drivers/block/drbd/drbd_req.c

   1 /*
   2    drbd_req.c
   3
   4    This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
   5
   6    Copyright (C) 2001-2008, LINBIT Information Technologies GmbH.
   7    Copyright (C) 1999-2008, Philipp Reisner <philipp.reisner@linbit.com>.
   8    Copyright (C) 2002-2008, 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/module.h>
  27
  28 #include <linux/slab.h>
  29 #include <linux/drbd.h>
  30 #include "drbd_int.h"
  31 #include "drbd_req.h"
  32
  33
  34 /* Update disk stats at start of I/O request */
  35 static void _drbd_start_io_acct(struct drbd_conf *mdev, struct drbd_request *req, struct bio *bio)
  36 {
  37         const int rw = bio_data_dir(bio);
  38         int cpu;
  39         cpu = part_stat_lock();
  40         part_stat_inc(cpu, &mdev->vdisk->part0, ios[rw]);
  41         part_stat_add(cpu, &mdev->vdisk->part0, sectors[rw], bio_sectors(bio));
  42         part_inc_in_flight(&mdev->vdisk->part0, rw);
  43         part_stat_unlock();
  44 }
  45
  46 /* Update disk stats when completing request upwards */
  47 static void _drbd_end_io_acct(struct drbd_conf *mdev, struct drbd_request *req)
  48 {
  49         int rw = bio_data_dir(req->master_bio);
  50         unsigned long duration = jiffies - req->start_time;
  51         int cpu;
  52         cpu = part_stat_lock();
  53         part_stat_add(cpu, &mdev->vdisk->part0, ticks[rw], duration);
  54         part_round_stats(cpu, &mdev->vdisk->part0);
  55         part_dec_in_flight(&mdev->vdisk->part0, rw);
  56         part_stat_unlock();
  57 }
  58
  59 static struct drbd_request *drbd_req_new(struct drbd_conf *mdev,
  60                                                struct bio *bio_src)
  61 {
  62         struct drbd_request *req;
  63
  64         req = mempool_alloc(drbd_request_mempool, GFP_NOIO);
  65         if (!req)
  66                 return NULL;
  67
  68         drbd_req_make_private_bio(req, bio_src);
  69         req->rq_state    = bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0;
  70         req->w.mdev      = mdev;
  71         req->master_bio  = bio_src;
  72         req->epoch       = 0;
  73
  74         drbd_clear_interval(&req->i);
  75         req->i.sector     = bio_src->bi_sector;
  76         req->i.size      = bio_src->bi_size;
  77         req->i.local = true;
  78         req->i.waiting = false;
  79
  80         INIT_LIST_HEAD(&req->tl_requests);
  81         INIT_LIST_HEAD(&req->w.list);
  82
  83         return req;
  84 }
  85
  86 static void drbd_req_free(struct drbd_request *req)
  87 {
  88         mempool_free(req, drbd_request_mempool);
  89 }
  90
  91 /* rw is bio_data_dir(), only READ or WRITE */
  92 static void _req_is_done(struct drbd_conf *mdev, struct drbd_request *req, const int rw)
  93 {
  94         const unsigned long s = req->rq_state;
  95
  96         /* remove it from the transfer log.
  97          * well, only if it had been there in the first
  98          * place... if it had not (local only or conflicting
  99          * and never sent), it should still be "empty" as
 100          * initialized in drbd_req_new(), so we can list_del() it
 101          * here unconditionally */
 102         list_del(&req->tl_requests);
 103
 104         /* if it was a write, we may have to set the corresponding
 105          * bit(s) out-of-sync first. If it had a local part, we need to
 106          * release the reference to the activity log. */
 107         if (rw == WRITE) {
 108                 /* Set out-of-sync unless both OK flags are set
 109                  * (local only or remote failed).
 110                  * Other places where we set out-of-sync:
 111                  * READ with local io-error */
 112                 if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
 113                         drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
 114
 115                 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
 116                         drbd_set_in_sync(mdev, req->i.sector, req->i.size);
 117
 118                 /* one might be tempted to move the drbd_al_complete_io
 119                  * to the local io completion callback drbd_request_endio.
 120                  * but, if this was a mirror write, we may only
 121                  * drbd_al_complete_io after this is RQ_NET_DONE,
 122                  * otherwise the extent could be dropped from the al
 123                  * before it has actually been written on the peer.
 124                  * if we crash before our peer knows about the request,
 125                  * but after the extent has been dropped from the al,
 126                  * we would forget to resync the corresponding extent.
 127                  */
 128                 if (s & RQ_LOCAL_MASK) {
 129                         if (get_ldev_if_state(mdev, D_FAILED)) {
 130                                 if (s & RQ_IN_ACT_LOG)
 131                                         drbd_al_complete_io(mdev, &req->i);
 132                                 put_ldev(mdev);
 133                         } else if (__ratelimit(&drbd_ratelimit_state)) {
 134                                 dev_warn(DEV, "Should have called drbd_al_complete_io(, %llu, %u), "
 135                                          "but my Disk seems to have failed :(\n",
 136                                          (unsigned long long) req->i.sector, req->i.size);
 137                         }
 138                 }
 139         }
 140
 141         drbd_req_free(req);
 142 }
 143
 144 static void queue_barrier(struct drbd_conf *mdev)
 145 {
 146         struct drbd_tl_epoch *b;
 147
 148         /* We are within the req_lock. Once we queued the barrier for sending,
 149          * we set the CREATE_BARRIER bit. It is cleared as soon as a new
 150          * barrier/epoch object is added. This is the only place this bit is
 151          * set. It indicates that the barrier for this epoch is already queued,
 152          * and no new epoch has been created yet. */
 153         if (test_bit(CREATE_BARRIER, &mdev->flags))
 154                 return;
 155
 156         b = mdev->tconn->newest_tle;
 157         b->w.cb = w_send_barrier;
 158         b->w.mdev = mdev;
 159         /* inc_ap_pending done here, so we won't
 160          * get imbalanced on connection loss.
