data io: %s [%d]:%s(,%llus,%s) %pS\n",
 206              current->comm, current->pid, __func__,
 207              (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ",
 208              (void*)_RET_IP_ );
 209
 210         if (sector < drbd_md_first_sector(bdev) ||
 211             sector + 7 > drbd_md_last_sector(bdev))
 212                 dev_alert(DEV, "%s [%d]:%s(,%llus,%s) out of range md access!\n",
 213                      current->comm, current->pid, __func__,
 214                      (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ");
 215
 216         /* we do all our meta data IO in aligned 4k blocks. */
 217         err = _drbd_md_sync_page_io(mdev, bdev, iop, sector, rw, 4096);
 218         if (err) {
 219                 dev_err(DEV, "drbd_md_sync_page_io(,%llus,%s) failed with error %d\n",
 220                     (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ", err);
 221         }
 222         return err;
 223 }
 224
 225 static struct bm_extent *find_active_resync_extent(struct drbd_conf *mdev, unsigned int enr)
 226 {
 227         struct lc_element *tmp;
 228         tmp = lc_find(mdev->resync, enr/AL_EXT_PER_BM_SECT);
 229         if (unlikely(tmp != NULL)) {
 230                 struct bm_extent  *bm_ext = lc_entry(tmp, struct bm_extent, lce);
 231                 if (test_bit(BME_NO_WRITES, &bm_ext->flags))
 232                         return bm_ext;
 233         }
 234         return NULL;
 235 }
 236
 237 static struct lc_element *_al_get(struct drbd_conf *mdev, unsigned int enr, bool nonblock)
 238 {
 239         struct lc_element *al_ext;
 240         struct bm_extent *bm_ext;
 241         int wake;
 242
 243         spin_lock_irq(&mdev->al_lock);
 244         bm_ext = find_active_resync_extent(mdev, enr);
 245         if (bm_ext) {
 246                 wake = !test_and_set_bit(BME_PRIORITY, &bm_ext->flags);
 247                 spin_unlock_irq(&mdev->al_lock);
 248                 if (wake)
 249                         wake_up(&mdev->al_wait);
 250                 return NULL;
 251         }
 252         if (nonblock)
 253                 al_ext = lc_try_get(mdev->act_log, enr);
 254         else
 255                 al_ext = lc_get(mdev->act_log, enr);
 256         spin_unlock_irq(&mdev->al_lock);
 257         return al_ext;
 258 }
 259
 260 bool drbd_al_begin_io_fastpath(struct drbd_conf *mdev, struct drbd_interval *i)
 261 {
 262         /* for bios crossing activity log extent boundaries,
 263          * we may need to activate two extents in one go */
 264         unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
 265         unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
 266
 267         D_ASSERT((unsigned)(last - first) <= 1);
 268         D_ASSERT(atomic_read(&mdev->local_cnt) > 0);
 269
 270         /* FIXME figure out a fast path for bios crossing AL extent boundaries */
 271         if (first != last)
 272                 return false;
 273
 274         return _al_get(mdev, first, true);
 275 }
 276
 277 bool drbd_al_begin_io_prepare(struct drbd_conf *mdev, struct drbd_interval *i)
 278 {
 279         /* for bios crossing activity log extent boundaries,
 280          * we may need to activate two extents in one go */
 281         unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
 282         unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
 283         unsigned enr;
 284         bool need_transaction = false;
 285
 286         D_ASSERT(first <= last);
 287         D_ASSERT(atomic_read(&mdev->local_cnt) > 0);
 288
 289         for (enr = first; enr <= last; enr++) {
 290                 struct lc_element *al_ext;
 291                 wait_event(mdev->al_wait,
 292                                 (al_ext = _al_get(mdev, enr, false)) != NULL);
 293                 if (al_ext->lc_number != enr)
 294                         need_transaction = true;
 295         }
 296         return need_transaction;
 297 }
 298
 299 static int al_write_transaction(struct drbd_conf *mdev, bool delegate);
 300
 301 /* When called through generic_make_request(), we must delegate
 302  * activity log I/O to the worker thread: a further request
 303  * submitted via generic_make_request() within the same task
 304  * would be queued on current->bio_list, and would only start
 305  * after this function returns (see generic_make_request()).
 306  *
 307  * However, if we *are* the worker, we must not delegate to ourselves.
 308  */
 309
 310 /*
 311  * @delegate:   delegate activity log I/O to the worker thread
 312  */
 313 void drbd_al_begin_io_commit(struct drbd_conf *mdev, bool delegate)
 314 {
 315         bool locked = false;
 316
 317         BUG_ON(delegate && current == mdev->tconn->worker.task);
 318
 319         /* Serialize multiple transactions.
 320          * This uses test_and_set_bit, memory barrier is implicit.
