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 "drbd_int.h"
  31
  32
  33 enum al_transaction_types {
  34         AL_TR_UPDATE = 0,
  35         AL_TR_INITIALIZED = 0xffff
  36 };
  37 /* all fields on disc in big endian */
  38 struct __packed al_transaction_on_disk {
  39         /* don't we all like magic */
  40         __be32  magic;
  41
  42         /* to identify the most recent transaction block
  43          * in the on disk ring buffer */
  44         __be32  tr_number;
  45
  46         /* checksum on the full 4k block, with this field set to 0. */
  47         __be32  crc32c;
  48
  49         /* type of transaction, special transaction types like:
  50          * purge-all, set-all-idle, set-all-active, ... to-be-defined
  51          * see also enum al_transaction_types */
  52         __be16  transaction_type;
  53
  54         /* we currently allow only a few thousand extents,
  55          * so 16bit will be enough for the slot number. */
  56
  57         /* how many updates in this transaction */
  58         __be16  n_updates;
  59
  60         /* maximum slot number, "al-extents" in drbd.conf speak.
  61          * Having this in each transaction should make reconfiguration
  62          * of that parameter easier. */
  63         __be16  context_size;
  64
  65         /* slot number the context starts with */
  66         __be16  context_start_slot_nr;
  67
  68         /* Some reserved bytes.  Expected usage is a 64bit counter of
  69          * sectors-written since device creation, and other data generation tag
  70          * supporting usage */
  71         __be32  __reserved[4];
  72
  73         /* --- 36 byte used --- */
  74
  75         /* Reserve space for up to AL_UPDATES_PER_TRANSACTION changes
  76          * in one transaction, then use the remaining byte in the 4k block for
  77          * context information.  "Flexible" number of updates per transaction
  78          * does not help, as we have to account for the case when all update
  79          * slots are used anyways, so it would only complicate code without
  80          * additional benefit.
  81          */
  82         __be16  update_slot_nr[AL_UPDATES_PER_TRANSACTION];
  83
  84         /* but the extent number is 32bit, which at an extent size of 4 MiB
  85          * allows to cover device sizes of up to 2**54 Byte (16 PiB) */
  86         __be32  update_extent_nr[AL_UPDATES_PER_TRANSACTION];
  87
  88         /* --- 420 bytes used (36 + 64*6) --- */
  89
  90         /* 4096 - 420 = 3676 = 919 * 4 */
  91         __be32  context[AL_CONTEXT_PER_TRANSACTION];
  92 };
  93
  94 void *drbd_md_get_buffer(struct drbd_device *device, const char *intent)
  95 {
  96         int r;
  97
  98         wait_event(device->misc_wait,
  99                    (r = atomic_cmpxchg(&device->md_io.in_use, 0, 1)) == 0 ||
 100                    device->state.disk <= D_FAILED);
 101
 102         if (r)
 103                 return NULL;
 104
 105         device->md_io.current_use = intent;
 106         device->md_io.start_jif = jiffies;
 107         device->md_io.submit_jif = device->md_io.start_jif - 1;
 108         return page_address(device->md_io.page);
 109 }
 110
 111 void drbd_md_put_buffer(struct drbd_device *device)
 112 {
 113         if (atomic_dec_and_test(&device->md_io.in_use))
 114                 wake_up(&device->misc_wait);
 115 }
 116
 117 void wait_until_done_or_force_detached(struct drbd_device *device, struct drbd_backing_dev *bdev,
 118                                      unsigned int *done)
 119 {
 120         long dt;
 121
 122         rcu_read_lock();
 123         dt = rcu_dereference(bdev->disk_conf)->disk_timeout;
 124         rcu_read_unlock();
 125         dt = dt * HZ / 10;
 126         if (dt == 0)
 127                 dt = MAX_SCHEDULE_TIMEOUT;
 128
 129         dt = wait_event_timeout(device->misc_wait,
 130                         *done || test_bit(FORCE_DETACH, &device->flags), dt);
 131         if (dt == 0) {
 132                 drbd_err(device, "meta-data IO operation timed out\n");
 133                 drbd_chk_io_error(device, 1, DRBD_FORCE_DETACH);
 134         }
 135 }
 136
 137 static int _drbd_md_sync_page_io(struct drbd_device *device,
 138                                  struct drbd_backing_dev *bdev,
 139                                  sector_t sector, int op)
 140 {
 141         struct bio *bio;
 142         /* we do all our meta data IO in aligned 4k blocks. */
 143         const int size = 4096;
 144         int err, op_flags = 0;
 145
 146         device->md_io.done = 0;
 147         device->md_io.error = -ENODEV;
 148
 149         if ((op == REQ_OP_WRITE) && !test_bit(MD_NO_FUA, &device->flags))
 150                 op_flags |= REQ_FUA | REQ_PREFLUSH;
 151         op_flags |= 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, device->md_io.page, size, 0) != size)
 158                 goto out;
 159         bio->bi_private = device;
 160         bio->bi_end_io = drbd_md_endio;
 161         bio_set_op_attrs(bio, op, op_flags);
 162
 163         if (op != REQ_OP_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_endio() */
 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, (op == REQ_OP_WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD))
 177                 bio_io_error(bio);
 178         else
 179                 submit_bio(bio);
 180         wait_until_done_or_force_detached(device, bdev, &device->md_io.done);
 181         if (!bio->bi_error)
 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 op)
 191 {
 192         int err;
 193         D_ASSERT(device, atomic_read(&device->md_io.in_use) == 1);
 194
 195         BUG_ON(!bdev->md_bdev);
 196
 197         dynamic_drbd_dbg(device, "meta_data io: %s [%d]:%s(,%llus,%s) %pS\n",
 198              current->comm, current->pid, __func__,
 199              (unsigned long long)sector, (op == REQ_OP_WRITE) ? "WRITE" : "READ",
 200              (void*)_RET_IP_ );
 201
 202         if (sector < drbd_md_first_sector(bdev) ||
 203             sector + 7 > drbd_md_last_sector(bdev))
 204                 drbd_alert(device, "%s [%d]:%s(,%llus,%s) out of range md access!\n",
 205                      current->comm, current->pid, __func__,
 206                      (unsigned long long)sector,
 207                      (op == REQ_OP_WRITE) ? "WRITE" : "READ");
 208
 209         err = _drbd_md_sync_page_io(device, bdev, sector, op);
 210         if (err) {