 161          * dec_ap_pending will be done in got_BarrierAck
 162          * or (on connection loss) in tl_clear.  */
 163         inc_ap_pending(mdev);
 164         drbd_queue_work(&mdev->tconn->data.work, &b->w);
 165         set_bit(CREATE_BARRIER, &mdev->flags);
 166 }
 167
 168 static void _about_to_complete_local_write(struct drbd_conf *mdev,
 169         struct drbd_request *req)
 170 {
 171         const unsigned long s = req->rq_state;
 172
 173         /* Before we can signal completion to the upper layers,
 174          * we may need to close the current epoch.
 175          * We can skip this, if this request has not even been sent, because we
 176          * did not have a fully established connection yet/anymore, during
 177          * bitmap exchange, or while we are C_AHEAD due to congestion policy.
 178          */
 179         if (mdev->state.conn >= C_CONNECTED &&
 180             (s & RQ_NET_SENT) != 0 &&
 181             req->epoch == mdev->tconn->newest_tle->br_number)
 182                 queue_barrier(mdev);
 183 }
 184
 185 void complete_master_bio(struct drbd_conf *mdev,
 186                 struct bio_and_error *m)
 187 {
 188         bio_endio(m->bio, m->error);
 189         dec_ap_bio(mdev);
 190 }
 191
 192
 193 static void drbd_remove_request_interval(struct rb_root *root,
 194                                          struct drbd_request *req)
 195 {
 196         struct drbd_conf *mdev = req->w.mdev;
 197         struct drbd_interval *i = &req->i;
 198
 199         drbd_remove_interval(root, i);
 200
 201         /* Wake up any processes waiting for this request to complete.  */
 202         if (i->waiting)
 203                 wake_up(&mdev->misc_wait);
 204 }
 205
 206 /* Helper for __req_mod().
 207  * Set m->bio to the master bio, if it is fit to be completed,
 208  * or leave it alone (it is initialized to NULL in __req_mod),
 209  * if it has already been completed, or cannot be completed yet.
 210  * If m->bio is set, the error status to be returned is placed in m->error.
 211  */
 212 void _req_may_be_done(struct drbd_request *req, struct bio_and_error *m)
 213 {
 214         const unsigned long s = req->rq_state;
 215         struct drbd_conf *mdev = req->w.mdev;
 216         int rw = req->rq_state & RQ_WRITE ? WRITE : READ;
 217
 218         /* we must not complete the master bio, while it is
 219          *      still being processed by _drbd_send_zc_bio (drbd_send_dblock)
 220          *      not yet acknowledged by the peer
 221          *      not yet completed by the local io subsystem
 222          * these flags may get cleared in any order by
 223          *      the worker,
 224          *      the receiver,
 225          *      the bio_endio completion callbacks.
 226          */
 227         if (s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED))
 228                 return;
 229         if (req->i.waiting) {
 230                 /* Retry all conflicting peer requests.  */
 231                 wake_up(&mdev->misc_wait);
 232         }
 233         if (s & RQ_NET_QUEUED)
 234                 return;
 235         if (s & RQ_NET_PENDING)
 236                 return;
 237
 238         if (req->master_bio) {
 239                 /* this is DATA_RECEIVED (remote read)
 240                  * or protocol C P_WRITE_ACK
 241                  * or protocol B P_RECV_ACK
 242                  * or protocol A "HANDED_OVER_TO_NETWORK" (SendAck)
 243                  * or canceled or failed,
 244                  * or killed from the transfer log due to connection loss.
 245                  */
 246
 247                 /*
 248                  * figure out whether to report success or failure.
 249                  *
 250                  * report success when at least one of the operations succeeded.
 251                  * or, to put the other way,
 252                  * only report failure, when both operations failed.
 253                  *
 254                  * what to do about the failures is handled elsewhere.
 255                  * what we need to do here is just: complete the master_bio.
 256                  *
 257                  * local completion error, if any, has been stored as ERR_PTR
 258                  * in private_bio within drbd_request_endio.
 259                  */
 260                 int ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
 261                 int error = PTR_ERR(req->private_bio);
 262
 263                 /* remove the request from the conflict detection
 264                  * respective block_id verification hash */
 265                 if (!drbd_interval_empty(&req->i)) {
 266                         struct rb_root *root;
 267
 268                         if (rw == WRITE)
 269                                 root = &mdev->write_requests;
 270                         else
 271                                 root = &mdev->read_requests;
 272                         drbd_remove_request_interval(root, req);
 273                 } else if (!(s & RQ_POSTPONED))
 274                         D_ASSERT((s & (RQ_NET_MASK & ~RQ_NET_DONE)) == 0);
 275
 276                 /* for writes we need to do some extra housekeeping */
 277                 if (rw == WRITE)
 278                         _about_to_complete_local_write(mdev, req);
 279
 280                 /* Update disk stats */
 281                 _drbd_end_io_acct(mdev, req);
 282
 283                 if (!(s & RQ_POSTPONED)) {
 284                         m->error = ok ? 0 : (error ?: -EIO);
 285                         m->bio = req->master_bio;
 286                 }
 287                 req->master_bio = NULL;
 288         }
 289
 290         if (s & RQ_LOCAL_PENDING)
 291                 return;
 292
 293         if ((s & RQ_NET_MASK) == 0 || (s & RQ_NET_DONE)) {
 294                 /* this is disconnected (local only) operation,
 295                  * or protocol C P_WRITE_ACK,
 296                  * or protocol A or B P_BARRIER_ACK,
 297                  * or killed from the transfer log due to connection loss. */
 298                 _req_is_done(mdev, req, rw);
 299         }
 300         /* else: network part and not DONE yet. that is
 301          * protocol A or B, barrier ack still pending... */
 302 }
 303
 304 static void _req_may_be_done_not_susp(struct drbd_request *req, struct bio_and_error *m)
 305 {
 306         struct drbd_conf *mdev = req->w.mdev;
 307
 308         if (!drbd_suspended(mdev))
 309                 _req_may_be_done(req, m);
 310 }
 311
 312 /* obviously this could be coded as many single functions
 313  * instead of one huge switch,
 314  * or by putting the code directly in the respective locations
 315  * (as it has been before).
