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 static bool drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size);
  35
  36 /* Update disk stats at start of I/O request */
  37 static void _drbd_start_io_acct(struct drbd_conf *mdev, struct drbd_request *req)
  38 {
  39         const int rw = bio_data_dir(req->master_bio);
  40         int cpu;
  41         cpu = part_stat_lock();
  42         part_round_stats(cpu, &mdev->vdisk->part0);
  43         part_stat_inc(cpu, &mdev->vdisk->part0, ios[rw]);
  44         part_stat_add(cpu, &mdev->vdisk->part0, sectors[rw], req->i.size >> 9);
  45         (void) cpu; /* The macro invocations above want the cpu argument, I do not like
  46                        the compiler warning about cpu only assigned but never used... */
  47         part_inc_in_flight(&mdev->vdisk->part0, rw);
  48         part_stat_unlock();
  49 }
  50
  51 /* Update disk stats when completing request upwards */
  52 static void _drbd_end_io_acct(struct drbd_conf *mdev, struct drbd_request *req)
  53 {
  54         int rw = bio_data_dir(req->master_bio);
  55         unsigned long duration = jiffies - req->start_time;
  56         int cpu;
  57         cpu = part_stat_lock();
  58         part_stat_add(cpu, &mdev->vdisk->part0, ticks[rw], duration);
  59         part_round_stats(cpu, &mdev->vdisk->part0);
  60         part_dec_in_flight(&mdev->vdisk->part0, rw);
  61         part_stat_unlock();
  62 }
  63
  64 static struct drbd_request *drbd_req_new(struct drbd_conf *mdev,
  65                                                struct bio *bio_src)
  66 {
  67         struct drbd_request *req;
  68
  69         req = mempool_alloc(drbd_request_mempool, GFP_NOIO);
  70         if (!req)
  71                 return NULL;
  72
  73         drbd_req_make_private_bio(req, bio_src);
  74         req->rq_state    = bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0;
  75         req->w.mdev      = mdev;
  76         req->master_bio  = bio_src;
  77         req->epoch       = 0;
  78
  79         drbd_clear_interval(&req->i);
  80         req->i.sector     = bio_src->bi_sector;
  81         req->i.size      = bio_src->bi_size;
  82         req->i.local = true;
  83         req->i.waiting = false;
  84
  85         INIT_LIST_HEAD(&req->tl_requests);
  86         INIT_LIST_HEAD(&req->w.list);
  87
  88         /* one reference to be put by __drbd_make_request */
  89         atomic_set(&req->completion_ref, 1);
  90         /* one kref as long as completion_ref > 0 */
  91         kref_init(&req->kref);
  92         return req;
  93 }
  94
  95 void drbd_req_destroy(struct kref *kref)
  96 {
  97         struct drbd_request *req = container_of(kref, struct drbd_request, kref);
  98         struct drbd_conf *mdev = req->w.mdev;
  99         const unsigned s = req->rq_state;
 100
 101         if ((req->master_bio && !(s & RQ_POSTPONED)) ||
 102                 atomic_read(&req->completion_ref) ||
 103                 (s & RQ_LOCAL_PENDING) ||
 104                 ((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) {
 105                 dev_err(DEV, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n",
 106                                 s, atomic_read(&req->completion_ref));
 107                 return;
 108         }
 109
 110         /* remove it from the transfer log.
 111          * well, only if it had been there in the first
 112          * place... if it had not (local only or conflicting
 113          * and never sent), it should still be "empty" as
 114          * initialized in drbd_req_new(), so we can list_del() it
 115          * here unconditionally */
 116         list_del_init(&req->tl_requests);
 117
 118         /* if it was a write, we may have to set the corresponding
 119          * bit(s) out-of-sync first. If it had a local part, we need to
 120          * release the reference to the activity log. */
 121         if (s & RQ_WRITE) {
 122                 /* Set out-of-sync unless both OK flags are set
 123                  * (local only or remote failed).
 124                  * Other places where we set out-of-sync:
 125                  * READ with local io-error */
 126
 127                 /* There is a special case:
 128                  * we may notice late that IO was suspended,
 129                  * and postpone, or schedule for retry, a write,
 130                  * before it even was submitted or sent.
 131                  * In that case we do not want to touch the bitmap at all.
 132                  */
 133                 if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) {
 134                         if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
 135                                 drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
 136
 137                         if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
 138                                 drbd_set_in_sync(mdev, req->i.sector, req->i.size);
 139                 }
 140
 141                 /* one might be tempted to move the drbd_al_complete_io
 142                  * to the local io completion callback drbd_request_endio.
 143                  * but, if this was a mirror write, we may only
 144                  * drbd_al_complete_io after this is RQ_NET_DONE,
 145                  * otherwise the extent could be dropped from the al
 146                  * before it has actually been written on the peer.
 147                  * if we crash before our peer knows about the request,
 148                  * but after the extent has been dropped from the al,
 149                  * we would forget to resync the corresponding extent.
 150                  */
 151                 if (s & RQ_IN_ACT_LOG) {
 152                         if (get_ldev_if_state(mdev, D_FAILED)) {
 153                                 drbd_al_complete_io(mdev, &req->i);
 154                                 put_ldev(mdev);
 155                         } else if (__ratelimit(&drbd_ratelimit_state)) {
 156                                 dev_warn(DEV, "Should have called drbd_al_complete_io(, %llu, %u), "
 157                                          "but my Disk seems to have failed :(\n",
 158                                          (unsigned long long) req->i.sector, req->i.size);
 159                         }
 160                 }
 161         }
 162
 163         mempool_free(req, drbd_request_mempool);
 164 }
 165
 166 static void wake_all_senders(struct drbd_tconn *tconn) {
 167         wake_up(&tconn->sender_work.q_wait);
 168 }
 169
 170 /* must hold resource->req_lock */
 171 void start_new_tl_epoch(struct drbd_tconn *tconn)
 172 {
 173         /* no point closing an epoch, if it is empty, anyways. */
 174         if (tconn->current_tle_writes == 0)
 175                 return;
 176
 177         tconn->current_tle_writes = 0;
 178         atomic_inc(&tconn->current_tle_nr);
 179         wake_all_senders(tconn);
 180 }
 181
 182 void complete_master_bio(struct drbd_conf *mdev,
 183                 struct bio_and_error *m)
 184 {
 185         bio_endio(m->bio, m->error);
 186         dec_ap_bio(mdev);
 187 }
 188
 189
 190 static void drbd_remove_request_interval(struct rb_root *root,
 191                                          struct drbd_request *req)
 192 {
 193         struct drbd_conf *mdev = req->w.mdev;
 194         struct drbd_interval *i = &req->i;
 195
 196         drbd_remove_interval(root, i);
 197
 198         /* Wake up any processes waiting for this request to complete.  */
 199         if (i->waiting)
 200                 wake_up(&mdev->misc_wait);
 201 }
 202
 203 /* Helper for __req_mod().
