drivers/block/drbd/drbd_actlog.c

   1 /*
   2    drbd_actlog.c
   3
   4    This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
   5
   6    Copyright (C) 2003-2008, LINBIT Information Technologies GmbH.
   7    Copyright (C) 2003-2008, Philipp Reisner <philipp.reisner@linbit.com>.
   8    Copyright (C) 2003-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
   9
  10    drbd is free software; you can redistribute it and/or modify
  11    it under the terms of the GNU General Public License as published by
  12    the Free Software Foundation; either version 2, or (at your option)
  13    any later version.
  14
  15    drbd is distributed in the hope that it will be useful,
  16    but WITHOUT ANY WARRANTY; without even the implied warranty of
  17    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18    GNU General Public License for more details.
  19
  20    You should have received a copy of the GNU General Public License
  21    along with drbd; see the file COPYING.  If not, write to
  22    the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
  23
  24  */
  25
  26 #include <linux/slab.h>
  27 #include <linux/crc32c.h>
  28 #include <linux/drbd.h>
  29 #include <linux/drbd_limits.h>
  30 #include <linux/dynamic_debug.h>
  31 #include "drbd_int.h"
  32
  33
  34 enum al_transaction_types {
  35         AL_TR_UPDATE = 0,
  36         AL_TR_INITIALIZED = 0xffff
  37 };
  38 /* all fields on disc in big endian */
  39 struct __packed al_transaction_on_disk {
  40         /* don't we all like magic */
  41         __be32  magic;
  42
  43         /* to identify the most recent transaction block
  44          * in the on disk ring buffer */
  45         __be32  tr_number;
  46
  47         /* checksum on the full 4k block, with this field set to 0. */
  48         __be32  crc32c;
  49
  50         /* type of transaction, special transaction types like:
  51          * purge-all, set-all-idle, set-all-active, ... to-be-defined
  52          * see also enum al_transaction_types */
  53         __be16  transaction_type;
  54
  55         /* we currently allow only a few thousand extents,
  56          * so 16bit will be enough for the slot number. */
  57
  58         /* how many updates in this transaction */
  59         __be16  n_updates;
  60
  61         /* maximum slot number, "al-extents" in drbd.conf speak.
  62          * Having this in each transaction should make reconfiguration
  63          * of that parameter easier. */
  64         __be16  context_size;
  65
  66         /* slot number the context starts with */
  67         __be16  context_start_slot_nr;
  68
  69         /* Some reserved bytes.  Expected usage is a 64bit counter of
  70          * sectors-written since device creation, and other data generation tag
  71          * supporting usage */
  72         __be32  __reserved[4];
  73
  74         /* --- 36 byte used --- */
  75
  76         /* Reserve space for up to AL_UPDATES_PER_TRANSACTION changes
  77          * in one transaction, then use the remaining byte in the 4k block for
  78          * context information.  "Flexible" number of updates per transaction
  79          * does not help, as we have to account for the case when all update
  80          * slots are used anyways, so it would only complicate code without
  81          * additional benefit.
  82          */
  83         __be16  update_slot_nr[AL_UPDATES_PER_TRANSACTION];
  84
  85         /* but the extent number is 32bit, which at an extent size of 4 MiB
  86          * allows to cover device sizes of up to 2**54 Byte (16 PiB) */
  87         __be32  update_extent_nr[AL_UPDATES_PER_TRANSACTION];
  88
  89         /* --- 420 bytes used (36 + 64*6) --- */
  90
  91         /* 4096 - 420 = 3676 = 919 * 4 */
  92         __be32  context[AL_CONTEXT_PER_TRANSACTION];
  93 };
  94
  95 void *drbd_md_get_buffer(struct drbd_device *device, const char *intent)
  96 {
  97         int r;
  98
  99         wait_event(device->misc_wait,
 100                    (r = atomic_cmpxchg(&device->md_io.in_use, 0, 1)) == 0 ||
 101                    device->state.disk <= D_FAILED);
 102
 103         if (r)
 104                 return NULL;
 105
 106         device->md_io.current_use = intent;
 107         device->md_io.start_jif = jiffies;
 108         device->md_io.submit_jif = device->md_io.start_jif - 1;
 109         return page_address(device->md_io.page);
 110 }
 111
 112 void drbd_md_put_buffer(struct drbd_device *device)
 113 {
 114         if (atomic_dec_and_test(&device->md_io.in_use))
 115                 wake_up(&device->misc_wait);
 116 }
 117
 118 void wait_until_done_or_force_detached(struct drbd_device *device, struct drbd_backing_dev *bdev,
 119                                      unsigned int *done)
 120 {
 121         long dt;
 122
 123         rcu_read_lock();
 124         dt = rcu_dereference(bdev->disk_conf)->disk_timeout;
 125         rcu_read_unlock();
 126         dt = dt * HZ / 10;
 127         if (dt == 0)
 128                 dt = MAX_SCHEDULE_TIMEOUT;
 129
 130         dt = wait_event_timeout(device->misc_wait,
 131                         *done || test_bit(FORCE_DETACH, &device->flags), dt);
 132         if (dt == 0) {
 133                 drbd_err(device, "meta-data IO operation timed out\n");
 134                 drbd_chk_io_error(device, 1, DRBD_FORCE_DETACH);
 135         }
 136 }
 137
 138 static int _drbd_md_sync_page_io(struct drbd_device *device,
 139                                  struct drbd_backing_dev *bdev,
 140                                  sector_t sector, int rw)
 141 {
 142         struct bio *bio;
 143         /* we do all our meta data IO in aligned 4k blocks. */
 144         const int size = 4096;
 145         int err;
 146
 147         device->md_io.done = 0;
 148         device->md_io.error = -ENODEV;
 149
 150         if ((rw & WRITE) && !test_bit(MD_NO_FUA, &device->flags))
 151                 rw |= REQ_FUA | REQ_FLUSH;
 152         rw |= REQ_SYNC | REQ_NOIDLE;
 153
 154         bio = bio_alloc_drbd(GFP_NOIO);
 155         bio->bi_bdev = bdev->md_bdev;