 316  *
 317  * but having it this way
 318  *  enforces that it is all in this one place, where it is easier to audit,
 319  *  it makes it obvious that whatever "event" "happens" to a request should
 320  *  happen "atomically" within the req_lock,
 321  *  and it enforces that we have to think in a very structured manner
 322  *  about the "events" that may happen to a request during its life time ...
 323  */
 324 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
 325                 struct bio_and_error *m)
 326 {
 327         struct drbd_conf *mdev = req->w.mdev;
 328         struct net_conf *nc;
 329         int p, rv = 0;
 330
 331         if (m)
 332                 m->bio = NULL;
 333
 334         switch (what) {
 335         default:
 336                 dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
 337                 break;
 338
 339         /* does not happen...
 340          * initialization done in drbd_req_new
 341         case CREATED:
 342                 break;
 343                 */
 344
 345         case TO_BE_SENT: /* via network */
 346                 /* reached via __drbd_make_request
 347                  * and from w_read_retry_remote */
 348                 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
 349                 req->rq_state |= RQ_NET_PENDING;
 350                 rcu_read_lock();
 351                 nc = rcu_dereference(mdev->tconn->net_conf);
 352                 p = nc->wire_protocol;
 353                 rcu_read_unlock();
 354                 req->rq_state |=
 355                         p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
 356                         p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
 357                 inc_ap_pending(mdev);
 358                 break;
 359
 360         case TO_BE_SUBMITTED: /* locally */
 361                 /* reached via __drbd_make_request */
 362                 D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK));
 363                 req->rq_state |= RQ_LOCAL_PENDING;
 364                 break;
 365
 366         case COMPLETED_OK:
 367                 if (req->rq_state & RQ_WRITE)
 368                         mdev->writ_cnt += req->i.size >> 9;
 369                 else
 370                         mdev->read_cnt += req->i.size >> 9;
 371
 372                 req->rq_state |= (RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
 373                 req->rq_state &= ~RQ_LOCAL_PENDING;
 374
 375                 _req_may_be_done_not_susp(req, m);
 376                 put_ldev(mdev);
 377                 break;
 378
 379         case ABORT_DISK_IO:
 380                 req->rq_state |= RQ_LOCAL_ABORTED;
 381                 if (req->rq_state & RQ_WRITE)
 382                         _req_may_be_done_not_susp(req, m);
 383                 else
 384                         goto goto_queue_for_net_read;
 385                 break;
 386
 387         case WRITE_COMPLETED_WITH_ERROR:
 388                 req->rq_state |= RQ_LOCAL_COMPLETED;
 389                 req->rq_state &= ~RQ_LOCAL_PENDING;
 390
 391                 __drbd_chk_io_error(mdev, false);
 392                 _req_may_be_done_not_susp(req, m);
 393                 put_ldev(mdev);
 394                 break;
 395
 396         case READ_AHEAD_COMPLETED_WITH_ERROR:
 397                 /* it is legal to fail READA */
 398                 req->rq_state |= RQ_LOCAL_COMPLETED;
 399                 req->rq_state &= ~RQ_LOCAL_PENDING;
 400                 _req_may_be_done_not_susp(req, m);
 401                 put_ldev(mdev);
 402                 break;
 403
 404         case READ_COMPLETED_WITH_ERROR:
 405                 drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
 406
 407                 req->rq_state |= RQ_LOCAL_COMPLETED;
 408                 req->rq_state &= ~RQ_LOCAL_PENDING;
 409
 410                 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
 411
 412                 __drbd_chk_io_error(mdev, false);
 413                 put_ldev(mdev);
 414
 415         goto_queue_for_net_read:
 416
 417                 /* no point in retrying if there is no good remote data,
 418                  * or we have no connection. */
 419                 if (mdev->state.pdsk != D_UP_TO_DATE) {
 420                         _req_may_be_done_not_susp(req, m);
 421                         break;
 422                 }
 423
 424                 /* _req_mod(req,TO_BE_SENT); oops, recursion... */
 425                 req->rq_state |= RQ_NET_PENDING;
 426                 inc_ap_pending(mdev);
 427                 /* fall through: _req_mod(req,QUEUE_FOR_NET_READ); */
 428
 429         case QUEUE_FOR_NET_READ:
 430                 /* READ or READA, and
 431                  * no local disk,
 432                  * or target area marked as invalid,
 433                  * or just got an io-error. */
 434                 /* from __drbd_make_request
 435                  * or from bio_endio during read io-error recovery */
 436
 437                 /* so we can verify the handle in the answer packet
 438                  * corresponding hlist_del is in _req_may_be_done() */
 439                 drbd_insert_interval(&mdev->read_requests, &req->i);
 440
 441                 set_bit(UNPLUG_REMOTE, &mdev->flags);
 442
 443                 D_ASSERT(req->rq_state & RQ_NET_PENDING);
 444                 req->rq_state |= RQ_NET_QUEUED;
 445                 req->w.cb = (req->rq_state & RQ_LOCAL_MASK)
 446                         ? w_read_retry_remote
 447                         : w_send_read_req;
 448                 drbd_queue_work(&mdev->tconn->data.work, &req->w);
 449                 break;
 450
 451         case QUEUE_FOR_NET_WRITE:
 452                 /* assert something? */
 453                 /* from __drbd_make_request only */
 454
 455                 /* corresponding hlist_del is in _req_may_be_done() */
 456                 drbd_insert_interval(&mdev->write_requests, &req->i);
 457
 458                 /* NOTE
 459                  * In case the req ended up on the transfer log before being
 460                  * queued on the worker, it could lead to this request being
 461                  * missed during cleanup after connection loss.