 204  * Set m->bio to the master bio, if it is fit to be completed,
 205  * or leave it alone (it is initialized to NULL in __req_mod),
 206  * if it has already been completed, or cannot be completed yet.
 207  * If m->bio is set, the error status to be returned is placed in m->error.
 208  */
 209 static
 210 void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
 211 {
 212         const unsigned s = req->rq_state;
 213         struct drbd_conf *mdev = req->w.mdev;
 214         int rw;
 215         int error, ok;
 216
 217         /* we must not complete the master bio, while it is
 218          *      still being processed by _drbd_send_zc_bio (drbd_send_dblock)
 219          *      not yet acknowledged by the peer
 220          *      not yet completed by the local io subsystem
 221          * these flags may get cleared in any order by
 222          *      the worker,
 223          *      the receiver,
 224          *      the bio_endio completion callbacks.
 225          */
 226         if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
 227             (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
 228             (s & RQ_COMPLETION_SUSP)) {
 229                 dev_err(DEV, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s);
 230                 return;
 231         }
 232
 233         if (!req->master_bio) {
 234                 dev_err(DEV, "drbd_req_complete: Logic BUG, master_bio == NULL!\n");
 235                 return;
 236         }
 237
 238         rw = bio_rw(req->master_bio);
 239
 240         /*
 241          * figure out whether to report success or failure.
 242          *
 243          * report success when at least one of the operations succeeded.
 244          * or, to put the other way,
 245          * only report failure, when both operations failed.
 246          *
 247          * what to do about the failures is handled elsewhere.
 248          * what we need to do here is just: complete the master_bio.
 249          *
 250          * local completion error, if any, has been stored as ERR_PTR
 251          * in private_bio within drbd_request_endio.
 252          */
 253         ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
 254         error = PTR_ERR(req->private_bio);
 255
 256         /* remove the request from the conflict detection
 257          * respective block_id verification hash */
 258         if (!drbd_interval_empty(&req->i)) {
 259                 struct rb_root *root;
 260
 261                 if (rw == WRITE)
 262                         root = &mdev->write_requests;
 263                 else
 264                         root = &mdev->read_requests;
 265                 drbd_remove_request_interval(root, req);
 266         }
 267
 268         /* Before we can signal completion to the upper layers,
 269          * we may need to close the current transfer log epoch.
 270          * We are within the request lock, so we can simply compare
 271          * the request epoch number with the current transfer log
 272          * epoch number.  If they match, increase the current_tle_nr,
 273          * and reset the transfer log epoch write_cnt.
 274          */
 275         if (rw == WRITE &&
 276             req->epoch == atomic_read(&mdev->tconn->current_tle_nr))
 277                 start_new_tl_epoch(mdev->tconn);
 278
 279         /* Update disk stats */
 280         _drbd_end_io_acct(mdev, req);
 281
 282         /* If READ failed,
 283          * have it be pushed back to the retry work queue,
 284          * so it will re-enter __drbd_make_request(),
 285          * and be re-assigned to a suitable local or remote path,
 286          * or failed if we do not have access to good data anymore.
 287          *
 288          * Unless it was failed early by __drbd_make_request(),
 289          * because no path was available, in which case
 290          * it was not even added to the transfer_log.
 291          *
 292          * READA may fail, and will not be retried.
 293          *
 294          * WRITE should have used all available paths already.
 295          */
 296         if (!ok && rw == READ && !list_empty(&req->tl_requests))
 297                 req->rq_state |= RQ_POSTPONED;
 298
 299         if (!(req->rq_state & RQ_POSTPONED)) {
 300                 m->error = ok ? 0 : (error ?: -EIO);
 301                 m->bio = req->master_bio;
 302                 req->master_bio = NULL;
 303         }
 304 }
 305
 306 static int drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
 307 {
 308         struct drbd_conf *mdev = req->w.mdev;
 309         D_ASSERT(m || (req->rq_state & RQ_POSTPONED));
 310
 311         if (!atomic_sub_and_test(put, &req->completion_ref))
 312                 return 0;
 313
 314         drbd_req_complete(req, m);
 315
 316         if (req->rq_state & RQ_POSTPONED) {
 317                 /* don't destroy the req object just yet,
 318                  * but queue it for retry */
 319                 drbd_restart_request(req);
 320                 return 0;
 321         }
 322
 323         return 1;
 324 }
 325
 326 /* I'd like this to be the only place that manipulates
 327  * req->completion_ref and req->kref. */
 328 static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m,
 329                 int clear, int set)
 330 {
 331         struct drbd_conf *mdev = req->w.mdev;
 332         unsigned s = req->rq_state;
 333         int c_put = 0;
 334         int k_put = 0;
 335
 336         if (drbd_suspended(mdev) && !((s | clear) & RQ_COMPLETION_SUSP))
 337                 set |= RQ_COMPLETION_SUSP;
 338
 339         /* apply */
 340
 341         req->rq_state &= ~clear;
 342         req->rq_state |= set;
 343
 344         /* no change? */
 345         if (req->rq_state == s)
 346                 return;
 347
 348         /* intent: get references */
 349
 350         if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING))
 351                 atomic_inc(&req->completion_ref);
 352
 353         if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) {
 354                 inc_ap_pending(mdev);
 355                 atomic_inc(&req->completion_ref);
 356         }
 357
 358         if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED))
 359                 atomic_inc(&req->completion_ref);
 360
 361         if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK))
 362                 kref_get(&req->kref); /* wait for the DONE */
 363
 364         if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT))
 365                 atomic_add(req->i.size >> 9, &mdev->ap_in_flight);
 366
 367         if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP))
 368                 atomic_inc(&req->completion_ref);
 369
 370         /* progress: put references */
 371
 372         if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP))
 373                 ++c_put;
 374
 375         if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) {
 376                 D_ASSERT(req->rq_state & RQ_LOCAL_PENDING);
 377                 /* local completion may still come in later,
 378                  * we need to keep the req object around. */
 379                 kref_get(&req->kref);
 380                 ++c_put;
 381         }
 382
 383         if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) {
 384                 if (req->rq_state & RQ_LOCAL_ABORTED)
 385                         ++k_put;
 386                 else
 387                         ++c_put;
 388         }
 389
 390         if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) {
 391                 dec_ap_pending(mdev);
 392                 ++c_put;
 393         }
 394
 395         if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED))
 396                 ++c_put;
 397
 398         if ((s & RQ_EXP_BARR_ACK) && !(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) {
 399                 if (req->rq_state & RQ_NET_SENT)
 400                         atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
 401                 ++k_put;
 402         }
 403
 404         /* potentially complete and destroy */
 405
 406         if (k_put || c_put) {
 407                 /* Completion does it's own kref_put.  If we are going to
 408                  * kref_sub below, we need req to be still around then. */
 409                 int at_least = k_put + !!c_put;
 410                 int refcount = atomic_read(&req->kref.refcount);
 411                 if (refcount < at_least)
 412                         dev_err(DEV,
 413                                 "mod_rq_state: Logic BUG: %x -> %x: refcount = %d, should be >= %d\n",
 414                                 s, req->rq_state, refcount, at_least);
 415         }
 416
 417         /* If we made progress, retry conflicting peer requests, if any. */
 418         if (req->i.waiting)
 419                 wake_up(&mdev->misc_wait);
 420
 421         if (c_put)
 422                 k_put += drbd_req_put_completion_ref(req, m, c_put);
 423         if (k_put)
 424                 kref_sub(&req->kref, k_put, drbd_req_destroy);
 425 }
 426
 427 static void drbd_report_io_error(struct drbd_conf *mdev, struct drbd_request *req)
 428 {
 429         char b[BDEVNAME_SIZE];
 430
 431         if (!__ratelimit(&drbd_ratelimit_state))
 432                 return;
 433
 434         dev_warn(DEV, "local %s IO error sector %llu+%u on %s\n",
 435                         (req->rq_state & RQ_WRITE) ? "WRITE" : "READ",
 436                         (unsigned long long)req->i.sector,
 437                         req->i.size >> 9,
 438                         bdevname(mdev->ldev->backing_bdev, b));
 439 }
 440
 441 /* obviously this could be coded as many single functions
 442  * instead of one huge switch,
 443  * or by putting the code directly in the respective locations
 444  * (as it has been before).