 211                 drbd_err(device, "drbd_md_sync_page_io(,%llus,%s) failed with error %d\n",
 212                     (unsigned long long)sector,
 213                     (op == REQ_OP_WRITE) ? "WRITE" : "READ", err);
 214         }
 215         return err;
 216 }
 217
 218 static struct bm_extent *find_active_resync_extent(struct drbd_device *device, unsigned int enr)
 219 {
 220         struct lc_element *tmp;
 221         tmp = lc_find(device->resync, enr/AL_EXT_PER_BM_SECT);
 222         if (unlikely(tmp != NULL)) {
 223                 struct bm_extent  *bm_ext = lc_entry(tmp, struct bm_extent, lce);
 224                 if (test_bit(BME_NO_WRITES, &bm_ext->flags))
 225                         return bm_ext;
 226         }
 227         return NULL;
 228 }
 229
 230 static struct lc_element *_al_get(struct drbd_device *device, unsigned int enr, bool nonblock)
 231 {
 232         struct lc_element *al_ext;
 233         struct bm_extent *bm_ext;
 234         int wake;
 235
 236         spin_lock_irq(&device->al_lock);
 237         bm_ext = find_active_resync_extent(device, enr);
 238         if (bm_ext) {
 239                 wake = !test_and_set_bit(BME_PRIORITY, &bm_ext->flags);
 240                 spin_unlock_irq(&device->al_lock);
 241                 if (wake)
 242                         wake_up(&device->al_wait);
 243                 return NULL;
 244         }
 245         if (nonblock)
 246                 al_ext = lc_try_get(device->act_log, enr);
 247         else
 248                 al_ext = lc_get(device->act_log, enr);
 249         spin_unlock_irq(&device->al_lock);
 250         return al_ext;
 251 }
 252
 253 bool drbd_al_begin_io_fastpath(struct drbd_device *device, struct drbd_interval *i)
 254 {
 255         /* for bios crossing activity log extent boundaries,
 256          * we may need to activate two extents in one go */
 257         unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
 258         unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
 259
 260         D_ASSERT(device, first <= last);
 261         D_ASSERT(device, atomic_read(&device->local_cnt) > 0);
 262
 263         /* FIXME figure out a fast path for bios crossing AL extent boundaries */
 264         if (first != last)
 265                 return false;
 266
 267         return _al_get(device, first, true);
 268 }
 269
 270 bool drbd_al_begin_io_prepare(struct drbd_device *device, struct drbd_interval *i)
 271 {
 272         /* for bios crossing activity log extent boundaries,
 273          * we may need to activate two extents in one go */
 274         unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
 275         unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
 276         unsigned enr;
 277         bool need_transaction = false;
 278
 279         D_ASSERT(device, first <= last);
 280         D_ASSERT(device, atomic_read(&device->local_cnt) > 0);
 281
 282         for (enr = first; enr <= last; enr++) {
 283                 struct lc_element *al_ext;
 284                 wait_event(device->al_wait,
 285                                 (al_ext = _al_get(device, enr, false)) != NULL);
 286                 if (al_ext->lc_number != enr)
 287                         need_transaction = true;
 288         }
 289         return need_transaction;
 290 }
 291
 292 #if (PAGE_SHIFT + 3) < (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT)
 293 /* Currently BM_BLOCK_SHIFT, BM_EXT_SHIFT and AL_EXTENT_SHIFT
 294  * are still coupled, or assume too much about their relation.
 295  * Code below will not work if this is violated.
 296  * Will be cleaned up with some followup patch.
 297  */
 298 # error FIXME
 299 #endif
 300
 301 static unsigned int al_extent_to_bm_page(unsigned int al_enr)
 302 {
 303         return al_enr >>
 304                 /* bit to page */
 305                 ((PAGE_SHIFT + 3) -
 306                 /* al extent number to bit */
 307                  (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT));
 308 }
 309
 310 static sector_t al_tr_number_to_on_disk_sector(struct drbd_device *device)
 311 {
 312         const unsigned int stripes = device->ldev->md.al_stripes;
 313         const unsigned int stripe_size_4kB = device->ldev->md.al_stripe_size_4k;
 314
 315         /* transaction number, modulo on-disk ring buffer wrap around */
 316         unsigned int t = device->al_tr_number % (device->ldev->md.al_size_4k);
 317
 318         /* ... to aligned 4k on disk block */
 319         t = ((t % stripes) * stripe_size_4kB) + t/stripes;
 320
 321         /* ... to 512 byte sector in activity log */
 322         t *= 8;
 323
 324         /* ... plus offset to the on disk position */
 325         return device->ldev->md.md_offset + device->ldev->md.al_offset + t;
 326 }
 327
 328 static int __al_write_transaction(struct drbd_device *device, struct al_transaction_on_disk *buffer)
 329 {
 330         struct lc_element *e;
 331         sector_t sector;
 332         int i, mx;
 333         unsigned extent_nr;
 334         unsigned crc = 0;
 335         int err = 0;
 336
 337         memset(buffer, 0, sizeof(*buffer));
 338         buffer->magic = cpu_to_be32(DRBD_AL_MAGIC);
 339         buffer->tr_number = cpu_to_be32(device->al_tr_number);
 340
 341         i = 0;
 342
 343         drbd_bm_reset_al_hints(device);
 344
 345         /* Even though no one can start to change this list
 346          * once we set the LC_LOCKED -- from drbd_al_begin_io(),
 347          * lc_try_lock_for_transaction() --, someone may still
 348          * be in the process of changing it. */
 349         spin_lock_irq(&device->al_lock);
 350         list_for_each_entry(e, &device->act_log->to_be_changed, list) {
 351                 if (i == AL_UPDATES_PER_TRANSACTION) {
 352                         i++;
 353                         break;
 354                 }
 355                 buffer->update_slot_nr[i] = cpu_to_be16(e->lc_index);
 356                 buffer->update_extent_nr[i] = cpu_to_be32(e->lc_new_number);
 357                 if (e->lc_number != LC_FREE)
 358                         drbd_bm_mark_for_writeout(device,
 359                                         al_extent_to_bm_page(e->lc_number));
 360                 i++;
 361         }
 362         spin_unlock_irq(&device->al_lock);
 363         BUG_ON(i > AL_UPDATES_PER_TRANSACTION);
 364
 365         buffer->n_updates = cpu_to_be16(i);
 366         for ( ; i < AL_UPDATES_PER_TRANSACTION; i++) {
