drivers/block/drbd/drbd_actlog.c

   1 /*
   2    drbd_actlog.c
   3
   4    This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
   5
   6    Copyright (C) 2003-2008, LINBIT Information Technologies GmbH.
   7    Copyright (C) 2003-2008, Philipp Reisner <philipp.reisner@linbit.com>.
   8    Copyright (C) 2003-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
   9
  10    drbd is free software; you can redistribute it and/or modify
  11    it under the terms of the GNU General Public License as published by
  12    the Free Software Foundation; either version 2, or (at your option)
  13    any later version.
  14
  15    drbd is distributed in the hope that it will be useful,
  16    but WITHOUT ANY WARRANTY; without even the implied warranty of
  17    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18    GNU General Public License for more details.
  19
  20    You should have received a copy of the GNU General Public License
  21    along with drbd; see the file COPYING.  If not, write to
  22    the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
  23
  24  */
  25
  26 #include <linux/slab.h>
  27 #include <linux/drbd.h>
  28 #include "drbd_int.h"
  29 #include "drbd_wrappers.h"
  30
  31 /* We maintain a trivial check sum in our on disk activity log.
  32  * With that we can ensure correct operation even when the storage
  33  * device might do a partial (last) sector write while loosing power.
  34  */
  35 struct __packed al_transaction {
  36         u32       magic;
  37         u32       tr_number;
  38         struct __packed {
  39                 u32 pos;
  40                 u32 extent; } updates[1 + AL_EXTENTS_PT];
  41         u32       xor_sum;
  42 };
  43
  44 struct update_odbm_work {
  45         struct drbd_work w;
  46         unsigned int enr;
  47 };
  48
  49 struct update_al_work {
  50         struct drbd_work w;
  51         struct lc_element *al_ext;
  52         struct completion event;
  53         unsigned int enr;
  54         /* if old_enr != LC_FREE, write corresponding bitmap sector, too */
  55         unsigned int old_enr;
  56 };
  57
  58 struct drbd_atodb_wait {
  59         atomic_t           count;
  60         struct completion  io_done;
  61         struct drbd_conf   *mdev;
  62         int                error;
  63 };
  64
  65
  66 int w_al_write_transaction(struct drbd_conf *, struct drbd_work *, int);
  67
  68 static int _drbd_md_sync_page_io(struct drbd_conf *mdev,
  69                                  struct drbd_backing_dev *bdev,
  70                                  struct page *page, sector_t sector,
  71                                  int rw, int size)
  72 {
  73         struct bio *bio;
  74         struct drbd_md_io md_io;
  75         int ok;
  76
  77         md_io.mdev = mdev;
  78         init_completion(&md_io.event);
  79         md_io.error = 0;
  80
  81         if ((rw & WRITE) && !test_bit(MD_NO_FUA, &mdev->flags))
  82                 rw |= REQ_FUA;
  83         rw |= REQ_UNPLUG | REQ_SYNC;
  84
  85         bio = bio_alloc(GFP_NOIO, 1);
  86         bio->bi_bdev = bdev->md_bdev;
  87         bio->bi_sector = sector;
  88         ok = (bio_add_page(bio, page, size, 0) == size);
  89         if (!ok)
  90                 goto out;
  91         bio->bi_private = &md_io;
  92         bio->bi_end_io = drbd_md_io_complete;
  93         bio->bi_rw = rw;
  94
  95         if (FAULT_ACTIVE(mdev, (rw & WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD))
  96                 bio_endio(bio, -EIO);
  97         else
  98                 submit_bio(rw, bio);
  99         wait_for_completion(&md_io.event);
 100         ok = bio_flagged(bio, BIO_UPTODATE) && md_io.error == 0;
 101
 102  out:
 103         bio_put(bio);
 104         return ok;
 105 }
 106
 107 int drbd_md_sync_page_io(struct drbd_conf *mdev, struct drbd_backing_dev *bdev,
 108                          sector_t sector, int rw)
 109 {
 110         int logical_block_size, mask, ok;
 111         int offset = 0;
 112         struct page *iop = mdev->md_io_page;
 113
 114         D_ASSERT(mutex_is_locked(&mdev->md_io_mutex));
 115
 116         BUG_ON(!bdev->md_bdev);
 117
 118         logical_block_size = bdev_logical_block_size(bdev->md_bdev);
 119         if (logical_block_size == 0)
 120                 logical_block_size = MD_SECTOR_SIZE;
 121
 122         /* in case logical_block_size != 512 [ s390 only? ] */
 123         if (logical_block_size != MD_SECTOR_SIZE) {
 124                 mask = (logical_block_size / MD_SECTOR_SIZE) - 1;
 125                 D_ASSERT(mask == 1 || mask == 3 || mask == 7);
 126                 D_ASSERT(logical_block_size == (mask+1) * MD_SECTOR_SIZE);
 127                 offset = sector & mask;
 128                 sector = sector & ~mask;
 129                 iop = mdev->md_io_tmpp;
 130
 131                 if (rw & WRITE) {
 132                         /* these are GFP_KERNEL pages, pre-allocated
 133                          * on device initialization */
 134                         void *p = page_address(mdev->md_io_page);
 135                         void *hp = page_address(mdev->md_io_tmpp);
 136
 137                         ok = _drbd_md_sync_page_io(mdev, bdev, iop, sector,
 138                                         READ, logical_block_size);
 139
 140                         if (unlikely(!ok)) {
 141                                 dev_err(DEV, "drbd_md_sync_page_io(,%llus,"
 142                                     "READ [logical_block_size!=512]) failed!\n",
 143                                     (unsigned long long)sector);
 144                                 return 0;
 145                         }
 146
 147                         memcpy(hp + offset*MD_SECTOR_SIZE, p, MD_SECTOR_SIZE);
 148                 }
 149         }
 150
 151         if (sector < drbd_md_first_sector(bdev) ||
 152             sector > drbd_md_last_sector(bdev))
 153                 dev_alert(DEV, "%s [%d]:%s(,%llus,%s) out of range md access!\n",
 154                      current->comm, current->pid, __func__,
 155                      (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ");
 156
 157         ok = _drbd_md_sync_page_io(mdev, bdev, iop, sector, rw, logical_block_size);
 158         if (unlikely(!ok)) {
 159                 dev_err(DEV, "drbd_md_sync_page_io(,%llus,%s) failed!\n",
 160                     (unsigned long long)sector, (rw & WRITE) ? "WRITE" : "READ");
 161                 return 0;
 162         }
 163
 164         if (logical_block_size != MD_SECTOR_SIZE && !(rw & WRITE)) {
 165                 void *p = page_address(mdev->md_io_page);
 166                 void *hp = page_address(mdev->md_io_tmpp);
 167
 168                 memcpy(p, hp + offset*MD_SECTOR_SIZE, MD_SECTOR_SIZE);
 169         }
 170
 171         return ok;
 172 }
 173
 174 static struct lc_element *_al_get(struct drbd_conf *mdev, unsigned int enr)
 175 {
 176         struct lc_element *al_ext;
 177         struct lc_element *tmp;
 178         unsigned long     al_flags = 0;
 179
 180         spin_lock_irq(&mdev->al_lock);
 181         tmp = lc_find(mdev->resync, enr/AL_EXT_PER_BM_SECT);
 182         if (unlikely(tmp != NULL)) {
 183                 struct bm_extent  *bm_ext = lc_entry(tmp, struct bm_extent, lce);
 184                 if (test_bit(BME_NO_WRITES, &bm_ext->flags)) {
 185                         spin_unlock_irq(&mdev->al_lock);
 186                         return NULL;
 187                 }
 188         }
 189         al_ext   = lc_get(mdev->act_log, enr);
 190         al_flags = mdev->act_log->flags;
 191         spin_unlock_irq(&mdev->al_lock);
 192
 193         /*
 194         if (!al_ext) {
 195                 if (al_flags & LC_STARVING)
 196                         dev_warn(DEV, "Have to wait for LRU element (AL too small?)\n");
 197                 if (al_flags & LC_DIRTY)
 198                         dev_warn(DEV, "Ongoing AL update (AL device too slow?)\n");
 199         }
 200         */
 201
 202         return al_ext;
 203 }
 204
 205 void drbd_al_begin_io(struct drbd_conf *mdev, sector_t sector)
 206 {
 207         unsigned int enr = (sector >> (AL_EXTENT_SHIFT-9));
 208         struct lc_element *al_ext;
 209         struct update_al_work al_work;
 210
 211         D_ASSERT(atomic_read(&mdev->local_cnt) > 0);
 212
 213         wait_event(mdev->al_wait, (al_ext = _al_get(mdev, enr)));
 214
 215         if (al_ext->lc_number != enr) {
 216                 /* drbd_al_write_transaction(mdev,al_ext,enr);
 217                  * recurses into generic_make_request(), which
 218                  * disallows recursion, bios being serialized on the
 219                  * current->bio_tail list now.