 445  *
 446  * but having it this way
 447  *  enforces that it is all in this one place, where it is easier to audit,
 448  *  it makes it obvious that whatever "event" "happens" to a request should
 449  *  happen "atomically" within the req_lock,
 450  *  and it enforces that we have to think in a very structured manner
 451  *  about the "events" that may happen to a request during its life time ...
 452  */
 453 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
 454                 struct bio_and_error *m)
 455 {
 456         struct drbd_conf *mdev = req->w.mdev;
 457         struct net_conf *nc;
 458         int p, rv = 0;
 459
 460         if (m)
 461                 m->bio = NULL;
 462
 463         switch (what) {
 464         default:
 465                 dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
 466                 break;
 467
 468         /* does not happen...
 469          * initialization done in drbd_req_new
 470         case CREATED:
 471                 break;
 472                 */
 473
 474         case TO_BE_SENT: /* via network */
 475                 /* reached via __drbd_make_request
 476                  * and from w_read_retry_remote */
 477                 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
 478                 rcu_read_lock();
 479                 nc = rcu_dereference(mdev->tconn->net_conf);
 480                 p = nc->wire_protocol;
 481                 rcu_read_unlock();
 482                 req->rq_state |=
 483                         p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
 484                         p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
 485                 mod_rq_state(req, m, 0, RQ_NET_PENDING);
 486                 break;
 487
 488         case TO_BE_SUBMITTED: /* locally */
 489                 /* reached via __drbd_make_request */
 490                 D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK));
 491                 mod_rq_state(req, m, 0, RQ_LOCAL_PENDING);
 492                 break;
 493
 494         case COMPLETED_OK:
 495                 if (req->rq_state & RQ_WRITE)
 496                         mdev->writ_cnt += req->i.size >> 9;
 497                 else
 498                         mdev->read_cnt += req->i.size >> 9;
 499
 500                 mod_rq_state(req, m, RQ_LOCAL_PENDING,
 501                                 RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
 502                 break;
 503
 504         case ABORT_DISK_IO:
 505                 mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED);
 506                 break;
 507
 508         case WRITE_COMPLETED_WITH_ERROR:
 509                 drbd_report_io_error(mdev, req);
 510                 __drbd_chk_io_error(mdev, DRBD_WRITE_ERROR);
 511                 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
 512                 break;
 513
 514         case READ_COMPLETED_WITH_ERROR:
 515                 drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
 516                 drbd_report_io_error(mdev, req);
 517                 __drbd_chk_io_error(mdev, DRBD_READ_ERROR);
 518                 /* fall through. */
 519         case READ_AHEAD_COMPLETED_WITH_ERROR:
 520                 /* it is legal to fail READA, no __drbd_chk_io_error in that case. */
 521                 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
 522                 break;
 523
 524         case QUEUE_FOR_NET_READ:
 525                 /* READ or READA, and
 526                  * no local disk,
 527                  * or target area marked as invalid,
 528                  * or just got an io-error. */
 529                 /* from __drbd_make_request
 530                  * or from bio_endio during read io-error recovery */
 531
 532                 /* So we can verify the handle in the answer packet.
 533                  * Corresponding drbd_remove_request_interval is in
 534                  * drbd_req_complete() */
 535                 D_ASSERT(drbd_interval_empty(&req->i));
 536                 drbd_insert_interval(&mdev->read_requests, &req->i);
 537
 538                 set_bit(UNPLUG_REMOTE, &mdev->flags);
 539
 540                 D_ASSERT(req->rq_state & RQ_NET_PENDING);
 541                 D_ASSERT((req->rq_state & RQ_LOCAL_MASK) == 0);
 542                 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
 543                 req->w.cb = w_send_read_req;
 544                 drbd_queue_work(&mdev->tconn->sender_work, &req->w);
 545                 break;
 546
 547         case QUEUE_FOR_NET_WRITE:
 548                 /* assert something? */
 549                 /* from __drbd_make_request only */
 550
 551                 /* Corresponding drbd_remove_request_interval is in
 552                  * drbd_req_complete() */
 553                 D_ASSERT(drbd_interval_empty(&req->i));
 554                 drbd_insert_interval(&mdev->write_requests, &req->i);
 555
 556                 /* NOTE
 557                  * In case the req ended up on the transfer log before being
 558                  * queued on the worker, it could lead to this request being
 559                  * missed during cleanup after connection loss.
 560                  * So we have to do both operations here,
 561                  * within the same lock that protects the transfer log.
 562                  *
 563                  * _req_add_to_epoch(req); this has to be after the
 564                  * _maybe_start_new_epoch(req); which happened in
 565                  * __drbd_make_request, because we now may set the bit
 566                  * again ourselves to close the current epoch.