 321          */
 322         wait_event(mdev->al_wait,
 323                         mdev->act_log->pending_changes == 0 ||
 324                         (locked = lc_try_lock_for_transaction(mdev->act_log)));
 325
 326         if (locked) {
 327                 /* Double check: it may have been committed by someone else,
 328                  * while we have been waiting for the lock. */
 329                 if (mdev->act_log->pending_changes) {
 330                         bool write_al_updates;
 331
 332                         rcu_read_lock();
 333                         write_al_updates = rcu_dereference(mdev->ldev->disk_conf)->al_updates;
 334                         rcu_read_unlock();
 335
 336                         if (write_al_updates)
 337                                 al_write_transaction(mdev, delegate);
 338                         spin_lock_irq(&mdev->al_lock);
 339                         /* FIXME
 340                         if (err)
 341                                 we need an "lc_cancel" here;
 342                         */
 343                         lc_committed(mdev->act_log);
 344                         spin_unlock_irq(&mdev->al_lock);
 345                 }
 346                 lc_unlock(mdev->act_log);
 347                 wake_up(&mdev->al_wait);
 348         }
 349 }
 350
 351 /*
 352  * @delegate:   delegate activity log I/O to the worker thread
 353  */
 354 void drbd_al_begin_io(struct drbd_conf *mdev, struct drbd_interval *i, bool delegate)
 355 {
 356         BUG_ON(delegate && current == mdev->tconn->worker.task);
 357
 358         if (drbd_al_begin_io_prepare(mdev, i))
 359                 drbd_al_begin_io_commit(mdev, delegate);
 360 }
 361
 362 int drbd_al_begin_io_nonblock(struct drbd_conf *mdev, struct drbd_interval *i)
 363 {
 364         struct lru_cache *al = mdev->act_log;
 365         /* for bios crossing activity log extent boundaries,
 366          * we may need to activate two extents in one go */
 367         unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
 368         unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
 369         unsigned nr_al_extents;
 370         unsigned available_update_slots;
 371         unsigned enr;
 372
 373         D_ASSERT(first <= last);
 374
 375         nr_al_extents = 1 + last - first; /* worst case: all touched extends are cold. */
 376         available_update_slots = min(al->nr_elements - al->used,
 377                                 al->max_pending_changes - al->pending_changes);
 378
 379         /* We want all necessary updates for a given request within the same transaction
 380          * We could first check how many updates are *actually* needed,
 381          * and use that instead of the worst-case nr_al_extents */
 382         if (available_update_slots < nr_al_extents)
 383                 return -EWOULDBLOCK;
 384
 385         /* Is resync active in this area? */
 386         for (enr = first; enr <= last; enr++) {
 387                 struct lc_element *tmp;
 388                 tmp = lc_find(mdev->resync, enr/AL_EXT_PER_BM_SECT);
 389                 if (unlikely(tmp != NULL)) {
 390                         struct bm_extent  *bm_ext = lc_entry(tmp, struct bm_extent, lce);
 391                         if (test_bit(BME_NO_WRITES, &bm_ext->flags)) {
 392                                 if (!test_and_set_bit(BME_PRIORITY, &bm_ext->flags))
 393                                         return -EBUSY;
 394                                 return -EWOULDBLOCK;
 395                         }
 396                 }
 397         }
 398
 399         /* Checkout the refcounts.
 400          * Given that we checked for available elements and update slots above,
 401          * this has to be successful. */
 402         for (enr = first; enr <= last; enr++) {
 403                 struct lc_element *al_ext;
 404                 al_ext = lc_get_cumulative(mdev->act_log, enr);
 405                 if (!al_ext)
 406                         dev_info(DEV, "LOGIC BUG for enr=%u\n", enr);
 407         }
 408         return 0;
 409 }
 410
 411 void drbd_al_complete_io(struct drbd_conf *mdev, struct drbd_interval *i)
 412 {
 413         /* for bios crossing activity log extent boundaries,
 414          * we may need to activate two extents in one go */
 415         unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
 416         unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
 417         unsigned enr;
 418         struct lc_element *extent;
 419         unsigned long flags;
 420
 421         D_ASSERT(first <= last);
 422         spin_lock_irqsave(&mdev->al_lock, flags);
 423
 424         for (enr = first; enr <= last; enr++) {
 425                 extent = lc_find(mdev->act_log, enr);
 426                 if (!extent) {
 427                         dev_err(DEV, "al_complete_io() called on inactive extent %u\n", enr);
 428                         continue;
 429                 }
 430                 lc_put(mdev->act_log, extent);
 431         }
 432         spin_unlock_irqrestore(&mdev->al_lock, flags);
 433         wake_up(&mdev->al_wait);
 434 }
 435
 436 #if (PAGE_SHIFT + 3) < (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT)
 437 /* Currently BM_BLOCK_SHIFT, BM_EXT_SHIFT and AL_EXTENT_SHIFT
 438  * are still coupled, or assume too much about their relation.
 439  * Code below will not work if this is violated.
 440  * Will be cleaned up with some followup patch.
 441  */
 442 # error FIXME
 443 #endif
 444
 445 static unsigned int al_extent_to_bm_page(unsigned int al_enr)
 446 {
 447         return al_enr >>
 448                 /* bit to page */
 449                 ((PAGE_SHIFT + 3) -
 450                 /* al extent number to bit */
 451                  (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT));
 452 }
 453
 454 static unsigned int rs_extent_to_bm_page(unsigned int rs_enr)
 455 {
 456         return rs_enr >>
 457                 /* bit to page */
 458                 ((PAGE_SHIFT + 3) -
 459                 /* resync extent number to bit */
 460                  (BM_EXT_SHIFT - BM_BLOCK_SHIFT));
 461 }
 462
 463 static sector_t al_tr_number_to_on_disk_sector(struct drbd_conf *mdev)
 464 {
 465         const unsigned int stripes = mdev->ldev->md.al_stripes;
 466         const unsigned int stripe_size_4kB = mdev->ldev->md.al_stripe_size_4k;
 467
 468         /* transaction number, modulo on-disk ring buffer wrap around */
 469         unsigned int t = mdev->al_tr_number % (mdev->ldev->md.al_size_4k);
 470
 471         /* ... to aligned 4k on disk block */