 220                  * we have to delegate updates to the activity log
 221                  * to the worker thread. */
 222                 init_completion(&al_work.event);
 223                 al_work.al_ext = al_ext;
 224                 al_work.enr = enr;
 225                 al_work.old_enr = al_ext->lc_number;
 226                 al_work.w.cb = w_al_write_transaction;
 227                 drbd_queue_work_front(&mdev->data.work, &al_work.w);
 228                 wait_for_completion(&al_work.event);
 229
 230                 mdev->al_writ_cnt++;
 231
 232                 spin_lock_irq(&mdev->al_lock);
 233                 lc_changed(mdev->act_log, al_ext);
 234                 spin_unlock_irq(&mdev->al_lock);
 235                 wake_up(&mdev->al_wait);
 236         }
 237 }
 238
 239 void drbd_al_complete_io(struct drbd_conf *mdev, sector_t sector)
 240 {
 241         unsigned int enr = (sector >> (AL_EXTENT_SHIFT-9));
 242         struct lc_element *extent;
 243         unsigned long flags;
 244
 245         spin_lock_irqsave(&mdev->al_lock, flags);
 246
 247         extent = lc_find(mdev->act_log, enr);
 248
 249         if (!extent) {
 250                 spin_unlock_irqrestore(&mdev->al_lock, flags);
 251                 dev_err(DEV, "al_complete_io() called on inactive extent %u\n", enr);
 252                 return;
 253         }
 254
 255         if (lc_put(mdev->act_log, extent) == 0)
 256                 wake_up(&mdev->al_wait);
 257
 258         spin_unlock_irqrestore(&mdev->al_lock, flags);
 259 }
 260
 261 int
 262 w_al_write_transaction(struct drbd_conf *mdev, struct drbd_work *w, int unused)
 263 {
 264         struct update_al_work *aw = container_of(w, struct update_al_work, w);
 265         struct lc_element *updated = aw->al_ext;
 266         const unsigned int new_enr = aw->enr;
 267         const unsigned int evicted = aw->old_enr;
 268         struct al_transaction *buffer;
 269         sector_t sector;
 270         int i, n, mx;
 271         unsigned int extent_nr;
 272         u32 xor_sum = 0;
 273
 274         if (!get_ldev(mdev)) {
 275                 dev_err(DEV,
 276                         "disk is %s, cannot start al transaction (-%d +%d)\n",
 277                         drbd_disk_str(mdev->state.disk), evicted, new_enr);
 278                 complete(&((struct update_al_work *)w)->event);
 279                 return 1;
 280         }
 281         /* do we have to do a bitmap write, first?
 282          * TODO reduce maximum latency:
 283          * submit both bios, then wait for both,
 284          * instead of doing two synchronous sector writes.
 285          * For now, we must not write the transaction,
 286          * if we cannot write out the bitmap of the evicted extent. */
 287         if (mdev->state.conn < C_CONNECTED && evicted != LC_FREE)
 288                 drbd_bm_write_sect(mdev, evicted/AL_EXT_PER_BM_SECT);
 289
 290         /* The bitmap write may have failed, causing a state change. */
 291         if (mdev->state.disk < D_INCONSISTENT) {
 292                 dev_err(DEV,
 293                         "disk is %s, cannot write al transaction (-%d +%d)\n",
 294                         drbd_disk_str(mdev->state.disk), evicted, new_enr);
 295                 complete(&((struct update_al_work *)w)->event);
 296                 put_ldev(mdev);
 297                 return 1;
 298         }
 299
 300         mutex_lock(&mdev->md_io_mutex); /* protects md_io_buffer, al_tr_cycle, ... */
 301         buffer = (struct al_transaction *)page_address(mdev->md_io_page);
 302
 303         buffer->magic = __constant_cpu_to_be32(DRBD_MAGIC);
 304         buffer->tr_number = cpu_to_be32(mdev->al_tr_number);
 305
 306         n = lc_index_of(mdev->act_log, updated);
 307
 308         buffer->updates[0].pos = cpu_to_be32(n);
 309         buffer->updates[0].extent = cpu_to_be32(new_enr);
 310
 311         xor_sum ^= new_enr;
 312
 313         mx = min_t(int, AL_EXTENTS_PT,
 314                    mdev->act_log->nr_elements - mdev->al_tr_cycle);
 315         for (i = 0; i < mx; i++) {
 316                 unsigned idx = mdev->al_tr_cycle + i;
 317                 extent_nr = lc_element_by_index(mdev->act_log, idx)->lc_number;
 318                 buffer->updates[i+1].pos = cpu_to_be32(idx);
 319                 buffer->updates[i+1].extent = cpu_to_be32(extent_nr);
 320                 xor_sum ^= extent_nr;
 321         }
 322         for (; i < AL_EXTENTS_PT; i++) {
 323                 buffer->updates[i+1].pos = __constant_cpu_to_be32(-1);
 324                 buffer->updates[i+1].extent = __constant_cpu_to_be32(LC_FREE);
 325                 xor_sum ^= LC_FREE;
 326         }
 327         mdev->al_tr_cycle += AL_EXTENTS_PT;
 328         if (mdev->al_tr_cycle >= mdev->act_log->nr_elements)
 329                 mdev->al_tr_cycle = 0;
 330
 331         buffer->xor_sum = cpu_to_be32(xor_sum);
 332
 333         sector =  mdev->ldev->md.md_offset
 334                 + mdev->ldev->md.al_offset + mdev->al_tr_pos;
 335
 336         if (!drbd_md_sync_page_io(mdev, mdev->ldev, sector, WRITE))
 337                 drbd_chk_io_error(mdev, 1, TRUE);
 338
 339         if (++mdev->al_tr_pos >
 340             div_ceil(mdev->act_log->nr_elements, AL_EXTENTS_PT))
 341                 mdev->al_tr_pos = 0;
 342
 343         D_ASSERT(mdev->al_tr_pos < MD_AL_MAX_SIZE);
 344         mdev->al_tr_number++;
 345
 346         mutex_unlock(&mdev->md_io_mutex);
 347
 348         complete(&((struct update_al_work *)w)->event);
 349         put_ldev(mdev);
 350
 351         return 1;
 352 }
 353
 354 /**
 355  * drbd_al_read_tr() - Read a single transaction from the on disk activity log
 356  * @mdev:       DRBD device.