 156         bio->bi_iter.bi_sector = sector;
 157         err = -EIO;
 158         if (bio_add_page(bio, device->md_io.page, size, 0) != size)
 159                 goto out;
 160         bio->bi_private = device;
 161         bio->bi_end_io = drbd_md_endio;
 162         bio->bi_rw = rw;
 163
 164         if (!(rw & WRITE) && device->state.disk == D_DISKLESS && device->ldev == NULL)
 165                 /* special case, drbd_md_read() during drbd_adm_attach(): no get_ldev */
 166                 ;
 167         else if (!get_ldev_if_state(device, D_ATTACHING)) {
 168                 /* Corresponding put_ldev in drbd_md_endio() */
 169                 drbd_err(device, "ASSERT FAILED: get_ldev_if_state() == 1 in _drbd_md_sync_page_io()\n");
 170                 err = -ENODEV;
 171                 goto out;
 172         }
 173
 174         bio_get(bio); /* one bio_put() is in the completion handler */
 175         atomic_inc(&device->md_io.in_use); /* drbd_md_put_buffer() is in the completion handler */
 176         device->md_io.submit_jif = jiffies;
 177         if (drbd_insert_fault(device, (rw & WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD))
 178                 bio_io_error(bio);
 179         else
 180                 submit_bio(rw, bio);
 181         wait_until_done_or_force_detached(device, bdev, &device->md_io.done);
 182         if (!bio->bi_error)
 183                 err = device->md_io.error;
 184
 185  out:
 186         bio_put(bio);
 187         return err;
 188 }
 189
 190 int drbd_md_sync_page_io(struct drbd_device *device, struct drbd_backing_dev *bdev,
 191                          sector_t sector, int rw)
 192 {
 193         int err;
 194         D_ASSERT(device, atomic_read(&device->md_io.in_use) == 1);
 195
 196         BUG_ON(!bdev->md_bdev);
 197
 198         dynamic_drbd_dbg(device, "meta_data io: %s [%d]:%s(,%llus,%s) %pS\n",
 199              current->comm, current->pid, __func__,
 200              (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ",
 201              (void*)_RET_IP_ );
 202
 203         if (sector < drbd_md_first_sector(bdev) ||
 204             sector + 7 > drbd_md_last_sector(bdev))
 205                 drbd_alert(device, "%s [%d]:%s(,%llus,%s) out of range md access!\n",
 206                      current->comm, current->pid, __func__,
 207                      (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ");
 208
 209         err = _drbd_md_sync_page_io(device, bdev, sector, rw);
 210         if (err) {
 211                 drbd_err(device, "drbd_md_sync_page_io(,%llus,%s) failed with error %d\n",
 212                     (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ", err);
 213         }
 214         return err;
 215 }
 216
 217 static struct bm_extent *find_active_resync_extent(struct drbd_device *device, unsigned int enr)
 218 {
 219         struct lc_element *tmp;
 220         tmp = lc_find(device->resync, enr/AL_EXT_PER_BM_SECT);
 221         if (unlikely(tmp != NULL)) {
 222                 struct bm_extent  *bm_ext = lc_entry(tmp, struct bm_extent, lce);
 223                 if (test_bit(BME_NO_WRITES, &bm_ext->flags))
 224                         return bm_ext;
 225         }
 226         return NULL;
 227 }
 228
 229 static struct lc_element *_al_get(struct drbd_device *device, unsigned int enr, bool nonblock)
 230 {
 231         struct lc_element *al_ext;
 232         struct bm_extent *bm_ext;
 233         int wake;
 234
 235         spin_lock_irq(&device->al_lock);
 236         bm_ext = find_active_resync_extent(device, enr);
 237         if (bm_ext) {
 238                 wake = !test_and_set_bit(BME_PRIORITY, &bm_ext->flags);
 239                 spin_unlock_irq(&device->al_lock);
 240                 if (wake)
 241                         wake_up(&device->al_wait);
 242                 return NULL;
 243         }
 244         if (nonblock)
 245                 al_ext = lc_try_get(device->act_log, enr);
 246         else
 247                 al_ext = lc_get(device->act_log, enr);
 248         spin_unlock_irq(&device->al_lock);
 249         return al_ext;
 250 }
 251
 252 bool drbd_al_begin_io_fastpath(struct drbd_device *device, struct drbd_interval *i)
 253 {
 254         /* for bios crossing activity log extent boundaries,
 255          * we may need to activate two extents in one go */
 256         unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
 257         unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
 258
 259         D_ASSERT(device, (unsigned)(last - first) <= 1);
 260         D_ASSERT(device, atomic_read(&device->local_cnt) > 0);
 261
 262         /* FIXME figure out a fast path for bios crossing AL extent boundaries */
 263         if (first != last)
 264                 return false;
 265
 266         return _al_get(device, first, true);
 267 }
 268
 269 bool drbd_al_begin_io_prepare(struct drbd_device *device, struct drbd_interval *i)
 270 {
 271         /* for bios crossing activity log extent boundaries,
 272          * we may need to activate two extents in one go */
 273         unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
 274         unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
 275         unsigned enr;
 276         bool need_transaction = false;
 277
 278         D_ASSERT(device, first <= last);
 279         D_ASSERT(device, atomic_read(&device->local_cnt) > 0);
 280
 281         for (enr = first; enr <= last; enr++) {
 282                 struct lc_element *al_ext;
 283                 wait_event(device->al_wait,
 284                                 (al_ext = _al_get(device, enr, false)) != NULL);
 285                 if (al_ext->lc_number != enr)
 286                         need_transaction = true;
 287         }
 288         return need_transaction;
 289 }
 290
 291 #if (PAGE_SHIFT + 3) < (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT)
 292 /* Currently BM_BLOCK_SHIFT, BM_EXT_SHIFT and AL_EXTENT_SHIFT
 293  * are still coupled, or assume too much about their relation.
 294  * Code below will not work if this is violated.
 295  * Will be cleaned up with some followup patch.