 567                  *
 568                  * Add req to the (now) current epoch (barrier). */
 569
 570                 /* otherwise we may lose an unplug, which may cause some remote
 571                  * io-scheduler timeout to expire, increasing maximum latency,
 572                  * hurting performance. */
 573                 set_bit(UNPLUG_REMOTE, &mdev->flags);
 574
 575                 /* queue work item to send data */
 576                 D_ASSERT(req->rq_state & RQ_NET_PENDING);
 577                 mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
 578                 req->w.cb =  w_send_dblock;
 579                 drbd_queue_work(&mdev->tconn->sender_work, &req->w);
 580
 581                 /* close the epoch, in case it outgrew the limit */
 582                 rcu_read_lock();
 583                 nc = rcu_dereference(mdev->tconn->net_conf);
 584                 p = nc->max_epoch_size;
 585                 rcu_read_unlock();
 586                 if (mdev->tconn->current_tle_writes >= p)
 587                         start_new_tl_epoch(mdev->tconn);
 588
 589                 break;
 590
 591         case QUEUE_FOR_SEND_OOS:
 592                 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
 593                 req->w.cb =  w_send_out_of_sync;
 594                 drbd_queue_work(&mdev->tconn->sender_work, &req->w);
 595                 break;
 596
 597         case READ_RETRY_REMOTE_CANCELED:
 598         case SEND_CANCELED:
 599         case SEND_FAILED:
 600                 /* real cleanup will be done from tl_clear.  just update flags
 601                  * so it is no longer marked as on the worker queue */
 602                 mod_rq_state(req, m, RQ_NET_QUEUED, 0);
 603                 break;
 604
 605         case HANDED_OVER_TO_NETWORK:
 606                 /* assert something? */
 607                 if (bio_data_dir(req->master_bio) == WRITE &&
 608                     !(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK))) {
 609                         /* this is what is dangerous about protocol A:
 610                          * pretend it was successfully written on the peer. */
 611                         if (req->rq_state & RQ_NET_PENDING)
 612                                 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
 613                         /* else: neg-ack was faster... */
 614                         /* it is still not yet RQ_NET_DONE until the
 615                          * corresponding epoch barrier got acked as well,
 616                          * so we know what to dirty on connection loss */
 617                 }
 618                 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT);
 619                 break;
 620
 621         case OOS_HANDED_TO_NETWORK:
 622                 /* Was not set PENDING, no longer QUEUED, so is now DONE
 623                  * as far as this connection is concerned. */
 624                 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
 625                 break;
 626
 627         case CONNECTION_LOST_WHILE_PENDING:
 628                 /* transfer log cleanup after connection loss */
 629                 mod_rq_state(req, m,
 630                                 RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
 631                                 RQ_NET_DONE);
 632                 break;
 633
 634         case CONFLICT_RESOLVED:
 635                 /* for superseded conflicting writes of multiple primaries,
 636                  * there is no need to keep anything in the tl, potential
 637                  * node crashes are covered by the activity log.
 638                  *
 639                  * If this request had been marked as RQ_POSTPONED before,
 640                  * it will actually not be completed, but "restarted",
 641                  * resubmitted from the retry worker context. */
 642                 D_ASSERT(req->rq_state & RQ_NET_PENDING);
 643                 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
 644                 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
 645                 break;
 646
 647         case WRITE_ACKED_BY_PEER_AND_SIS:
 648                 req->rq_state |= RQ_NET_SIS;
 649         case WRITE_ACKED_BY_PEER:
 650                 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
 651                 /* protocol C; successfully written on peer.
 652                  * Nothing more to do here.
 653                  * We want to keep the tl in place for all protocols, to cater
 654                  * for volatile write-back caches on lower level devices. */
 655
 656                 goto ack_common;
 657         case RECV_ACKED_BY_PEER:
 658                 D_ASSERT(req->rq_state & RQ_EXP_RECEIVE_ACK);
 659                 /* protocol B; pretends to be successfully written on peer.
 660                  * see also notes above in HANDED_OVER_TO_NETWORK about
 661                  * protocol != C */
 662         ack_common:
 663                 D_ASSERT(req->rq_state & RQ_NET_PENDING);
 664                 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
 665                 break;
 666
 667         case POSTPONE_WRITE:
 668                 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
 669                 /* If this node has already detected the write conflict, the
 670                  * worker will be waiting on misc_wait.  Wake it up once this
 671                  * request has completed locally.
 672                  */
 673                 D_ASSERT(req->rq_state & RQ_NET_PENDING);
 674                 req->rq_state |= RQ_POSTPONED;
 675                 if (req->i.waiting)
 676                         wake_up(&mdev->misc_wait);
 677                 /* Do not clear RQ_NET_PENDING. This request will make further
 678                  * progress via restart_conflicting_writes() or
 679                  * fail_postponed_requests(). Hopefully. */
 680                 break;
 681
 682         case NEG_ACKED:
 683                 mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0);
 684                 break;
 685
 686         case FAIL_FROZEN_DISK_IO:
 687                 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
 688                         break;
 689                 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
 690                 break;
 691
 692         case RESTART_FROZEN_DISK_IO:
 693                 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
 694                         break;
 695
 696                 mod_rq_state(req, m,
 697                                 RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
 698                                 RQ_LOCAL_PENDING);
 699
 700                 rv = MR_READ;
 701                 if (bio_data_dir(req->master_bio) == WRITE)
 702                         rv = MR_WRITE;
 703
 704                 get_ldev(mdev); /* always succeeds in this call path */
 705                 req->w.cb = w_restart_disk_io;
 706                 drbd_queue_work(&mdev->tconn->sender_work, &req->w);
 707                 break;
 708
 709         case RESEND:
 710                 /* Simply complete (local only) READs. */
 711                 if (!(req->rq_state & RQ_WRITE) && !req->w.cb) {
 712                         mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
 713                         break;
 714                 }
 715
 716                 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
 717                    before the connection loss (B&C only); only P_BARRIER_ACK
 718                    (or the local completion?) was missing when we suspended.
 719                    Throwing them out of the TL here by pretending we got a BARRIER_ACK.
 720                    During connection handshake, we ensure that the peer was not rebooted. */
 721                 if (!(req->rq_state & RQ_NET_OK)) {
 722                         /* FIXME could this possibly be a req->w.cb == w_send_out_of_sync?
 723                          * in that case we must not set RQ_NET_PENDING. */
 724
 725                         mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING);
 726                         if (req->w.cb) {
 727                                 drbd_queue_work(&mdev->tconn->sender_work, &req->w);
 728                                 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
 729                         } /* else: FIXME can this happen? */
 730                         break;
 731                 }
 732                 /* else, fall through to BARRIER_ACKED */
 733
 734         case BARRIER_ACKED:
 735                 /* barrier ack for READ requests does not make sense */
 736                 if (!(req->rq_state & RQ_WRITE))
 737                         break;
 738
 739                 if (req->rq_state & RQ_NET_PENDING) {
 740                         /* barrier came in before all requests were acked.