 357  * @bdev:       Block device to read form.
 358  * @b:          pointer to an al_transaction.
 359  * @index:      On disk slot of the transaction to read.
 360  *
 361  * Returns -1 on IO error, 0 on checksum error and 1 upon success.
 362  */
 363 static int drbd_al_read_tr(struct drbd_conf *mdev,
 364                            struct drbd_backing_dev *bdev,
 365                            struct al_transaction *b,
 366                            int index)
 367 {
 368         sector_t sector;
 369         int rv, i;
 370         u32 xor_sum = 0;
 371
 372         sector = bdev->md.md_offset + bdev->md.al_offset + index;
 373
 374         /* Dont process error normally,
 375          * as this is done before disk is attached! */
 376         if (!drbd_md_sync_page_io(mdev, bdev, sector, READ))
 377                 return -1;
 378
 379         rv = (be32_to_cpu(b->magic) == DRBD_MAGIC);
 380
 381         for (i = 0; i < AL_EXTENTS_PT + 1; i++)
 382                 xor_sum ^= be32_to_cpu(b->updates[i].extent);
 383         rv &= (xor_sum == be32_to_cpu(b->xor_sum));
 384
 385         return rv;
 386 }
 387
 388 /**
 389  * drbd_al_read_log() - Restores the activity log from its on disk representation.
 390  * @mdev:       DRBD device.
 391  * @bdev:       Block device to read form.
 392  *
 393  * Returns 1 on success, returns 0 when reading the log failed due to IO errors.
 394  */
 395 int drbd_al_read_log(struct drbd_conf *mdev, struct drbd_backing_dev *bdev)
 396 {
 397         struct al_transaction *buffer;
 398         int i;
 399         int rv;
 400         int mx;
 401         int active_extents = 0;
 402         int transactions = 0;
 403         int found_valid = 0;
 404         int from = 0;
 405         int to = 0;
 406         u32 from_tnr = 0;
 407         u32 to_tnr = 0;
 408         u32 cnr;
 409
 410         mx = div_ceil(mdev->act_log->nr_elements, AL_EXTENTS_PT);
 411
 412         /* lock out all other meta data io for now,
 413          * and make sure the page is mapped.
 414          */
 415         mutex_lock(&mdev->md_io_mutex);
 416         buffer = page_address(mdev->md_io_page);
 417
 418         /* Find the valid transaction in the log */
 419         for (i = 0; i <= mx; i++) {
 420                 rv = drbd_al_read_tr(mdev, bdev, buffer, i);
 421                 if (rv == 0)
 422                         continue;
 423                 if (rv == -1) {
 424                         mutex_unlock(&mdev->md_io_mutex);
 425                         return 0;
 426                 }
 427                 cnr = be32_to_cpu(buffer->tr_number);
 428
 429                 if (++found_valid == 1) {
 430                         from = i;
 431                         to = i;
 432                         from_tnr = cnr;
 433                         to_tnr = cnr;
 434                         continue;
 435                 }
 436                 if ((int)cnr - (int)from_tnr < 0) {
 437                         D_ASSERT(from_tnr - cnr + i - from == mx+1);
 438                         from = i;
 439                         from_tnr = cnr;
 440                 }
 441                 if ((int)cnr - (int)to_tnr > 0) {
 442                         D_ASSERT(cnr - to_tnr == i - to);
 443                         to = i;
 444                         to_tnr = cnr;
 445                 }
 446         }
 447
 448         if (!found_valid) {
 449                 dev_warn(DEV, "No usable activity log found.\n");
 450                 mutex_unlock(&mdev->md_io_mutex);
 451                 return 1;
 452         }
 453
 454         /* Read the valid transactions.
 455          * dev_info(DEV, "Reading from %d to %d.\n",from,to); */
 456         i = from;
 457         while (1) {
 458                 int j, pos;
 459                 unsigned int extent_nr;
 460                 unsigned int trn;
 461
 462                 rv = drbd_al_read_tr(mdev, bdev, buffer, i);
 463                 ERR_IF(rv == 0) goto cancel;
 464                 if (rv == -1) {
 465                         mutex_unlock(&mdev->md_io_mutex);
 466                         return 0;
 467                 }
 468
 469                 trn = be32_to_cpu(buffer->tr_number);
 470
 471                 spin_lock_irq(&mdev->al_lock);
 472
 473                 /* This loop runs backwards because in the cyclic
 474                    elements there might be an old version of the
 475                    updated element (in slot 0). So the element in slot 0
 476                    can overwrite old versions. */
 477                 for (j = AL_EXTENTS_PT; j >= 0; j--) {
 478                         pos = be32_to_cpu(buffer->updates[j].pos);
 479                         extent_nr = be32_to_cpu(buffer->updates[j].extent);
 480
 481                         if (extent_nr == LC_FREE)
 482                                 continue;
 483
 484                         lc_set(mdev->act_log, extent_nr, pos);
 485                         active_extents++;
 486                 }
 487                 spin_unlock_irq(&mdev->al_lock);
 488
 489                 transactions++;
 490
 491 cancel:
 492                 if (i == to)
 493                         break;
 494                 i++;
 495                 if (i > mx)
 496                         i = 0;
 497         }
 498
 499         mdev->al_tr_number = to_tnr+1;
 500         mdev->al_tr_pos = to;
 501         if (++mdev->al_tr_pos >
 502             div_ceil(mdev->act_log->nr_elements, AL_EXTENTS_PT))
 503                 mdev->al_tr_pos = 0;
 504
 505         /* ok, we are done with it */
 506         mutex_unlock(&mdev->md_io_mutex);
 507
 508         dev_info(DEV, "Found %d transactions (%d active extents) in activity log.\n",
 509              transactions, active_extents);
 510
 511         return 1;
 512 }
 513
 514 static void atodb_endio(struct bio *bio, int error)
 515 {
 516         struct drbd_atodb_wait *wc = bio->bi_private;
 517         struct drbd_conf *mdev = wc->mdev;
 518         struct page *page;
 519         int uptodate = bio_flagged(bio, BIO_UPTODATE);
 520
 521         /* strange behavior of some lower level drivers...
 522          * fail the request by clearing the uptodate flag,
 523          * but do not return any error?! */
 524         if (!error && !uptodate)
 525                 error = -EIO;
 526
 527         drbd_chk_io_error(mdev, error, TRUE);
 528         if (error && wc->error == 0)
 529                 wc->error = error;
 530
 531         if (atomic_dec_and_test(&wc->count))
 532                 complete(&wc->io_done);
 533
 534         page = bio->bi_io_vec[0].bv_page;
 535         put_page(page);
 536         bio_put(bio);
 537         mdev->bm_writ_cnt++;
 538         put_ldev(mdev);
 539 }
 540
 541 /* sector to word */
 542 #define S2W(s)  ((s)<<(BM_EXT_SHIFT-BM_BLOCK_SHIFT-LN2_BPL))
 543
 544 /* activity log to on disk bitmap -- prepare bio unless that sector
 545  * is already covered by previously prepared bios */
 546 static int atodb_prepare_unless_covered(struct drbd_conf *mdev,
 547                                         struct bio **bios,
 548                                         unsigned int enr,
 549                                         struct drbd_atodb_wait *wc) __must_hold(local)
 550 {
 551         struct bio *bio;
 552         struct page *page;
 553         sector_t on_disk_sector;
 554         unsigned int page_offset = PAGE_SIZE;
 555         int offset;
 556         int i = 0;
 557         int err = -ENOMEM;
 558
 559         /* We always write aligned, full 4k blocks,
 560          * so we can ignore the logical_block_size (for now) */
 561         enr &= ~7U;
 562         on_disk_sector = enr + mdev->ldev->md.md_offset
 563                              + mdev->ldev->md.bm_offset;
 564
 565         D_ASSERT(!(on_disk_sector & 7U));
 566
 567         /* Check if that enr is already covered by an already created bio.
