linux/drivers/block/drbd/drbd_req.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3   drbd_req.c
   4
   5   This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
   6
   7   Copyright (C) 2001-2008, LINBIT Information Technologies GmbH.
   8   Copyright (C) 1999-2008, Philipp Reisner <philipp.reisner@linbit.com>.
   9   Copyright (C) 2002-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
  10
  11
  12 */
  13
  14#include <linux/module.h>
  15
  16#include <linux/slab.h>
  17#include <linux/drbd.h>
  18#include "drbd_int.h"
  19#include "drbd_req.h"
  20
  21
  22static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size);
  23
  24static struct drbd_request *drbd_req_new(struct drbd_device *device, struct bio *bio_src)
  25{
  26        struct drbd_request *req;
  27
  28        req = mempool_alloc(&drbd_request_mempool, GFP_NOIO);
  29        if (!req)
  30                return NULL;
  31        memset(req, 0, sizeof(*req));
  32
  33        req->private_bio = bio_clone_fast(bio_src, GFP_NOIO, &drbd_io_bio_set);
  34        req->private_bio->bi_private = req;
  35        req->private_bio->bi_end_io = drbd_request_endio;
  36
  37        req->rq_state = (bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0)
  38                      | (bio_op(bio_src) == REQ_OP_WRITE_SAME ? RQ_WSAME : 0)
  39                      | (bio_op(bio_src) == REQ_OP_WRITE_ZEROES ? RQ_ZEROES : 0)
  40                      | (bio_op(bio_src) == REQ_OP_DISCARD ? RQ_UNMAP : 0);
  41        req->device = device;
  42        req->master_bio = bio_src;
  43        req->epoch = 0;
  44
  45        drbd_clear_interval(&req->i);
  46        req->i.sector     = bio_src->bi_iter.bi_sector;
  47        req->i.size      = bio_src->bi_iter.bi_size;
  48        req->i.local = true;
  49        req->i.waiting = false;
  50
  51        INIT_LIST_HEAD(&req->tl_requests);
  52        INIT_LIST_HEAD(&req->w.list);
  53        INIT_LIST_HEAD(&req->req_pending_master_completion);
  54        INIT_LIST_HEAD(&req->req_pending_local);
  55
  56        /* one reference to be put by __drbd_make_request */
  57        atomic_set(&req->completion_ref, 1);
  58        /* one kref as long as completion_ref > 0 */
  59        kref_init(&req->kref);
  60        return req;
  61}
  62
  63static void drbd_remove_request_interval(struct rb_root *root,
  64                                         struct drbd_request *req)
  65{
  66        struct drbd_device *device = req->device;
  67        struct drbd_interval *i = &req->i;
  68
  69        drbd_remove_interval(root, i);
  70
  71        /* Wake up any processes waiting for this request to complete.  */
  72        if (i->waiting)
  73                wake_up(&device->misc_wait);
  74}
  75
  76void drbd_req_destroy(struct kref *kref)
  77{
  78        struct drbd_request *req = container_of(kref, struct drbd_request, kref);
  79        struct drbd_device *device = req->device;
  80        const unsigned s = req->rq_state;
  81
  82        if ((req->master_bio && !(s & RQ_POSTPONED)) ||
  83                atomic_read(&req->completion_ref) ||
  84                (s & RQ_LOCAL_PENDING) ||
  85                ((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) {
  86                drbd_err(device, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n",
  87                                s, atomic_read(&req->completion_ref));
  88                return;
  89        }
  90
  91        /* If called from mod_rq_state (expected normal case) or
  92         * drbd_send_and_submit (the less likely normal path), this holds the
  93         * req_lock, and req->tl_requests will typicaly be on ->transfer_log,
  94         * though it may be still empty (never added to the transfer log).
  95         *
  96         * If called from do_retry(), we do NOT hold the req_lock, but we are
  97         * still allowed to unconditionally list_del(&req->tl_requests),
  98         * because it will be on a local on-stack list only. */
  99        list_del_init(&req->tl_requests);
 100
 101        /* finally remove the request from the conflict detection
 102         * respective block_id verification interval tree. */
 103        if (!drbd_interval_empty(&req->i)) {
 104                struct rb_root *root;
 105
 106                if (s & RQ_WRITE)
 107                        root = &device->write_requests;
 108                else
 109                        root = &device->read_requests;
 110                drbd_remove_request_interval(root, req);
 111        } else if (s & (RQ_NET_MASK & ~RQ_NET_DONE) && req->i.size != 0)
 112                drbd_err(device, "drbd_req_destroy: Logic BUG: interval empty, but: rq_state=0x%x, sect=%llu, size=%u\n",
 113                        s, (unsigned long long)req->i.sector, req->i.size);
 114
 115        /* if it was a write, we may have to set the corresponding
 116         * bit(s) out-of-sync first. If it had a local part, we need to
 117         * release the reference to the activity log. */
 118        if (s & RQ_WRITE) {
 119                /* Set out-of-sync unless both OK flags are set
 120                 * (local only or remote failed).
 121                 * Other places where we set out-of-sync:
 122                 * READ with local io-error */
 123
 124                /* There is a special case:
 125                 * we may notice late that IO was suspended,
 126                 * and postpone, or schedule for retry, a write,
 127                 * before it even was submitted or sent.
 128                 * In that case we do not want to touch the bitmap at all.
 129                 */
 130                if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) {
 131                        if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
 132                                drbd_set_out_of_sync(device, req->i.sector, req->i.size);
 133
 134                        if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
 135                                drbd_set_in_sync(device, req->i.sector, req->i.size);
 136                }
 137
 138                /* one might be tempted to move the drbd_al_complete_io
 139                 * to the local io completion callback drbd_request_endio.
 140                 * but, if this was a mirror write, we may only
 141                 * drbd_al_complete_io after this is RQ_NET_DONE,
 142                 * otherwise the extent could be dropped from the al
 143                 * before it has actually been written on the peer.
 144                 * if we crash before our peer knows about the request,
 145                 * but after the extent has been dropped from the al,
 146                 * we would forget to resync the corresponding extent.
 147                 */
 148                if (s & RQ_IN_ACT_LOG) {
 149                        if (get_ldev_if_state(device, D_FAILED)) {
 150                                drbd_al_complete_io(device, &req->i);
 151                                put_ldev(device);
 152                        } else if (__ratelimit(&drbd_ratelimit_state)) {
 153                                drbd_warn(device, "Should have called drbd_al_complete_io(, %llu, %u), "
 154                                         "but my Disk seems to have failed :(\n",
 155                                         (unsigned long long) req->i.sector, req->i.size);
 156                        }
 157                }
 158        }
 159
 160        mempool_free(req, &drbd_request_mempool);
 161}
 162
 163static void wake_all_senders(struct drbd_connection *connection)
 164{
 165        wake_up(&connection->sender_work.q_wait);
 166}
 167
 168/* must hold resource->req_lock */
 169void start_new_tl_epoch(struct drbd_connection *connection)
 170{
 171        /* no point closing an epoch, if it is empty, anyways. */
 172        if (connection->current_tle_writes == 0)
 173                return;
 174
 175        connection->current_tle_writes = 0;
 176        atomic_inc(&connection->current_tle_nr);
 177        wake_all_senders(connection);
 178}
 179
 180void complete_master_bio(struct drbd_device *device,
 181                struct bio_and_error *m)
 182{
 183        m->bio->bi_status = errno_to_blk_status(m->error);
 184        bio_endio(m->bio);
 185        dec_ap_bio(device);
 186}
 187
 188
 189/* Helper for __req_mod().
 190 * Set m->bio to the master bio, if it is fit to be completed,
 191 * or leave it alone (it is initialized to NULL in __req_mod),
 192 * if it has already been completed, or cannot be completed yet.
 193 * If m->bio is set, the error status to be returned is placed in m->error.
 194 */
 195static
 196void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
 197{
 198        const unsigned s = req->rq_state;
 199        struct drbd_device *device = req->device;
 200        int error, ok;
 201
 202        /* we must not complete the master bio, while it is
 203         *      still being processed by _drbd_send_zc_bio (drbd_send_dblock)
 204         *      not yet acknowledged by the peer
 205         *      not yet completed by the local io subsystem
 206         * these flags may get cleared in any order by
 207         *      the worker,
 208         *      the receiver,
 209         *      the bio_endio completion callbacks.