 741                          * this is bad, because if the connection is lost now,
 742                          * we won't be able to clean them up... */
 743                         dev_err(DEV, "FIXME (BARRIER_ACKED but pending)\n");
 744                 }
 745                 /* Allowed to complete requests, even while suspended.
 746                  * As this is called for all requests within a matching epoch,
 747                  * we need to filter, and only set RQ_NET_DONE for those that
 748                  * have actually been on the wire. */
 749                 mod_rq_state(req, m, RQ_COMPLETION_SUSP,
 750                                 (req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0);
 751                 break;
 752
 753         case DATA_RECEIVED:
 754                 D_ASSERT(req->rq_state & RQ_NET_PENDING);
 755                 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
 756                 break;
 757
 758         case QUEUE_AS_DRBD_BARRIER:
 759                 start_new_tl_epoch(mdev->tconn);
 760                 mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE);
 761                 break;
 762         };
 763
 764         return rv;
 765 }
 766
 767 /* we may do a local read if:
 768  * - we are consistent (of course),
 769  * - or we are generally inconsistent,
 770  *   BUT we are still/already IN SYNC for this area.
 771  *   since size may be bigger than BM_BLOCK_SIZE,
 772  *   we may need to check several bits.
 773  */
 774 static bool drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size)
 775 {
 776         unsigned long sbnr, ebnr;
 777         sector_t esector, nr_sectors;
 778
 779         if (mdev->state.disk == D_UP_TO_DATE)
 780                 return true;
 781         if (mdev->state.disk != D_INCONSISTENT)
 782                 return false;
 783         esector = sector + (size >> 9) - 1;
 784         nr_sectors = drbd_get_capacity(mdev->this_bdev);
 785         D_ASSERT(sector  < nr_sectors);
 786         D_ASSERT(esector < nr_sectors);
 787
 788         sbnr = BM_SECT_TO_BIT(sector);
 789         ebnr = BM_SECT_TO_BIT(esector);
 790
 791         return drbd_bm_count_bits(mdev, sbnr, ebnr) == 0;
 792 }
 793
 794 static bool remote_due_to_read_balancing(struct drbd_conf *mdev, sector_t sector,
 795                 enum drbd_read_balancing rbm)
 796 {
 797         struct backing_dev_info *bdi;
 798         int stripe_shift;
 799
 800         switch (rbm) {
 801         case RB_CONGESTED_REMOTE:
 802                 bdi = &mdev->ldev->backing_bdev->bd_disk->queue->backing_dev_info;
 803                 return bdi_read_congested(bdi);
 804         case RB_LEAST_PENDING:
 805                 return atomic_read(&mdev->local_cnt) >
 806                         atomic_read(&mdev->ap_pending_cnt) + atomic_read(&mdev->rs_pending_cnt);
 807         case RB_32K_STRIPING:  /* stripe_shift = 15 */
 808         case RB_64K_STRIPING:
 809         case RB_128K_STRIPING:
 810         case RB_256K_STRIPING:
 811         case RB_512K_STRIPING:
 812         case RB_1M_STRIPING:   /* stripe_shift = 20 */
 813                 stripe_shift = (rbm - RB_32K_STRIPING + 15);
 814                 return (sector >> (stripe_shift - 9)) & 1;
 815         case RB_ROUND_ROBIN:
 816                 return test_and_change_bit(READ_BALANCE_RR, &mdev->flags);
 817         case RB_PREFER_REMOTE:
 818                 return true;
 819         case RB_PREFER_LOCAL:
 820         default:
 821                 return false;
 822         }
 823 }
 824
 825 /*
 826  * complete_conflicting_writes  -  wait for any conflicting write requests
 827  *
 828  * The write_requests tree contains all active write requests which we
 829  * currently know about.  Wait for any requests to complete which conflict with
 830  * the new one.
 831  *
 832  * Only way out: remove the conflicting intervals from the tree.
 833  */
 834 static void complete_conflicting_writes(struct drbd_request *req)
 835 {
 836         DEFINE_WAIT(wait);
 837         struct drbd_conf *mdev = req->w.mdev;
 838         struct drbd_interval *i;
 839         sector_t sector = req->i.sector;
 840         int size = req->i.size;
 841
 842         i = drbd_find_overlap(&mdev->write_requests, sector, size);
 843         if (!i)
 844                 return;
 845
 846         for (;;) {
 847                 prepare_to_wait(&mdev->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
 848                 i = drbd_find_overlap(&mdev->write_requests, sector, size);
 849                 if (!i)
 850                         break;
 851                 /* Indicate to wake up device->misc_wait on progress.  */
 852                 i->waiting = true;
 853                 spin_unlock_irq(&mdev->tconn->req_lock);
 854                 schedule();
 855                 spin_lock_irq(&mdev->tconn->req_lock);
 856         }
 857         finish_wait(&mdev->misc_wait, &wait);
 858 }
 859
 860 /* called within req_lock and rcu_read_lock() */
 861 static void maybe_pull_ahead(struct drbd_conf *mdev)
 862 {
 863         struct drbd_tconn *tconn = mdev->tconn;
 864         struct net_conf *nc;
 865         bool congested = false;
 866         enum drbd_on_congestion on_congestion;
 867
 868         rcu_read_lock();
 869         nc = rcu_dereference(tconn->net_conf);
 870         on_congestion = nc ? nc->on_congestion : OC_BLOCK;
 871         rcu_read_unlock();
 872         if (on_congestion == OC_BLOCK ||
 873             tconn->agreed_pro_version < 96)
 874                 return;
 875
 876         /* If I don't even have good local storage, we can not reasonably try
 877          * to pull ahead of the peer. We also need the local reference to make
 878          * sure mdev->act_log is there.
 879          */
 880         if (!get_ldev_if_state(mdev, D_UP_TO_DATE))
 881                 return;
 882
 883         if (nc->cong_fill &&
 884             atomic_read(&mdev->ap_in_flight) >= nc->cong_fill) {
 885                 dev_info(DEV, "Congestion-fill threshold reached\n");
 886                 congested = true;
 887         }
 888
 889         if (mdev->act_log->used >= nc->cong_extents) {
 890                 dev_info(DEV, "Congestion-extents threshold reached\n");
 891                 congested = true;
 892         }
 893
 894         if (congested) {
 895                 /* start a new epoch for non-mirrored writes */
 896                 start_new_tl_epoch(mdev->tconn);
 897
 898                 if (on_congestion == OC_PULL_AHEAD)
 899                         _drbd_set_state(_NS(mdev, conn, C_AHEAD), 0, NULL);
 900                 else  /*nc->on_congestion == OC_DISCONNECT */
 901                         _drbd_set_state(_NS(mdev, conn, C_DISCONNECTING), 0, NULL);
 902         }
 903         put_ldev(mdev);
 904 }
 905
 906 /* If this returns false, and req->private_bio is still set,
 907  * this should be submitted locally.