 472         t = ((t % stripes) * stripe_size_4kB) + t/stripes;
 473
 474         /* ... to 512 byte sector in activity log */
 475         t *= 8;
 476
 477         /* ... plus offset to the on disk position */
 478         return mdev->ldev->md.md_offset + mdev->ldev->md.al_offset + t;
 479 }
 480
 481 static int
 482 _al_write_transaction(struct drbd_conf *mdev)
 483 {
 484         struct al_transaction_on_disk *buffer;
 485         struct lc_element *e;
 486         sector_t sector;
 487         int i, mx;
 488         unsigned extent_nr;
 489         unsigned crc = 0;
 490         int err = 0;
 491
 492         if (!get_ldev(mdev)) {
 493                 dev_err(DEV, "disk is %s, cannot start al transaction\n",
 494                         drbd_disk_str(mdev->state.disk));
 495                 return -EIO;
 496         }
 497
 498         /* The bitmap write may have failed, causing a state change. */
 499         if (mdev->state.disk < D_INCONSISTENT) {
 500                 dev_err(DEV,
 501                         "disk is %s, cannot write al transaction\n",
 502                         drbd_disk_str(mdev->state.disk));
 503                 put_ldev(mdev);
 504                 return -EIO;
 505         }
 506
 507         buffer = drbd_md_get_buffer(mdev); /* protects md_io_buffer, al_tr_cycle, ... */
 508         if (!buffer) {
 509                 dev_err(DEV, "disk failed while waiting for md_io buffer\n");
 510                 put_ldev(mdev);
 511                 return -ENODEV;
 512         }
 513
 514         memset(buffer, 0, sizeof(*buffer));
 515         buffer->magic = cpu_to_be32(DRBD_AL_MAGIC);
 516         buffer->tr_number = cpu_to_be32(mdev->al_tr_number);
 517
 518         i = 0;
 519
 520         /* Even though no one can start to change this list
 521          * once we set the LC_LOCKED -- from drbd_al_begin_io(),
 522          * lc_try_lock_for_transaction() --, someone may still
 523          * be in the process of changing it. */
 524         spin_lock_irq(&mdev->al_lock);
 525         list_for_each_entry(e, &mdev->act_log->to_be_changed, list) {
 526                 if (i == AL_UPDATES_PER_TRANSACTION) {
 527                         i++;
 528                         break;
 529                 }
 530                 buffer->update_slot_nr[i] = cpu_to_be16(e->lc_index);
 531                 buffer->update_extent_nr[i] = cpu_to_be32(e->lc_new_number);
 532                 if (e->lc_number != LC_FREE)
 533                         drbd_bm_mark_for_writeout(mdev,
 534                                         al_extent_to_bm_page(e->lc_number));
 535                 i++;
 536         }
 537         spin_unlock_irq(&mdev->al_lock);
 538         BUG_ON(i > AL_UPDATES_PER_TRANSACTION);
 539
 540         buffer->n_updates = cpu_to_be16(i);
 541         for ( ; i < AL_UPDATES_PER_TRANSACTION; i++) {
 542                 buffer->update_slot_nr[i] = cpu_to_be16(-1);
 543                 buffer->update_extent_nr[i] = cpu_to_be32(LC_FREE);
 544         }
 545
 546         buffer->context_size = cpu_to_be16(mdev->act_log->nr_elements);
 547         buffer->context_start_slot_nr = cpu_to_be16(mdev->al_tr_cycle);
 548
 549         mx = min_t(int, AL_CONTEXT_PER_TRANSACTION,
 550                    mdev->act_log->nr_elements - mdev->al_tr_cycle);
 551         for (i = 0; i < mx; i++) {
 552                 unsigned idx = mdev->al_tr_cycle + i;
 553                 extent_nr = lc_element_by_index(mdev->act_log, idx)->lc_number;
 554                 buffer->context[i] = cpu_to_be32(extent_nr);
 555         }
 556         for (; i < AL_CONTEXT_PER_TRANSACTION; i++)
 557                 buffer->context[i] = cpu_to_be32(LC_FREE);
 558
 559         mdev->al_tr_cycle += AL_CONTEXT_PER_TRANSACTION;
 560         if (mdev->al_tr_cycle >= mdev->act_log->nr_elements)
 561                 mdev->al_tr_cycle = 0;
 562
 563         sector = al_tr_number_to_on_disk_sector(mdev);
 564
 565         crc = crc32c(0, buffer, 4096);
 566         buffer->crc32c = cpu_to_be32(crc);
 567
 568         if (drbd_bm_write_hinted(mdev))
 569                 err = -EIO;
 570         else {
 571                 bool write_al_updates;
 572                 rcu_read_lock();
 573                 write_al_updates = rcu_dereference(mdev->ldev->disk_conf)->al_updates;
 574                 rcu_read_unlock();
 575                 if (write_al_updates) {
 576                         if (drbd_md_sync_page_io(mdev, mdev->ldev, sector, WRITE)) {
 577                                 err = -EIO;
 578                                 drbd_chk_io_error(mdev, 1, DRBD_META_IO_ERROR);
 579                         } else {
 580                                 mdev->al_tr_number++;
 581                                 mdev->al_writ_cnt++;
 582                         }
 583                 }
 584         }
 585
 586         drbd_md_put_buffer(mdev);
 587         put_ldev(mdev);
 588
 589         return err;
 590 }
 591
 592
 593 static int w_al_write_transaction(struct drbd_work *w, int unused)
 594 {
 595         struct update_al_work *aw = container_of(w, struct update_al_work, w);
 596         struct drbd_conf *mdev = w->mdev;
 597         int err;
 598
 599         err = _al_write_transaction(mdev);
 600         aw->err = err;
 601         complete(&aw->event);
 602
 603         return err != -EIO ? err : 0;
 604 }
 605
 606 /* Calls from worker context (see w_restart_disk_io()) need to write the
 607    transaction directly. Others came through generic_make_request(),
 608    those need to delegate it to the worker. */
 609 static int al_write_transaction(struct drbd_conf *mdev, bool delegate)
 610 {
 611         if (delegate) {
 612                 struct update_al_work al_work;
 613                 init_completion(&al_work.event);
 614                 al_work.w.cb = w_al_write_transaction;
 615                 al_work.w.mdev = mdev;
 616                 drbd_queue_work_front(&mdev->tconn->sender_work, &al_work.w);
 617                 wait_for_completion(&al_work.event);
 618                 return al_work.err;
 619         } else
 620                 return _al_write_transaction(mdev);
 621 }
 622
 623 static int _try_lc_del(struct drbd_conf *mdev, struct lc_element *al_ext)
 624 {
 625         int rv;
 626
 627         spin_lock_irq(&mdev->al_lock);
 628         rv = (al_ext->refcnt == 0);
 629         if (likely(rv))
 630                 lc_del(mdev->act_log, al_ext);
 631         spin_unlock_irq(&mdev->al_lock);
 632
 633         return rv;
 634 }
 635
 636 /**
 637  * drbd_al_shrink() - Removes all active extents form the activity log
 638  * @mdev:       DRBD device.