 210         */
 211        if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
 212            (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
 213            (s & RQ_COMPLETION_SUSP)) {
 214                drbd_err(device, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s);
 215                return;
 216        }
 217
 218        if (!req->master_bio) {
 219                drbd_err(device, "drbd_req_complete: Logic BUG, master_bio == NULL!\n");
 220                return;
 221        }
 222
 223        /*
 224         * figure out whether to report success or failure.
 225         *
 226         * report success when at least one of the operations succeeded.
 227         * or, to put the other way,
 228         * only report failure, when both operations failed.
 229         *
 230         * what to do about the failures is handled elsewhere.
 231         * what we need to do here is just: complete the master_bio.
 232         *
 233         * local completion error, if any, has been stored as ERR_PTR
 234         * in private_bio within drbd_request_endio.
 235         */
 236        ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
 237        error = PTR_ERR(req->private_bio);
 238
 239        /* Before we can signal completion to the upper layers,
 240         * we may need to close the current transfer log epoch.
 241         * We are within the request lock, so we can simply compare
 242         * the request epoch number with the current transfer log
 243         * epoch number.  If they match, increase the current_tle_nr,
 244         * and reset the transfer log epoch write_cnt.
 245         */
 246        if (op_is_write(bio_op(req->master_bio)) &&
 247            req->epoch == atomic_read(&first_peer_device(device)->connection->current_tle_nr))
 248                start_new_tl_epoch(first_peer_device(device)->connection);
 249
 250        /* Update disk stats */
 251        bio_end_io_acct(req->master_bio, req->start_jif);
 252
 253        /* If READ failed,
 254         * have it be pushed back to the retry work queue,
 255         * so it will re-enter __drbd_make_request(),
 256         * and be re-assigned to a suitable local or remote path,
 257         * or failed if we do not have access to good data anymore.
 258         *
 259         * Unless it was failed early by __drbd_make_request(),
 260         * because no path was available, in which case
 261         * it was not even added to the transfer_log.
 262         *
 263         * read-ahead may fail, and will not be retried.
 264         *
 265         * WRITE should have used all available paths already.
 266         */
 267        if (!ok &&
 268            bio_op(req->master_bio) == REQ_OP_READ &&
 269            !(req->master_bio->bi_opf & REQ_RAHEAD) &&
 270            !list_empty(&req->tl_requests))
 271                req->rq_state |= RQ_POSTPONED;
 272
 273        if (!(req->rq_state & RQ_POSTPONED)) {
 274                m->error = ok ? 0 : (error ?: -EIO);
 275                m->bio = req->master_bio;
 276                req->master_bio = NULL;
 277                /* We leave it in the tree, to be able to verify later
 278                 * write-acks in protocol != C during resync.
 279                 * But we mark it as "complete", so it won't be counted as
 280                 * conflict in a multi-primary setup. */
 281                req->i.completed = true;
 282        }
 283
 284        if (req->i.waiting)
 285                wake_up(&device->misc_wait);
 286
 287        /* Either we are about to complete to upper layers,
 288         * or we will restart this request.
 289         * In either case, the request object will be destroyed soon,
 290         * so better remove it from all lists. */
 291        list_del_init(&req->req_pending_master_completion);
 292}
 293
 294/* still holds resource->req_lock */
 295static void drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
 296{
 297        struct drbd_device *device = req->device;
 298        D_ASSERT(device, m || (req->rq_state & RQ_POSTPONED));
 299
 300        if (!put)
 301                return;
 302
 303        if (!atomic_sub_and_test(put, &req->completion_ref))
 304                return;
 305
 306        drbd_req_complete(req, m);
 307
 308        /* local completion may still come in later,
 309         * we need to keep the req object around. */
 310        if (req->rq_state & RQ_LOCAL_ABORTED)
 311                return;
 312
 313        if (req->rq_state & RQ_POSTPONED) {
 314                /* don't destroy the req object just yet,
 315                 * but queue it for retry */
 316                drbd_restart_request(req);
 317                return;
 318        }
 319
 320        kref_put(&req->kref, drbd_req_destroy);
 321}
 322
 323static void set_if_null_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
 324{
 325        struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 326        if (!connection)
 327                return;
 328        if (connection->req_next == NULL)
 329                connection->req_next = req;
 330}
 331
 332static void advance_conn_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
 333{
 334        struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 335        if (!connection)
 336                return;
 337        if (connection->req_next != req)
 338                return;
 339        list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
 340                const unsigned s = req->rq_state;
 341                if (s & RQ_NET_QUEUED)
 342                        break;
 343        }
 344        if (&req->tl_requests == &connection->transfer_log)
 345                req = NULL;
 346        connection->req_next = req;
 347}
 348
 349static void set_if_null_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
 350{
 351        struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 352        if (!connection)
 353                return;
 354        if (connection->req_ack_pending == NULL)
 355                connection->req_ack_pending = req;
 356}
 357
 358static void advance_conn_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
 359{
 360        struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 361        if (!connection)
 362                return;
 363        if (connection->req_ack_pending != req)
 364                return;
 365        list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
 366                const unsigned s = req->rq_state;
 367                if ((s & RQ_NET_SENT) && (s & RQ_NET_PENDING))
 368                        break;
 369        }
 370        if (&req->tl_requests == &connection->transfer_log)
 371                req = NULL;
 372        connection->req_ack_pending = req;
 373}
 374
 375static void set_if_null_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
 376{
 377        struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 378        if (!connection)
 379                return;
 380        if (connection->req_not_net_done == NULL)
 381                connection->req_not_net_done = req;
 382}
 383
 384static void advance_conn_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
 385{
 386        struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 387        if (!connection)
 388                return;
 389        if (connection->req_not_net_done != req)
 390                return;
 391        list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
 392                const unsigned s = req->rq_state;
 393                if ((s & RQ_NET_SENT) && !(s & RQ_NET_DONE))
 394                        break;
 395        }
 396        if (&req->tl_requests == &connection->transfer_log)
 397                req = NULL;
 398        connection->req_not_net_done = req;
 399}
 400
 401/* I'd like this to be the only place that manipulates
 402 * req->completion_ref and req->kref. */
 403static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m,
 404                int clear, int set)
 405{
 406        struct drbd_device *device = req->device;
 407        struct drbd_peer_device *peer_device = first_peer_device(device);
 408        unsigned s = req->rq_state;
 409        int c_put = 0;
 410
 411        if (drbd_suspended(device) && !((s | clear) & RQ_COMPLETION_SUSP))
 412                set |= RQ_COMPLETION_SUSP;
 413
 414        /* apply */
 415
 416        req->rq_state &= ~clear;
 417        req->rq_state |= set;
 418
 419        /* no change? */
 420        if (req->rq_state == s)
 421                return;
 422
 423        /* intent: get references */
 424
 425        kref_get(&req->kref);
 426
 427        if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING))
 428                atomic_inc(&req->completion_ref);
 429
 430        if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) {
 431                inc_ap_pending(device);
 432                atomic_inc(&req->completion_ref);
 433        }
 434
 435        if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED)) {
 436                atomic_inc(&req->completion_ref);
 437                set_if_null_req_next(peer_device, req);
 438        }
 439
 440        if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK))
 441                kref_get(&req->kref); /* wait for the DONE */
 442
 443        if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT)) {
 444                /* potentially already completed in the ack_receiver thread */
 445                if (!(s & RQ_NET_DONE)) {
 446                        atomic_add(req->i.size >> 9, &device->ap_in_flight);
 447                        set_if_null_req_not_net_done(peer_device, req);
 448                }
 449                if (req->rq_state & RQ_NET_PENDING)
 450                        set_if_null_req_ack_pending(peer_device, req);
 451        }
 452
 453        if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP))
 454                atomic_inc(&req->completion_ref);
 455
 456        /* progress: put references */
 457
 458        if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP))
 459                ++c_put;
 460
 461        if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) {
 462                D_ASSERT(device, req->rq_state & RQ_LOCAL_PENDING);
 463                ++c_put;
 464        }
 465
 466        if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) {
 467                if (req->rq_state & RQ_LOCAL_ABORTED)
 468                        kref_put(&req->kref, drbd_req_destroy);
 469                else
 470                        ++c_put;
 471                list_del_init(&req->req_pending_local);
 472        }
 473
 474        if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) {
 475                dec_ap_pending(device);
 476                ++c_put;
 477                req->acked_jif = jiffies;
 478                advance_conn_req_ack_pending(peer_device, req);
 479        }
 480
 481        if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED)) {
 482                ++c_put;
 483                advance_conn_req_next(peer_device, req);
 484        }
 485
 486        if (!(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) {
 487                if (s & RQ_NET_SENT)
 488                        atomic_sub(req->i.size >> 9, &device->ap_in_flight);
 489                if (s & RQ_EXP_BARR_ACK)
 490                        kref_put(&req->kref, drbd_req_destroy);
 491                req->net_done_jif = jiffies;
 492
 493                /* in ahead/behind mode, or just in case,
 494                 * before we finally destroy this request,
 495                 * the caching pointers must not reference it anymore */
 496                advance_conn_req_next(peer_device, req);
 497                advance_conn_req_ack_pending(peer_device, req);
 498                advance_conn_req_not_net_done(peer_device, req);
 499        }
 500
 501        /* potentially complete and destroy */
 502
 503        /* If we made progress, retry conflicting peer requests, if any. */
 504        if (req->i.waiting)
 505                wake_up(&device->misc_wait);
 506
 507        drbd_req_put_completion_ref(req, m, c_put);
 508        kref_put(&req->kref, drbd_req_destroy);
 509}
 510
 511static void drbd_report_io_error(struct drbd_device *device, struct drbd_request *req)
 512{
 513        char b[BDEVNAME_SIZE];
 514
 515        if (!__ratelimit(&drbd_ratelimit_state))
 516                return;
 517
 518        drbd_warn(device, "local %s IO error sector %llu+%u on %s\n",
 519                        (req->rq_state & RQ_WRITE) ? "WRITE" : "READ",
 520                        (unsigned long long)req->i.sector,
 521                        req->i.size >> 9,
 522                        bdevname(device->ldev->backing_bdev, b));
 523}
 524
 525/* Helper for HANDED_OVER_TO_NETWORK.