 296  */
 297 # error FIXME
 298 #endif
 299
 300 static unsigned int al_extent_to_bm_page(unsigned int al_enr)
 301 {
 302         return al_enr >>
 303                 /* bit to page */
 304                 ((PAGE_SHIFT + 3) -
 305                 /* al extent number to bit */
 306                  (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT));
 307 }
 308
 309 static sector_t al_tr_number_to_on_disk_sector(struct drbd_device *device)
 310 {
 311         const unsigned int stripes = device->ldev->md.al_stripes;
 312         const unsigned int stripe_size_4kB = device->ldev->md.al_stripe_size_4k;
 313
 314         /* transaction number, modulo on-disk ring buffer wrap around */
 315         unsigned int t = device->al_tr_number % (device->ldev->md.al_size_4k);
 316
 317         /* ... to aligned 4k on disk block */
 318         t = ((t % stripes) * stripe_size_4kB) + t/stripes;
 319
 320         /* ... to 512 byte sector in activity log */
 321         t *= 8;
 322
 323         /* ... plus offset to the on disk position */
 324         return device->ldev->md.md_offset + device->ldev->md.al_offset + t;
 325 }
 326
 327 static int __al_write_transaction(struct drbd_device *device, struct al_transaction_on_disk *buffer)
 328 {
 329         struct lc_element *e;
 330         sector_t sector;
 331         int i, mx;
 332         unsigned extent_nr;
 333         unsigned crc = 0;
 334         int err = 0;
 335
 336         memset(buffer, 0, sizeof(*buffer));
 337         buffer->magic = cpu_to_be32(DRBD_AL_MAGIC);
 338         buffer->tr_number = cpu_to_be32(device->al_tr_number);
 339
 340         i = 0;
 341
 342         /* Even though no one can start to change this list
 343          * once we set the LC_LOCKED -- from drbd_al_begin_io(),
 344          * lc_try_lock_for_transaction() --, someone may still
 345          * be in the process of changing it. */
 346         spin_lock_irq(&device->al_lock);
 347         list_for_each_entry(e, &device->act_log->to_be_changed, list) {
 348                 if (i == AL_UPDATES_PER_TRANSACTION) {
 349                         i++;
 350                         break;
 351                 }
 352                 buffer->update_slot_nr[i] = cpu_to_be16(e->lc_index);
 353                 buffer->update_extent_nr[i] = cpu_to_be32(e->lc_new_number);
 354                 if (e->lc_number != LC_FREE)
 355                         drbd_bm_mark_for_writeout(device,
 356                                         al_extent_to_bm_page(e->lc_number));
 357                 i++;
 358         }
 359         spin_unlock_irq(&device->al_lock);
 360         BUG_ON(i > AL_UPDATES_PER_TRANSACTION);
 361
 362         buffer->n_updates = cpu_to_be16(i);
 363         for ( ; i < AL_UPDATES_PER_TRANSACTION; i++) {
 364                 buffer->update_slot_nr[i] = cpu_to_be16(-1);
 365                 buffer->update_extent_nr[i] = cpu_to_be32(LC_FREE);
 366         }
 367
 368         buffer->context_size = cpu_to_be16(device->act_log->nr_elements);
 369         buffer->context_start_slot_nr = cpu_to_be16(device->al_tr_cycle);
 370
 371         mx = min_t(int, AL_CONTEXT_PER_TRANSACTION,
 372                    device->act_log->nr_elements - device->al_tr_cycle);
 373         for (i = 0; i < mx; i++) {
 374                 unsigned idx = device->al_tr_cycle + i;
 375                 extent_nr = lc_element_by_index(device->act_log, idx)->lc_number;
 376                 buffer->context[i] = cpu_to_be32(extent_nr);
 377         }
 378         for (; i < AL_CONTEXT_PER_TRANSACTION; i++)
 379                 buffer->context[i] = cpu_to_be32(LC_FREE);
 380
 381         device->al_tr_cycle += AL_CONTEXT_PER_TRANSACTION;
 382         if (device->al_tr_cycle >= device->act_log->nr_elements)
 383                 device->al_tr_cycle = 0;
 384
 385         sector = al_tr_number_to_on_disk_sector(device);
 386
 387         crc = crc32c(0, buffer, 4096);
 388         buffer->crc32c = cpu_to_be32(crc);
 389
 390         if (drbd_bm_write_hinted(device))
 391                 err = -EIO;
 392         else {
 393                 bool write_al_updates;
 394                 rcu_read_lock();
 395                 write_al_updates = rcu_dereference(device->ldev->disk_conf)->al_updates;
 396                 rcu_read_unlock();
 397                 if (write_al_updates) {
 398                         if (drbd_md_sync_page_io(device, device->ldev, sector, WRITE)) {
 399                                 err = -EIO;
 400                                 drbd_chk_io_error(device, 1, DRBD_META_IO_ERROR);
 401                         } else {
 402                                 device->al_tr_number++;
 403                                 device->al_writ_cnt++;
 404                         }
 405                 }
 406         }
 407
 408         return err;
 409 }
 410
 411 static int al_write_transaction(struct drbd_device *device)
 412 {
 413         struct al_transaction_on_disk *buffer;
 414         int err;
 415
 416         if (!get_ldev(device)) {
 417                 drbd_err(device, "disk is %s, cannot start al transaction\n",
 418                         drbd_disk_str(device->state.disk));
 419                 return -EIO;
 420         }
 421
 422         /* The bitmap write may have failed, causing a state change. */
 423         if (device->state.disk < D_INCONSISTENT) {
 424                 drbd_err(device,
 425                         "disk is %s, cannot write al transaction\n",
 426                         drbd_disk_str(device->state.disk));
 427                 put_ldev(device);
 428                 return -EIO;
 429         }
 430
 431         /* protects md_io_buffer, al_tr_cycle, ... */
 432         buffer = drbd_md_get_buffer(device, __func__);
 433         if (!buffer) {
 434                 drbd_err(device, "disk failed while waiting for md_io buffer\n");
 435                 put_ldev(device);
 436                 return -ENODEV;
 437         }
 438
 439         err = __al_write_transaction(device, buffer);
 440
 441         drbd_md_put_buffer(device);
 442         put_ldev(device);
 443
 444         return err;
 445 }
 446
 447
 448 void drbd_al_begin_io_commit(struct drbd_device *device)
 449 {
 450         bool locked = false;
 451
 452         /* Serialize multiple transactions.
 453          * This uses test_and_set_bit, memory barrier is implicit.