 568          * Caution, bios[] is not NULL terminated,
 569          * but only initialized to all NULL.
 570          * For completely scattered activity log,
 571          * the last invocation iterates over all bios,
 572          * and finds the last NULL entry.
 573          */
 574         while ((bio = bios[i])) {
 575                 if (bio->bi_sector == on_disk_sector)
 576                         return 0;
 577                 i++;
 578         }
 579         /* bios[i] == NULL, the next not yet used slot */
 580
 581         /* GFP_KERNEL, we are not in the write-out path */
 582         bio = bio_alloc(GFP_KERNEL, 1);
 583         if (bio == NULL)
 584                 return -ENOMEM;
 585
 586         if (i > 0) {
 587                 const struct bio_vec *prev_bv = bios[i-1]->bi_io_vec;
 588                 page_offset = prev_bv->bv_offset + prev_bv->bv_len;
 589                 page = prev_bv->bv_page;
 590         }
 591         if (page_offset == PAGE_SIZE) {
 592                 page = alloc_page(__GFP_HIGHMEM);
 593                 if (page == NULL)
 594                         goto out_bio_put;
 595                 page_offset = 0;
 596         } else {
 597                 get_page(page);
 598         }
 599
 600         offset = S2W(enr);
 601         drbd_bm_get_lel(mdev, offset,
 602                         min_t(size_t, S2W(8), drbd_bm_words(mdev) - offset),
 603                         kmap(page) + page_offset);
 604         kunmap(page);
 605
 606         bio->bi_private = wc;
 607         bio->bi_end_io = atodb_endio;
 608         bio->bi_bdev = mdev->ldev->md_bdev;
 609         bio->bi_sector = on_disk_sector;
 610
 611         if (bio_add_page(bio, page, 4096, page_offset) != 4096)
 612                 goto out_put_page;
 613
 614         atomic_inc(&wc->count);
 615         /* we already know that we may do this...
 616          * get_ldev_if_state(mdev,D_ATTACHING);
 617          * just get the extra reference, so that the local_cnt reflects
 618          * the number of pending IO requests DRBD at its backing device.
 619          */
 620         atomic_inc(&mdev->local_cnt);
 621
 622         bios[i] = bio;
 623
 624         return 0;
 625
 626 out_put_page:
 627         err = -EINVAL;
 628         put_page(page);
 629 out_bio_put:
 630         bio_put(bio);
 631         return err;
 632 }
 633
 634 /**
 635  * drbd_al_to_on_disk_bm() -  * Writes bitmap parts covered by active AL extents
 636  * @mdev:       DRBD device.
 637  *
 638  * Called when we detach (unconfigure) local storage,
 639  * or when we go from R_PRIMARY to R_SECONDARY role.
 640  */
 641 void drbd_al_to_on_disk_bm(struct drbd_conf *mdev)
 642 {
 643         int i, nr_elements;
 644         unsigned int enr;
 645         struct bio **bios;
 646         struct drbd_atodb_wait wc;
 647
 648         ERR_IF (!get_ldev_if_state(mdev, D_ATTACHING))
 649                 return; /* sorry, I don't have any act_log etc... */
 650
 651         wait_event(mdev->al_wait, lc_try_lock(mdev->act_log));
 652
 653         nr_elements = mdev->act_log->nr_elements;
 654
 655         /* GFP_KERNEL, we are not in anyone's write-out path */
 656         bios = kzalloc(sizeof(struct bio *) * nr_elements, GFP_KERNEL);
 657         if (!bios)
 658                 goto submit_one_by_one;
 659
 660         atomic_set(&wc.count, 0);
 661         init_completion(&wc.io_done);
 662         wc.mdev = mdev;
 663         wc.error = 0;
 664
 665         for (i = 0; i < nr_elements; i++) {
 666                 enr = lc_element_by_index(mdev->act_log, i)->lc_number;
 667                 if (enr == LC_FREE)
 668                         continue;
 669                 /* next statement also does atomic_inc wc.count and local_cnt */
 670                 if (atodb_prepare_unless_covered(mdev, bios,
 671                                                 enr/AL_EXT_PER_BM_SECT,
 672                                                 &wc))
 673                         goto free_bios_submit_one_by_one;
 674         }
 675
 676         /* unnecessary optimization? */
 677         lc_unlock(mdev->act_log);
 678         wake_up(&mdev->al_wait);
 679
 680         /* all prepared, submit them */
 681         for (i = 0; i < nr_elements; i++) {
 682                 if (bios[i] == NULL)
 683                         break;
 684                 if (FAULT_ACTIVE(mdev, DRBD_FAULT_MD_WR)) {
 685                         bios[i]->bi_rw = WRITE;
 686                         bio_endio(bios[i], -EIO);
 687                 } else {
 688                         submit_bio(WRITE, bios[i]);
 689                 }
 690         }
 691
 692         drbd_blk_run_queue(bdev_get_queue(mdev->ldev->md_bdev));
 693
 694         /* always (try to) flush bitmap to stable storage */
 695         drbd_md_flush(mdev);
 696
 697         /* In case we did not submit a single IO do not wait for
 698          * them to complete. ( Because we would wait forever here. )
 699          *
 700          * In case we had IOs and they are already complete, there
 701          * is not point in waiting anyways.
 702          * Therefore this if () ... */
 703         if (atomic_read(&wc.count))
 704                 wait_for_completion(&wc.io_done);
 705
 706         put_ldev(mdev);
 707
 708         kfree(bios);
 709         return;
 710
 711  free_bios_submit_one_by_one:
 712         /* free everything by calling the endio callback directly. */
 713         for (i = 0; i < nr_elements && bios[i]; i++)
 714                 bio_endio(bios[i], 0);
 715
 716         kfree(bios);
 717
 718  submit_one_by_one:
 719         dev_warn(DEV, "Using the slow drbd_al_to_on_disk_bm()\n");
 720
 721         for (i = 0; i < mdev->act_log->nr_elements; i++) {
 722                 enr = lc_element_by_index(mdev->act_log, i)->lc_number;
 723                 if (enr == LC_FREE)
 724                         continue;
 725                 /* Really slow: if we have al-extents 16..19 active,
 726                  * sector 4 will be written four times! Synchronous! */
 727                 drbd_bm_write_sect(mdev, enr/AL_EXT_PER_BM_SECT);
 728         }
 729
 730         lc_unlock(mdev->act_log);
 731         wake_up(&mdev->al_wait);
 732         put_ldev(mdev);
 733 }
 734
 735 /**
 736  * drbd_al_apply_to_bm() - Sets the bitmap to diry(1) where covered ba active AL extents
 737  * @mdev:       DRBD device.