 526 * Is this a protocol A write (neither WRITE_ACK nor RECEIVE_ACK expected)?
 527 * Is it also still "PENDING"?
 528 * --> If so, clear PENDING and set NET_OK below.
 529 * If it is a protocol A write, but not RQ_PENDING anymore, neg-ack was faster
 530 * (and we must not set RQ_NET_OK) */
 531static inline bool is_pending_write_protocol_A(struct drbd_request *req)
 532{
 533        return (req->rq_state &
 534                   (RQ_WRITE|RQ_NET_PENDING|RQ_EXP_WRITE_ACK|RQ_EXP_RECEIVE_ACK))
 535                == (RQ_WRITE|RQ_NET_PENDING);
 536}
 537
 538/* obviously this could be coded as many single functions
 539 * instead of one huge switch,
 540 * or by putting the code directly in the respective locations
 541 * (as it has been before).
 542 *
 543 * but having it this way
 544 *  enforces that it is all in this one place, where it is easier to audit,
 545 *  it makes it obvious that whatever "event" "happens" to a request should
 546 *  happen "atomically" within the req_lock,
 547 *  and it enforces that we have to think in a very structured manner
 548 *  about the "events" that may happen to a request during its life time ...
 549 */
 550int __req_mod(struct drbd_request *req, enum drbd_req_event what,
 551                struct bio_and_error *m)
 552{
 553        struct drbd_device *const device = req->device;
 554        struct drbd_peer_device *const peer_device = first_peer_device(device);
 555        struct drbd_connection *const connection = peer_device ? peer_device->connection : NULL;
 556        struct net_conf *nc;
 557        int p, rv = 0;
 558
 559        if (m)
 560                m->bio = NULL;
 561
 562        switch (what) {
 563        default:
 564                drbd_err(device, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
 565                break;
 566
 567        /* does not happen...
 568         * initialization done in drbd_req_new
 569        case CREATED:
 570                break;
 571                */
 572
 573        case TO_BE_SENT: /* via network */
 574                /* reached via __drbd_make_request
 575                 * and from w_read_retry_remote */
 576                D_ASSERT(device, !(req->rq_state & RQ_NET_MASK));
 577                rcu_read_lock();
 578                nc = rcu_dereference(connection->net_conf);
 579                p = nc->wire_protocol;
 580                rcu_read_unlock();
 581                req->rq_state |=
 582                        p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
 583                        p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
 584                mod_rq_state(req, m, 0, RQ_NET_PENDING);
 585                break;
 586
 587        case TO_BE_SUBMITTED: /* locally */
 588                /* reached via __drbd_make_request */
 589                D_ASSERT(device, !(req->rq_state & RQ_LOCAL_MASK));
 590                mod_rq_state(req, m, 0, RQ_LOCAL_PENDING);
 591                break;
 592
 593        case COMPLETED_OK:
 594                if (req->rq_state & RQ_WRITE)
 595                        device->writ_cnt += req->i.size >> 9;
 596                else
 597                        device->read_cnt += req->i.size >> 9;
 598
 599                mod_rq_state(req, m, RQ_LOCAL_PENDING,
 600                                RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
 601                break;
 602
 603        case ABORT_DISK_IO:
 604                mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED);
 605                break;
 606
 607        case WRITE_COMPLETED_WITH_ERROR:
 608                drbd_report_io_error(device, req);
 609                __drbd_chk_io_error(device, DRBD_WRITE_ERROR);
 610                mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
 611                break;
 612
 613        case READ_COMPLETED_WITH_ERROR:
 614                drbd_set_out_of_sync(device, req->i.sector, req->i.size);
 615                drbd_report_io_error(device, req);
 616                __drbd_chk_io_error(device, DRBD_READ_ERROR);
 617                fallthrough;
 618        case READ_AHEAD_COMPLETED_WITH_ERROR:
 619                /* it is legal to fail read-ahead, no __drbd_chk_io_error in that case. */
 620                mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
 621                break;
 622
 623        case DISCARD_COMPLETED_NOTSUPP:
 624        case DISCARD_COMPLETED_WITH_ERROR:
 625                /* I'd rather not detach from local disk just because it
 626                 * failed a REQ_OP_DISCARD. */
 627                mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
 628                break;
 629
 630        case QUEUE_FOR_NET_READ:
 631                /* READ, and
 632                 * no local disk,
 633                 * or target area marked as invalid,
 634                 * or just got an io-error. */
 635                /* from __drbd_make_request
 636                 * or from bio_endio during read io-error recovery */
 637
 638                /* So we can verify the handle in the answer packet.
 639                 * Corresponding drbd_remove_request_interval is in
 640                 * drbd_req_complete() */
 641                D_ASSERT(device, drbd_interval_empty(&req->i));
 642                drbd_insert_interval(&device->read_requests, &req->i);
 643
 644                set_bit(UNPLUG_REMOTE, &device->flags);
 645
 646                D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 647                D_ASSERT(device, (req->rq_state & RQ_LOCAL_MASK) == 0);
 648                mod_rq_state(req, m, 0, RQ_NET_QUEUED);
 649                req->w.cb = w_send_read_req;
 650                drbd_queue_work(&connection->sender_work,
 651                                &req->w);
 652                break;
 653
 654        case QUEUE_FOR_NET_WRITE:
 655                /* assert something? */
 656                /* from __drbd_make_request only */
 657
 658                /* Corresponding drbd_remove_request_interval is in
 659                 * drbd_req_complete() */
 660                D_ASSERT(device, drbd_interval_empty(&req->i));
 661                drbd_insert_interval(&device->write_requests, &req->i);
 662
 663                /* NOTE
 664                 * In case the req ended up on the transfer log before being
 665                 * queued on the worker, it could lead to this request being
 666                 * missed during cleanup after connection loss.