 462                  * So we have to do both operations here,
 463                  * within the same lock that protects the transfer log.
 464                  *
 465                  * _req_add_to_epoch(req); this has to be after the
 466                  * _maybe_start_new_epoch(req); which happened in
 467                  * __drbd_make_request, because we now may set the bit
 468                  * again ourselves to close the current epoch.
 469                  *
 470                  * Add req to the (now) current epoch (barrier). */
 471
 472                 /* otherwise we may lose an unplug, which may cause some remote
 473                  * io-scheduler timeout to expire, increasing maximum latency,
 474                  * hurting performance. */
 475                 set_bit(UNPLUG_REMOTE, &mdev->flags);
 476
 477                 /* see __drbd_make_request,
 478                  * just after it grabs the req_lock */
 479                 D_ASSERT(test_bit(CREATE_BARRIER, &mdev->flags) == 0);
 480
 481                 req->epoch = mdev->tconn->newest_tle->br_number;
 482
 483                 /* increment size of current epoch */
 484                 mdev->tconn->newest_tle->n_writes++;
 485
 486                 /* queue work item to send data */
 487                 D_ASSERT(req->rq_state & RQ_NET_PENDING);
 488                 req->rq_state |= RQ_NET_QUEUED;
 489                 req->w.cb =  w_send_dblock;
 490                 drbd_queue_work(&mdev->tconn->data.work, &req->w);
 491
 492                 /* close the epoch, in case it outgrew the limit */
 493                 rcu_read_lock();
 494                 nc = rcu_dereference(mdev->tconn->net_conf);
 495                 p = nc->max_epoch_size;
 496                 rcu_read_unlock();
 497                 if (mdev->tconn->newest_tle->n_writes >= p)
 498                         queue_barrier(mdev);
 499
 500                 break;
 501
 502         case QUEUE_FOR_SEND_OOS:
 503                 req->rq_state |= RQ_NET_QUEUED;
 504                 req->w.cb =  w_send_out_of_sync;
 505                 drbd_queue_work(&mdev->tconn->data.work, &req->w);
 506                 break;
 507
 508         case OOS_HANDED_TO_NETWORK:
 509                 /* actually the same */
 510         case SEND_CANCELED:
 511                 /* treat it the same */
 512         case SEND_FAILED:
 513                 /* real cleanup will be done from tl_clear.  just update flags
 514                  * so it is no longer marked as on the worker queue */
 515                 req->rq_state &= ~RQ_NET_QUEUED;
 516                 /* if we did it right, tl_clear should be scheduled only after
 517                  * this, so this should not be necessary! */
 518                 _req_may_be_done_not_susp(req, m);
 519                 break;
 520
 521         case HANDED_OVER_TO_NETWORK:
 522                 /* assert something? */
 523                 if (bio_data_dir(req->master_bio) == WRITE)
 524                         atomic_add(req->i.size >> 9, &mdev->ap_in_flight);
 525
 526                 if (bio_data_dir(req->master_bio) == WRITE &&
 527                     !(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK))) {
 528                         /* this is what is dangerous about protocol A:
 529                          * pretend it was successfully written on the peer. */
 530                         if (req->rq_state & RQ_NET_PENDING) {
 531                                 dec_ap_pending(mdev);
 532                                 req->rq_state &= ~RQ_NET_PENDING;
 533                                 req->rq_state |= RQ_NET_OK;
 534                         } /* else: neg-ack was faster... */
 535                         /* it is still not yet RQ_NET_DONE until the
 536                          * corresponding epoch barrier got acked as well,
 537                          * so we know what to dirty on connection loss */
 538                 }
 539                 req->rq_state &= ~RQ_NET_QUEUED;
 540                 req->rq_state |= RQ_NET_SENT;
 541                 /* because _drbd_send_zc_bio could sleep, and may want to
 542                  * dereference the bio even after the "WRITE_ACKED_BY_PEER" and
 543                  * "COMPLETED_OK" events came in, once we return from
 544                  * _drbd_send_zc_bio (drbd_send_dblock), we have to check
 545                  * whether it is done already, and end it.  */
 546                 _req_may_be_done_not_susp(req, m);
 547                 break;
 548
 549         case READ_RETRY_REMOTE_CANCELED:
 550                 req->rq_state &= ~RQ_NET_QUEUED;
 551                 /* fall through, in case we raced with drbd_disconnect */
 552         case CONNECTION_LOST_WHILE_PENDING:
 553                 /* transfer log cleanup after connection loss */
 554                 /* assert something? */
 555                 if (req->rq_state & RQ_NET_PENDING)
 556                         dec_ap_pending(mdev);
 557                 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
 558                 req->rq_state |= RQ_NET_DONE;
 559                 if (req->rq_state & RQ_NET_SENT && req->rq_state & RQ_WRITE)
 560                         atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
 561
 562                 /* if it is still queued, we may not complete it here.
 563                  * it will be canceled soon. */
 564                 if (!(req->rq_state & RQ_NET_QUEUED))
 565                         _req_may_be_done(req, m); /* Allowed while state.susp */
 566                 break;
 567
 568         case WRITE_ACKED_BY_PEER_AND_SIS:
 569                 req->rq_state |= RQ_NET_SIS;
 570         case DISCARD_WRITE:
 571                 /* for discarded conflicting writes of multiple primaries,
 572                  * there is no need to keep anything in the tl, potential
 573                  * node crashes are covered by the activity log. */
 574                 req->rq_state |= RQ_NET_DONE;
 575                 /* fall through */
 576         case WRITE_ACKED_BY_PEER:
 577                 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
 578                 /* protocol C; successfully written on peer.