 908  *
 909  * If it returns false, but req->private_bio is not set,
 910  * we do not have access to good data :(
 911  *
 912  * Otherwise, this destroys req->private_bio, if any,
 913  * and returns true.
 914  */
 915 static bool do_remote_read(struct drbd_request *req)
 916 {
 917         struct drbd_conf *mdev = req->w.mdev;
 918         enum drbd_read_balancing rbm;
 919
 920         if (req->private_bio) {
 921                 if (!drbd_may_do_local_read(mdev,
 922                                         req->i.sector, req->i.size)) {
 923                         bio_put(req->private_bio);
 924                         req->private_bio = NULL;
 925                         put_ldev(mdev);
 926                 }
 927         }
 928
 929         if (mdev->state.pdsk != D_UP_TO_DATE)
 930                 return false;
 931
 932         if (req->private_bio == NULL)
 933                 return true;
 934
 935         /* TODO: improve read balancing decisions, take into account drbd
 936          * protocol, pending requests etc. */
 937
 938         rcu_read_lock();
 939         rbm = rcu_dereference(mdev->ldev->disk_conf)->read_balancing;
 940         rcu_read_unlock();
 941
 942         if (rbm == RB_PREFER_LOCAL && req->private_bio)
 943                 return false; /* submit locally */
 944
 945         if (remote_due_to_read_balancing(mdev, req->i.sector, rbm)) {
 946                 if (req->private_bio) {
 947                         bio_put(req->private_bio);
 948                         req->private_bio = NULL;
 949                         put_ldev(mdev);
 950                 }
 951                 return true;
 952         }
 953
 954         return false;
 955 }
 956
 957 /* returns number of connections (== 1, for drbd 8.4)
 958  * expected to actually write this data,
 959  * which does NOT include those that we are L_AHEAD for. */
 960 static int drbd_process_write_request(struct drbd_request *req)
 961 {
 962         struct drbd_conf *mdev = req->w.mdev;
 963         int remote, send_oos;
 964
 965         remote = drbd_should_do_remote(mdev->state);
 966         send_oos = drbd_should_send_out_of_sync(mdev->state);
 967
 968         /* Need to replicate writes.  Unless it is an empty flush,
 969          * which is better mapped to a DRBD P_BARRIER packet,
 970          * also for drbd wire protocol compatibility reasons.
 971          * If this was a flush, just start a new epoch.
 972          * Unless the current epoch was empty anyways, or we are not currently
 973          * replicating, in which case there is no point. */
 974         if (unlikely(req->i.size == 0)) {
 975                 /* The only size==0 bios we expect are empty flushes. */
 976                 D_ASSERT(req->master_bio->bi_rw & REQ_FLUSH);
 977                 if (remote)
 978                         _req_mod(req, QUEUE_AS_DRBD_BARRIER);
 979                 return remote;
 980         }
 981
 982         if (!remote && !send_oos)
 983                 return 0;
 984
 985         D_ASSERT(!(remote && send_oos));
 986
 987         if (remote) {
 988                 _req_mod(req, TO_BE_SENT);
 989                 _req_mod(req, QUEUE_FOR_NET_WRITE);
 990         } else if (drbd_set_out_of_sync(mdev, req->i.sector, req->i.size))
 991                 _req_mod(req, QUEUE_FOR_SEND_OOS);
 992
 993         return remote;
 994 }
 995
 996 static void
 997 drbd_submit_req_private_bio(struct drbd_request *req)
 998 {
 999         struct drbd_conf *mdev = req->w.mdev;
1000         struct bio *bio = req->private_bio;
1001         const int rw = bio_rw(bio);
1002
1003         bio->bi_bdev = mdev->ldev->backing_bdev;
1004
1005         /* State may have changed since we grabbed our reference on the
1006          * ->ldev member. Double check, and short-circuit to endio.
1007          * In case the last activity log transaction failed to get on
1008          * stable storage, and this is a WRITE, we may not even submit
1009          * this bio. */
1010         if (get_ldev(mdev)) {
1011                 if (drbd_insert_fault(mdev,
1012                                       rw == WRITE ? DRBD_FAULT_DT_WR
1013                                     : rw == READ  ? DRBD_FAULT_DT_RD
1014                                     :               DRBD_FAULT_DT_RA))
1015                         bio_endio(bio, -EIO);
1016                 else
1017                         generic_make_request(bio);
1018                 put_ldev(mdev);
1019         } else
1020                 bio_endio(bio, -EIO);
1021 }
1022
1023 static void drbd_queue_write(struct drbd_conf *mdev, struct drbd_request *req)
1024 {
1025         spin_lock(&mdev->submit.lock);
1026         list_add_tail(&req->tl_requests, &mdev->submit.writes);
1027         spin_unlock(&mdev->submit.lock);
1028         queue_work(mdev->submit.wq, &mdev->submit.worker);
1029 }
1030
1031 /* returns the new drbd_request pointer, if the caller is expected to
1032  * drbd_send_and_submit() it (to save latency), or NULL if we queued the
1033  * request on the submitter thread.
1034  * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request.
1035  */
1036 struct drbd_request *
1037 drbd_request_prepare(struct drbd_conf *mdev, struct bio *bio, unsigned long start_time)
1038 {
1039         const int rw = bio_data_dir(bio);
1040         struct drbd_request *req;
1041
1042         /* allocate outside of all locks; */
1043         req = drbd_req_new(mdev, bio);
1044         if (!req) {
1045                 dec_ap_bio(mdev);
1046                 /* only pass the error to the upper layers.
1047                  * if user cannot handle io errors, that's not our business. */
1048                 dev_err(DEV, "could not kmalloc() req\n");
1049                 bio_endio(bio, -ENOMEM);
1050                 return ERR_PTR(-ENOMEM);
1051         }
1052         req->start_time = start_time;
1053
1054         if (!get_ldev(mdev)) {
1055                 bio_put(req->private_bio);
1056                 req->private_bio = NULL;
1057         }
1058
1059         /* Update disk stats */
1060         _drbd_start_io_acct(mdev, req);
1061
1062         if (rw == WRITE && req->private_bio && req->i.size
1063         && !test_bit(AL_SUSPENDED, &mdev->flags)) {
1064                 if (!drbd_al_begin_io_fastpath(mdev, &req->i)) {
1065                         drbd_queue_write(mdev, req);
1066                         return NULL;
1067                 }
1068                 req->rq_state |= RQ_IN_ACT_LOG;
1069         }
1070
1071         return req;
1072 }
1073
1074 static void drbd_send_and_submit(struct drbd_conf *mdev, struct drbd_request *req)
1075 {
1076         const int rw = bio_rw(req->master_bio);
1077         struct bio_and_error m = { NULL, };
1078         bool no_remote = false;
1079
1080         spin_lock_irq(&mdev->tconn->req_lock);
1081         if (rw == WRITE) {
1082                 /* This may temporarily give up the req_lock,
1083                  * but will re-aquire it before it returns here.