 454          */
 455         wait_event(device->al_wait,
 456                         device->act_log->pending_changes == 0 ||
 457                         (locked = lc_try_lock_for_transaction(device->act_log)));
 458
 459         if (locked) {
 460                 /* Double check: it may have been committed by someone else,
 461                  * while we have been waiting for the lock. */
 462                 if (device->act_log->pending_changes) {
 463                         bool write_al_updates;
 464
 465                         rcu_read_lock();
 466                         write_al_updates = rcu_dereference(device->ldev->disk_conf)->al_updates;
 467                         rcu_read_unlock();
 468
 469                         if (write_al_updates)
 470                                 al_write_transaction(device);
 471                         spin_lock_irq(&device->al_lock);
 472                         /* FIXME
 473                         if (err)
 474                                 we need an "lc_cancel" here;
 475                         */
 476                         lc_committed(device->act_log);
 477                         spin_unlock_irq(&device->al_lock);
 478                 }
 479                 lc_unlock(device->act_log);
 480                 wake_up(&device->al_wait);
 481         }
 482 }
 483
 484 /*
 485  * @delegate:   delegate activity log I/O to the worker thread
 486  */
 487 void drbd_al_begin_io(struct drbd_device *device, struct drbd_interval *i)
 488 {
 489         if (drbd_al_begin_io_prepare(device, i))
 490                 drbd_al_begin_io_commit(device);
 491 }
 492
 493 int drbd_al_begin_io_nonblock(struct drbd_device *device, struct drbd_interval *i)
 494 {
 495         struct lru_cache *al = device->act_log;
 496         /* for bios crossing activity log extent boundaries,
 497          * we may need to activate two extents in one go */
 498         unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
 499         unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
 500         unsigned nr_al_extents;
 501         unsigned available_update_slots;
 502         unsigned enr;
 503
 504         D_ASSERT(device, first <= last);
 505
 506         nr_al_extents = 1 + last - first; /* worst case: all touched extends are cold. */
 507         available_update_slots = min(al->nr_elements - al->used,
 508                                 al->max_pending_changes - al->pending_changes);
 509
 510         /* We want all necessary updates for a given request within the same transaction
 511          * We could first check how many updates are *actually* needed,
 512          * and use that instead of the worst-case nr_al_extents */
 513         if (available_update_slots < nr_al_extents) {
 514                 /* Too many activity log extents are currently "hot".
 515                  *
 516                  * If we have accumulated pending changes already,
 517                  * we made progress.
 518                  *
 519                  * If we cannot get even a single pending change through,
 520                  * stop the fast path until we made some progress,
 521                  * or requests to "cold" extents could be starved. */
 522                 if (!al->pending_changes)
 523                         __set_bit(__LC_STARVING, &device->act_log->flags);
 524                 return -ENOBUFS;
 525         }
 526
 527         /* Is resync active in this area? */
 528         for (enr = first; enr <= last; enr++) {
 529                 struct lc_element *tmp;
 530                 tmp = lc_find(device->resync, enr/AL_EXT_PER_BM_SECT);
 531                 if (unlikely(tmp != NULL)) {
 532                         struct bm_extent  *bm_ext = lc_entry(tmp, struct bm_extent, lce);
 533                         if (test_bit(BME_NO_WRITES, &bm_ext->flags)) {
 534                                 if (!test_and_set_bit(BME_PRIORITY, &bm_ext->flags))
 535                                         return -EBUSY;
 536                                 return -EWOULDBLOCK;
 537                         }
 538                 }
 539         }
 540
 541         /* Checkout the refcounts.
 542          * Given that we checked for available elements and update slots above,
 543          * this has to be successful. */
 544         for (enr = first; enr <= last; enr++) {
 545                 struct lc_element *al_ext;
 546                 al_ext = lc_get_cumulative(device->act_log, enr);
 547                 if (!al_ext)
 548                         drbd_info(device, "LOGIC BUG for enr=%u\n", enr);
 549         }
 550         return 0;
 551 }
 552
 553 void drbd_al_complete_io(struct drbd_device *device, struct drbd_interval *i)
 554 {
 555         /* for bios crossing activity log extent boundaries,
 556          * we may need to activate two extents in one go */
 557         unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
 558         unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
 559         unsigned enr;
 560         struct lc_element *extent;
 561         unsigned long flags;
 562
 563         D_ASSERT(device, first <= last);
 564         spin_lock_irqsave(&device->al_lock, flags);
 565
 566         for (enr = first; enr <= last; enr++) {
 567                 extent = lc_find(device->act_log, enr);
 568                 if (!extent) {
 569                         drbd_err(device, "al_complete_io() called on inactive extent %u\n", enr);
 570                         continue;
 571                 }
 572                 lc_put(device->act_log, extent);
 573         }
 574         spin_unlock_irqrestore(&device->al_lock, flags);
 575         wake_up(&device->al_wait);
 576 }
 577
 578 static int _try_lc_del(struct drbd_device *device, struct lc_element *al_ext)
 579 {
 580         int rv;
 581
 582         spin_lock_irq(&device->al_lock);
 583         rv = (al_ext->refcnt == 0);
 584         if (likely(rv))
 585                 lc_del(device->act_log, al_ext);
 586         spin_unlock_irq(&device->al_lock);
 587
 588         return rv;
 589 }
 590
 591 /**
 592  * drbd_al_shrink() - Removes all active extents form the activity log
 593  * @device:     DRBD device.
 594  *
 595  * Removes all active extents form the activity log, waiting until
 596  * the reference count of each entry dropped to 0 first, of course.
 597  *
 598  * You need to lock device->act_log with lc_try_lock() / lc_unlock()
 599  */
 600 void drbd_al_shrink(struct drbd_device *device)
 601 {
 602         struct lc_element *al_ext;
 603         int i;
 604
 605         D_ASSERT(device, test_bit(__LC_LOCKED, &device->act_log->flags));
 606
 607         for (i = 0; i < device->act_log->nr_elements; i++) {
 608                 al_ext = lc_element_by_index(device->act_log, i);
 609                 if (al_ext->lc_number == LC_FREE)
 610                         continue;
 611                 wait_event(device->al_wait, _try_lc_del(device, al_ext));
 612         }
 613
 614         wake_up(&device->al_wait);
 615 }
 616
 617 int drbd_al_initialize(struct drbd_device *device, void *buffer)
 618 {
 619         struct al_transaction_on_disk *al = buffer;
 620         struct drbd_md *md = &device->ldev->md;
 621         int al_size_4k = md->al_stripes * md->al_stripe_size_4k;
 622         int i;
 623
 624         __al_write_transaction(device, al);
 625         /* There may or may not have been a pending transaction. */
 626         spin_lock_irq(&device->al_lock);
 627         lc_committed(device->act_log);
 628         spin_unlock_irq(&device->al_lock);
 629
 630         /* The rest of the transactions will have an empty "updates" list, and
 631          * are written out only to provide the context, and to initialize the
 632          * on-disk ring buffer. */
 633         for (i = 1; i < al_size_4k; i++) {
 634                 int err = __al_write_transaction(device, al);
 635                 if (err)
 636                         return err;
 637         }
 638         return 0;
 639 }
 640
 641 static const char *drbd_change_sync_fname[] = {
 642         [RECORD_RS_FAILED] = "drbd_rs_failed_io",
 643         [SET_IN_SYNC] = "drbd_set_in_sync",
 644         [SET_OUT_OF_SYNC] = "drbd_set_out_of_sync"
 645 };
 646
 647 /* ATTENTION. The AL's extents are 4MB each, while the extents in the
 648  * resync LRU-cache are 16MB each.