 738  */
 739 void drbd_al_apply_to_bm(struct drbd_conf *mdev)
 740 {
 741         unsigned int enr;
 742         unsigned long add = 0;
 743         char ppb[10];
 744         int i, tmp;
 745
 746         wait_event(mdev->al_wait, lc_try_lock(mdev->act_log));
 747
 748         for (i = 0; i < mdev->act_log->nr_elements; i++) {
 749                 enr = lc_element_by_index(mdev->act_log, i)->lc_number;
 750                 if (enr == LC_FREE)
 751                         continue;
 752                 tmp = drbd_bm_ALe_set_all(mdev, enr);
 753                 dynamic_dev_dbg(DEV, "AL: set %d bits in extent %u\n", tmp, enr);
 754                 add += tmp;
 755         }
 756
 757         lc_unlock(mdev->act_log);
 758         wake_up(&mdev->al_wait);
 759
 760         dev_info(DEV, "Marked additional %s as out-of-sync based on AL.\n",
 761              ppsize(ppb, Bit2KB(add)));
 762 }
 763
 764 static int _try_lc_del(struct drbd_conf *mdev, struct lc_element *al_ext)
 765 {
 766         int rv;
 767
 768         spin_lock_irq(&mdev->al_lock);
 769         rv = (al_ext->refcnt == 0);
 770         if (likely(rv))
 771                 lc_del(mdev->act_log, al_ext);
 772         spin_unlock_irq(&mdev->al_lock);
 773
 774         return rv;
 775 }
 776
 777 /**
 778  * drbd_al_shrink() - Removes all active extents form the activity log
 779  * @mdev:       DRBD device.
 780  *
 781  * Removes all active extents form the activity log, waiting until
 782  * the reference count of each entry dropped to 0 first, of course.
 783  *
 784  * You need to lock mdev->act_log with lc_try_lock() / lc_unlock()
 785  */
 786 void drbd_al_shrink(struct drbd_conf *mdev)
 787 {
 788         struct lc_element *al_ext;
 789         int i;
 790
 791         D_ASSERT(test_bit(__LC_DIRTY, &mdev->act_log->flags));
 792
 793         for (i = 0; i < mdev->act_log->nr_elements; i++) {
 794                 al_ext = lc_element_by_index(mdev->act_log, i);
 795                 if (al_ext->lc_number == LC_FREE)
 796                         continue;
 797                 wait_event(mdev->al_wait, _try_lc_del(mdev, al_ext));
 798         }
 799
 800         wake_up(&mdev->al_wait);
 801 }
 802
 803 static int w_update_odbm(struct drbd_conf *mdev, struct drbd_work *w, int unused)
 804 {
 805         struct update_odbm_work *udw = container_of(w, struct update_odbm_work, w);
 806
 807         if (!get_ldev(mdev)) {
 808                 if (__ratelimit(&drbd_ratelimit_state))
 809                         dev_warn(DEV, "Can not update on disk bitmap, local IO disabled.\n");
 810                 kfree(udw);
 811                 return 1;
 812         }
 813
 814         drbd_bm_write_sect(mdev, udw->enr);
 815         put_ldev(mdev);
 816
 817         kfree(udw);
 818
 819         if (drbd_bm_total_weight(mdev) <= mdev->rs_failed) {
 820                 switch (mdev->state.conn) {
 821                 case C_SYNC_SOURCE:  case C_SYNC_TARGET:
 822                 case C_PAUSED_SYNC_S: case C_PAUSED_SYNC_T:
 823                         drbd_resync_finished(mdev);
 824                 default:
 825                         /* nothing to do */
 826                         break;
 827                 }
 828         }
 829         drbd_bcast_sync_progress(mdev);
 830
 831         return 1;
 832 }
 833
 834
 835 /* ATTENTION. The AL's extents are 4MB each, while the extents in the
 836  * resync LRU-cache are 16MB each.
 837  * The caller of this function has to hold an get_ldev() reference.
 838  *
 839  * TODO will be obsoleted once we have a caching lru of the on disk bitmap
 840  */
 841 static void drbd_try_clear_on_disk_bm(struct drbd_conf *mdev, sector_t sector,
 842                                       int count, int success)
 843 {
 844         struct lc_element *e;
 845         struct update_odbm_work *udw;
 846
 847         unsigned int enr;
 848
 849         D_ASSERT(atomic_read(&mdev->local_cnt));
 850
 851         /* I simply assume that a sector/size pair never crosses
 852          * a 16 MB extent border. (Currently this is true...) */
 853         enr = BM_SECT_TO_EXT(sector);
 854
 855         e = lc_get(mdev->resync, enr);
 856         if (e) {
 857                 struct bm_extent *ext = lc_entry(e, struct bm_extent, lce);
 858                 if (ext->lce.lc_number == enr) {
 859                         if (success)
 860                                 ext->rs_left -= count;
 861                         else
 862                                 ext->rs_failed += count;
 863                         if (ext->rs_left < ext->rs_failed) {
 864                                 dev_err(DEV, "BAD! sector=%llus enr=%u rs_left=%d "
 865                                     "rs_failed=%d count=%d\n",
 866                                      (unsigned long long)sector,
 867                                      ext->lce.lc_number, ext->rs_left,
 868                                      ext->rs_failed, count);
 869                                 dump_stack();
 870
 871                                 lc_put(mdev->resync, &ext->lce);
 872                                 drbd_force_state(mdev, NS(conn, C_DISCONNECTING));
 873                                 return;
 874                         }
 875                 } else {
 876                         /* Normally this element should be in the cache,
 877                          * since drbd_rs_begin_io() pulled it already in.
 878                          *
 879                          * But maybe an application write finished, and we set
 880                          * something outside the resync lru_cache in sync.
 881                          */
 882                         int rs_left = drbd_bm_e_weight(mdev, enr);
 883                         if (ext->flags != 0) {
 884                                 dev_warn(DEV, "changing resync lce: %d[%u;%02lx]"
 885                                      " -> %d[%u;00]\n",
 886                                      ext->lce.lc_number, ext->rs_left,
 887                                      ext->flags, enr, rs_left);
 888                                 ext->flags = 0;
 889                         }
 890                         if (ext->rs_failed) {
 891                                 dev_warn(DEV, "Kicking resync_lru element enr=%u "
 892                                      "out with rs_failed=%d\n",
 893                                      ext->lce.lc_number, ext->rs_failed);
 894                                 set_bit(WRITE_BM_AFTER_RESYNC, &mdev->flags);
 895                         }
 896                         ext->rs_left = rs_left;
 897                         ext->rs_failed = success ? 0 : count;
 898                         lc_changed(mdev->resync, &ext->lce);
 899                 }
 900                 lc_put(mdev->resync, &ext->lce);
 901                 /* no race, we are within the al_lock! */
 902
 903                 if (ext->rs_left == ext->rs_failed) {
 904                         ext->rs_failed = 0;
 905
 906                         udw = kmalloc(sizeof(*udw), GFP_ATOMIC);
 907                         if (udw) {
 908                                 udw->enr = ext->lce.lc_number;
 909                                 udw->w.cb = w_update_odbm;
 910                                 drbd_queue_work_front(&mdev->data.work, &udw->w);
 911                         } else {
 912                                 dev_warn(DEV, "Could not kmalloc an udw\n");
 913                                 set_bit(WRITE_BM_AFTER_RESYNC, &mdev->flags);
 914                         }
 915                 }
 916         } else {
 917                 dev_err(DEV, "lc_get() failed! locked=%d/%d flags=%lu\n",
 918                     mdev->resync_locked,
 919                     mdev->resync->nr_elements,
 920                     mdev->resync->flags);
 921         }
 922 }
 923
 924 /* clear the bit corresponding to the piece of storage in question:
 925  * size byte of data starting from sector.  Only clear a bits of the affected
 926  * one ore more _aligned_ BM_BLOCK_SIZE blocks.