1084                  * Needs to be before the check on drbd_suspended() */
1085                 complete_conflicting_writes(req);
1086                 /* no more giving up req_lock from now on! */
1087
1088                 /* check for congestion, and potentially stop sending
1089                  * full data updates, but start sending "dirty bits" only. */
1090                 maybe_pull_ahead(mdev);
1091         }
1092
1093
1094         if (drbd_suspended(mdev)) {
1095                 /* push back and retry: */
1096                 req->rq_state |= RQ_POSTPONED;
1097                 if (req->private_bio) {
1098                         bio_put(req->private_bio);
1099                         req->private_bio = NULL;
1100                         put_ldev(mdev);
1101                 }
1102                 goto out;
1103         }
1104
1105         /* We fail READ/READA early, if we can not serve it.
1106          * We must do this before req is registered on any lists.
1107          * Otherwise, drbd_req_complete() will queue failed READ for retry. */
1108         if (rw != WRITE) {
1109                 if (!do_remote_read(req) && !req->private_bio)
1110                         goto nodata;
1111         }
1112
1113         /* which transfer log epoch does this belong to? */
1114         req->epoch = atomic_read(&mdev->tconn->current_tle_nr);
1115
1116         /* no point in adding empty flushes to the transfer log,
1117          * they are mapped to drbd barriers already. */
1118         if (likely(req->i.size!=0)) {
1119                 if (rw == WRITE)
1120                         mdev->tconn->current_tle_writes++;
1121
1122                 list_add_tail(&req->tl_requests, &mdev->tconn->transfer_log);
1123         }
1124
1125         if (rw == WRITE) {
1126                 if (!drbd_process_write_request(req))
1127                         no_remote = true;
1128         } else {
1129                 /* We either have a private_bio, or we can read from remote.
1130                  * Otherwise we had done the goto nodata above. */
1131                 if (req->private_bio == NULL) {
1132                         _req_mod(req, TO_BE_SENT);
1133                         _req_mod(req, QUEUE_FOR_NET_READ);
1134                 } else
1135                         no_remote = true;
1136         }
1137
1138         if (req->private_bio) {
1139                 /* needs to be marked within the same spinlock */
1140                 _req_mod(req, TO_BE_SUBMITTED);
1141                 /* but we need to give up the spinlock to submit */
1142                 spin_unlock_irq(&mdev->tconn->req_lock);
1143                 drbd_submit_req_private_bio(req);
1144                 spin_lock_irq(&mdev->tconn->req_lock);
1145         } else if (no_remote) {
1146 nodata:
1147                 if (__ratelimit(&drbd_ratelimit_state))
1148                         dev_err(DEV, "IO ERROR: neither local nor remote data, sector %llu+%u\n",
1149                                         (unsigned long long)req->i.sector, req->i.size >> 9);
1150                 /* A write may have been queued for send_oos, however.
1151                  * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
1152         }
1153
1154 out:
1155         if (drbd_req_put_completion_ref(req, &m, 1))
1156                 kref_put(&req->kref, drbd_req_destroy);
1157         spin_unlock_irq(&mdev->tconn->req_lock);
1158
1159         if (m.bio)
1160                 complete_master_bio(mdev, &m);
1161 }
1162
1163 void __drbd_make_request(struct drbd_conf *mdev, struct bio *bio, unsigned long start_time)
1164 {
1165         struct drbd_request *req = drbd_request_prepare(mdev, bio, start_time);
1166         if (IS_ERR_OR_NULL(req))
1167                 return;
1168         drbd_send_and_submit(mdev, req);
1169 }
1170
1171 static void submit_fast_path(struct drbd_conf *mdev, struct list_head *incoming)
1172 {
1173         struct drbd_request *req, *tmp;
1174         list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1175                 const int rw = bio_data_dir(req->master_bio);
1176
1177                 if (rw == WRITE /* rw != WRITE should not even end up here! */
1178                 && req->private_bio && req->i.size
1179                 && !test_bit(AL_SUSPENDED, &mdev->flags)) {
1180                         if (!drbd_al_begin_io_fastpath(mdev, &req->i))
1181                                 continue;
1182
1183                         req->rq_state |= RQ_IN_ACT_LOG;
1184                 }
1185
1186                 list_del_init(&req->tl_requests);
1187                 drbd_send_and_submit(mdev, req);
1188         }
1189 }
1190
1191 static bool prepare_al_transaction_nonblock(struct drbd_conf *mdev,
1192                                             struct list_head *incoming,
1193                                             struct list_head *pending)
1194 {
1195         struct drbd_request *req, *tmp;
1196         int wake = 0;
1197         int err;
1198
1199         spin_lock_irq(&mdev->al_lock);
1200         list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1201                 err = drbd_al_begin_io_nonblock(mdev, &req->i);
1202                 if (err == -EBUSY)
1203                         wake = 1;
1204                 if (err)
1205                         continue;
1206                 req->rq_state |= RQ_IN_ACT_LOG;
1207                 list_move_tail(&req->tl_requests, pending);
1208         }
1209         spin_unlock_irq(&mdev->al_lock);
1210         if (wake)
1211                 wake_up(&mdev->al_wait);
1212
1213         return !list_empty(pending);
1214 }
1215
1216 void do_submit(struct work_struct *ws)
1217 {
1218         struct drbd_conf *mdev = container_of(ws, struct drbd_conf, submit.worker);
1219         LIST_HEAD(incoming);
1220         LIST_HEAD(pending);
1221         struct drbd_request *req, *tmp;
1222
1223         for (;;) {
1224                 spin_lock(&mdev->submit.lock);
1225                 list_splice_tail_init(&mdev->submit.writes, &incoming);
1226                 spin_unlock(&mdev->submit.lock);
1227
1228                 submit_fast_path(mdev, &incoming);
1229                 if (list_empty(&incoming))
1230                         break;
1231
1232                 wait_event(mdev->al_wait, prepare_al_transaction_nonblock(mdev, &incoming, &pending));
1233                 /* Maybe more was queued, while we prepared the transaction?
1234                  * Try to stuff them into this transaction as well.
1235                  * Be strictly non-blocking here, no wait_event, we already
1236                  * have something to commit.
1237                  * Stop if we don't make any more progres.