 667                 * So we have to do both operations here,
 668                 * within the same lock that protects the transfer log.
 669                 *
 670                 * _req_add_to_epoch(req); this has to be after the
 671                 * _maybe_start_new_epoch(req); which happened in
 672                 * __drbd_make_request, because we now may set the bit
 673                 * again ourselves to close the current epoch.
 674                 *
 675                 * Add req to the (now) current epoch (barrier). */
 676
 677                /* otherwise we may lose an unplug, which may cause some remote
 678                 * io-scheduler timeout to expire, increasing maximum latency,
 679                 * hurting performance. */
 680                set_bit(UNPLUG_REMOTE, &device->flags);
 681
 682                /* queue work item to send data */
 683                D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 684                mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
 685                req->w.cb =  w_send_dblock;
 686                drbd_queue_work(&connection->sender_work,
 687                                &req->w);
 688
 689                /* close the epoch, in case it outgrew the limit */
 690                rcu_read_lock();
 691                nc = rcu_dereference(connection->net_conf);
 692                p = nc->max_epoch_size;
 693                rcu_read_unlock();
 694                if (connection->current_tle_writes >= p)
 695                        start_new_tl_epoch(connection);
 696
 697                break;
 698
 699        case QUEUE_FOR_SEND_OOS:
 700                mod_rq_state(req, m, 0, RQ_NET_QUEUED);
 701                req->w.cb =  w_send_out_of_sync;
 702                drbd_queue_work(&connection->sender_work,
 703                                &req->w);
 704                break;
 705
 706        case READ_RETRY_REMOTE_CANCELED:
 707        case SEND_CANCELED:
 708        case SEND_FAILED:
 709                /* real cleanup will be done from tl_clear.  just update flags
 710                 * so it is no longer marked as on the worker queue */
 711                mod_rq_state(req, m, RQ_NET_QUEUED, 0);
 712                break;
 713
 714        case HANDED_OVER_TO_NETWORK:
 715                /* assert something? */
 716                if (is_pending_write_protocol_A(req))
 717                        /* this is what is dangerous about protocol A:
 718                         * pretend it was successfully written on the peer. */
 719                        mod_rq_state(req, m, RQ_NET_QUEUED|RQ_NET_PENDING,
 720                                                RQ_NET_SENT|RQ_NET_OK);
 721                else
 722                        mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT);
 723                /* It is still not yet RQ_NET_DONE until the
 724                 * corresponding epoch barrier got acked as well,
 725                 * so we know what to dirty on connection loss. */
 726                break;
 727
 728        case OOS_HANDED_TO_NETWORK:
 729                /* Was not set PENDING, no longer QUEUED, so is now DONE
 730                 * as far as this connection is concerned. */
 731                mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
 732                break;
 733
 734        case CONNECTION_LOST_WHILE_PENDING:
 735                /* transfer log cleanup after connection loss */
 736                mod_rq_state(req, m,
 737                                RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
 738                                RQ_NET_DONE);
 739                break;
 740
 741        case CONFLICT_RESOLVED:
 742                /* for superseded conflicting writes of multiple primaries,
 743                 * there is no need to keep anything in the tl, potential
 744                 * node crashes are covered by the activity log.
 745                 *
 746                 * If this request had been marked as RQ_POSTPONED before,
 747                 * it will actually not be completed, but "restarted",
 748                 * resubmitted from the retry worker context. */
 749                D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 750                D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
 751                mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
 752                break;
 753
 754        case WRITE_ACKED_BY_PEER_AND_SIS:
 755                req->rq_state |= RQ_NET_SIS;
 756                fallthrough;
 757        case WRITE_ACKED_BY_PEER:
 758                /* Normal operation protocol C: successfully written on peer.
 759                 * During resync, even in protocol != C,
 760                 * we requested an explicit write ack anyways.
 761                 * Which means we cannot even assert anything here.
 762                 * Nothing more to do here.
 763                 * We want to keep the tl in place for all protocols, to cater
 764                 * for volatile write-back caches on lower level devices. */
 765                goto ack_common;
 766        case RECV_ACKED_BY_PEER:
 767                D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK);
 768                /* protocol B; pretends to be successfully written on peer.
 769                 * see also notes above in HANDED_OVER_TO_NETWORK about
 770                 * protocol != C */
 771        ack_common:
 772                mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
 773                break;
 774
 775        case POSTPONE_WRITE:
 776                D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
 777                /* If this node has already detected the write conflict, the
 778                 * worker will be waiting on misc_wait.  Wake it up once this
 779                 * request has completed locally.
 780                 */
 781                D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 782                req->rq_state |= RQ_POSTPONED;
 783                if (req->i.waiting)
 784                        wake_up(&device->misc_wait);
 785                /* Do not clear RQ_NET_PENDING. This request will make further
 786                 * progress via restart_conflicting_writes() or
 787                 * fail_postponed_requests(). Hopefully. */
 788                break;
 789
 790        case NEG_ACKED:
 791                mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0);
 792                break;
 793
 794        case FAIL_FROZEN_DISK_IO:
 795                if (!(req->rq_state & RQ_LOCAL_COMPLETED))
 796                        break;
 797                mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
 798                break;
 799
 800        case RESTART_FROZEN_DISK_IO:
 801                if (!(req->rq_state & RQ_LOCAL_COMPLETED))
 802                        break;
 803
 804                mod_rq_state(req, m,
 805                                RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
 806                                RQ_LOCAL_PENDING);
 807
 808                rv = MR_READ;
 809                if (bio_data_dir(req->master_bio) == WRITE)
 810                        rv = MR_WRITE;
 811
 812                get_ldev(device); /* always succeeds in this call path */
 813                req->w.cb = w_restart_disk_io;
 814                drbd_queue_work(&connection->sender_work,
 815                                &req->w);
 816                break;
 817
 818        case RESEND:
 819                /* Simply complete (local only) READs. */
 820                if (!(req->rq_state & RQ_WRITE) && !req->w.cb) {
 821                        mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
 822                        break;
 823                }
 824
 825                /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
 826                   before the connection loss (B&C only); only P_BARRIER_ACK
 827                   (or the local completion?) was missing when we suspended.
 828                   Throwing them out of the TL here by pretending we got a BARRIER_ACK.
 829                   During connection handshake, we ensure that the peer was not rebooted. */
 830                if (!(req->rq_state & RQ_NET_OK)) {
 831                        /* FIXME could this possibly be a req->dw.cb == w_send_out_of_sync?
 832                         * in that case we must not set RQ_NET_PENDING. */
 833
 834                        mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING);
 835                        if (req->w.cb) {
 836                                /* w.cb expected to be w_send_dblock, or w_send_read_req */
 837                                drbd_queue_work(&connection->sender_work,
 838                                                &req->w);
 839                                rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
 840                        } /* else: FIXME can this happen? */
 841                        break;
 842                }
 843                fallthrough;    /* to BARRIER_ACKED */
 844
 845        case BARRIER_ACKED:
 846                /* barrier ack for READ requests does not make sense */
 847                if (!(req->rq_state & RQ_WRITE))
 848                        break;
 849
 850                if (req->rq_state & RQ_NET_PENDING) {
 851                        /* barrier came in before all requests were acked.
 852                         * this is bad, because if the connection is lost now,
 853                         * we won't be able to clean them up... */
 854                        drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n");
 855                }
 856                /* Allowed to complete requests, even while suspended.
 857                 * As this is called for all requests within a matching epoch,
 858                 * we need to filter, and only set RQ_NET_DONE for those that
 859                 * have actually been on the wire. */
 860                mod_rq_state(req, m, RQ_COMPLETION_SUSP,
 861                                (req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0);
 862                break;
 863
 864        case DATA_RECEIVED:
 865                D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 866                mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
 867                break;
 868
 869        case QUEUE_AS_DRBD_BARRIER:
 870                start_new_tl_epoch(connection);
 871                mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE);
 872                break;
 873        }
 874
 875        return rv;
 876}
 877
 878/* we may do a local read if:
 879 * - we are consistent (of course),
 880 * - or we are generally inconsistent,
 881 *   BUT we are still/already IN SYNC for this area.