 639  *
 640  * Removes all active extents form the activity log, waiting until
 641  * the reference count of each entry dropped to 0 first, of course.
 642  *
 643  * You need to lock mdev->act_log with lc_try_lock() / lc_unlock()
 644  */
 645 void drbd_al_shrink(struct drbd_conf *mdev)
 646 {
 647         struct lc_element *al_ext;
 648         int i;
 649
 650         D_ASSERT(test_bit(__LC_LOCKED, &mdev->act_log->flags));
 651
 652         for (i = 0; i < mdev->act_log->nr_elements; i++) {
 653                 al_ext = lc_element_by_index(mdev->act_log, i);
 654                 if (al_ext->lc_number == LC_FREE)
 655                         continue;
 656                 wait_event(mdev->al_wait, _try_lc_del(mdev, al_ext));
 657         }
 658
 659         wake_up(&mdev->al_wait);
 660 }
 661
 662 static int w_update_odbm(struct drbd_work *w, int unused)
 663 {
 664         struct update_odbm_work *udw = container_of(w, struct update_odbm_work, w);
 665         struct drbd_conf *mdev = w->mdev;
 666         struct sib_info sib = { .sib_reason = SIB_SYNC_PROGRESS, };
 667
 668         if (!get_ldev(mdev)) {
 669                 if (__ratelimit(&drbd_ratelimit_state))
 670                         dev_warn(DEV, "Can not update on disk bitmap, local IO disabled.\n");
 671                 kfree(udw);
 672                 return 0;
 673         }
 674
 675         drbd_bm_write_page(mdev, rs_extent_to_bm_page(udw->enr));
 676         put_ldev(mdev);
 677
 678         kfree(udw);
 679
 680         if (drbd_bm_total_weight(mdev) <= mdev->rs_failed) {
 681                 switch (mdev->state.conn) {
 682                 case C_SYNC_SOURCE:  case C_SYNC_TARGET:
 683                 case C_PAUSED_SYNC_S: case C_PAUSED_SYNC_T:
 684                         drbd_resync_finished(mdev);
 685                 default:
 686                         /* nothing to do */
 687                         break;
 688                 }
 689         }
 690         drbd_bcast_event(mdev, &sib);
 691
 692         return 0;
 693 }
 694
 695
 696 /* ATTENTION. The AL's extents are 4MB each, while the extents in the
 697  * resync LRU-cache are 16MB each.
 698  * The caller of this function has to hold an get_ldev() reference.
 699  *
 700  * TODO will be obsoleted once we have a caching lru of the on disk bitmap
 701  */
 702 static void drbd_try_clear_on_disk_bm(struct drbd_conf *mdev, sector_t sector,
 703                                       int count, int success)
 704 {
 705         struct lc_element *e;
 706         struct update_odbm_work *udw;
 707
 708         unsigned int enr;
 709
 710         D_ASSERT(atomic_read(&mdev->local_cnt));
 711
 712         /* I simply assume that a sector/size pair never crosses
 713          * a 16 MB extent border. (Currently this is true...) */
 714         enr = BM_SECT_TO_EXT(sector);
 715
 716         e = lc_get(mdev->resync, enr);
 717         if (e) {
 718                 struct bm_extent *ext = lc_entry(e, struct bm_extent, lce);
 719                 if (ext->lce.lc_number == enr) {
 720                         if (success)
 721                                 ext->rs_left -= count;
 722                         else
 723                                 ext->rs_failed += count;
 724                         if (ext->rs_left < ext->rs_failed) {
 725                                 dev_warn(DEV, "BAD! sector=%llus enr=%u rs_left=%d "
 726                                     "rs_failed=%d count=%d cstate=%s\n",
 727                                      (unsigned long long)sector,
 728                                      ext->lce.lc_number, ext->rs_left,
 729                                      ext->rs_failed, count,
 730                                      drbd_conn_str(mdev->state.conn));
 731
 732                                 /* We don't expect to be able to clear more bits
 733                                  * than have been set when we originally counted
 734                                  * the set bits to cache that value in ext->rs_left.
 735                                  * Whatever the reason (disconnect during resync,
 736                                  * delayed local completion of an application write),
 737                                  * try to fix it up by recounting here. */
 738                                 ext->rs_left = drbd_bm_e_weight(mdev, enr);
 739                         }
 740                 } else {
 741                         /* Normally this element should be in the cache,
 742                          * since drbd_rs_begin_io() pulled it already in.
 743                          *
 744                          * But maybe an application write finished, and we set
 745                          * something outside the resync lru_cache in sync.