 579                  * Nothing to do here.
 580                  * We want to keep the tl in place for all protocols, to cater
 581                  * for volatile write-back caches on lower level devices.
 582                  *
 583                  * A barrier request is expected to have forced all prior
 584                  * requests onto stable storage, so completion of a barrier
 585                  * request could set NET_DONE right here, and not wait for the
 586                  * P_BARRIER_ACK, but that is an unnecessary optimization. */
 587
 588                 goto ack_common;
 589                 /* this makes it effectively the same as for: */
 590         case RECV_ACKED_BY_PEER:
 591                 D_ASSERT(req->rq_state & RQ_EXP_RECEIVE_ACK);
 592                 /* protocol B; pretends to be successfully written on peer.
 593                  * see also notes above in HANDED_OVER_TO_NETWORK about
 594                  * protocol != C */
 595         ack_common:
 596                 req->rq_state |= RQ_NET_OK;
 597                 D_ASSERT(req->rq_state & RQ_NET_PENDING);
 598                 dec_ap_pending(mdev);
 599                 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
 600                 req->rq_state &= ~RQ_NET_PENDING;
 601                 _req_may_be_done_not_susp(req, m);
 602                 break;
 603
 604         case POSTPONE_WRITE:
 605                 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
 606                 /* If this node has already detected the write conflict, the
 607                  * worker will be waiting on misc_wait.  Wake it up once this
 608                  * request has completed locally.
 609                  */
 610                 D_ASSERT(req->rq_state & RQ_NET_PENDING);
 611                 req->rq_state |= RQ_POSTPONED;
 612                 _req_may_be_done_not_susp(req, m);
 613                 break;
 614
 615         case NEG_ACKED:
 616                 /* assert something? */
 617                 if (req->rq_state & RQ_NET_PENDING) {
 618                         dec_ap_pending(mdev);
 619                         atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
 620                 }
 621                 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
 622
 623                 req->rq_state |= RQ_NET_DONE;
 624                 _req_may_be_done_not_susp(req, m);
 625                 /* else: done by HANDED_OVER_TO_NETWORK */
 626                 break;
 627
 628         case FAIL_FROZEN_DISK_IO:
 629                 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
 630                         break;
 631
 632                 _req_may_be_done(req, m); /* Allowed while state.susp */
 633                 break;
 634
 635         case RESTART_FROZEN_DISK_IO:
 636                 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
 637                         break;
 638
 639                 req->rq_state &= ~RQ_LOCAL_COMPLETED;
 640
 641                 rv = MR_READ;
 642                 if (bio_data_dir(req->master_bio) == WRITE)
 643                         rv = MR_WRITE;
 644
 645                 get_ldev(mdev);
 646                 req->w.cb = w_restart_disk_io;
 647                 drbd_queue_work(&mdev->tconn->data.work, &req->w);
 648                 break;
 649
 650         case RESEND:
 651                 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
 652                    before the connection loss (B&C only); only P_BARRIER_ACK was missing.
 653                    Trowing them out of the TL here by pretending we got a BARRIER_ACK
 654                    We ensure that the peer was not rebooted */
 655                 if (!(req->rq_state & RQ_NET_OK)) {
 656                         if (req->w.cb) {
 657                                 drbd_queue_work(&mdev->tconn->data.work, &req->w);
 658                                 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
 659                         }
 660                         break;
 661                 }
 662                 /* else, fall through to BARRIER_ACKED */
 663
 664         case BARRIER_ACKED:
 665                 if (!(req->rq_state & RQ_WRITE))
 666                         break;
 667
 668                 if (req->rq_state & RQ_NET_PENDING) {
 669                         /* barrier came in before all requests have been acked.
 670                          * this is bad, because if the connection is lost now,
 671                          * we won't be able to clean them up... */
 672                         dev_err(DEV, "FIXME (BARRIER_ACKED but pending)\n");
 673                         list_move(&req->tl_requests, &mdev->tconn->out_of_sequence_requests);
 674                 }
 675                 if ((req->rq_state & RQ_NET_MASK) != 0) {
 676                         req->rq_state |= RQ_NET_DONE;
 677                         if (!(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK)))
 678                                 atomic_sub(req->i.size>>9, &mdev->ap_in_flight);
 679                 }
 680                 _req_may_be_done(req, m); /* Allowed while state.susp */
 681                 break;
 682
 683         case DATA_RECEIVED:
 684                 D_ASSERT(req->rq_state & RQ_NET_PENDING);
 685                 dec_ap_pending(mdev);
 686                 req->rq_state &= ~RQ_NET_PENDING;
 687                 req->rq_state |= (RQ_NET_OK|RQ_NET_DONE);
 688                 _req_may_be_done_not_susp(req, m);
 689                 break;
 690         };
 691
 692         return rv;
 693 }
 694
 695 /* we may do a local read if:
 696  * - we are consistent (of course),
 697  * - or we are generally inconsistent,
 698  *   BUT we are still/already IN SYNC for this area.
 699  *   since size may be bigger than BM_BLOCK_SIZE,
 700  *   we may need to check several bits.