 649  * The caller of this function has to hold an get_ldev() reference.
 650  *
 651  * Adjusts the caching members ->rs_left (success) or ->rs_failed (!success),
 652  * potentially pulling in (and recounting the corresponding bits)
 653  * this resync extent into the resync extent lru cache.
 654  *
 655  * Returns whether all bits have been cleared for this resync extent,
 656  * precisely: (rs_left <= rs_failed)
 657  *
 658  * TODO will be obsoleted once we have a caching lru of the on disk bitmap
 659  */
 660 static bool update_rs_extent(struct drbd_device *device,
 661                 unsigned int enr, int count,
 662                 enum update_sync_bits_mode mode)
 663 {
 664         struct lc_element *e;
 665
 666         D_ASSERT(device, atomic_read(&device->local_cnt));
 667
 668         /* When setting out-of-sync bits,
 669          * we don't need it cached (lc_find).
 670          * But if it is present in the cache,
 671          * we should update the cached bit count.
 672          * Otherwise, that extent should be in the resync extent lru cache
 673          * already -- or we want to pull it in if necessary -- (lc_get),
 674          * then update and check rs_left and rs_failed. */
 675         if (mode == SET_OUT_OF_SYNC)
 676                 e = lc_find(device->resync, enr);
 677         else
 678                 e = lc_get(device->resync, enr);
 679         if (e) {
 680                 struct bm_extent *ext = lc_entry(e, struct bm_extent, lce);
 681                 if (ext->lce.lc_number == enr) {
 682                         if (mode == SET_IN_SYNC)
 683                                 ext->rs_left -= count;
 684                         else if (mode == SET_OUT_OF_SYNC)
 685                                 ext->rs_left += count;
 686                         else
 687                                 ext->rs_failed += count;
 688                         if (ext->rs_left < ext->rs_failed) {
 689                                 drbd_warn(device, "BAD! enr=%u rs_left=%d "
 690                                     "rs_failed=%d count=%d cstate=%s\n",
 691                                      ext->lce.lc_number, ext->rs_left,
 692                                      ext->rs_failed, count,
 693                                      drbd_conn_str(device->state.conn));
 694
 695                                 /* We don't expect to be able to clear more bits
 696                                  * than have been set when we originally counted
 697                                  * the set bits to cache that value in ext->rs_left.
 698                                  * Whatever the reason (disconnect during resync,
 699                                  * delayed local completion of an application write),
 700                                  * try to fix it up by recounting here. */
 701                                 ext->rs_left = drbd_bm_e_weight(device, enr);
 702                         }
 703                 } else {
 704                         /* Normally this element should be in the cache,
 705                          * since drbd_rs_begin_io() pulled it already in.
 706                          *
 707                          * But maybe an application write finished, and we set
 708                          * something outside the resync lru_cache in sync.
 709                          */
 710                         int rs_left = drbd_bm_e_weight(device, enr);
 711                         if (ext->flags != 0) {
 712                                 drbd_warn(device, "changing resync lce: %d[%u;%02lx]"
 713                                      " -> %d[%u;00]\n",
 714                                      ext->lce.lc_number, ext->rs_left,
 715                                      ext->flags, enr, rs_left);
 716                                 ext->flags = 0;
 717                         }
 718                         if (ext->rs_failed) {
 719                                 drbd_warn(device, "Kicking resync_lru element enr=%u "
 720                                      "out with rs_failed=%d\n",
 721                                      ext->lce.lc_number, ext->rs_failed);
 722                         }
 723                         ext->rs_left = rs_left;
 724                         ext->rs_failed = (mode == RECORD_RS_FAILED) ? count : 0;
 725                         /* we don't keep a persistent log of the resync lru,
 726                          * we can commit any change right away. */
 727                         lc_committed(device->resync);
 728                 }
 729                 if (mode != SET_OUT_OF_SYNC)
 730                         lc_put(device->resync, &ext->lce);
 731                 /* no race, we are within the al_lock! */
 732
 733                 if (ext->rs_left <= ext->rs_failed) {
 734                         ext->rs_failed = 0;
 735                         return true;
 736                 }
 737         } else if (mode != SET_OUT_OF_SYNC) {
 738                 /* be quiet if lc_find() did not find it. */
 739                 drbd_err(device, "lc_get() failed! locked=%d/%d flags=%lu\n",
 740                     device->resync_locked,
 741                     device->resync->nr_elements,
 742                     device->resync->flags);
 743         }
 744         return false;
 745 }
 746
 747 void drbd_advance_rs_marks(struct drbd_device *device, unsigned long still_to_go)
 748 {
 749         unsigned long now = jiffies;
 750         unsigned long last = device->rs_mark_time[device->rs_last_mark];
 751         int next = (device->rs_last_mark + 1) % DRBD_SYNC_MARKS;
 752         if (time_after_eq(now, last + DRBD_SYNC_MARK_STEP)) {
 753                 if (device->rs_mark_left[device->rs_last_mark] != still_to_go &&
 754                     device->state.conn != C_PAUSED_SYNC_T &&
 755                     device->state.conn != C_PAUSED_SYNC_S) {
 756                         device->rs_mark_time[next] = now;
 757                         device->rs_mark_left[next] = still_to_go;
 758                         device->rs_last_mark = next;
 759                 }
 760         }
 761 }
 762
 763 /* It is called lazy update, so don't do write-out too often. */
 764 static bool lazy_bitmap_update_due(struct drbd_device *device)
 765 {
 766         return time_after(jiffies, device->rs_last_bcast + 2*HZ);
 767 }
 768
 769 static void maybe_schedule_on_disk_bitmap_update(struct drbd_device *device, bool rs_done)
 770 {
 771         if (rs_done)
 772                 set_bit(RS_DONE, &device->flags);
 773                 /* and also set RS_PROGRESS below */
 774         else if (!lazy_bitmap_update_due(device))
 775                 return;
 776
 777         drbd_device_post_work(device, RS_PROGRESS);
 778 }
 779
 780 static int update_sync_bits(struct drbd_device *device,
 781                 unsigned long sbnr, unsigned long ebnr,
 782                 enum update_sync_bits_mode mode)
 783 {
 784         /*
 785          * We keep a count of set bits per resync-extent in the ->rs_left
 786          * caching member, so we need to loop and work within the resync extent
 787          * alignment. Typically this loop will execute exactly once.