 746                          */
 747                         int rs_left = drbd_bm_e_weight(mdev, enr);
 748                         if (ext->flags != 0) {
 749                                 dev_warn(DEV, "changing resync lce: %d[%u;%02lx]"
 750                                      " -> %d[%u;00]\n",
 751                                      ext->lce.lc_number, ext->rs_left,
 752                                      ext->flags, enr, rs_left);
 753                                 ext->flags = 0;
 754                         }
 755                         if (ext->rs_failed) {
 756                                 dev_warn(DEV, "Kicking resync_lru element enr=%u "
 757                                      "out with rs_failed=%d\n",
 758                                      ext->lce.lc_number, ext->rs_failed);
 759                         }
 760                         ext->rs_left = rs_left;
 761                         ext->rs_failed = success ? 0 : count;
 762                         /* we don't keep a persistent log of the resync lru,
 763                          * we can commit any change right away. */
 764                         lc_committed(mdev->resync);
 765                 }
 766                 lc_put(mdev->resync, &ext->lce);
 767                 /* no race, we are within the al_lock! */
 768
 769                 if (ext->rs_left == ext->rs_failed) {
 770                         ext->rs_failed = 0;
 771
 772                         udw = kmalloc(sizeof(*udw), GFP_ATOMIC);
 773                         if (udw) {
 774                                 udw->enr = ext->lce.lc_number;
 775                                 udw->w.cb = w_update_odbm;
 776                                 udw->w.mdev = mdev;
 777                                 drbd_queue_work_front(&mdev->tconn->sender_work, &udw->w);
 778                         } else {
 779                                 dev_warn(DEV, "Could not kmalloc an udw\n");
 780                         }
 781                 }
 782         } else {
 783                 dev_err(DEV, "lc_get() failed! locked=%d/%d flags=%lu\n",
 784                     mdev->resync_locked,
 785                     mdev->resync->nr_elements,
 786                     mdev->resync->flags);
 787         }
 788 }
 789
 790 void drbd_advance_rs_marks(struct drbd_conf *mdev, unsigned long still_to_go)
 791 {
 792         unsigned long now = jiffies;
 793         unsigned long last = mdev->rs_mark_time[mdev->rs_last_mark];
 794         int next = (mdev->rs_last_mark + 1) % DRBD_SYNC_MARKS;
 795         if (time_after_eq(now, last + DRBD_SYNC_MARK_STEP)) {
 796                 if (mdev->rs_mark_left[mdev->rs_last_mark] != still_to_go &&
 797                     mdev->state.conn != C_PAUSED_SYNC_T &&
 798                     mdev->state.conn != C_PAUSED_SYNC_S) {
 799                         mdev->rs_mark_time[next] = now;
 800                         mdev->rs_mark_left[next] = still_to_go;
 801                         mdev->rs_last_mark = next;
 802                 }
 803         }
 804 }
 805
 806 /* clear the bit corresponding to the piece of storage in question:
 807  * size byte of data starting from sector.  Only clear a bits of the affected
 808  * one ore more _aligned_ BM_BLOCK_SIZE blocks.
 809  *
 810  * called by worker on C_SYNC_TARGET and receiver on SyncSource.
 811  *
 812  */
 813 void __drbd_set_in_sync(struct drbd_conf *mdev, sector_t sector, int size,
 814                        const char *file, const unsigned int line)
 815 {
 816         /* Is called from worker and receiver context _only_ */
 817         unsigned long sbnr, ebnr, lbnr;
 818         unsigned long count = 0;
 819         sector_t esector, nr_sectors;
 820         int wake_up = 0;
 821         unsigned long flags;
 822
 823         if (size <= 0 || !IS_ALIGNED(size, 512) || size > DRBD_MAX_BIO_SIZE) {
 824                 dev_err(DEV, "drbd_set_in_sync: sector=%llus size=%d nonsense!\n",
 825                                 (unsigned long long)sector, size);
 826                 return;
 827         }
 828
 829         if (!get_ldev(mdev))
 830                 return; /* no disk, no metadata, no bitmap to clear bits in */
 831
 832         nr_sectors = drbd_get_capacity(mdev->this_bdev);
 833         esector = sector + (size >> 9) - 1;
 834
 835         if (!expect(sector < nr_sectors))
 836                 goto out;
 837         if (!expect(esector < nr_sectors))
 838                 esector = nr_sectors - 1;
 839
 840         lbnr = BM_SECT_TO_BIT(nr_sectors-1);
 841
 842         /* we clear it (in sync).
 843          * round up start sector, round down end sector.  we make sure we only
 844          * clear full, aligned, BM_BLOCK_SIZE (4K) blocks */
 845         if (unlikely(esector < BM_SECT_PER_BIT-1))
 846                 goto out;
 847         if (unlikely(esector == (nr_sectors-1)))
 848                 ebnr = lbnr;
 849         else
 850                 ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1));
 851         sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1);
 852
 853         if (sbnr > ebnr)
 854                 goto out;
 855
 856         /*
 857          * ok, (capacity & 7) != 0 sometimes, but who cares...
 858          * we count rs_{total,left} in bits, not sectors.
 859          */
 860         count = drbd_bm_clear_bits(mdev, sbnr, ebnr);
 861         if (count) {
 862                 drbd_advance_rs_marks(mdev, drbd_bm_total_weight(mdev));
 863                 spin_lock_irqsave(&mdev->al_lock, flags);
 864                 drbd_try_clear_on_disk_bm(mdev, sector, count, true);
 865                 spin_unlock_irqrestore(&mdev->al_lock, flags);
 866
 867                 /* just wake_up unconditional now, various lc_chaged(),
 868                  * lc_put() in drbd_try_clear_on_disk_bm(). */
 869                 wake_up = 1;
 870         }
 871 out:
 872         put_ldev(mdev);
 873         if (wake_up)
 874                 wake_up(&mdev->al_wait);
 875 }
 876
 877 /*
 878  * this is intended to set one request worth of data out of sync.
 879  * affects at least 1 bit,
 880  * and at most 1+DRBD_MAX_BIO_SIZE/BM_BLOCK_SIZE bits.
 881  *
 882  * called by tl_clear and drbd_send_dblock (==drbd_make_request).
 883  * so this can be _any_ process.
 884  */
 885 int __drbd_set_out_of_sync(struct drbd_conf *mdev, sector_t sector, int size,
 886                             const char *file, const unsigned int line)
 887 {
 888         unsigned long sbnr, ebnr, flags;
 889         sector_t esector, nr_sectors;
 890         unsigned int enr, count = 0;
 891         struct lc_element *e;
 892
 893         /* this should be an empty REQ_FLUSH */
 894         if (size == 0)
 895                 return 0;
 896
 897         if (size < 0 || !IS_ALIGNED(size, 512) || size > DRBD_MAX_BIO_SIZE) {
 898                 dev_err(DEV, "sector: %llus, size: %d\n",
 899                         (unsigned long long)sector, size);
 900                 return 0;
 901         }
 902
 903         if (!get_ldev(mdev))
 904                 return 0; /* no disk, no metadata, no bitmap to set bits in */
 905
 906         nr_sectors = drbd_get_capacity(mdev->this_bdev);
 907         esector = sector + (size >> 9) - 1;
 908
 909         if (!expect(sector < nr_sectors))
 910                 goto out;
 911         if (!expect(esector < nr_sectors))
 912                 esector = nr_sectors - 1;
 913
 914         /* we set it out of sync,
 915          * we do not need to round anything here */
 916         sbnr = BM_SECT_TO_BIT(sector);
 917         ebnr = BM_SECT_TO_BIT(esector);
 918
 919         /* ok, (capacity & 7) != 0 sometimes, but who cares...