 701  */
 702 static bool drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size)
 703 {
 704         unsigned long sbnr, ebnr;
 705         sector_t esector, nr_sectors;
 706
 707         if (mdev->state.disk == D_UP_TO_DATE)
 708                 return true;
 709         if (mdev->state.disk != D_INCONSISTENT)
 710                 return false;
 711         esector = sector + (size >> 9) - 1;
 712         nr_sectors = drbd_get_capacity(mdev->this_bdev);
 713         D_ASSERT(sector  < nr_sectors);
 714         D_ASSERT(esector < nr_sectors);
 715
 716         sbnr = BM_SECT_TO_BIT(sector);
 717         ebnr = BM_SECT_TO_BIT(esector);
 718
 719         return drbd_bm_count_bits(mdev, sbnr, ebnr) == 0;
 720 }
 721
 722 /*
 723  * complete_conflicting_writes  -  wait for any conflicting write requests
 724  *
 725  * The write_requests tree contains all active write requests which we
 726  * currently know about.  Wait for any requests to complete which conflict with
 727  * the new one.
 728  */
 729 static int complete_conflicting_writes(struct drbd_conf *mdev,
 730                                        sector_t sector, int size)
 731 {
 732         for(;;) {
 733                 struct drbd_interval *i;
 734                 int err;
 735
 736                 i = drbd_find_overlap(&mdev->write_requests, sector, size);
 737                 if (!i)
 738                         return 0;
 739                 err = drbd_wait_misc(mdev, i);
 740                 if (err)
 741                         return err;
 742         }
 743 }
 744
 745 int __drbd_make_request(struct drbd_conf *mdev, struct bio *bio, unsigned long start_time)
 746 {
 747         const int rw = bio_rw(bio);
 748         const int size = bio->bi_size;
 749         const sector_t sector = bio->bi_sector;
 750         struct drbd_tl_epoch *b = NULL;
 751         struct drbd_request *req;
 752         struct net_conf *nc;
 753         int local, remote, send_oos = 0;
 754         int err;
 755         int ret = 0;
 756
 757         /* allocate outside of all locks; */
 758         req = drbd_req_new(mdev, bio);
 759         if (!req) {
 760                 dec_ap_bio(mdev);
 761                 /* only pass the error to the upper layers.
 762                  * if user cannot handle io errors, that's not our business. */
 763                 dev_err(DEV, "could not kmalloc() req\n");
 764                 bio_endio(bio, -ENOMEM);
 765                 return 0;
 766         }
 767         req->start_time = start_time;
 768
 769         local = get_ldev(mdev);
 770         if (!local) {
 771                 bio_put(req->private_bio); /* or we get a bio leak */
 772                 req->private_bio = NULL;
 773         }
 774         if (rw == WRITE) {
 775                 remote = 1;
 776         } else {
 777                 /* READ || READA */
 778                 if (local) {
 779                         if (!drbd_may_do_local_read(mdev, sector, size)) {
 780                                 /* we could kick the syncer to
 781                                  * sync this extent asap, wait for
 782                                  * it, then continue locally.
 783                                  * Or just issue the request remotely.
 784                                  */
 785                                 local = 0;
 786                                 bio_put(req->private_bio);
 787                                 req->private_bio = NULL;
 788                                 put_ldev(mdev);
 789                         }
 790                 }
 791                 remote = !local && mdev->state.pdsk >= D_UP_TO_DATE;
 792         }
 793
 794         /* If we have a disk, but a READA request is mapped to remote,
 795          * we are R_PRIMARY, D_INCONSISTENT, SyncTarget.
 796          * Just fail that READA request right here.
 797          *
 798          * THINK: maybe fail all READA when not local?
 799          *        or make this configurable...
 800          *        if network is slow, READA won't do any good.
 801          */
 802         if (rw == READA && mdev->state.disk >= D_INCONSISTENT && !local) {
 803                 err = -EWOULDBLOCK;
 804                 goto fail_and_free_req;
 805         }
 806
 807         /* For WRITES going to the local disk, grab a reference on the target
 808          * extent.  This waits for any resync activity in the corresponding
 809          * resync extent to finish, and, if necessary, pulls in the target
 810          * extent into the activity log, which involves further disk io because
 811          * of transactional on-disk meta data updates. */
 812         if (rw == WRITE && local && !test_bit(AL_SUSPENDED, &mdev->flags)) {
 813                 req->rq_state |= RQ_IN_ACT_LOG;
 814                 drbd_al_begin_io(mdev, &req->i);
 815         }
 816
 817         remote = remote && drbd_should_do_remote(mdev->state);
 818         send_oos = rw == WRITE && drbd_should_send_out_of_sync(mdev->state);
 819         D_ASSERT(!(remote && send_oos));
 820
 821         if (!(local || remote) && !drbd_suspended(mdev)) {
 822                 if (__ratelimit(&drbd_ratelimit_state))
 823                         dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
 824                 err = -EIO;
 825                 goto fail_free_complete;
 826         }
 827
 828         /* For WRITE request, we have to make sure that we have an
 829          * unused_spare_tle, in case we need to start a new epoch.
 830          * I try to be smart and avoid to pre-allocate always "just in case",
 831          * but there is a race between testing the bit and pointer outside the
 832          * spinlock, and grabbing the spinlock.