 367                 buffer->update_slot_nr[i] = cpu_to_be16(-1);
 368                 buffer->update_extent_nr[i] = cpu_to_be32(LC_FREE);
 369         }
 370
 371         buffer->context_size = cpu_to_be16(device->act_log->nr_elements);
 372         buffer->context_start_slot_nr = cpu_to_be16(device->al_tr_cycle);
 373
 374         mx = min_t(int, AL_CONTEXT_PER_TRANSACTION,
 375                    device->act_log->nr_elements - device->al_tr_cycle);
 376         for (i = 0; i < mx; i++) {
 377                 unsigned idx = device->al_tr_cycle + i;
 378                 extent_nr = lc_element_by_index(device->act_log, idx)->lc_number;
 379                 buffer->context[i] = cpu_to_be32(extent_nr);
 380         }
 381         for (; i < AL_CONTEXT_PER_TRANSACTION; i++)
 382                 buffer->context[i] = cpu_to_be32(LC_FREE);
 383
 384         device->al_tr_cycle += AL_CONTEXT_PER_TRANSACTION;
 385         if (device->al_tr_cycle >= device->act_log->nr_elements)
 386                 device->al_tr_cycle = 0;
 387
 388         sector = al_tr_number_to_on_disk_sector(device);
 389
 390         crc = crc32c(0, buffer, 4096);
 391         buffer->crc32c = cpu_to_be32(crc);
 392
 393         if (drbd_bm_write_hinted(device))
 394                 err = -EIO;
 395         else {
 396                 bool write_al_updates;
 397                 rcu_read_lock();
 398                 write_al_updates = rcu_dereference(device->ldev->disk_conf)->al_updates;
 399                 rcu_read_unlock();
 400                 if (write_al_updates) {
 401                         if (drbd_md_sync_page_io(device, device->ldev, sector, WRITE)) {
 402                                 err = -EIO;
 403                                 drbd_chk_io_error(device, 1, DRBD_META_IO_ERROR);
 404                         } else {
 405                                 device->al_tr_number++;
 406                                 device->al_writ_cnt++;
 407                         }
 408                 }
 409         }
 410
 411         return err;
 412 }
 413
 414 static int al_write_transaction(struct drbd_device *device)
 415 {
 416         struct al_transaction_on_disk *buffer;
 417         int err;
 418
 419         if (!get_ldev(device)) {
 420                 drbd_err(device, "disk is %s, cannot start al transaction\n",
 421                         drbd_disk_str(device->state.disk));
 422                 return -EIO;
 423         }
 424
 425         /* The bitmap write may have failed, causing a state change. */
 426         if (device->state.disk < D_INCONSISTENT) {
 427                 drbd_err(device,
 428                         "disk is %s, cannot write al transaction\n",
 429                         drbd_disk_str(device->state.disk));
 430                 put_ldev(device);
 431                 return -EIO;
 432         }
 433
 434         /* protects md_io_buffer, al_tr_cycle, ... */
 435         buffer = drbd_md_get_buffer(device, __func__);
 436         if (!buffer) {
 437                 drbd_err(device, "disk failed while waiting for md_io buffer\n");
 438                 put_ldev(device);
 439                 return -ENODEV;
 440         }
 441
 442         err = __al_write_transaction(device, buffer);
 443
 444         drbd_md_put_buffer(device);
 445         put_ldev(device);
 446
 447         return err;
 448 }
 449
 450
 451 void drbd_al_begin_io_commit(struct drbd_device *device)
 452 {
 453         bool locked = false;
 454
 455         /* Serialize multiple transactions.
 456          * This uses test_and_set_bit, memory barrier is implicit.
 457          */
 458         wait_event(device->al_wait,
 459                         device->act_log->pending_changes == 0 ||
 460                         (locked = lc_try_lock_for_transaction(device->act_log)));
 461
 462         if (locked) {
 463                 /* Double check: it may have been committed by someone else,
 464                  * while we have been waiting for the lock. */
 465                 if (device->act_log->pending_changes) {
 466                         bool write_al_updates;
 467
 468                         rcu_read_lock();
 469                         write_al_updates = rcu_dereference(device->ldev->disk_conf)->al_updates;
 470                         rcu_read_unlock();
 471
 472                         if (write_al_updates)
 473                                 al_write_transaction(device);
 474                         spin_lock_irq(&device->al_lock);
 475                         /* FIXME
 476                         if (err)
 477                                 we need an "lc_cancel" here;
 478                         */
 479                         lc_committed(device->act_log);
 480                         spin_unlock_irq(&device->al_lock);
 481                 }
 482                 lc_unlock(device->act_log);
 483                 wake_up(&device->al_wait);
 484         }
 485 }
 486
 487 /*
 488  * @delegate:   delegate activity log I/O to the worker thread
 489  */
 490 void drbd_al_begin_io(struct drbd_device *device, struct drbd_interval *i)
 491 {
 492         if (drbd_al_begin_io_prepare(device, i))
 493                 drbd_al_begin_io_commit(device);
 494 }
 495
 496 int drbd_al_begin_io_nonblock(struct drbd_device *device, struct drbd_interval *i)
 497 {
 498         struct lru_cache *al = device->act_log;
 499         /* for bios crossing activity log extent boundaries,
 500          * we may need to activate two extents in one go */
 501         unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
 502         unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
 503         unsigned nr_al_extents;
 504         unsigned available_update_slots;
 505         unsigned enr;
 506
 507         D_ASSERT(device, first <= last);
 508
 509         nr_al_extents = 1 + last - first; /* worst case: all touched extends are cold. */
 510         available_update_slots = min(al->nr_elements - al->used,
 511                                 al->max_pending_changes - al->pending_changes);
 512
 513         /* We want all necessary updates for a given request within the same transaction
 514          * We could first check how many updates are *actually* needed,
 515          * and use that instead of the worst-case nr_al_extents */
 516         if (available_update_slots < nr_al_extents) {
 517                 /* Too many activity log extents are currently "hot".