 882 *   since size may be bigger than BM_BLOCK_SIZE,
 883 *   we may need to check several bits.
 884 */
 885static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size)
 886{
 887        unsigned long sbnr, ebnr;
 888        sector_t esector, nr_sectors;
 889
 890        if (device->state.disk == D_UP_TO_DATE)
 891                return true;
 892        if (device->state.disk != D_INCONSISTENT)
 893                return false;
 894        esector = sector + (size >> 9) - 1;
 895        nr_sectors = get_capacity(device->vdisk);
 896        D_ASSERT(device, sector  < nr_sectors);
 897        D_ASSERT(device, esector < nr_sectors);
 898
 899        sbnr = BM_SECT_TO_BIT(sector);
 900        ebnr = BM_SECT_TO_BIT(esector);
 901
 902        return drbd_bm_count_bits(device, sbnr, ebnr) == 0;
 903}
 904
 905static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector,
 906                enum drbd_read_balancing rbm)
 907{
 908        int stripe_shift;
 909
 910        switch (rbm) {
 911        case RB_CONGESTED_REMOTE:
 912                return bdi_read_congested(
 913                        device->ldev->backing_bdev->bd_disk->bdi);
 914        case RB_LEAST_PENDING:
 915                return atomic_read(&device->local_cnt) >
 916                        atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt);
 917        case RB_32K_STRIPING:  /* stripe_shift = 15 */
 918        case RB_64K_STRIPING:
 919        case RB_128K_STRIPING:
 920        case RB_256K_STRIPING:
 921        case RB_512K_STRIPING:
 922        case RB_1M_STRIPING:   /* stripe_shift = 20 */
 923                stripe_shift = (rbm - RB_32K_STRIPING + 15);
 924                return (sector >> (stripe_shift - 9)) & 1;
 925        case RB_ROUND_ROBIN:
 926                return test_and_change_bit(READ_BALANCE_RR, &device->flags);
 927        case RB_PREFER_REMOTE:
 928                return true;
 929        case RB_PREFER_LOCAL:
 930        default:
 931                return false;
 932        }
 933}
 934
 935/*
 936 * complete_conflicting_writes  -  wait for any conflicting write requests
 937 *
 938 * The write_requests tree contains all active write requests which we
 939 * currently know about.  Wait for any requests to complete which conflict with
 940 * the new one.
 941 *
 942 * Only way out: remove the conflicting intervals from the tree.
 943 */
 944static void complete_conflicting_writes(struct drbd_request *req)
 945{
 946        DEFINE_WAIT(wait);
 947        struct drbd_device *device = req->device;
 948        struct drbd_interval *i;
 949        sector_t sector = req->i.sector;
 950        int size = req->i.size;
 951
 952        for (;;) {
 953                drbd_for_each_overlap(i, &device->write_requests, sector, size) {
 954                        /* Ignore, if already completed to upper layers. */
 955                        if (i->completed)
 956                                continue;
 957                        /* Handle the first found overlap.  After the schedule
 958                         * we have to restart the tree walk. */
 959                        break;
 960                }
 961                if (!i) /* if any */
 962                        break;
 963
 964                /* Indicate to wake up device->misc_wait on progress.  */
 965                prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
 966                i->waiting = true;
 967                spin_unlock_irq(&device->resource->req_lock);
 968                schedule();
 969                spin_lock_irq(&device->resource->req_lock);
 970        }
 971        finish_wait(&device->misc_wait, &wait);
 972}
 973
 974/* called within req_lock */
 975static void maybe_pull_ahead(struct drbd_device *device)
 976{
 977        struct drbd_connection *connection = first_peer_device(device)->connection;
 978        struct net_conf *nc;
 979        bool congested = false;
 980        enum drbd_on_congestion on_congestion;
 981
 982        rcu_read_lock();
 983        nc = rcu_dereference(connection->net_conf);
 984        on_congestion = nc ? nc->on_congestion : OC_BLOCK;
 985        rcu_read_unlock();
 986        if (on_congestion == OC_BLOCK ||
 987            connection->agreed_pro_version < 96)
 988                return;
 989
 990        if (on_congestion == OC_PULL_AHEAD && device->state.conn == C_AHEAD)
 991                return; /* nothing to do ... */
 992
 993        /* If I don't even have good local storage, we can not reasonably try
 994         * to pull ahead of the peer. We also need the local reference to make
 995         * sure device->act_log is there.
 996         */
 997        if (!get_ldev_if_state(device, D_UP_TO_DATE))
 998                return;
 999
1000        if (nc->cong_fill &&
1001            atomic_read(&device->ap_in_flight) >= nc->cong_fill) {
1002                drbd_info(device, "Congestion-fill threshold reached\n");
1003                congested = true;
1004        }
1005
1006        if (device->act_log->used >= nc->cong_extents) {
1007                drbd_info(device, "Congestion-extents threshold reached\n");
1008                congested = true;
1009        }
1010
1011        if (congested) {
1012                /* start a new epoch for non-mirrored writes */
1013                start_new_tl_epoch(first_peer_device(device)->connection);
1014
1015                if (on_congestion == OC_PULL_AHEAD)
1016                        _drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL);
1017                else  /*nc->on_congestion == OC_DISCONNECT */
1018                        _drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL);
1019        }
1020        put_ldev(device);
1021}
1022
1023/* If this returns false, and req->private_bio is still set,
1024 * this should be submitted locally.
1025 *
1026 * If it returns false, but req->private_bio is not set,
1027 * we do not have access to good data :(
1028 *
1029 * Otherwise, this destroys req->private_bio, if any,
1030 * and returns true.
1031 */
1032static bool do_remote_read(struct drbd_request *req)
1033{
1034        struct drbd_device *device = req->device;
1035        enum drbd_read_balancing rbm;
1036
1037        if (req->private_bio) {
1038                if (!drbd_may_do_local_read(device,
1039                                        req->i.sector, req->i.size)) {
1040                        bio_put(req->private_bio);
1041                        req->private_bio = NULL;
1042                        put_ldev(device);
1043                }
1044        }
1045
1046        if (device->state.pdsk != D_UP_TO_DATE)
1047                return false;
1048
1049        if (req->private_bio == NULL)
1050                return true;
1051
1052        /* TODO: improve read balancing decisions, take into account drbd
1053         * protocol, pending requests etc. */
1054
1055        rcu_read_lock();
1056        rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing;
1057        rcu_read_unlock();
1058
1059        if (rbm == RB_PREFER_LOCAL && req->private_bio)
1060                return false; /* submit locally */
1061
1062        if (remote_due_to_read_balancing(device, req->i.sector, rbm)) {
1063                if (req->private_bio) {
1064                        bio_put(req->private_bio);
1065                        req->private_bio = NULL;
1066                        put_ldev(device);
1067                }
1068                return true;
1069        }
1070
1071        return false;
1072}
1073
1074bool drbd_should_do_remote(union drbd_dev_state s)
1075{
1076        return s.pdsk == D_UP_TO_DATE ||
1077                (s.pdsk >= D_INCONSISTENT &&
1078                 s.conn >= C_WF_BITMAP_T &&
1079                 s.conn < C_AHEAD);
1080        /* Before proto 96 that was >= CONNECTED instead of >= C_WF_BITMAP_T.
1081           That is equivalent since before 96 IO was frozen in the C_WF_BITMAP*
1082           states. */
1083}
1084
1085static bool drbd_should_send_out_of_sync(union drbd_dev_state s)
1086{
1087        return s.conn == C_AHEAD || s.conn == C_WF_BITMAP_S;
1088        /* pdsk = D_INCONSISTENT as a consequence. Protocol 96 check not necessary
1089           since we enter state C_AHEAD only if proto >= 96 */
1090}
1091
1092/* returns number of connections (== 1, for drbd 8.4)
1093 * expected to actually write this data,
1094 * which does NOT include those that we are L_AHEAD for. */
1095static int drbd_process_write_request(struct drbd_request *req)
1096{
1097        struct drbd_device *device = req->device;
1098        int remote, send_oos;
1099
1100        remote = drbd_should_do_remote(device->state);
1101        send_oos = drbd_should_send_out_of_sync(device->state);
1102
1103        /* Need to replicate writes.  Unless it is an empty flush,
1104         * which is better mapped to a DRBD P_BARRIER packet,
1105         * also for drbd wire protocol compatibility reasons.