 920          * we count rs_{total,left} in bits, not sectors.  */
 921         spin_lock_irqsave(&mdev->al_lock, flags);
 922         count = drbd_bm_set_bits(mdev, sbnr, ebnr);
 923
 924         enr = BM_SECT_TO_EXT(sector);
 925         e = lc_find(mdev->resync, enr);
 926         if (e)
 927                 lc_entry(e, struct bm_extent, lce)->rs_left += count;
 928         spin_unlock_irqrestore(&mdev->al_lock, flags);
 929
 930 out:
 931         put_ldev(mdev);
 932
 933         return count;
 934 }
 935
 936 static
 937 struct bm_extent *_bme_get(struct drbd_conf *mdev, unsigned int enr)
 938 {
 939         struct lc_element *e;
 940         struct bm_extent *bm_ext;
 941         int wakeup = 0;
 942         unsigned long rs_flags;
 943
 944         spin_lock_irq(&mdev->al_lock);
 945         if (mdev->resync_locked > mdev->resync->nr_elements/2) {
 946                 spin_unlock_irq(&mdev->al_lock);
 947                 return NULL;
 948         }
 949         e = lc_get(mdev->resync, enr);
 950         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
 951         if (bm_ext) {
 952                 if (bm_ext->lce.lc_number != enr) {
 953                         bm_ext->rs_left = drbd_bm_e_weight(mdev, enr);
 954                         bm_ext->rs_failed = 0;
 955                         lc_committed(mdev->resync);
 956                         wakeup = 1;
 957                 }
 958                 if (bm_ext->lce.refcnt == 1)
 959                         mdev->resync_locked++;
 960                 set_bit(BME_NO_WRITES, &bm_ext->flags);
 961         }
 962         rs_flags = mdev->resync->flags;
 963         spin_unlock_irq(&mdev->al_lock);
 964         if (wakeup)
 965                 wake_up(&mdev->al_wait);
 966
 967         if (!bm_ext) {
 968                 if (rs_flags & LC_STARVING)
 969                         dev_warn(DEV, "Have to wait for element"
 970                              " (resync LRU too small?)\n");
 971                 BUG_ON(rs_flags & LC_LOCKED);
 972         }
 973
 974         return bm_ext;
 975 }
 976
 977 static int _is_in_al(struct drbd_conf *mdev, unsigned int enr)
 978 {
 979         int rv;
 980
 981         spin_lock_irq(&mdev->al_lock);
 982         rv = lc_is_used(mdev->act_log, enr);
 983         spin_unlock_irq(&mdev->al_lock);
 984
 985         return rv;
 986 }
 987
 988 /**
 989  * drbd_rs_begin_io() - Gets an extent in the resync LRU cache and sets it to BME_LOCKED
 990  * @mdev:       DRBD device.
 991  * @sector:     The sector number.
 992  *
 993  * This functions sleeps on al_wait. Returns 0 on success, -EINTR if interrupted.
 994  */
 995 int drbd_rs_begin_io(struct drbd_conf *mdev, sector_t sector)
 996 {
 997         unsigned int enr = BM_SECT_TO_EXT(sector);
 998         struct bm_extent *bm_ext;
 999         int i, sig;
1000         int sa = 200; /* Step aside 200 times, then grab the extent and let app-IO wait.
1001                          200 times -> 20 seconds. */
1002
1003 retry:
1004         sig = wait_event_interruptible(mdev->al_wait,
1005                         (bm_ext = _bme_get(mdev, enr)));
1006         if (sig)
1007                 return -EINTR;
1008
1009         if (test_bit(BME_LOCKED, &bm_ext->flags))
1010                 return 0;
1011
1012         for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
1013                 sig = wait_event_interruptible(mdev->al_wait,
1014                                                !_is_in_al(mdev, enr * AL_EXT_PER_BM_SECT + i) ||
1015                                                test_bit(BME_PRIORITY, &bm_ext->flags));
1016
1017                 if (sig || (test_bit(BME_PRIORITY, &bm_ext->flags) && sa)) {
1018                         spin_lock_irq(&mdev->al_lock);
1019                         if (lc_put(mdev->resync, &bm_ext->lce) == 0) {
1020                                 bm_ext->flags = 0; /* clears BME_NO_WRITES and eventually BME_PRIORITY */
1021                                 mdev->resync_locked--;
1022                                 wake_up(&mdev->al_wait);
1023                         }
1024                         spin_unlock_irq(&mdev->al_lock);
1025                         if (sig)
1026                                 return -EINTR;
1027                         if (schedule_timeout_interruptible(HZ/10))
1028                                 return -EINTR;
1029                         if (sa && --sa == 0)
1030                                 dev_warn(DEV,"drbd_rs_begin_io() stepped aside for 20sec."
1031                                          "Resync stalled?\n");
1032                         goto retry;
1033                 }
1034         }
1035         set_bit(BME_LOCKED, &bm_ext->flags);
1036         return 0;
1037 }
1038
1039 /**
1040  * drbd_try_rs_begin_io() - Gets an extent in the resync LRU cache, does not sleep
1041  * @mdev:       DRBD device.
1042  * @sector:     The sector number.