 518                  *
 519                  * If we have accumulated pending changes already,
 520                  * we made progress.
 521                  *
 522                  * If we cannot get even a single pending change through,
 523                  * stop the fast path until we made some progress,
 524                  * or requests to "cold" extents could be starved. */
 525                 if (!al->pending_changes)
 526                         __set_bit(__LC_STARVING, &device->act_log->flags);
 527                 return -ENOBUFS;
 528         }
 529
 530         /* Is resync active in this area? */
 531         for (enr = first; enr <= last; enr++) {
 532                 struct lc_element *tmp;
 533                 tmp = lc_find(device->resync, enr/AL_EXT_PER_BM_SECT);
 534                 if (unlikely(tmp != NULL)) {
 535                         struct bm_extent  *bm_ext = lc_entry(tmp, struct bm_extent, lce);
 536                         if (test_bit(BME_NO_WRITES, &bm_ext->flags)) {
 537                                 if (!test_and_set_bit(BME_PRIORITY, &bm_ext->flags))
 538                                         return -EBUSY;
 539                                 return -EWOULDBLOCK;
 540                         }
 541                 }
 542         }
 543
 544         /* Checkout the refcounts.
 545          * Given that we checked for available elements and update slots above,
 546          * this has to be successful. */
 547         for (enr = first; enr <= last; enr++) {
 548                 struct lc_element *al_ext;
 549                 al_ext = lc_get_cumulative(device->act_log, enr);
 550                 if (!al_ext)
 551                         drbd_info(device, "LOGIC BUG for enr=%u\n", enr);
 552         }
 553         return 0;
 554 }
 555
 556 void drbd_al_complete_io(struct drbd_device *device, struct drbd_interval *i)
 557 {
 558         /* for bios crossing activity log extent boundaries,
 559          * we may need to activate two extents in one go */
 560         unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
 561         unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
 562         unsigned enr;
 563         struct lc_element *extent;
 564         unsigned long flags;
 565
 566         D_ASSERT(device, first <= last);
 567         spin_lock_irqsave(&device->al_lock, flags);
 568
 569         for (enr = first; enr <= last; enr++) {
 570                 extent = lc_find(device->act_log, enr);
 571                 if (!extent) {
 572                         drbd_err(device, "al_complete_io() called on inactive extent %u\n", enr);
 573                         continue;
 574                 }
 575                 lc_put(device->act_log, extent);
 576         }
 577         spin_unlock_irqrestore(&device->al_lock, flags);
 578         wake_up(&device->al_wait);
 579 }
 580
 581 static int _try_lc_del(struct drbd_device *device, struct lc_element *al_ext)
 582 {
 583         int rv;
 584
 585         spin_lock_irq(&device->al_lock);
 586         rv = (al_ext->refcnt == 0);
 587         if (likely(rv))
 588                 lc_del(device->act_log, al_ext);
 589         spin_unlock_irq(&device->al_lock);
 590
 591         return rv;
 592 }
 593
 594 /**
 595  * drbd_al_shrink() - Removes all active extents form the activity log
 596  * @device:     DRBD device.
 597  *
 598  * Removes all active extents form the activity log, waiting until
 599  * the reference count of each entry dropped to 0 first, of course.
 600  *
 601  * You need to lock device->act_log with lc_try_lock() / lc_unlock()
 602  */
 603 void drbd_al_shrink(struct drbd_device *device)
 604 {
 605         struct lc_element *al_ext;
 606         int i;
 607
 608         D_ASSERT(device, test_bit(__LC_LOCKED, &device->act_log->flags));
 609
 610         for (i = 0; i < device->act_log->nr_elements; i++) {
 611                 al_ext = lc_element_by_index(device->act_log, i);
 612                 if (al_ext->lc_number == LC_FREE)
 613                         continue;
 614                 wait_event(device->al_wait, _try_lc_del(device, al_ext));
 615         }
 616
 617         wake_up(&device->al_wait);
 618 }
 619
 620 int drbd_al_initialize(struct drbd_device *device, void *buffer)
 621 {
 622         struct al_transaction_on_disk *al = buffer;
 623         struct drbd_md *md = &device->ldev->md;
 624         int al_size_4k = md->al_stripes * md->al_stripe_size_4k;
 625         int i;
 626
 627         __al_write_transaction(device, al);
 628         /* There may or may not have been a pending transaction. */
 629         spin_lock_irq(&device->al_lock);
 630         lc_committed(device->act_log);
 631         spin_unlock_irq(&device->al_lock);
 632
 633         /* The rest of the transactions will have an empty "updates" list, and
 634          * are written out only to provide the context, and to initialize the
 635          * on-disk ring buffer. */
 636         for (i = 1; i < al_size_4k; i++) {
 637                 int err = __al_write_transaction(device, al);
 638                 if (err)
 639                         return err;
 640         }
 641         return 0;
 642 }
 643
 644 static const char *drbd_change_sync_fname[] = {
 645         [RECORD_RS_FAILED] = "drbd_rs_failed_io",
 646         [SET_IN_SYNC] = "drbd_set_in_sync",
 647         [SET_OUT_OF_SYNC] = "drbd_set_out_of_sync"
 648 };
 649
 650 /* ATTENTION. The AL's extents are 4MB each, while the extents in the
 651  * resync LRU-cache are 16MB each.
 652  * The caller of this function has to hold an get_ldev() reference.
 653  *
 654  * Adjusts the caching members ->rs_left (success) or ->rs_failed (!success),
 655  * potentially pulling in (and recounting the corresponding bits)
 656  * this resync extent into the resync extent lru cache.
 657  *
 658  * Returns whether all bits have been cleared for this resync extent,
 659  * precisely: (rs_left <= rs_failed)
 660  *
 661  * TODO will be obsoleted once we have a caching lru of the on disk bitmap
 662  */
 663 static bool update_rs_extent(struct drbd_device *device,
 664                 unsigned int enr, int count,
 665                 enum update_sync_bits_mode mode)
 666 {
 667         struct lc_element *e;
 668
 669         D_ASSERT(device, atomic_read(&device->local_cnt));
 670
 671         /* When setting out-of-sync bits,
 672          * we don't need it cached (lc_find).