1106         * If this was a flush, just start a new epoch.
1107         * Unless the current epoch was empty anyways, or we are not currently
1108         * replicating, in which case there is no point. */
1109        if (unlikely(req->i.size == 0)) {
1110                /* The only size==0 bios we expect are empty flushes. */
1111                D_ASSERT(device, req->master_bio->bi_opf & REQ_PREFLUSH);
1112                if (remote)
1113                        _req_mod(req, QUEUE_AS_DRBD_BARRIER);
1114                return remote;
1115        }
1116
1117        if (!remote && !send_oos)
1118                return 0;
1119
1120        D_ASSERT(device, !(remote && send_oos));
1121
1122        if (remote) {
1123                _req_mod(req, TO_BE_SENT);
1124                _req_mod(req, QUEUE_FOR_NET_WRITE);
1125        } else if (drbd_set_out_of_sync(device, req->i.sector, req->i.size))
1126                _req_mod(req, QUEUE_FOR_SEND_OOS);
1127
1128        return remote;
1129}
1130
1131static void drbd_process_discard_or_zeroes_req(struct drbd_request *req, int flags)
1132{
1133        int err = drbd_issue_discard_or_zero_out(req->device,
1134                                req->i.sector, req->i.size >> 9, flags);
1135        if (err)
1136                req->private_bio->bi_status = BLK_STS_IOERR;
1137        bio_endio(req->private_bio);
1138}
1139
1140static void
1141drbd_submit_req_private_bio(struct drbd_request *req)
1142{
1143        struct drbd_device *device = req->device;
1144        struct bio *bio = req->private_bio;
1145        unsigned int type;
1146
1147        if (bio_op(bio) != REQ_OP_READ)
1148                type = DRBD_FAULT_DT_WR;
1149        else if (bio->bi_opf & REQ_RAHEAD)
1150                type = DRBD_FAULT_DT_RA;
1151        else
1152                type = DRBD_FAULT_DT_RD;
1153
1154        bio_set_dev(bio, device->ldev->backing_bdev);
1155
1156        /* State may have changed since we grabbed our reference on the
1157         * ->ldev member. Double check, and short-circuit to endio.
1158         * In case the last activity log transaction failed to get on
1159         * stable storage, and this is a WRITE, we may not even submit
1160         * this bio. */
1161        if (get_ldev(device)) {
1162                if (drbd_insert_fault(device, type))
1163                        bio_io_error(bio);
1164                else if (bio_op(bio) == REQ_OP_WRITE_ZEROES)
1165                        drbd_process_discard_or_zeroes_req(req, EE_ZEROOUT |
1166                            ((bio->bi_opf & REQ_NOUNMAP) ? 0 : EE_TRIM));
1167                else if (bio_op(bio) == REQ_OP_DISCARD)
1168                        drbd_process_discard_or_zeroes_req(req, EE_TRIM);
1169                else
1170                        submit_bio_noacct(bio);
1171                put_ldev(device);
1172        } else
1173                bio_io_error(bio);
1174}
1175
1176static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req)
1177{
1178        spin_lock_irq(&device->resource->req_lock);
1179        list_add_tail(&req->tl_requests, &device->submit.writes);
1180        list_add_tail(&req->req_pending_master_completion,
1181                        &device->pending_master_completion[1 /* WRITE */]);
1182        spin_unlock_irq(&device->resource->req_lock);
1183        queue_work(device->submit.wq, &device->submit.worker);
1184        /* do_submit() may sleep internally on al_wait, too */
1185        wake_up(&device->al_wait);
1186}
1187
1188/* returns the new drbd_request pointer, if the caller is expected to
1189 * drbd_send_and_submit() it (to save latency), or NULL if we queued the
1190 * request on the submitter thread.
1191 * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request.
1192 */
1193static struct drbd_request *
1194drbd_request_prepare(struct drbd_device *device, struct bio *bio)
1195{
1196        const int rw = bio_data_dir(bio);
1197        struct drbd_request *req;
1198
1199        /* allocate outside of all locks; */
1200        req = drbd_req_new(device, bio);
1201        if (!req) {
1202                dec_ap_bio(device);
1203                /* only pass the error to the upper layers.
1204                 * if user cannot handle io errors, that's not our business. */
1205                drbd_err(device, "could not kmalloc() req\n");
1206                bio->bi_status = BLK_STS_RESOURCE;
1207                bio_endio(bio);
1208                return ERR_PTR(-ENOMEM);
1209        }
1210
1211        /* Update disk stats */
1212        req->start_jif = bio_start_io_acct(req->master_bio);
1213
1214        if (!get_ldev(device)) {
1215                bio_put(req->private_bio);
1216                req->private_bio = NULL;
1217        }
1218
1219        /* process discards always from our submitter thread */
1220        if (bio_op(bio) == REQ_OP_WRITE_ZEROES ||
1221            bio_op(bio) == REQ_OP_DISCARD)
1222                goto queue_for_submitter_thread;
1223
1224        if (rw == WRITE && req->private_bio && req->i.size
1225        && !test_bit(AL_SUSPENDED, &device->flags)) {
1226                if (!drbd_al_begin_io_fastpath(device, &req->i))
1227                        goto queue_for_submitter_thread;
1228                req->rq_state |= RQ_IN_ACT_LOG;
1229                req->in_actlog_jif = jiffies;
1230        }
1231        return req;
1232
1233 queue_for_submitter_thread:
1234        atomic_inc(&device->ap_actlog_cnt);
1235        drbd_queue_write(device, req);
1236        return NULL;
1237}
1238
1239/* Require at least one path to current data.
1240 * We don't want to allow writes on C_STANDALONE D_INCONSISTENT:
1241 * We would not allow to read what was written,
1242 * we would not have bumped the data generation uuids,
1243 * we would cause data divergence for all the wrong reasons.
1244 *
1245 * If we don't see at least one D_UP_TO_DATE, we will fail this request,
1246 * which either returns EIO, or, if OND_SUSPEND_IO is set, suspends IO,
1247 * and queues for retry later.
1248 */
1249static bool may_do_writes(struct drbd_device *device)
1250{
1251        const union drbd_dev_state s = device->state;
1252        return s.disk == D_UP_TO_DATE || s.pdsk == D_UP_TO_DATE;
1253}
1254
1255struct drbd_plug_cb {
1256        struct blk_plug_cb cb;
1257        struct drbd_request *most_recent_req;
1258        /* do we need more? */
1259};
1260
1261static void drbd_unplug(struct blk_plug_cb *cb, bool from_schedule)
1262{
1263        struct drbd_plug_cb *plug = container_of(cb, struct drbd_plug_cb, cb);
1264        struct drbd_resource *resource = plug->cb.data;
1265        struct drbd_request *req = plug->most_recent_req;
1266
1267        kfree(cb);
1268        if (!req)
1269                return;
1270
1271        spin_lock_irq(&resource->req_lock);
1272        /* In case the sender did not process it yet, raise the flag to
1273         * have it followed with P_UNPLUG_REMOTE just after. */
1274        req->rq_state |= RQ_UNPLUG;
1275        /* but also queue a generic unplug */
1276        drbd_queue_unplug(req->device);
1277        kref_put(&req->kref, drbd_req_destroy);
1278        spin_unlock_irq(&resource->req_lock);
1279}
1280
1281static struct drbd_plug_cb* drbd_check_plugged(struct drbd_resource *resource)
1282{
1283        /* A lot of text to say
1284         * return (struct drbd_plug_cb*)blk_check_plugged(); */
1285        struct drbd_plug_cb *plug;
1286        struct blk_plug_cb *cb = blk_check_plugged(drbd_unplug, resource, sizeof(*plug));
1287
1288        if (cb)
1289                plug = container_of(cb, struct drbd_plug_cb, cb);
1290        else
1291                plug = NULL;
1292        return plug;
1293}
1294
1295static void drbd_update_plug(struct drbd_plug_cb *plug, struct drbd_request *req)
1296{
1297        struct drbd_request *tmp = plug->most_recent_req;
1298        /* Will be sent to some peer.