 927  *
 928  * called by worker on C_SYNC_TARGET and receiver on SyncSource.
 929  *
 930  */
 931 void __drbd_set_in_sync(struct drbd_conf *mdev, sector_t sector, int size,
 932                        const char *file, const unsigned int line)
 933 {
 934         /* Is called from worker and receiver context _only_ */
 935         unsigned long sbnr, ebnr, lbnr;
 936         unsigned long count = 0;
 937         sector_t esector, nr_sectors;
 938         int wake_up = 0;
 939         unsigned long flags;
 940
 941         if (size <= 0 || (size & 0x1ff) != 0 || size > DRBD_MAX_SEGMENT_SIZE) {
 942                 dev_err(DEV, "drbd_set_in_sync: sector=%llus size=%d nonsense!\n",
 943                                 (unsigned long long)sector, size);
 944                 return;
 945         }
 946         nr_sectors = drbd_get_capacity(mdev->this_bdev);
 947         esector = sector + (size >> 9) - 1;
 948
 949         ERR_IF(sector >= nr_sectors) return;
 950         ERR_IF(esector >= nr_sectors) esector = (nr_sectors-1);
 951
 952         lbnr = BM_SECT_TO_BIT(nr_sectors-1);
 953
 954         /* we clear it (in sync).
 955          * round up start sector, round down end sector.  we make sure we only
 956          * clear full, aligned, BM_BLOCK_SIZE (4K) blocks */
 957         if (unlikely(esector < BM_SECT_PER_BIT-1))
 958                 return;
 959         if (unlikely(esector == (nr_sectors-1)))
 960                 ebnr = lbnr;
 961         else
 962                 ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1));
 963         sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1);
 964
 965         if (sbnr > ebnr)
 966                 return;
 967
 968         /*
 969          * ok, (capacity & 7) != 0 sometimes, but who cares...
 970          * we count rs_{total,left} in bits, not sectors.
 971          */
 972         count = drbd_bm_clear_bits(mdev, sbnr, ebnr);
 973         if (count && get_ldev(mdev)) {
 974                 unsigned long now = jiffies;
 975                 unsigned long last = mdev->rs_mark_time[mdev->rs_last_mark];
 976                 int next = (mdev->rs_last_mark + 1) % DRBD_SYNC_MARKS;
 977                 if (time_after_eq(now, last + DRBD_SYNC_MARK_STEP)) {
 978                         unsigned long tw = drbd_bm_total_weight(mdev);
 979                         if (mdev->rs_mark_left[mdev->rs_last_mark] != tw &&
 980                             mdev->state.conn != C_PAUSED_SYNC_T &&
 981                             mdev->state.conn != C_PAUSED_SYNC_S) {
 982                                 mdev->rs_mark_time[next] = now;
 983                                 mdev->rs_mark_left[next] = tw;
 984                                 mdev->rs_last_mark = next;
 985                         }
 986                 }
 987                 spin_lock_irqsave(&mdev->al_lock, flags);
 988                 drbd_try_clear_on_disk_bm(mdev, sector, count, TRUE);
 989                 spin_unlock_irqrestore(&mdev->al_lock, flags);
 990
 991                 /* just wake_up unconditional now, various lc_chaged(),
 992                  * lc_put() in drbd_try_clear_on_disk_bm(). */
 993                 wake_up = 1;
 994                 put_ldev(mdev);
 995         }
 996         if (wake_up)
 997                 wake_up(&mdev->al_wait);
 998 }
 999
1000 /*
1001  * this is intended to set one request worth of data out of sync.
1002  * affects at least 1 bit,
1003  * and at most 1+DRBD_MAX_SEGMENT_SIZE/BM_BLOCK_SIZE bits.
1004  *
1005  * called by tl_clear and drbd_send_dblock (==drbd_make_request).
1006  * so this can be _any_ process.
1007  */
1008 void __drbd_set_out_of_sync(struct drbd_conf *mdev, sector_t sector, int size,
1009                             const char *file, const unsigned int line)
1010 {
1011         unsigned long sbnr, ebnr, lbnr, flags;
1012         sector_t esector, nr_sectors;
1013         unsigned int enr, count;
1014         struct lc_element *e;
1015
1016         if (size <= 0 || (size & 0x1ff) != 0 || size > DRBD_MAX_SEGMENT_SIZE) {
1017                 dev_err(DEV, "sector: %llus, size: %d\n",
1018                         (unsigned long long)sector, size);
1019                 return;
1020         }
1021
1022         if (!get_ldev(mdev))
1023                 return; /* no disk, no metadata, no bitmap to set bits in */
1024
1025         nr_sectors = drbd_get_capacity(mdev->this_bdev);
1026         esector = sector + (size >> 9) - 1;
1027
1028         ERR_IF(sector >= nr_sectors)
1029                 goto out;
1030         ERR_IF(esector >= nr_sectors)
1031                 esector = (nr_sectors-1);
1032
1033         lbnr = BM_SECT_TO_BIT(nr_sectors-1);
1034
1035         /* we set it out of sync,
1036          * we do not need to round anything here */
1037         sbnr = BM_SECT_TO_BIT(sector);
1038         ebnr = BM_SECT_TO_BIT(esector);
1039
1040         /* ok, (capacity & 7) != 0 sometimes, but who cares...
1041          * we count rs_{total,left} in bits, not sectors.  */
1042         spin_lock_irqsave(&mdev->al_lock, flags);
1043         count = drbd_bm_set_bits(mdev, sbnr, ebnr);
1044
1045         enr = BM_SECT_TO_EXT(sector);
1046         e = lc_find(mdev->resync, enr);
1047         if (e)
1048                 lc_entry(e, struct bm_extent, lce)->rs_left += count;
1049         spin_unlock_irqrestore(&mdev->al_lock, flags);
1050
1051 out:
1052         put_ldev(mdev);
1053 }
1054
1055 static
1056 struct bm_extent *_bme_get(struct drbd_conf *mdev, unsigned int enr)
1057 {
1058         struct lc_element *e;
1059         struct bm_extent *bm_ext;
1060         int wakeup = 0;
1061         unsigned long rs_flags;
1062
1063         spin_lock_irq(&mdev->al_lock);
1064         if (mdev->resync_locked > mdev->resync->nr_elements/2) {
1065                 spin_unlock_irq(&mdev->al_lock);
1066                 return NULL;
1067         }
1068         e = lc_get(mdev->resync, enr);
1069         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1070         if (bm_ext) {
1071                 if (bm_ext->lce.lc_number != enr) {
1072                         bm_ext->rs_left = drbd_bm_e_weight(mdev, enr);
1073                         bm_ext->rs_failed = 0;
1074                         lc_changed(mdev->resync, &bm_ext->lce);
1075                         wakeup = 1;
1076                 }
1077                 if (bm_ext->lce.refcnt == 1)
1078                         mdev->resync_locked++;
1079                 set_bit(BME_NO_WRITES, &bm_ext->flags);
1080         }
1081         rs_flags = mdev->resync->flags;
1082         spin_unlock_irq(&mdev->al_lock);
1083         if (wakeup)
1084                 wake_up(&mdev->al_wait);
1085
1086         if (!bm_ext) {
1087                 if (rs_flags & LC_STARVING)
1088                         dev_warn(DEV, "Have to wait for element"
1089                              " (resync LRU too small?)\n");
1090                 BUG_ON(rs_flags & LC_DIRTY);
1091         }
1092
1093         return bm_ext;
1094 }
1095
1096 static int _is_in_al(struct drbd_conf *mdev, unsigned int enr)
1097 {
1098         struct lc_element *al_ext;
1099         int rv = 0;
1100
1101         spin_lock_irq(&mdev->al_lock);
1102         if (unlikely(enr == mdev->act_log->new_number))
1103                 rv = 1;
1104         else {
1105                 al_ext = lc_find(mdev->act_log, enr);
1106                 if (al_ext) {
1107                         if (al_ext->refcnt)
1108                                 rv = 1;
1109                 }
1110         }
1111         spin_unlock_irq(&mdev->al_lock);
1112
1113         /*
1114         if (unlikely(rv)) {
1115                 dev_info(DEV, "Delaying sync read until app's write is done\n");
1116         }
1117         */
1118         return rv;
1119 }
1120
1121 /**
1122  * drbd_rs_begin_io() - Gets an extent in the resync LRU cache and sets it to BME_LOCKED
1123  * @mdev:       DRBD device.