1238                  */
1239                 for (;;) {
1240                         LIST_HEAD(more_pending);
1241                         LIST_HEAD(more_incoming);
1242                         bool made_progress;
1243
1244                         /* It is ok to look outside the lock,
1245                          * it's only an optimization anyways */
1246                         if (list_empty(&mdev->submit.writes))
1247                                 break;
1248
1249                         spin_lock(&mdev->submit.lock);
1250                         list_splice_tail_init(&mdev->submit.writes, &more_incoming);
1251                         spin_unlock(&mdev->submit.lock);
1252
1253                         if (list_empty(&more_incoming))
1254                                 break;
1255
1256                         made_progress = prepare_al_transaction_nonblock(mdev, &more_incoming, &more_pending);
1257
1258                         list_splice_tail_init(&more_pending, &pending);
1259                         list_splice_tail_init(&more_incoming, &incoming);
1260
1261                         if (!made_progress)
1262                                 break;
1263                 }
1264                 drbd_al_begin_io_commit(mdev, false);
1265
1266                 list_for_each_entry_safe(req, tmp, &pending, tl_requests) {
1267                         list_del_init(&req->tl_requests);
1268                         drbd_send_and_submit(mdev, req);
1269                 }
1270         }
1271 }
1272
1273 void drbd_make_request(struct request_queue *q, struct bio *bio)
1274 {
1275         struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1276         unsigned long start_time;
1277
1278         start_time = jiffies;
1279
1280         /*
1281          * what we "blindly" assume:
1282          */
1283         D_ASSERT(IS_ALIGNED(bio->bi_size, 512));
1284
1285         inc_ap_bio(mdev);
1286         __drbd_make_request(mdev, bio, start_time);
1287 }
1288
1289 /* This is called by bio_add_page().
1290  *
1291  * q->max_hw_sectors and other global limits are already enforced there.
1292  *
1293  * We need to call down to our lower level device,
1294  * in case it has special restrictions.
1295  *
1296  * We also may need to enforce configured max-bio-bvecs limits.
1297  *
1298  * As long as the BIO is empty we have to allow at least one bvec,
1299  * regardless of size and offset, so no need to ask lower levels.
1300  */
1301 int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec)
1302 {
1303         struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1304         unsigned int bio_size = bvm->bi_size;
1305         int limit = DRBD_MAX_BIO_SIZE;
1306         int backing_limit;
1307
1308         if (bio_size && get_ldev(mdev)) {
1309                 unsigned int max_hw_sectors = queue_max_hw_sectors(q);
1310                 struct request_queue * const b =
1311                         mdev->ldev->backing_bdev->bd_disk->queue;
1312                 if (b->merge_bvec_fn) {
1313                         backing_limit = b->merge_bvec_fn(b, bvm, bvec);
1314                         limit = min(limit, backing_limit);
1315                 }
1316                 put_ldev(mdev);
1317                 if ((limit >> 9) > max_hw_sectors)
1318                         limit = max_hw_sectors << 9;
1319         }
1320         return limit;
1321 }
1322
1323 struct drbd_request *find_oldest_request(struct drbd_tconn *tconn)
1324 {
1325         /* Walk the transfer log,
1326          * and find the oldest not yet completed request */
1327         struct drbd_request *r;
1328         list_for_each_entry(r, &tconn->transfer_log, tl_requests) {
1329                 if (atomic_read(&r->completion_ref))
1330                         return r;
1331         }
1332         return NULL;
1333 }
1334
1335 void request_timer_fn(unsigned long data)
1336 {
1337         struct drbd_conf *mdev = (struct drbd_conf *) data;
1338         struct drbd_tconn *tconn = mdev->tconn;
1339         struct drbd_request *req; /* oldest request */
1340         struct net_conf *nc;
1341         unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1342         unsigned long now;
1343
1344         rcu_read_lock();
1345         nc = rcu_dereference(tconn->net_conf);
1346         if (nc && mdev->state.conn >= C_WF_REPORT_PARAMS)
1347                 ent = nc->timeout * HZ/10 * nc->ko_count;
1348
1349         if (get_ldev(mdev)) { /* implicit state.disk >= D_INCONSISTENT */
1350                 dt = rcu_dereference(mdev->ldev->disk_conf)->disk_timeout * HZ / 10;
1351                 put_ldev(mdev);
1352         }
1353         rcu_read_unlock();
1354
1355         et = min_not_zero(dt, ent);
1356
1357         if (!et)
1358                 return; /* Recurring timer stopped */
1359
1360         now = jiffies;
1361
1362         spin_lock_irq(&tconn->req_lock);
1363         req = find_oldest_request(tconn);
1364         if (!req) {
1365                 spin_unlock_irq(&tconn->req_lock);
1366                 mod_timer(&mdev->request_timer, now + et);
1367                 return;
1368         }
1369
1370         /* The request is considered timed out, if
1371          * - we have some effective timeout from the configuration,
1372          *   with above state restrictions applied,
1373          * - the oldest request is waiting for a response from the network
1374          *   resp. the local disk,
1375          * - the oldest request is in fact older than the effective timeout,
1376          * - the connection was established (resp. disk was attached)
1377          *   for longer than the timeout already.
1378          * Note that for 32bit jiffies and very stable connections/disks,
1379          * we may have a wrap around, which is catched by
1380          *   !time_in_range(now, last_..._jif, last_..._jif + timeout).
1381          *
1382          * Side effect: once per 32bit wrap-around interval, which means every
1383          * ~198 days with 250 HZ, we have a window where the timeout would need
1384          * to expire twice (worst case) to become effective. Good enough.
1385          */
1386         if (ent && req->rq_state & RQ_NET_PENDING &&
1387                  time_after(now, req->start_time + ent) &&
1388                 !time_in_range(now, tconn->last_reconnect_jif, tconn->last_reconnect_jif + ent)) {
1389                 dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n");
1390                 _drbd_set_state(_NS(mdev, conn, C_TIMEOUT), CS_VERBOSE | CS_HARD, NULL);
1391         }
1392         if (dt && req->rq_state & RQ_LOCAL_PENDING && req->w.mdev == mdev &&
1393                  time_after(now, req->start_time + dt) &&
1394                 !time_in_range(now, mdev->last_reattach_jif, mdev->last_reattach_jif + dt)) {
1395                 dev_warn(DEV, "Local backing device failed to meet the disk-timeout\n");
1396                 __drbd_chk_io_error(mdev, DRBD_FORCE_DETACH);
1397         }
1398         nt = (time_after(now, req->start_time + et) ? now : req->start_time) + et;
1399         spin_unlock_irq(&tconn->req_lock);
1400         mod_timer(&mdev->request_timer, nt);
1401 }