1299         * Remember to tag it with UNPLUG_REMOTE on unplug */
1300        kref_get(&req->kref);
1301        plug->most_recent_req = req;
1302        if (tmp)
1303                kref_put(&tmp->kref, drbd_req_destroy);
1304}
1305
1306static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req)
1307{
1308        struct drbd_resource *resource = device->resource;
1309        const int rw = bio_data_dir(req->master_bio);
1310        struct bio_and_error m = { NULL, };
1311        bool no_remote = false;
1312        bool submit_private_bio = false;
1313
1314        spin_lock_irq(&resource->req_lock);
1315        if (rw == WRITE) {
1316                /* This may temporarily give up the req_lock,
1317                 * but will re-aquire it before it returns here.
1318                 * Needs to be before the check on drbd_suspended() */
1319                complete_conflicting_writes(req);
1320                /* no more giving up req_lock from now on! */
1321
1322                /* check for congestion, and potentially stop sending
1323                 * full data updates, but start sending "dirty bits" only. */
1324                maybe_pull_ahead(device);
1325        }
1326
1327
1328        if (drbd_suspended(device)) {
1329                /* push back and retry: */
1330                req->rq_state |= RQ_POSTPONED;
1331                if (req->private_bio) {
1332                        bio_put(req->private_bio);
1333                        req->private_bio = NULL;
1334                        put_ldev(device);
1335                }
1336                goto out;
1337        }
1338
1339        /* We fail READ early, if we can not serve it.
1340         * We must do this before req is registered on any lists.
1341         * Otherwise, drbd_req_complete() will queue failed READ for retry. */
1342        if (rw != WRITE) {
1343                if (!do_remote_read(req) && !req->private_bio)
1344                        goto nodata;
1345        }
1346
1347        /* which transfer log epoch does this belong to? */
1348        req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr);
1349
1350        /* no point in adding empty flushes to the transfer log,
1351         * they are mapped to drbd barriers already. */
1352        if (likely(req->i.size!=0)) {
1353                if (rw == WRITE)
1354                        first_peer_device(device)->connection->current_tle_writes++;
1355
1356                list_add_tail(&req->tl_requests, &first_peer_device(device)->connection->transfer_log);
1357        }
1358
1359        if (rw == WRITE) {
1360                if (req->private_bio && !may_do_writes(device)) {
1361                        bio_put(req->private_bio);
1362                        req->private_bio = NULL;
1363                        put_ldev(device);
1364                        goto nodata;
1365                }
1366                if (!drbd_process_write_request(req))
1367                        no_remote = true;
1368        } else {
1369                /* We either have a private_bio, or we can read from remote.
1370                 * Otherwise we had done the goto nodata above. */
1371                if (req->private_bio == NULL) {
1372                        _req_mod(req, TO_BE_SENT);
1373                        _req_mod(req, QUEUE_FOR_NET_READ);
1374                } else
1375                        no_remote = true;
1376        }
1377
1378        if (no_remote == false) {
1379                struct drbd_plug_cb *plug = drbd_check_plugged(resource);
1380                if (plug)
1381                        drbd_update_plug(plug, req);
1382        }
1383
1384        /* If it took the fast path in drbd_request_prepare, add it here.
1385         * The slow path has added it already. */
1386        if (list_empty(&req->req_pending_master_completion))
1387                list_add_tail(&req->req_pending_master_completion,
1388                        &device->pending_master_completion[rw == WRITE]);
1389        if (req->private_bio) {
1390                /* needs to be marked within the same spinlock */
1391                req->pre_submit_jif = jiffies;
1392                list_add_tail(&req->req_pending_local,
1393                        &device->pending_completion[rw == WRITE]);
1394                _req_mod(req, TO_BE_SUBMITTED);
1395                /* but we need to give up the spinlock to submit */
1396                submit_private_bio = true;
1397        } else if (no_remote) {
1398nodata:
1399                if (__ratelimit(&drbd_ratelimit_state))
1400                        drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n",
1401                                        (unsigned long long)req->i.sector, req->i.size >> 9);
1402                /* A write may have been queued for send_oos, however.
1403                 * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
1404        }
1405
1406out:
1407        drbd_req_put_completion_ref(req, &m, 1);
1408        spin_unlock_irq(&resource->req_lock);
1409
1410        /* Even though above is a kref_put(), this is safe.
1411         * As long as we still need to submit our private bio,
1412         * we hold a completion ref, and the request cannot disappear.
1413         * If however this request did not even have a private bio to submit
1414         * (e.g. remote read), req may already be invalid now.
1415         * That's why we cannot check on req->private_bio. */
1416        if (submit_private_bio)
1417                drbd_submit_req_private_bio(req);
1418        if (m.bio)
1419                complete_master_bio(device, &m);
1420}
1421
1422void __drbd_make_request(struct drbd_device *device, struct bio *bio)
1423{
1424        struct drbd_request *req = drbd_request_prepare(device, bio);
1425        if (IS_ERR_OR_NULL(req))
1426                return;
1427        drbd_send_and_submit(device, req);
1428}
1429
1430static void submit_fast_path(struct drbd_device *device, struct list_head *incoming)
1431{
1432        struct blk_plug plug;
1433        struct drbd_request *req, *tmp;
1434
1435        blk_start_plug(&plug);
1436        list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1437                const int rw = bio_data_dir(req->master_bio);
1438
1439                if (rw == WRITE /* rw != WRITE should not even end up here! */
1440                && req->private_bio && req->i.size
1441                && !test_bit(AL_SUSPENDED, &device->flags)) {
1442                        if (!drbd_al_begin_io_fastpath(device, &req->i))
1443                                continue;
1444
1445                        req->rq_state |= RQ_IN_ACT_LOG;
1446                        req->in_actlog_jif = jiffies;
1447                        atomic_dec(&device->ap_actlog_cnt);
1448                }
1449
1450                list_del_init(&req->tl_requests);
1451                drbd_send_and_submit(device, req);
1452        }
1453        blk_finish_plug(&plug);
1454}
1455
1456static bool prepare_al_transaction_nonblock(struct drbd_device *device,
1457                                            struct list_head *incoming,
1458                                            struct list_head *pending,
1459                                            struct list_head *later)
1460{
1461        struct drbd_request *req;
1462        int wake = 0;
1463        int err;
1464
1465        spin_lock_irq(&device->al_lock);
1466        while ((req = list_first_entry_or_null(incoming, struct drbd_request, tl_requests))) {
1467                err = drbd_al_begin_io_nonblock(device, &req->i);
1468                if (err == -ENOBUFS)
1469                        break;
1470                if (err == -EBUSY)
1471                        wake = 1;
1472                if (err)
1473                        list_move_tail(&req->tl_requests, later);
1474                else
1475                        list_move_tail(&req->tl_requests, pending);
1476        }
1477        spin_unlock_irq(&device->al_lock);
1478        if (wake)
1479                wake_up(&device->al_wait);
1480        return !list_empty(pending);
1481}
1482
1483static void send_and_submit_pending(struct drbd_device *device, struct list_head *pending)
1484{
1485        struct blk_plug plug;
1486        struct drbd_request *req;
1487
1488        blk_start_plug(&plug);
1489        while ((req = list_first_entry_or_null(pending, struct drbd_request, tl_requests))) {
1490                req->rq_state |= RQ_IN_ACT_LOG;
1491                req->in_actlog_jif = jiffies;
1492                atomic_dec(&device->ap_actlog_cnt);
1493                list_del_init(&req->tl_requests);
1494                drbd_send_and_submit(device, req);
1495        }
1496        blk_finish_plug(&plug);
1497}
1498
1499void do_submit(struct work_struct *ws)
1500{
1501        struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker);
1502        LIST_HEAD(incoming);    /* from drbd_make_request() */
1503        LIST_HEAD(pending);     /* to be submitted after next AL-transaction commit */
1504        LIST_HEAD(busy);        /* blocked by resync requests */
1505
1506        /* grab new incoming requests */
1507        spin_lock_irq(&device->resource->req_lock);
1508        list_splice_tail_init(&device->submit.writes, &incoming);
1509        spin_unlock_irq(&device->resource->req_lock);
1510
1511        for (;;) {
1512                DEFINE_WAIT(wait);
1513
1514                /* move used-to-be-busy back to front of incoming */
1515                list_splice_init(&busy, &incoming);
1516                submit_fast_path(device, &incoming);
1517                if (list_empty(&incoming))
1518                        break;
1519
1520                for (;;) {
1521                        prepare_to_wait(&device->al_wait, &wait, TASK_UNINTERRUPTIBLE);
1522
1523                        list_splice_init(&busy, &incoming);
1524                        prepare_al_transaction_nonblock(device, &incoming, &pending, &busy);
1525                        if (!list_empty(&pending))
1526                                break;
1527
1528                        schedule();
1529
1530                        /* If all currently "hot" activity log extents are kept busy by
1531                         * incoming requests, we still must not totally starve new
1532                         * requests to "cold" extents.