 788          */
 789         unsigned long flags;
 790         unsigned long count = 0;
 791         unsigned int cleared = 0;
 792         while (sbnr <= ebnr) {
 793                 /* set temporary boundary bit number to last bit number within
 794                  * the resync extent of the current start bit number,
 795                  * but cap at provided end bit number */
 796                 unsigned long tbnr = min(ebnr, sbnr | BM_BLOCKS_PER_BM_EXT_MASK);
 797                 unsigned long c;
 798
 799                 if (mode == RECORD_RS_FAILED)
 800                         /* Only called from drbd_rs_failed_io(), bits
 801                          * supposedly still set.  Recount, maybe some
 802                          * of the bits have been successfully cleared
 803                          * by application IO meanwhile.
 804                          */
 805                         c = drbd_bm_count_bits(device, sbnr, tbnr);
 806                 else if (mode == SET_IN_SYNC)
 807                         c = drbd_bm_clear_bits(device, sbnr, tbnr);
 808                 else /* if (mode == SET_OUT_OF_SYNC) */
 809                         c = drbd_bm_set_bits(device, sbnr, tbnr);
 810
 811                 if (c) {
 812                         spin_lock_irqsave(&device->al_lock, flags);
 813                         cleared += update_rs_extent(device, BM_BIT_TO_EXT(sbnr), c, mode);
 814                         spin_unlock_irqrestore(&device->al_lock, flags);
 815                         count += c;
 816                 }
 817                 sbnr = tbnr + 1;
 818         }
 819         if (count) {
 820                 if (mode == SET_IN_SYNC) {
 821                         unsigned long still_to_go = drbd_bm_total_weight(device);
 822                         bool rs_is_done = (still_to_go <= device->rs_failed);
 823                         drbd_advance_rs_marks(device, still_to_go);
 824                         if (cleared || rs_is_done)
 825                                 maybe_schedule_on_disk_bitmap_update(device, rs_is_done);
 826                 } else if (mode == RECORD_RS_FAILED)
 827                         device->rs_failed += count;
 828                 wake_up(&device->al_wait);
 829         }
 830         return count;
 831 }
 832
 833 /* clear the bit corresponding to the piece of storage in question:
 834  * size byte of data starting from sector.  Only clear a bits of the affected
 835  * one ore more _aligned_ BM_BLOCK_SIZE blocks.
 836  *
 837  * called by worker on C_SYNC_TARGET and receiver on SyncSource.
 838  *
 839  */
 840 int __drbd_change_sync(struct drbd_device *device, sector_t sector, int size,
 841                 enum update_sync_bits_mode mode)
 842 {
 843         /* Is called from worker and receiver context _only_ */
 844         unsigned long sbnr, ebnr, lbnr;
 845         unsigned long count = 0;
 846         sector_t esector, nr_sectors;
 847
 848         /* This would be an empty REQ_FLUSH, be silent. */
 849         if ((mode == SET_OUT_OF_SYNC) && size == 0)
 850                 return 0;
 851
 852         if (size <= 0 || !IS_ALIGNED(size, 512) || size > DRBD_MAX_DISCARD_SIZE) {
 853                 drbd_err(device, "%s: sector=%llus size=%d nonsense!\n",
 854                                 drbd_change_sync_fname[mode],
 855                                 (unsigned long long)sector, size);
 856                 return 0;
 857         }
 858
 859         if (!get_ldev(device))
 860                 return 0; /* no disk, no metadata, no bitmap to manipulate bits in */
 861
 862         nr_sectors = drbd_get_capacity(device->this_bdev);
 863         esector = sector + (size >> 9) - 1;
 864
 865         if (!expect(sector < nr_sectors))
 866                 goto out;
 867         if (!expect(esector < nr_sectors))
 868                 esector = nr_sectors - 1;
 869
 870         lbnr = BM_SECT_TO_BIT(nr_sectors-1);
 871
 872         if (mode == SET_IN_SYNC) {
 873                 /* Round up start sector, round down end sector.  We make sure
 874                  * we only clear full, aligned, BM_BLOCK_SIZE blocks. */
 875                 if (unlikely(esector < BM_SECT_PER_BIT-1))
 876                         goto out;
 877                 if (unlikely(esector == (nr_sectors-1)))
 878                         ebnr = lbnr;
 879                 else
 880                         ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1));
 881                 sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1);
 882         } else {
 883                 /* We set it out of sync, or record resync failure.
 884                  * Should not round anything here. */
 885                 sbnr = BM_SECT_TO_BIT(sector);
 886                 ebnr = BM_SECT_TO_BIT(esector);
 887         }
 888
 889         count = update_sync_bits(device, sbnr, ebnr, mode);
 890 out:
 891         put_ldev(device);
 892         return count;
 893 }
 894
 895 static
 896 struct bm_extent *_bme_get(struct drbd_device *device, unsigned int enr)
 897 {
 898         struct lc_element *e;
 899         struct bm_extent *bm_ext;
 900         int wakeup = 0;
 901         unsigned long rs_flags;
 902
 903         spin_lock_irq(&device->al_lock);
 904         if (device->resync_locked > device->resync->nr_elements/2) {
 905                 spin_unlock_irq(&device->al_lock);
 906                 return NULL;
 907         }
 908         e = lc_get(device->resync, enr);
 909         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
 910         if (bm_ext) {
 911                 if (bm_ext->lce.lc_number != enr) {
 912                         bm_ext->rs_left = drbd_bm_e_weight(device, enr);
 913                         bm_ext->rs_failed = 0;
 914                         lc_committed(device->resync);
 915                         wakeup = 1;
 916                 }
 917                 if (bm_ext->lce.refcnt == 1)
 918                         device->resync_locked++;
 919                 set_bit(BME_NO_WRITES, &bm_ext->flags);
 920         }
 921         rs_flags = device->resync->flags;
 922         spin_unlock_irq(&device->al_lock);
 923         if (wakeup)
 924                 wake_up(&device->al_wait);
 925
 926         if (!bm_ext) {
 927                 if (rs_flags & LC_STARVING)
 928                         drbd_warn(device, "Have to wait for element"
 929                              " (resync LRU too small?)\n");
 930                 BUG_ON(rs_flags & LC_LOCKED);
 931         }
 932
 933         return bm_ext;
 934 }
 935
 936 static int _is_in_al(struct drbd_device *device, unsigned int enr)
 937 {
 938         int rv;
 939
 940         spin_lock_irq(&device->al_lock);
 941         rv = lc_is_used(device->act_log, enr);
 942         spin_unlock_irq(&device->al_lock);
 943
 944         return rv;
 945 }
 946
 947 /**
 948  * drbd_rs_begin_io() - Gets an extent in the resync LRU cache and sets it to BME_LOCKED
 949  * @device:     DRBD device.