 833          * if we lost that race, we retry.  */
 834         if (rw == WRITE && (remote || send_oos) &&
 835             mdev->tconn->unused_spare_tle == NULL &&
 836             test_bit(CREATE_BARRIER, &mdev->flags)) {
 837 allocate_barrier:
 838                 b = kmalloc(sizeof(struct drbd_tl_epoch), GFP_NOIO);
 839                 if (!b) {
 840                         dev_err(DEV, "Failed to alloc barrier.\n");
 841                         err = -ENOMEM;
 842                         goto fail_free_complete;
 843                 }
 844         }
 845
 846         /* GOOD, everything prepared, grab the spin_lock */
 847         spin_lock_irq(&mdev->tconn->req_lock);
 848
 849         if (rw == WRITE) {
 850                 err = complete_conflicting_writes(mdev, sector, size);
 851                 if (err) {
 852                         if (err != -ERESTARTSYS)
 853                                 _conn_request_state(mdev->tconn,
 854                                                     NS(conn, C_TIMEOUT),
 855                                                     CS_HARD);
 856                         spin_unlock_irq(&mdev->tconn->req_lock);
 857                         err = -EIO;
 858                         goto fail_free_complete;
 859                 }
 860         }
 861
 862         if (drbd_suspended(mdev)) {
 863                 /* If we got suspended, use the retry mechanism of
 864                    generic_make_request() to restart processing of this
 865                    bio. In the next call to drbd_make_request
 866                    we sleep in inc_ap_bio() */
 867                 ret = 1;
 868                 spin_unlock_irq(&mdev->tconn->req_lock);
 869                 goto fail_free_complete;
 870         }
 871
 872         if (remote || send_oos) {
 873                 remote = drbd_should_do_remote(mdev->state);
 874                 send_oos = rw == WRITE && drbd_should_send_out_of_sync(mdev->state);
 875                 D_ASSERT(!(remote && send_oos));
 876
 877                 if (!(remote || send_oos))
 878                         dev_warn(DEV, "lost connection while grabbing the req_lock!\n");
 879                 if (!(local || remote)) {
 880                         dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
 881                         spin_unlock_irq(&mdev->tconn->req_lock);
 882                         err = -EIO;
 883                         goto fail_free_complete;
 884                 }
 885         }
 886
 887         if (b && mdev->tconn->unused_spare_tle == NULL) {
 888                 mdev->tconn->unused_spare_tle = b;
 889                 b = NULL;
 890         }
 891         if (rw == WRITE && (remote || send_oos) &&
 892             mdev->tconn->unused_spare_tle == NULL &&
 893             test_bit(CREATE_BARRIER, &mdev->flags)) {
 894                 /* someone closed the current epoch
 895                  * while we were grabbing the spinlock */
 896                 spin_unlock_irq(&mdev->tconn->req_lock);
 897                 goto allocate_barrier;
 898         }
 899
 900
 901         /* Update disk stats */
 902         _drbd_start_io_acct(mdev, req, bio);
 903
 904         /* _maybe_start_new_epoch(mdev);
 905          * If we need to generate a write barrier packet, we have to add the
 906          * new epoch (barrier) object, and queue the barrier packet for sending,
 907          * and queue the req's data after it _within the same lock_, otherwise
 908          * we have race conditions were the reorder domains could be mixed up.
 909          *
 910          * Even read requests may start a new epoch and queue the corresponding
 911          * barrier packet.  To get the write ordering right, we only have to
 912          * make sure that, if this is a write request and it triggered a
 913          * barrier packet, this request is queued within the same spinlock. */
 914         if ((remote || send_oos) && mdev->tconn->unused_spare_tle &&
 915             test_and_clear_bit(CREATE_BARRIER, &mdev->flags)) {
 916                 _tl_add_barrier(mdev->tconn, mdev->tconn->unused_spare_tle);
 917                 mdev->tconn->unused_spare_tle = NULL;
 918         } else {
 919                 D_ASSERT(!(remote && rw == WRITE &&
 920                            test_bit(CREATE_BARRIER, &mdev->flags)));
 921         }
 922
 923         /* NOTE
 924          * Actually, 'local' may be wrong here already, since we may have failed
 925          * to write to the meta data, and may become wrong anytime because of
 926          * local io-error for some other request, which would lead to us
 927          * "detaching" the local disk.
 928          *
 929          * 'remote' may become wrong any time because the network could fail.
 930          *
 931          * This is a harmless race condition, though, since it is handled
 932          * correctly at the appropriate places; so it just defers the failure
 933          * of the respective operation.
 934          */
 935
 936         /* mark them early for readability.
 937          * this just sets some state flags. */
 938         if (remote)
 939                 _req_mod(req, TO_BE_SENT);
 940         if (local)
 941                 _req_mod(req, TO_BE_SUBMITTED);
 942
 943         list_add_tail(&req->tl_requests, &mdev->tconn->newest_tle->requests);
 944
 945         /* NOTE remote first: to get the concurrent write detection right,
 946          * we must register the request before start of local IO.  */
 947         if (remote) {
 948                 /* either WRITE and C_CONNECTED,
 949                  * or READ, and no local disk,
 950                  * or READ, but not in sync.
 951                  */
 952                 _req_mod(req, (rw == WRITE)
 953                                 ? QUEUE_FOR_NET_WRITE
 954                                 : QUEUE_FOR_NET_READ);
 955         }
 956         if (send_oos && drbd_set_out_of_sync(mdev, sector, size))
 957                 _req_mod(req, QUEUE_FOR_SEND_OOS);
 958
 959         rcu_read_lock();
 960         nc = rcu_dereference(mdev->tconn->net_conf);
 961         if (remote &&
 962             nc->on_congestion != OC_BLOCK && mdev->tconn->agreed_pro_version >= 96) {
 963                 int congested = 0;
 964
 965                 if (nc->cong_fill &&
 966                     atomic_read(&mdev->ap_in_flight) >= nc->cong_fill) {
 967                         dev_info(DEV, "Congestion-fill threshold reached\n");
 968                         congested = 1;
 969                 }
 970
 971                 if (mdev->act_log->used >= nc->cong_extents) {
 972                         dev_info(DEV, "Congestion-extents threshold reached\n");
 973                         congested = 1;
 974                 }
 975
 976                 if (congested) {
 977                         queue_barrier(mdev); /* last barrier, after mirrored writes */
 978
 979                         if (nc->on_congestion == OC_PULL_AHEAD)
 980                                 _drbd_set_state(_NS(mdev, conn, C_AHEAD), 0, NULL);
 981                         else  /*nc->on_congestion == OC_DISCONNECT */
 982                                 _drbd_set_state(_NS(mdev, conn, C_DISCONNECTING), 0, NULL);
 983                 }
 984         }
 985         rcu_read_unlock();
 986
 987         spin_unlock_irq(&mdev->tconn->req_lock);
 988         kfree(b); /* if someone else has beaten us to it... */
 989
 990         if (local) {
 991                 req->private_bio->bi_bdev = mdev->ldev->backing_bdev;
 992
 993                 /* State may have changed since we grabbed our reference on the
 994                  * mdev->ldev member. Double check, and short-circuit to endio.