 673          * But if it is present in the cache,
 674          * we should update the cached bit count.
 675          * Otherwise, that extent should be in the resync extent lru cache
 676          * already -- or we want to pull it in if necessary -- (lc_get),
 677          * then update and check rs_left and rs_failed. */
 678         if (mode == SET_OUT_OF_SYNC)
 679                 e = lc_find(device->resync, enr);
 680         else
 681                 e = lc_get(device->resync, enr);
 682         if (e) {
 683                 struct bm_extent *ext = lc_entry(e, struct bm_extent, lce);
 684                 if (ext->lce.lc_number == enr) {
 685                         if (mode == SET_IN_SYNC)
 686                                 ext->rs_left -= count;
 687                         else if (mode == SET_OUT_OF_SYNC)
 688                                 ext->rs_left += count;
 689                         else
 690                                 ext->rs_failed += count;
 691                         if (ext->rs_left < ext->rs_failed) {
 692                                 drbd_warn(device, "BAD! enr=%u rs_left=%d "
 693                                     "rs_failed=%d count=%d cstate=%s\n",
 694                                      ext->lce.lc_number, ext->rs_left,
 695                                      ext->rs_failed, count,
 696                                      drbd_conn_str(device->state.conn));
 697
 698                                 /* We don't expect to be able to clear more bits
 699                                  * than have been set when we originally counted
 700                                  * the set bits to cache that value in ext->rs_left.
 701                                  * Whatever the reason (disconnect during resync,
 702                                  * delayed local completion of an application write),
 703                                  * try to fix it up by recounting here. */
 704                                 ext->rs_left = drbd_bm_e_weight(device, enr);
 705                         }
 706                 } else {
 707                         /* Normally this element should be in the cache,
 708                          * since drbd_rs_begin_io() pulled it already in.
 709                          *
 710                          * But maybe an application write finished, and we set
 711                          * something outside the resync lru_cache in sync.
 712                          */
 713                         int rs_left = drbd_bm_e_weight(device, enr);
 714                         if (ext->flags != 0) {
 715                                 drbd_warn(device, "changing resync lce: %d[%u;%02lx]"
 716                                      " -> %d[%u;00]\n",
 717                                      ext->lce.lc_number, ext->rs_left,
 718                                      ext->flags, enr, rs_left);
 719                                 ext->flags = 0;
 720                         }
 721                         if (ext->rs_failed) {
 722                                 drbd_warn(device, "Kicking resync_lru element enr=%u "
 723                                      "out with rs_failed=%d\n",
 724                                      ext->lce.lc_number, ext->rs_failed);
 725                         }
 726                         ext->rs_left = rs_left;
 727                         ext->rs_failed = (mode == RECORD_RS_FAILED) ? count : 0;
 728                         /* we don't keep a persistent log of the resync lru,
 729                          * we can commit any change right away. */
 730                         lc_committed(device->resync);
 731                 }
 732                 if (mode != SET_OUT_OF_SYNC)
 733                         lc_put(device->resync, &ext->lce);
 734                 /* no race, we are within the al_lock! */
 735
 736                 if (ext->rs_left <= ext->rs_failed) {
 737                         ext->rs_failed = 0;
 738                         return true;
 739                 }
 740         } else if (mode != SET_OUT_OF_SYNC) {
 741                 /* be quiet if lc_find() did not find it. */
 742                 drbd_err(device, "lc_get() failed! locked=%d/%d flags=%lu\n",
 743                     device->resync_locked,
 744                     device->resync->nr_elements,
 745                     device->resync->flags);
 746         }
 747         return false;
 748 }
 749
 750 void drbd_advance_rs_marks(struct drbd_device *device, unsigned long still_to_go)
 751 {
 752         unsigned long now = jiffies;
 753         unsigned long last = device->rs_mark_time[device->rs_last_mark];
 754         int next = (device->rs_last_mark + 1) % DRBD_SYNC_MARKS;
 755         if (time_after_eq(now, last + DRBD_SYNC_MARK_STEP)) {
 756                 if (device->rs_mark_left[device->rs_last_mark] != still_to_go &&
 757                     device->state.conn != C_PAUSED_SYNC_T &&
 758                     device->state.conn != C_PAUSED_SYNC_S) {
 759                         device->rs_mark_time[next] = now;
 760                         device->rs_mark_left[next] = still_to_go;
 761                         device->rs_last_mark = next;
 762                 }
 763         }
 764 }
 765
 766 /* It is called lazy update, so don't do write-out too often. */
 767 static bool lazy_bitmap_update_due(struct drbd_device *device)
 768 {
 769         return time_after(jiffies, device->rs_last_bcast + 2*HZ);
 770 }
 771
 772 static void maybe_schedule_on_disk_bitmap_update(struct drbd_device *device, bool rs_done)
 773 {
 774         if (rs_done) {
 775                 struct drbd_connection *connection = first_peer_device(device)->connection;
 776                 if (connection->agreed_pro_version <= 95 ||
 777                     is_sync_target_state(device->state.conn))
 778                         set_bit(RS_DONE, &device->flags);
 779                         /* and also set RS_PROGRESS below */
 780
 781                 /* Else: rather wait for explicit notification via receive_state,
 782                  * to avoid uuids-rotated-too-fast causing full resync
 783                  * in next handshake, in case the replication link breaks
 784                  * at the most unfortunate time... */
 785         } else if (!lazy_bitmap_update_due(device))
 786                 return;
 787
 788         drbd_device_post_work(device, RS_PROGRESS);
 789 }
 790
 791 static int update_sync_bits(struct drbd_device *device,
 792                 unsigned long sbnr, unsigned long ebnr,
 793                 enum update_sync_bits_mode mode)
 794 {
 795         /*
 796          * We keep a count of set bits per resync-extent in the ->rs_left
 797          * caching member, so we need to loop and work within the resync extent
 798          * alignment. Typically this loop will execute exactly once.