1043  *
1044  * Gets an extent in the resync LRU cache, sets it to BME_NO_WRITES, then
1045  * tries to set it to BME_LOCKED. Returns 0 upon success, and -EAGAIN
1046  * if there is still application IO going on in this area.
1047  */
1048 int drbd_try_rs_begin_io(struct drbd_conf *mdev, sector_t sector)
1049 {
1050         unsigned int enr = BM_SECT_TO_EXT(sector);
1051         const unsigned int al_enr = enr*AL_EXT_PER_BM_SECT;
1052         struct lc_element *e;
1053         struct bm_extent *bm_ext;
1054         int i;
1055
1056         spin_lock_irq(&mdev->al_lock);
1057         if (mdev->resync_wenr != LC_FREE && mdev->resync_wenr != enr) {
1058                 /* in case you have very heavy scattered io, it may
1059                  * stall the syncer undefined if we give up the ref count
1060                  * when we try again and requeue.
1061                  *
1062                  * if we don't give up the refcount, but the next time
1063                  * we are scheduled this extent has been "synced" by new
1064                  * application writes, we'd miss the lc_put on the
1065                  * extent we keep the refcount on.
1066                  * so we remembered which extent we had to try again, and
1067                  * if the next requested one is something else, we do
1068                  * the lc_put here...
1069                  * we also have to wake_up
1070                  */
1071                 e = lc_find(mdev->resync, mdev->resync_wenr);
1072                 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1073                 if (bm_ext) {
1074                         D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags));
1075                         D_ASSERT(test_bit(BME_NO_WRITES, &bm_ext->flags));
1076                         clear_bit(BME_NO_WRITES, &bm_ext->flags);
1077                         mdev->resync_wenr = LC_FREE;
1078                         if (lc_put(mdev->resync, &bm_ext->lce) == 0)
1079                                 mdev->resync_locked--;
1080                         wake_up(&mdev->al_wait);
1081                 } else {
1082                         dev_alert(DEV, "LOGIC BUG\n");
1083                 }
1084         }
1085         /* TRY. */
1086         e = lc_try_get(mdev->resync, enr);
1087         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1088         if (bm_ext) {
1089                 if (test_bit(BME_LOCKED, &bm_ext->flags))
1090                         goto proceed;
1091                 if (!test_and_set_bit(BME_NO_WRITES, &bm_ext->flags)) {
1092                         mdev->resync_locked++;
1093                 } else {
1094                         /* we did set the BME_NO_WRITES,
1095                          * but then could not set BME_LOCKED,
1096                          * so we tried again.
1097                          * drop the extra reference. */
1098                         bm_ext->lce.refcnt--;
1099                         D_ASSERT(bm_ext->lce.refcnt > 0);
1100                 }
1101                 goto check_al;
1102         } else {
1103                 /* do we rather want to try later? */
1104                 if (mdev->resync_locked > mdev->resync->nr_elements-3)
1105                         goto try_again;
1106                 /* Do or do not. There is no try. -- Yoda */
1107                 e = lc_get(mdev->resync, enr);
1108                 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1109                 if (!bm_ext) {
1110                         const unsigned long rs_flags = mdev->resync->flags;
1111                         if (rs_flags & LC_STARVING)
1112                                 dev_warn(DEV, "Have to wait for element"
1113                                      " (resync LRU too small?)\n");
1114                         BUG_ON(rs_flags & LC_LOCKED);
1115                         goto try_again;
1116                 }
1117                 if (bm_ext->lce.lc_number != enr) {
1118                         bm_ext->rs_left = drbd_bm_e_weight(mdev, enr);
1119                         bm_ext->rs_failed = 0;
1120                         lc_committed(mdev->resync);
1121                         wake_up(&mdev->al_wait);
1122                         D_ASSERT(test_bit(BME_LOCKED, &bm_ext->flags) == 0);
1123                 }
1124                 set_bit(BME_NO_WRITES, &bm_ext->flags);
1125                 D_ASSERT(bm_ext->lce.refcnt == 1);
1126                 mdev->resync_locked++;
1127                 goto check_al;
1128         }
1129 check_al:
1130         for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
1131                 if (lc_is_used(mdev->act_log, al_enr+i))
1132                         goto try_again;
1133         }
1134         set_bit(BME_LOCKED, &bm_ext->flags);
1135 proceed:
1136         mdev->resync_wenr = LC_FREE;
1137         spin_unlock_irq(&mdev->al_lock);
1138         return 0;
1139
1140 try_again:
1141         if (bm_ext)
1142                 mdev->resync_wenr = enr;
1143         spin_unlock_irq(&mdev->al_lock);
1144         return -EAGAIN;
1145 }
1146
1147 void drbd_rs_complete_io(struct drbd_conf *mdev, sector_t sector)
1148 {
1149         unsigned int enr = BM_SECT_TO_EXT(sector);
1150         struct lc_element *e;
1151         struct bm_extent *bm_ext;
1152         unsigned long flags;
1153
1154         spin_lock_irqsave(&mdev->al_lock, flags);
1155         e = lc_find(mdev->resync, enr);
1156         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1157         if (!bm_ext) {
1158                 spin_unlock_irqrestore(&mdev->al_lock, flags);
1159                 if (__ratelimit(&drbd_ratelimit_state))
1160                         dev_err(DEV, "drbd_rs_complete_io() called, but extent not found\n");
1161                 return;
1162         }
1163
1164         if (bm_ext->lce.refcnt == 0) {
1165                 spin_unlock_irqrestore(&mdev->al_lock, flags);
1166                 dev_err(DEV, "drbd_rs_complete_io(,%llu [=%u]) called, "
1167                     "but refcnt is 0!?\n",
1168                     (unsigned long long)sector, enr);
1169                 return;
1170         }
1171
1172         if (lc_put(mdev->resync, &bm_ext->lce) == 0) {
1173                 bm_ext->flags = 0; /* clear BME_LOCKED, BME_NO_WRITES and BME_PRIORITY */
1174                 mdev->resync_locked--;
1175                 wake_up(&mdev->al_wait);
1176         }
1177
1178         spin_unlock_irqrestore(&mdev->al_lock, flags);
1179 }
1180
1181 /**
1182  * drbd_rs_cancel_all() - Removes all extents from the resync LRU (even BME_LOCKED)
1183  * @mdev:       DRBD device.