 950  * @sector:     The sector number.
 951  *
 952  * This functions sleeps on al_wait. Returns 0 on success, -EINTR if interrupted.
 953  */
 954 int drbd_rs_begin_io(struct drbd_device *device, sector_t sector)
 955 {
 956         unsigned int enr = BM_SECT_TO_EXT(sector);
 957         struct bm_extent *bm_ext;
 958         int i, sig;
 959         bool sa;
 960
 961 retry:
 962         sig = wait_event_interruptible(device->al_wait,
 963                         (bm_ext = _bme_get(device, enr)));
 964         if (sig)
 965                 return -EINTR;
 966
 967         if (test_bit(BME_LOCKED, &bm_ext->flags))
 968                 return 0;
 969
 970         /* step aside only while we are above c-min-rate; unless disabled. */
 971         sa = drbd_rs_c_min_rate_throttle(device);
 972
 973         for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
 974                 sig = wait_event_interruptible(device->al_wait,
 975                                                !_is_in_al(device, enr * AL_EXT_PER_BM_SECT + i) ||
 976                                                (sa && test_bit(BME_PRIORITY, &bm_ext->flags)));
 977
 978                 if (sig || (sa && test_bit(BME_PRIORITY, &bm_ext->flags))) {
 979                         spin_lock_irq(&device->al_lock);
 980                         if (lc_put(device->resync, &bm_ext->lce) == 0) {
 981                                 bm_ext->flags = 0; /* clears BME_NO_WRITES and eventually BME_PRIORITY */
 982                                 device->resync_locked--;
 983                                 wake_up(&device->al_wait);
 984                         }
 985                         spin_unlock_irq(&device->al_lock);
 986                         if (sig)
 987                                 return -EINTR;
 988                         if (schedule_timeout_interruptible(HZ/10))
 989                                 return -EINTR;
 990                         goto retry;
 991                 }
 992         }
 993         set_bit(BME_LOCKED, &bm_ext->flags);
 994         return 0;
 995 }
 996
 997 /**
 998  * drbd_try_rs_begin_io() - Gets an extent in the resync LRU cache, does not sleep
 999  * @device:     DRBD device.
1000  * @sector:     The sector number.
1001  *
1002  * Gets an extent in the resync LRU cache, sets it to BME_NO_WRITES, then
1003  * tries to set it to BME_LOCKED. Returns 0 upon success, and -EAGAIN
1004  * if there is still application IO going on in this area.
1005  */
1006 int drbd_try_rs_begin_io(struct drbd_device *device, sector_t sector)
1007 {
1008         unsigned int enr = BM_SECT_TO_EXT(sector);
1009         const unsigned int al_enr = enr*AL_EXT_PER_BM_SECT;
1010         struct lc_element *e;
1011         struct bm_extent *bm_ext;
1012         int i;
1013         bool throttle = drbd_rs_should_slow_down(device, sector, true);
1014
1015         /* If we need to throttle, a half-locked (only marked BME_NO_WRITES,
1016          * not yet BME_LOCKED) extent needs to be kicked out explicitly if we
1017          * need to throttle. There is at most one such half-locked extent,
1018          * which is remembered in resync_wenr. */
1019
1020         if (throttle && device->resync_wenr != enr)
1021                 return -EAGAIN;
1022
1023         spin_lock_irq(&device->al_lock);
1024         if (device->resync_wenr != LC_FREE && device->resync_wenr != enr) {
1025                 /* in case you have very heavy scattered io, it may
1026                  * stall the syncer undefined if we give up the ref count
1027                  * when we try again and requeue.
1028                  *
1029                  * if we don't give up the refcount, but the next time
1030                  * we are scheduled this extent has been "synced" by new
1031                  * application writes, we'd miss the lc_put on the
1032                  * extent we keep the refcount on.
1033                  * so we remembered which extent we had to try again, and
1034                  * if the next requested one is something else, we do
1035                  * the lc_put here...
1036                  * we also have to wake_up
1037                  */
1038                 e = lc_find(device->resync, device->resync_wenr);
1039                 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1040                 if (bm_ext) {
1041                         D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1042                         D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1043                         clear_bit(BME_NO_WRITES, &bm_ext->flags);
1044                         device->resync_wenr = LC_FREE;
1045                         if (lc_put(device->resync, &bm_ext->lce) == 0) {
1046                                 bm_ext->flags = 0;
1047                                 device->resync_locked--;
1048                         }
1049                         wake_up(&device->al_wait);
1050                 } else {
1051                         drbd_alert(device, "LOGIC BUG\n");
1052                 }
1053         }
1054         /* TRY. */
1055         e = lc_try_get(device->resync, enr);
1056         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1057         if (bm_ext) {
1058                 if (test_bit(BME_LOCKED, &bm_ext->flags))
1059                         goto proceed;
1060                 if (!test_and_set_bit(BME_NO_WRITES, &bm_ext->flags)) {
1061                         device->resync_locked++;
1062                 } else {
1063                         /* we did set the BME_NO_WRITES,
1064                          * but then could not set BME_LOCKED,
1065                          * so we tried again.