 799          */
 800         unsigned long flags;
 801         unsigned long count = 0;
 802         unsigned int cleared = 0;
 803         while (sbnr <= ebnr) {
 804                 /* set temporary boundary bit number to last bit number within
 805                  * the resync extent of the current start bit number,
 806                  * but cap at provided end bit number */
 807                 unsigned long tbnr = min(ebnr, sbnr | BM_BLOCKS_PER_BM_EXT_MASK);
 808                 unsigned long c;
 809
 810                 if (mode == RECORD_RS_FAILED)
 811                         /* Only called from drbd_rs_failed_io(), bits
 812                          * supposedly still set.  Recount, maybe some
 813                          * of the bits have been successfully cleared
 814                          * by application IO meanwhile.
 815                          */
 816                         c = drbd_bm_count_bits(device, sbnr, tbnr);
 817                 else if (mode == SET_IN_SYNC)
 818                         c = drbd_bm_clear_bits(device, sbnr, tbnr);
 819                 else /* if (mode == SET_OUT_OF_SYNC) */
 820                         c = drbd_bm_set_bits(device, sbnr, tbnr);
 821
 822                 if (c) {
 823                         spin_lock_irqsave(&device->al_lock, flags);
 824                         cleared += update_rs_extent(device, BM_BIT_TO_EXT(sbnr), c, mode);
 825                         spin_unlock_irqrestore(&device->al_lock, flags);
 826                         count += c;
 827                 }
 828                 sbnr = tbnr + 1;
 829         }
 830         if (count) {
 831                 if (mode == SET_IN_SYNC) {
 832                         unsigned long still_to_go = drbd_bm_total_weight(device);
 833                         bool rs_is_done = (still_to_go <= device->rs_failed);
 834                         drbd_advance_rs_marks(device, still_to_go);
 835                         if (cleared || rs_is_done)
 836                                 maybe_schedule_on_disk_bitmap_update(device, rs_is_done);
 837                 } else if (mode == RECORD_RS_FAILED)
 838                         device->rs_failed += count;
 839                 wake_up(&device->al_wait);
 840         }
 841         return count;
 842 }
 843
 844 static bool plausible_request_size(int size)
 845 {
 846         return size > 0
 847                 && size <= DRBD_MAX_BATCH_BIO_SIZE
 848                 && IS_ALIGNED(size, 512);
 849 }
 850
 851 /* clear the bit corresponding to the piece of storage in question:
 852  * size byte of data starting from sector.  Only clear a bits of the affected
 853  * one ore more _aligned_ BM_BLOCK_SIZE blocks.
 854  *
 855  * called by worker on C_SYNC_TARGET and receiver on SyncSource.
 856  *
 857  */
 858 int __drbd_change_sync(struct drbd_device *device, sector_t sector, int size,
 859                 enum update_sync_bits_mode mode)
 860 {
 861         /* Is called from worker and receiver context _only_ */
 862         unsigned long sbnr, ebnr, lbnr;
 863         unsigned long count = 0;
 864         sector_t esector, nr_sectors;
 865
 866         /* This would be an empty REQ_PREFLUSH, be silent. */
 867         if ((mode == SET_OUT_OF_SYNC) && size == 0)
 868                 return 0;
 869
 870         if (!plausible_request_size(size)) {
 871                 drbd_err(device, "%s: sector=%llus size=%d nonsense!\n",
 872                                 drbd_change_sync_fname[mode],
 873                                 (unsigned long long)sector, size);
 874                 return 0;
 875         }
 876
 877         if (!get_ldev(device))
 878                 return 0; /* no disk, no metadata, no bitmap to manipulate bits in */
 879
 880         nr_sectors = drbd_get_capacity(device->this_bdev);
 881         esector = sector + (size >> 9) - 1;
 882
 883         if (!expect(sector < nr_sectors))
 884                 goto out;
 885         if (!expect(esector < nr_sectors))
 886                 esector = nr_sectors - 1;
 887
 888         lbnr = BM_SECT_TO_BIT(nr_sectors-1);
 889
 890         if (mode == SET_IN_SYNC) {
 891                 /* Round up start sector, round down end sector.  We make sure
 892                  * we only clear full, aligned, BM_BLOCK_SIZE blocks. */
 893                 if (unlikely(esector < BM_SECT_PER_BIT-1))
 894                         goto out;
 895                 if (unlikely(esector == (nr_sectors-1)))
 896                         ebnr = lbnr;
 897                 else
 898                         ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1));
 899                 sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1);
 900         } else {
 901                 /* We set it out of sync, or record resync failure.
 902                  * Should not round anything here. */
 903                 sbnr = BM_SECT_TO_BIT(sector);
 904                 ebnr = BM_SECT_TO_BIT(esector);
 905         }
 906
 907         count = update_sync_bits(device, sbnr, ebnr, mode);
 908 out:
 909         put_ldev(device);
 910         return count;
 911 }
 912
 913 static
 914 struct bm_extent *_bme_get(struct drbd_device *device, unsigned int enr)
 915 {
 916         struct lc_element *e;
 917         struct bm_extent *bm_ext;
 918         int wakeup = 0;
 919         unsigned long rs_flags;
 920
 921         spin_lock_irq(&device->al_lock);
 922         if (device->resync_locked > device->resync->nr_elements/2) {
 923                 spin_unlock_irq(&device->al_lock);
 924                 return NULL;
 925         }
 926         e = lc_get(device->resync, enr);
 927         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
 928         if (bm_ext) {
 929                 if (bm_ext->lce.lc_number != enr) {
 930                         bm_ext->rs_left = drbd_bm_e_weight(device, enr);
 931                         bm_ext->rs_failed = 0;
 932                         lc_committed(device->resync);
 933                         wakeup = 1;
 934                 }
 935                 if (bm_ext->lce.refcnt == 1)
 936                         device->resync_locked++;
 937                 set_bit(BME_NO_WRITES, &bm_ext->flags);
 938         }
 939         rs_flags = device->resync->flags;
 940         spin_unlock_irq(&device->al_lock);
 941         if (wakeup)
 942                 wake_up(&device->al_wait);
 943
 944         if (!bm_ext) {
 945                 if (rs_flags & LC_STARVING)
 946                         drbd_warn(device, "Have to wait for element"
 947                              " (resync LRU too small?)\n");
 948                 BUG_ON(rs_flags & LC_LOCKED);
 949         }
 950
 951         return bm_ext;
 952 }
 953
 954 static int _is_in_al(struct drbd_device *device, unsigned int enr)
 955 {
 956         int rv;
 957
 958         spin_lock_irq(&device->al_lock);
 959         rv = lc_is_used(device->act_log, enr);
 960         spin_unlock_irq(&device->al_lock);
 961
 962         return rv;
 963 }
 964
 965 /**
 966  * drbd_rs_begin_io() - Gets an extent in the resync LRU cache and sets it to BME_LOCKED
 967  * @device:     DRBD device.