1124  * @sector:     The sector number.
1125  *
1126  * This functions sleeps on al_wait. Returns 0 on success, -EINTR if interrupted.
1127  */
1128 int drbd_rs_begin_io(struct drbd_conf *mdev, sector_t sector)
1129 {
1130         unsigned int enr = BM_SECT_TO_EXT(sector);
1131         struct bm_extent *bm_ext;
1132         int i, sig;
1133
1134         sig = wait_event_interruptible(mdev->al_wait,
1135                         (bm_ext = _bme_get(mdev, enr)));
1136         if (sig)
1137                 return -EINTR;
1138
1139         if (test_bit(BME_LOCKED, &bm_ext->flags))
1140                 return 0;
1141
1142         for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
1143                 sig = wait_event_interruptible(mdev->al_wait,
1144                                 !_is_in_al(mdev, enr * AL_EXT_PER_BM_SECT + i));
1145                 if (sig) {
1146                         spin_lock_irq(&mdev->al_lock);
1147                         if (lc_put(mdev->resync, &bm_ext->lce) == 0) {
1148                                 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1149                                 mdev->resync_locked--;
1150                                 wake_up(&mdev->al_wait);
1151                         }
1152                         spin_unlock_irq(&mdev->al_lock);
1153                         return -EINTR;
1154                 }
1155         }
1156         set_bit(BME_LOCKED, &bm_ext->flags);
1157         return 0;
1158 }
1159
1160 /**
1161  * drbd_try_rs_begin_io() - Gets an extent in the resync LRU cache, does not sleep
1162  * @mdev:       DRBD device.
1163  * @sector:     The sector number.
1164  *
1165  * Gets an extent in the resync LRU cache, sets it to BME_NO_WRITES, then
1166  * tries to set it to BME_LOCKED. Returns 0 upon success, and -EAGAIN
1167  * if there is still application IO going on in this area.
1168  */
1169 int drbd_try_rs_begin_io(struct drbd_conf *mdev, sector_t sector)
1170 {
1171         unsigned int enr = BM_SECT_TO_EXT(sector);
1172         const unsigned int al_enr = enr*AL_EXT_PER_BM_SECT;
1173         struct lc_element *e;
1174         struct bm_extent *bm_ext;
1175         int i;
1176
1177         spin_lock_irq(&mdev->al_lock);
1178         if (mdev->resync_wenr != LC_FREE && mdev->resync_wenr != enr) {
1179                 /* in case you have very heavy scattered io, it may
1180                  * stall the syncer undefined if we give up the ref count
1181                  * when we try again and requeue.
1182                  *
1183                  * if we don't give up the refcount, but the next time
1184                  * we are scheduled this extent has been "synced" by new
1185                  * application writes, we'd miss the lc_put on the
1186                  * extent we keep the refcount on.
1187                  * so we remembered which extent we had to try again, and
1188                  * if the next requested one is something else, we do
1189                  * the lc_put here...
1190                  * we also have to wake_up
1191                  */
1192                 e = lc_find(mdev->resync, mdev->resync_wenr);
1193                 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1194                 if (bm_ext) {
1195                         D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags));
1196                         D_ASSERT(test_bit(BME_NO_WRITES, &bm_ext->flags));
1197                         clear_bit(BME_NO_WRITES, &bm_ext->flags);
1198                         mdev->resync_wenr = LC_FREE;
1199                         if (lc_put(mdev->resync, &bm_ext->lce) == 0)
1200                                 mdev->resync_locked--;
1201                         wake_up(&mdev->al_wait);
1202                 } else {
1203                         dev_alert(DEV, "LOGIC BUG\n");
1204                 }
1205         }
1206         /* TRY. */
1207         e = lc_try_get(mdev->resync, enr);
1208         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1209         if (bm_ext) {
1210                 if (test_bit(BME_LOCKED, &bm_ext->flags))
1211                         goto proceed;
1212                 if (!test_and_set_bit(BME_NO_WRITES, &bm_ext->flags)) {
1213                         mdev->resync_locked++;
1214                 } else {
1215                         /* we did set the BME_NO_WRITES,
1216                          * but then could not set BME_LOCKED,
1217                          * so we tried again.
1218                          * drop the extra reference. */
1219                         bm_ext->lce.refcnt--;
1220                         D_ASSERT(bm_ext->lce.refcnt > 0);
1221                 }
1222                 goto check_al;
1223         } else {
1224                 /* do we rather want to try later? */
1225                 if (mdev->resync_locked > mdev->resync->nr_elements-3)
1226                         goto try_again;
1227                 /* Do or do not. There is no try. -- Yoda */
1228                 e = lc_get(mdev->resync, enr);
1229                 bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1230                 if (!bm_ext) {
1231                         const unsigned long rs_flags = mdev->resync->flags;
1232                         if (rs_flags & LC_STARVING)
1233                                 dev_warn(DEV, "Have to wait for element"
1234                                      " (resync LRU too small?)\n");
1235                         BUG_ON(rs_flags & LC_DIRTY);
1236                         goto try_again;
1237                 }
1238                 if (bm_ext->lce.lc_number != enr) {
1239                         bm_ext->rs_left = drbd_bm_e_weight(mdev, enr);
1240                         bm_ext->rs_failed = 0;
1241                         lc_changed(mdev->resync, &bm_ext->lce);
1242                         wake_up(&mdev->al_wait);
1243                         D_ASSERT(test_bit(BME_LOCKED, &bm_ext->flags) == 0);
1244                 }
1245                 set_bit(BME_NO_WRITES, &bm_ext->flags);
1246                 D_ASSERT(bm_ext->lce.refcnt == 1);
1247                 mdev->resync_locked++;
1248                 goto check_al;
1249         }
1250 check_al:
1251         for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
1252                 if (unlikely(al_enr+i == mdev->act_log->new_number))
1253                         goto try_again;
1254                 if (lc_is_used(mdev->act_log, al_enr+i))
1255                         goto try_again;
1256         }
1257         set_bit(BME_LOCKED, &bm_ext->flags);
1258 proceed:
1259         mdev->resync_wenr = LC_FREE;
1260         spin_unlock_irq(&mdev->al_lock);
1261         return 0;
1262
1263 try_again:
1264         if (bm_ext)
1265                 mdev->resync_wenr = enr;
1266         spin_unlock_irq(&mdev->al_lock);
1267         return -EAGAIN;
1268 }
1269
1270 void drbd_rs_complete_io(struct drbd_conf *mdev, sector_t sector)
1271 {
1272         unsigned int enr = BM_SECT_TO_EXT(sector);
1273         struct lc_element *e;
1274         struct bm_extent *bm_ext;
1275         unsigned long flags;
1276
1277         spin_lock_irqsave(&mdev->al_lock, flags);
1278         e = lc_find(mdev->resync, enr);
1279         bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
1280         if (!bm_ext) {
1281                 spin_unlock_irqrestore(&mdev->al_lock, flags);
1282                 if (__ratelimit(&drbd_ratelimit_state))
1283                         dev_err(DEV, "drbd_rs_complete_io() called, but extent not found\n");
1284                 return;
1285         }
1286
1287         if (bm_ext->lce.refcnt == 0) {
1288                 spin_unlock_irqrestore(&mdev->al_lock, flags);
1289                 dev_err(DEV, "drbd_rs_complete_io(,%llu [=%u]) called, "
1290                     "but refcnt is 0!?\n",
1291                     (unsigned long long)sector, enr);
1292                 return;
1293         }
1294
1295         if (lc_put(mdev->resync, &bm_ext->lce) == 0) {
1296                 clear_bit(BME_LOCKED, &bm_ext->flags);
1297                 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1298                 mdev->resync_locked--;
1299                 wake_up(&mdev->al_wait);
1300         }
1301
1302         spin_unlock_irqrestore(&mdev->al_lock, flags);
1303 }
1304
1305 /**
1306  * drbd_rs_cancel_all() - Removes all extents from the resync LRU (even BME_LOCKED)
1307  * @mdev:       DRBD device.