 995                  * In case the last activity log transaction failed to get on
 996                  * stable storage, and this is a WRITE, we may not even submit
 997                  * this bio. */
 998                 if (get_ldev(mdev)) {
 999                         if (drbd_insert_fault(mdev,   rw == WRITE ? DRBD_FAULT_DT_WR
1000                                                     : rw == READ  ? DRBD_FAULT_DT_RD
1001                                                     :               DRBD_FAULT_DT_RA))
1002                                 bio_endio(req->private_bio, -EIO);
1003                         else
1004                                 generic_make_request(req->private_bio);
1005                         put_ldev(mdev);
1006                 } else
1007                         bio_endio(req->private_bio, -EIO);
1008         }
1009
1010         return 0;
1011
1012 fail_free_complete:
1013         if (req->rq_state & RQ_IN_ACT_LOG)
1014                 drbd_al_complete_io(mdev, &req->i);
1015 fail_and_free_req:
1016         if (local) {
1017                 bio_put(req->private_bio);
1018                 req->private_bio = NULL;
1019                 put_ldev(mdev);
1020         }
1021         if (!ret)
1022                 bio_endio(bio, err);
1023
1024         drbd_req_free(req);
1025         dec_ap_bio(mdev);
1026         kfree(b);
1027
1028         return ret;
1029 }
1030
1031 int drbd_make_request(struct request_queue *q, struct bio *bio)
1032 {
1033         struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1034         unsigned long start_time;
1035
1036         start_time = jiffies;
1037
1038         /*
1039          * what we "blindly" assume:
1040          */
1041         D_ASSERT(bio->bi_size > 0);
1042         D_ASSERT(IS_ALIGNED(bio->bi_size, 512));
1043
1044         inc_ap_bio(mdev);
1045         return __drbd_make_request(mdev, bio, start_time);
1046 }
1047
1048 /* This is called by bio_add_page().
1049  *
1050  * q->max_hw_sectors and other global limits are already enforced there.
1051  *
1052  * We need to call down to our lower level device,
1053  * in case it has special restrictions.
1054  *
1055  * We also may need to enforce configured max-bio-bvecs limits.
1056  *
1057  * As long as the BIO is empty we have to allow at least one bvec,
1058  * regardless of size and offset, so no need to ask lower levels.
1059  */
1060 int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec)
1061 {
1062         struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1063         unsigned int bio_size = bvm->bi_size;
1064         int limit = DRBD_MAX_BIO_SIZE;
1065         int backing_limit;
1066
1067         if (bio_size && get_ldev(mdev)) {
1068                 struct request_queue * const b =
1069                         mdev->ldev->backing_bdev->bd_disk->queue;
1070                 if (b->merge_bvec_fn) {
1071                         backing_limit = b->merge_bvec_fn(b, bvm, bvec);
1072                         limit = min(limit, backing_limit);
1073                 }
1074                 put_ldev(mdev);
1075         }
1076         return limit;
1077 }
1078
1079 void request_timer_fn(unsigned long data)
1080 {
1081         struct drbd_conf *mdev = (struct drbd_conf *) data;
1082         struct drbd_tconn *tconn = mdev->tconn;
1083         struct drbd_request *req; /* oldest request */
1084         struct list_head *le;
1085         struct net_conf *nc;
1086         unsigned long ent = 0, dt = 0, et; /* effective timeout = ko_count * timeout */
1087
1088         rcu_read_lock();
1089         nc = rcu_dereference(tconn->net_conf);
1090         ent = nc ? nc->timeout * HZ/10 * nc->ko_count : 0;
1091
1092         if (get_ldev(mdev)) {
1093                 dt = rcu_dereference(mdev->ldev->disk_conf)->disk_timeout * HZ / 10;
1094                 put_ldev(mdev);
1095         }
1096         rcu_read_unlock();
1097
1098         et = min_not_zero(dt, ent);
1099
1100         if (!et || (mdev->state.conn < C_WF_REPORT_PARAMS && mdev->state.disk <= D_FAILED))
1101                 return; /* Recurring timer stopped */
1102
1103         spin_lock_irq(&tconn->req_lock);
1104         le = &tconn->oldest_tle->requests;
1105         if (list_empty(le)) {
1106                 spin_unlock_irq(&tconn->req_lock);
1107                 mod_timer(&mdev->request_timer, jiffies + et);
1108                 return;
1109         }
1110
1111         le = le->prev;
1112         req = list_entry(le, struct drbd_request, tl_requests);
1113         if (ent && req->rq_state & RQ_NET_PENDING) {
1114                 if (time_is_before_eq_jiffies(req->start_time + ent)) {
1115                         dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n");
1116                         _drbd_set_state(_NS(mdev, conn, C_TIMEOUT), CS_VERBOSE | CS_HARD, NULL);
1117                 }
1118         }
1119         if (dt && req->rq_state & RQ_LOCAL_PENDING) {
1120                 if (time_is_before_eq_jiffies(req->start_time + dt)) {
1121                         dev_warn(DEV, "Local backing device failed to meet the disk-timeout\n");
1122                         __drbd_chk_io_error(mdev, 1);
1123                 }
1124         }
1125         spin_unlock_irq(&tconn->req_lock);
1126         mod_timer(&mdev->request_timer, req->start_time + et);
1127 }