1184  */
1185 void drbd_rs_cancel_all(struct drbd_conf *mdev)
1186 {
1187         spin_lock_irq(&mdev->al_lock);
1188
1189         if (get_ldev_if_state(mdev, D_FAILED)) { /* Makes sure ->resync is there. */
1190                 lc_reset(mdev->resync);
1191                 put_ldev(mdev);
1192         }
1193         mdev->resync_locked = 0;
1194         mdev->resync_wenr = LC_FREE;
1195         spin_unlock_irq(&mdev->al_lock);
1196         wake_up(&mdev->al_wait);
1197 }
1198
1199 /**
1200  * drbd_rs_del_all() - Gracefully remove all extents from the resync LRU
1201  * @mdev:       DRBD device.
1202  *
1203  * Returns 0 upon success, -EAGAIN if at least one reference count was
1204  * not zero.
1205  */
1206 int drbd_rs_del_all(struct drbd_conf *mdev)
1207 {
1208         struct lc_element *e;
1209         struct bm_extent *bm_ext;
1210         int i;
1211
1212         spin_lock_irq(&mdev->al_lock);
1213
1214         if (get_ldev_if_state(mdev, D_FAILED)) {
1215                 /* ok, ->resync is there. */
1216                 for (i = 0; i < mdev->resync->nr_elements; i++) {
1217                         e = lc_element_by_index(mdev->resync, i);
1218                         bm_ext = lc_entry(e, struct bm_extent, lce);
1219                         if (bm_ext->lce.lc_number == LC_FREE)
1220                                 continue;
1221                         if (bm_ext->lce.lc_number == mdev->resync_wenr) {
1222                                 dev_info(DEV, "dropping %u in drbd_rs_del_all, apparently"
1223                                      " got 'synced' by application io\n",
1224                                      mdev->resync_wenr);
1225                                 D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags));
1226                                 D_ASSERT(test_bit(BME_NO_WRITES, &bm_ext->flags));
1227                                 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1228                                 mdev->resync_wenr = LC_FREE;
1229                                 lc_put(mdev->resync, &bm_ext->lce);
1230                         }
1231                         if (bm_ext->lce.refcnt != 0) {
1232                                 dev_info(DEV, "Retrying drbd_rs_del_all() later. "
1233                                      "refcnt=%d\n", bm_ext->lce.refcnt);
1234                                 put_ldev(mdev);
1235                                 spin_unlock_irq(&mdev->al_lock);
1236                                 return -EAGAIN;
1237                         }
1238                         D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags));
1239                         D_ASSERT(!test_bit(BME_NO_WRITES, &bm_ext->flags));
1240                         lc_del(mdev->resync, &bm_ext->lce);
1241                 }
1242                 D_ASSERT(mdev->resync->used == 0);
1243                 put_ldev(mdev);
1244         }
1245         spin_unlock_irq(&mdev->al_lock);
1246         wake_up(&mdev->al_wait);
1247
1248         return 0;
1249 }
1250
1251 /**
1252  * drbd_rs_failed_io() - Record information on a failure to resync the specified blocks
1253  * @mdev:       DRBD device.
1254  * @sector:     The sector number.
1255  * @size:       Size of failed IO operation, in byte.
1256  */
1257 void drbd_rs_failed_io(struct drbd_conf *mdev, sector_t sector, int size)
1258 {
1259         /* Is called from worker and receiver context _only_ */
1260         unsigned long sbnr, ebnr, lbnr;
1261         unsigned long count;
1262         sector_t esector, nr_sectors;
1263         int wake_up = 0;
1264
1265         if (size <= 0 || !IS_ALIGNED(size, 512) || size > DRBD_MAX_BIO_SIZE) {
1266                 dev_err(DEV, "drbd_rs_failed_io: sector=%llus size=%d nonsense!\n",
1267                                 (unsigned long long)sector, size);
1268                 return;
1269         }
1270         nr_sectors = drbd_get_capacity(mdev->this_bdev);
1271         esector = sector + (size >> 9) - 1;
1272
1273         if (!expect(sector < nr_sectors))
1274                 return;
1275         if (!expect(esector < nr_sectors))
1276                 esector = nr_sectors - 1;
1277
1278         lbnr = BM_SECT_TO_BIT(nr_sectors-1);
1279
1280         /*
1281          * round up start sector, round down end sector.  we make sure we only
1282          * handle full, aligned, BM_BLOCK_SIZE (4K) blocks */
1283         if (unlikely(esector < BM_SECT_PER_BIT-1))
1284                 return;
1285         if (unlikely(esector == (nr_sectors-1)))
1286                 ebnr = lbnr;
1287         else
1288                 ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1));
1289         sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1);
1290
1291         if (sbnr > ebnr)
1292                 return;
1293
1294         /*
1295          * ok, (capacity & 7) != 0 sometimes, but who cares...
1296          * we count rs_{total,left} in bits, not sectors.
1297          */
1298         spin_lock_irq(&mdev->al_lock);
1299         count = drbd_bm_count_bits(mdev, sbnr, ebnr);
1300         if (count) {
1301                 mdev->rs_failed += count;
1302
1303                 if (get_ldev(mdev)) {
1304                         drbd_try_clear_on_disk_bm(mdev, sector, count, false);
1305                         put_ldev(mdev);
1306                 }
1307
1308                 /* just wake_up unconditional now, various lc_chaged(),
1309                  * lc_put() in drbd_try_clear_on_disk_bm(). */
1310                 wake_up = 1;
1311         }
1312         spin_unlock_irq(&mdev->al_lock);
1313         if (wake_up)
1314                 wake_up(&mdev->al_wait);
1315 }