1066                          * drop the extra reference. */
1067                         bm_ext->lce.refcnt--;
1068                         D_ASSERT(device, bm_ext->lce.refcnt > 0);
1069                 }
1070                 goto check_al;
1071         } else {
1072                 /* do we rather want to try later? */
1073                 if (device->resync_locked > device->resync->nr_elements-3)
1074                         goto try_again;
1075                 /* Do or do not. There is no try. -- Yoda */
1076                 e = lc_get(device->resync, enr);
1077                 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1078                 if (!bm_ext) {
1079                         const unsigned long rs_flags = device->resync->flags;
1080                         if (rs_flags & LC_STARVING)
1081                                 drbd_warn(device, "Have to wait for element"
1082                                      " (resync LRU too small?)\n");
1083                         BUG_ON(rs_flags & LC_LOCKED);
1084                         goto try_again;
1085                 }
1086                 if (bm_ext->lce.lc_number != enr) {
1087                         bm_ext->rs_left = drbd_bm_e_weight(device, enr);
1088                         bm_ext->rs_failed = 0;
1089                         lc_committed(device->resync);
1090                         wake_up(&device->al_wait);
1091                         D_ASSERT(device, test_bit(BME_LOCKED, &bm_ext->flags) == 0);
1092                 }
1093                 set_bit(BME_NO_WRITES, &bm_ext->flags);
1094                 D_ASSERT(device, bm_ext->lce.refcnt == 1);
1095                 device->resync_locked++;
1096                 goto check_al;
1097         }
1098 check_al:
1099         for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
1100                 if (lc_is_used(device->act_log, al_enr+i))
1101                         goto try_again;
1102         }
1103         set_bit(BME_LOCKED, &bm_ext->flags);
1104 proceed:
1105         device->resync_wenr = LC_FREE;
1106         spin_unlock_irq(&device->al_lock);
1107         return 0;
1108
1109 try_again:
1110         if (bm_ext) {
1111                 if (throttle) {
1112                         D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1113                         D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1114                         clear_bit(BME_NO_WRITES, &bm_ext->flags);
1115                         device->resync_wenr = LC_FREE;
1116                         if (lc_put(device->resync, &bm_ext->lce) == 0) {
1117                                 bm_ext->flags = 0;
1118                                 device->resync_locked--;
1119                         }
1120                         wake_up(&device->al_wait);
1121                 } else
1122                         device->resync_wenr = enr;
1123         }
1124         spin_unlock_irq(&device->al_lock);
1125         return -EAGAIN;
1126 }
1127
1128 void drbd_rs_complete_io(struct drbd_device *device, sector_t sector)
1129 {
1130         unsigned int enr = BM_SECT_TO_EXT(sector);
1131         struct lc_element *e;
1132         struct bm_extent *bm_ext;
1133         unsigned long flags;
1134
1135         spin_lock_irqsave(&device->al_lock, flags);
1136         e = lc_find(device->resync, enr);
1137         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1138         if (!bm_ext) {
1139                 spin_unlock_irqrestore(&device->al_lock, flags);
1140                 if (__ratelimit(&drbd_ratelimit_state))
1141                         drbd_err(device, "drbd_rs_complete_io() called, but extent not found\n");
1142                 return;
1143         }
1144
1145         if (bm_ext->lce.refcnt == 0) {
1146                 spin_unlock_irqrestore(&device->al_lock, flags);
1147                 drbd_err(device, "drbd_rs_complete_io(,%llu [=%u]) called, "
1148                     "but refcnt is 0!?\n",
1149                     (unsigned long long)sector, enr);
1150                 return;
1151         }
1152
1153         if (lc_put(device->resync, &bm_ext->lce) == 0) {
1154                 bm_ext->flags = 0; /* clear BME_LOCKED, BME_NO_WRITES and BME_PRIORITY */
1155                 device->resync_locked--;
1156                 wake_up(&device->al_wait);
1157         }
1158
1159         spin_unlock_irqrestore(&device->al_lock, flags);
1160 }
1161
1162 /**
1163  * drbd_rs_cancel_all() - Removes all extents from the resync LRU (even BME_LOCKED)
1164  * @device:     DRBD device.
1165  */
1166 void drbd_rs_cancel_all(struct drbd_device *device)
1167 {
1168         spin_lock_irq(&device->al_lock);
1169
1170         if (get_ldev_if_state(device, D_FAILED)) { /* Makes sure ->resync is there. */
1171                 lc_reset(device->resync);
1172                 put_ldev(device);
1173         }
1174         device->resync_locked = 0;
1175         device->resync_wenr = LC_FREE;
1176         spin_unlock_irq(&device->al_lock);
1177         wake_up(&device->al_wait);
1178 }
1179
1180 /**
1181  * drbd_rs_del_all() - Gracefully remove all extents from the resync LRU
1182  * @device:     DRBD device.
1183  *
1184  * Returns 0 upon success, -EAGAIN if at least one reference count was
1185  * not zero.
1186  */
1187 int drbd_rs_del_all(struct drbd_device *device)
1188 {
1189         struct lc_element *e;
1190         struct bm_extent *bm_ext;
1191         int i;
1192
1193         spin_lock_irq(&device->al_lock);
1194
1195         if (get_ldev_if_state(device, D_FAILED)) {
1196                 /* ok, ->resync is there. */
1197                 for (i = 0; i < device->resync->nr_elements; i++) {
1198                         e = lc_element_by_index(device->resync, i);
1199                         bm_ext = lc_entry(e, struct bm_extent, lce);
1200                         if (bm_ext->lce.lc_number == LC_FREE)
1201                                 continue;
1202                         if (bm_ext->lce.lc_number == device->resync_wenr) {
1203                                 drbd_info(device, "dropping %u in drbd_rs_del_all, apparently"
1204                                      " got 'synced' by application io\n",
1205                                      device->resync_wenr);
1206                                 D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1207                                 D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
1208                                 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1209                                 device->resync_wenr = LC_FREE;
1210                                 lc_put(device->resync, &bm_ext->lce);
1211                         }
1212                         if (bm_ext->lce.refcnt != 0) {
1213                                 drbd_info(device, "Retrying drbd_rs_del_all() later. "
1214                                      "refcnt=%d\n", bm_ext->lce.refcnt);
1215                                 put_ldev(device);
1216                                 spin_unlock_irq(&device->al_lock);
1217                                 return -EAGAIN;
1218                         }
1219                         D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
1220                         D_ASSERT(device, !test_bit(BME_NO_WRITES, &bm_ext->flags));
1221                         lc_del(device->resync, &bm_ext->lce);
1222                 }
1223                 D_ASSERT(device, device->resync->used == 0);
1224                 put_ldev(device);
1225         }
1226         spin_unlock_irq(&device->al_lock);
1227         wake_up(&device->al_wait);
1228
1229         return 0;
1230 }