 968  * @sector:     The sector number.
 969  *
 970  * This functions sleeps on al_wait. Returns 0 on success, -EINTR if interrupted.
 971  */
 972 int drbd_rs_begin_io(struct drbd_device *device, sector_t sector)
 973 {
 974         unsigned int enr = BM_SECT_TO_EXT(sector);
 975         struct bm_extent *bm_ext;
 976         int i, sig;
 977         bool sa;
 978
 979 retry:
 980         sig = wait_event_interruptible(device->al_wait,
 981                         (bm_ext = _bme_get(device, enr)));
 982         if (sig)
 983                 return -EINTR;
 984
 985         if (test_bit(BME_LOCKED, &bm_ext->flags))
 986                 return 0;
 987
 988         /* step aside only while we are above c-min-rate; unless disabled. */
 989         sa = drbd_rs_c_min_rate_throttle(device);
 990
 991         for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
 992                 sig = wait_event_interruptible(device->al_wait,
 993                                                !_is_in_al(device, enr * AL_EXT_PER_BM_SECT + i) ||
 994                                                (sa && test_bit(BME_PRIORITY, &bm_ext->flags)));
 995
 996                 if (sig || (sa && test_bit(BME_PRIORITY, &bm_ext->flags))) {
 997                         spin_lock_irq(&device->al_lock);
 998                         if (lc_put(device->resync, &bm_ext->lce) == 0) {
 999                                 bm_ext->flags = 0; /* clears BME_NO_WRITES and eventually BME_PRIORITY */
1000                                 device->resync_locked--;
1001                                 wake_up(&device->al_wait);
1002                         }
1003                         spin_unlock_irq(&device->al_lock);
1004                         if (sig)
1005                                 return -EINTR;
1006                         if (schedule_timeout_interruptible(HZ/10))
1007                                 return -EINTR;
1008                         goto retry;
1009                 }
1010         }
1011         set_bit(BME_LOCKED, &bm_ext->flags);
1012         return 0;
1013 }
1014
1015 /**
1016  * drbd_try_rs_begin_io() - Gets an extent in the resync LRU cache, does not sleep
1017  * @device:     DRBD device.
1018  * @sector:     The sector number.
1019  *
1020  * Gets an extent in the resync LRU cache, sets it to BME_NO_WRITES, then
1021  * tries to set it to BME_LOCKED. Returns 0 upon success, and -EAGAIN
1022  * if there is still application IO going on in this area.
1023  */
1024 int drbd_try_rs_begin_io(struct drbd_device *device, sector_t sector)
1025 {
1026         unsigned int enr = BM_SECT_TO_EXT(sector);
1027         const unsigned int al_enr = enr*AL_EXT_PER_BM_SECT;
1028         struct lc_element *e;
1029         struct bm_extent *bm_ext;
1030         int i;
1031         bool throttle = drbd_rs_should_slow_down(device, sector, true);
1032
1033         /* If we need to throttle, a half-locked (only marked BME_NO_WRITES,
1034          * not yet BME_LOCKED) extent needs to be kicked out explicitly if we
1035          * need to throttle. There is at most one such half-locked extent,
1036          * which is remembered in resync_wenr. */
1037
1038         if (throttle && device->resync_wenr != enr)
1039                 return -EAGAIN;
1040
1041         spin_lock_irq(&device->al_lock);
1042         if (device->resync_wenr != LC_FREE && device->resync_wenr != enr) {
1043                 /* in case you have very heavy scattered io, it may
1044                  * stall the syncer undefined if we give up the ref count
1045                  * when we try again and requeue.
1046                  *
1047                  * if we don't give up the refcount, but the next time
1048                  * we are scheduled this extent has been "synced" by new
1049                  * application writes, we'd miss the lc_put on the
1050                  * extent we keep the refcount on.
1051                  * so we remembered which extent we had to try again, and
1052                  * if the next requested one is something else, we do
1053                  * the lc_put here...
1054                  * we also have to wake_up
1055                  */
1056                 e = lc_find(device->resync, device->resync_wenr);
1057                 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1058                 if (bm_ext) {
1059                         D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1060                         D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1061                         clear_bit(BME_NO_WRITES, &bm_ext->flags);
1062                         device->resync_wenr = LC_FREE;
1063                         if (lc_put(device->resync, &bm_ext->lce) == 0) {
1064                                 bm_ext->flags = 0;
1065                                 device->resync_locked--;
1066                         }
1067                         wake_up(&device->al_wait);
1068                 } else {
1069                         drbd_alert(device, "LOGIC BUG\n");
1070                 }
1071         }
1072         /* TRY. */
1073         e = lc_try_get(device->resync, enr);
1074         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1075         if (bm_ext) {
1076                 if (test_bit(BME_LOCKED, &bm_ext->flags))
1077                         goto proceed;
1078                 if (!test_and_set_bit(BME_NO_WRITES, &bm_ext->flags)) {
1079                         device->resync_locked++;
1080                 } else {
1081                         /* we did set the BME_NO_WRITES,
1082                          * but then could not set BME_LOCKED,
1083                          * so we tried again.