1308  */
1309 void drbd_rs_cancel_all(struct drbd_conf *mdev)
1310 {
1311         spin_lock_irq(&mdev->al_lock);
1312
1313         if (get_ldev_if_state(mdev, D_FAILED)) { /* Makes sure ->resync is there. */
1314                 lc_reset(mdev->resync);
1315                 put_ldev(mdev);
1316         }
1317         mdev->resync_locked = 0;
1318         mdev->resync_wenr = LC_FREE;
1319         spin_unlock_irq(&mdev->al_lock);
1320         wake_up(&mdev->al_wait);
1321 }
1322
1323 /**
1324  * drbd_rs_del_all() - Gracefully remove all extents from the resync LRU
1325  * @mdev:       DRBD device.
1326  *
1327  * Returns 0 upon success, -EAGAIN if at least one reference count was
1328  * not zero.
1329  */
1330 int drbd_rs_del_all(struct drbd_conf *mdev)
1331 {
1332         struct lc_element *e;
1333         struct bm_extent *bm_ext;
1334         int i;
1335
1336         spin_lock_irq(&mdev->al_lock);
1337
1338         if (get_ldev_if_state(mdev, D_FAILED)) {
1339                 /* ok, ->resync is there. */
1340                 for (i = 0; i < mdev->resync->nr_elements; i++) {
1341                         e = lc_element_by_index(mdev->resync, i);
1342                         bm_ext = lc_entry(e, struct bm_extent, lce);
1343                         if (bm_ext->lce.lc_number == LC_FREE)
1344                                 continue;
1345                         if (bm_ext->lce.lc_number == mdev->resync_wenr) {
1346                                 dev_info(DEV, "dropping %u in drbd_rs_del_all, apparently"
1347                                      " got 'synced' by application io\n",
1348                                      mdev->resync_wenr);
1349                                 D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags));
1350                                 D_ASSERT(test_bit(BME_NO_WRITES, &bm_ext->flags));
1351                                 clear_bit(BME_NO_WRITES, &bm_ext->flags);
1352                                 mdev->resync_wenr = LC_FREE;
1353                                 lc_put(mdev->resync, &bm_ext->lce);
1354                         }
1355                         if (bm_ext->lce.refcnt != 0) {
1356                                 dev_info(DEV, "Retrying drbd_rs_del_all() later. "
1357                                      "refcnt=%d\n", bm_ext->lce.refcnt);
1358                                 put_ldev(mdev);
1359                                 spin_unlock_irq(&mdev->al_lock);
1360                                 return -EAGAIN;
1361                         }
1362                         D_ASSERT(!test_bit(BME_LOCKED, &bm_ext->flags));
1363                         D_ASSERT(!test_bit(BME_NO_WRITES, &bm_ext->flags));
1364                         lc_del(mdev->resync, &bm_ext->lce);
1365                 }
1366                 D_ASSERT(mdev->resync->used == 0);
1367                 put_ldev(mdev);
1368         }
1369         spin_unlock_irq(&mdev->al_lock);
1370
1371         return 0;
1372 }
1373
1374 /**
1375  * drbd_rs_failed_io() - Record information on a failure to resync the specified blocks
1376  * @mdev:       DRBD device.
1377  * @sector:     The sector number.
1378  * @size:       Size of failed IO operation, in byte.
1379  */
1380 void drbd_rs_failed_io(struct drbd_conf *mdev, sector_t sector, int size)
1381 {
1382         /* Is called from worker and receiver context _only_ */
1383         unsigned long sbnr, ebnr, lbnr;
1384         unsigned long count;
1385         sector_t esector, nr_sectors;
1386         int wake_up = 0;
1387
1388         if (size <= 0 || (size & 0x1ff) != 0 || size > DRBD_MAX_SEGMENT_SIZE) {
1389                 dev_err(DEV, "drbd_rs_failed_io: sector=%llus size=%d nonsense!\n",
1390                                 (unsigned long long)sector, size);
1391                 return;
1392         }
1393         nr_sectors = drbd_get_capacity(mdev->this_bdev);
1394         esector = sector + (size >> 9) - 1;
1395
1396         ERR_IF(sector >= nr_sectors) return;
1397         ERR_IF(esector >= nr_sectors) esector = (nr_sectors-1);
1398
1399         lbnr = BM_SECT_TO_BIT(nr_sectors-1);
1400
1401         /*
1402          * round up start sector, round down end sector.  we make sure we only
1403          * handle full, aligned, BM_BLOCK_SIZE (4K) blocks */
1404         if (unlikely(esector < BM_SECT_PER_BIT-1))
1405                 return;
1406         if (unlikely(esector == (nr_sectors-1)))
1407                 ebnr = lbnr;
1408         else
1409                 ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1));
1410         sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1);
1411
1412         if (sbnr > ebnr)
1413                 return;
1414
1415         /*
1416          * ok, (capacity & 7) != 0 sometimes, but who cares...
1417          * we count rs_{total,left} in bits, not sectors.
1418          */
1419         spin_lock_irq(&mdev->al_lock);
1420         count = drbd_bm_count_bits(mdev, sbnr, ebnr);
1421         if (count) {
1422                 mdev->rs_failed += count;
1423
1424                 if (get_ldev(mdev)) {
1425                         drbd_try_clear_on_disk_bm(mdev, sector, count, FALSE);
1426                         put_ldev(mdev);
1427                 }
1428
1429                 /* just wake_up unconditional now, various lc_chaged(),
1430                  * lc_put() in drbd_try_clear_on_disk_bm(). */
1431                 wake_up = 1;
1432         }
1433         spin_unlock_irq(&mdev->al_lock);
1434         if (wake_up)
1435                 wake_up(&mdev->al_wait);
1436 }