1084                          * drop the extra reference. */
1085                         bm_ext->lce.refcnt--;
1086                         D_ASSERT(device, bm_ext->lce.refcnt > 0);
1087                 }
1088                 goto check_al;
1089         } else {
1090                 /* do we rather want to try later? */
1091                 if (device->resync_locked > device->resync->nr_elements-3)
1092                         goto try_again;
1093                 /* Do or do not. There is no try. -- Yoda */
1094                 e = lc_get(device->resync, enr);
1095                 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1096                 if (!bm_ext) {
1097                         const unsigned long rs_flags = device->resync->flags;
1098                         if (rs_flags & LC_STARVING)
1099                                 drbd_warn(device, "Have to wait for element"
1100                                      " (resync LRU too small?)\n");
1101                         BUG_ON(rs_flags & LC_LOCKED);
1102                         goto try_again;
1103                 }
1104                 if (bm_ext->lce.lc_number != enr) {
1105                         bm_ext->rs_left = drbd_bm_e_weight(device, enr);
1106                         bm_ext->rs_failed = 0;
1107                         lc_committed(device->resync);
1108                         wake_up(&device->al_wait);
1109                         D_ASSERT(device, test_bit(BME_LOCKED, &bm_ext->flags) == 0);
1110                 }
1111                 set_bit(BME_NO_WRITES, &bm_ext->flags);
1112                 D_ASSERT(device, bm_ext->lce.refcnt == 1);
1113                 device->resync_locked++;
1114                 goto check_al;
1115         }
1116 check_al:
1117         for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
1118                 if (lc_is_used(device->act_log, al_enr+i))
1119                         goto try_again;
1120         }
1121         set_bit(BME_LOCKED, &bm_ext->flags);
1122 proceed:
1123         device->resync_wenr = LC_FREE;
1124         spin_unlock_irq(&device->al_lock);
1125         return 0;
1126
1127 try_again:
1128         if (bm_ext) {
1129                 if (throttle) {
1130                         D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1131                         D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1132                         clear_bit(BME_NO_WRITES, &bm_ext->flags);
1133                         device->resync_wenr = LC_FREE;
1134                         if (lc_put(device->resync, &bm_ext->lce) == 0) {
1135                                 bm_ext->flags = 0;
1136                                 device->resync_locked--;
1137                         }
1138                         wake_up(&device->al_wait);
1139                 } else
1140                         device->resync_wenr = enr;
1141         }
1142         spin_unlock_irq(&device->al_lock);
1143         return -EAGAIN;
1144 }
1145
1146 void drbd_rs_complete_io(struct drbd_device *device, sector_t sector)
1147 {
1148         unsigned int enr = BM_SECT_TO_EXT(sector);
1149         struct lc_element *e;
1150         struct bm_extent *bm_ext;
1151         unsigned long flags;
1152
1153         spin_lock_irqsave(&device->al_lock, flags);
1154         e = lc_find(device->resync, enr);
1155         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1156         if (!bm_ext) {
1157                 spin_unlock_irqrestore(&device->al_lock, flags);
1158                 if (__ratelimit(&drbd_ratelimit_state))
1159                         drbd_err(device, "drbd_rs_complete_io() called, but extent not found\n");
1160                 return;
1161         }
1162
1163         if (bm_ext->lce.refcnt == 0) {
1164                 spin_unlock_irqrestore(&device->al_lock, flags);
1165                 drbd_err(device, "drbd_rs_complete_io(,%llu [=%u]) called, "
1166                     "but refcnt is 0!?\n",
1167                     (unsigned long long)sector, enr);
1168                 return;
1169         }
1170
1171         if (lc_put(device->resync, &bm_ext->lce) == 0) {
1172                 bm_ext->flags = 0; /* clear BME_LOCKED, BME_NO_WRITES and BME_PRIORITY */
1173                 device->resync_locked--;
1174                 wake_up(&device->al_wait);
1175         }
1176
1177         spin_unlock_irqrestore(&device->al_lock, flags);
1178 }
1179
1180 /**
1181  * drbd_rs_cancel_all() - Removes all extents from the resync LRU (even BME_LOCKED)
1182  * @device:     DRBD device.
1183  */
1184 void drbd_rs_cancel_all(struct drbd_device *device)
1185 {
1186         spin_lock_irq(&device->al_lock);
1187
1188         if (get_ldev_if_state(device, D_FAILED)) { /* Makes sure ->resync is there. */
1189                 lc_reset(device->resync);
1190                 put_ldev(device);
1191         }
1192         device->resync_locked = 0;
1193         device->resync_wenr = LC_FREE;
1194         spin_unlock_irq(&device->al_lock);
1195         wake_up(&device->al_wait);
1196 }
1197
1198 /**
1199  * drbd_rs_del_all() - Gracefully remove all extents from the resync LRU
1200  * @device:     DRBD device.
1201  *
1202  * Returns 0 upon success, -EAGAIN if at least one reference count was
1203  * not zero.
1204  */
1205 int drbd_rs_del_all(struct drbd_device *device)
1206 {
1207         struct lc_element *e;
1208         struct bm_extent *bm_ext;
1209         int i;
1210
1211         spin_lock_irq(&device->al_lock);
1212
1213         if (get_ldev_if_state(device, D_FAILED)) {
1214                 /* ok, ->resync is there. */
1215                 for (i = 0; i < device->resync->nr_elements; i++) {
1216                         e = lc_element_by_index(device->resync, i);
1217                         bm_ext = lc_entry(e, struct bm_extent, lce);
1218                         if (bm_ext->lce.lc_number == LC_FREE)
1219                                 continue;
1220                         if (bm_ext->lce.lc_number == device->resync_wenr) {
1221                                 drbd_info(device, "dropping %u in drbd_rs_del_all, apparently"
1222                                      " got 'synced' by application io\n",
1223                                      device->resync_wenr);
1224                                 D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1225                                 D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1226                                 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1227                                 device->resync_wenr = LC_FREE;
1228                                 lc_put(device->resync, &bm_ext->lce);
1229                         }
1230                         if (bm_ext->lce.refcnt != 0) {
1231                                 drbd_info(device, "Retrying drbd_rs_del_all() later. "
1232                                      "refcnt=%d\n", bm_ext->lce.refcnt);
1233                                 put_ldev(device);
1234                                 spin_unlock_irq(&device->al_lock);
1235                                 return -EAGAIN;
1236                         }
1237                         D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1238                         D_ASSERT(device, !test_bit(BME_NO_WRITES, &bm_ext->flags));
1239                         lc_del(device->resync, &bm_ext->lce);
1240                 }
1241                 D_ASSERT(device, device->resync->used == 0);
1242                 put_ldev(device);
1243         }
1244         spin_unlock_irq(&device->al_lock);
1245         wake_up(&device->al_wait);
1246
1247         return 0;
1248 }