1533                         * Something left on &incoming means there had not been
1534                         * enough update slots available, and the activity log
1535                         * has been marked as "starving".
1536                         *
1537                         * Try again now, without looking for new requests,
1538                         * effectively blocking all new requests until we made
1539                         * at least _some_ progress with what we currently have.
1540                         */
1541                        if (!list_empty(&incoming))
1542                                continue;
1543
1544                        /* Nothing moved to pending, but nothing left
1545                         * on incoming: all moved to busy!
1546                         * Grab new and iterate. */
1547                        spin_lock_irq(&device->resource->req_lock);
1548                        list_splice_tail_init(&device->submit.writes, &incoming);
1549                        spin_unlock_irq(&device->resource->req_lock);
1550                }
1551                finish_wait(&device->al_wait, &wait);
1552
1553                /* If the transaction was full, before all incoming requests
1554                 * had been processed, skip ahead to commit, and iterate
1555                 * without splicing in more incoming requests from upper layers.
1556                 *
1557                 * Else, if all incoming have been processed,
1558                 * they have become either "pending" (to be submitted after
1559                 * next transaction commit) or "busy" (blocked by resync).
1560                 *
1561                 * Maybe more was queued, while we prepared the transaction?
1562                 * Try to stuff those into this transaction as well.
1563                 * Be strictly non-blocking here,
1564                 * we already have something to commit.
1565                 *
1566                 * Commit if we don't make any more progres.
1567                 */
1568
1569                while (list_empty(&incoming)) {
1570                        LIST_HEAD(more_pending);
1571                        LIST_HEAD(more_incoming);
1572                        bool made_progress;
1573
1574                        /* It is ok to look outside the lock,
1575                         * it's only an optimization anyways */
1576                        if (list_empty(&device->submit.writes))
1577                                break;
1578
1579                        spin_lock_irq(&device->resource->req_lock);
1580                        list_splice_tail_init(&device->submit.writes, &more_incoming);
1581                        spin_unlock_irq(&device->resource->req_lock);
1582
1583                        if (list_empty(&more_incoming))
1584                                break;
1585
1586                        made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending, &busy);
1587
1588                        list_splice_tail_init(&more_pending, &pending);
1589                        list_splice_tail_init(&more_incoming, &incoming);
1590                        if (!made_progress)
1591                                break;
1592                }
1593
1594                drbd_al_begin_io_commit(device);
1595                send_and_submit_pending(device, &pending);
1596        }
1597}
1598
1599blk_qc_t drbd_submit_bio(struct bio *bio)
1600{
1601        struct drbd_device *device = bio->bi_bdev->bd_disk->private_data;
1602
1603        blk_queue_split(&bio);
1604
1605        /*
1606         * what we "blindly" assume:
1607         */
1608        D_ASSERT(device, IS_ALIGNED(bio->bi_iter.bi_size, 512));
1609
1610        inc_ap_bio(device);
1611        __drbd_make_request(device, bio);
1612        return BLK_QC_T_NONE;
1613}
1614
1615static bool net_timeout_reached(struct drbd_request *net_req,
1616                struct drbd_connection *connection,
1617                unsigned long now, unsigned long ent,
1618                unsigned int ko_count, unsigned int timeout)
1619{
1620        struct drbd_device *device = net_req->device;
1621
1622        if (!time_after(now, net_req->pre_send_jif + ent))
1623                return false;
1624
1625        if (time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent))
1626                return false;
1627
1628        if (net_req->rq_state & RQ_NET_PENDING) {
1629                drbd_warn(device, "Remote failed to finish a request within %ums > ko-count (%u) * timeout (%u * 0.1s)\n",
1630                        jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout);
1631                return true;
1632        }
1633
1634        /* We received an ACK already (or are using protocol A),
1635         * but are waiting for the epoch closing barrier ack.
1636         * Check if we sent the barrier already.  We should not blame the peer
1637         * for being unresponsive, if we did not even ask it yet. */
1638        if (net_req->epoch == connection->send.current_epoch_nr) {
1639                drbd_warn(device,
1640                        "We did not send a P_BARRIER for %ums > ko-count (%u) * timeout (%u * 0.1s); drbd kernel thread blocked?\n",
1641                        jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout);
1642                return false;
1643        }
1644
1645        /* Worst case: we may have been blocked for whatever reason, then
1646         * suddenly are able to send a lot of requests (and epoch separating
1647         * barriers) in quick succession.
1648         * The timestamp of the net_req may be much too old and not correspond
1649         * to the sending time of the relevant unack'ed barrier packet, so
1650         * would trigger a spurious timeout.  The latest barrier packet may
1651         * have a too recent timestamp to trigger the timeout, potentially miss
1652         * a timeout.  Right now we don't have a place to conveniently store
1653         * these timestamps.
1654         * But in this particular situation, the application requests are still
1655         * completed to upper layers, DRBD should still "feel" responsive.
1656         * No need yet to kill this connection, it may still recover.
1657         * If not, eventually we will have queued enough into the network for
1658         * us to block. From that point of view, the timestamp of the last sent
1659         * barrier packet is relevant enough.
1660         */
1661        if (time_after(now, connection->send.last_sent_barrier_jif + ent)) {
1662                drbd_warn(device, "Remote failed to answer a P_BARRIER (sent at %lu jif; now=%lu jif) within %ums > ko-count (%u) * timeout (%u * 0.1s)\n",
1663                        connection->send.last_sent_barrier_jif, now,
1664                        jiffies_to_msecs(now - connection->send.last_sent_barrier_jif), ko_count, timeout);
1665                return true;
1666        }
1667        return false;
1668}
1669
1670/* A request is considered timed out, if
1671 * - we have some effective timeout from the configuration,
1672 *   with some state restrictions applied,
1673 * - the oldest request is waiting for a response from the network
1674 *   resp. the local disk,
1675 * - the oldest request is in fact older than the effective timeout,
1676 * - the connection was established (resp. disk was attached)
1677 *   for longer than the timeout already.
1678 * Note that for 32bit jiffies and very stable connections/disks,
1679 * we may have a wrap around, which is catched by
1680 *   !time_in_range(now, last_..._jif, last_..._jif + timeout).
1681 *
1682 * Side effect: once per 32bit wrap-around interval, which means every
1683 * ~198 days with 250 HZ, we have a window where the timeout would need
1684 * to expire twice (worst case) to become effective. Good enough.
1685 */
1686
1687void request_timer_fn(struct timer_list *t)
1688{
1689        struct drbd_device *device = from_timer(device, t, request_timer);
1690        struct drbd_connection *connection = first_peer_device(device)->connection;
1691        struct drbd_request *req_read, *req_write, *req_peer; /* oldest request */
1692        struct net_conf *nc;
1693        unsigned long oldest_submit_jif;
1694        unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1695        unsigned long now;
1696        unsigned int ko_count = 0, timeout = 0;
1697
1698        rcu_read_lock();
1699        nc = rcu_dereference(connection->net_conf);
1700        if (nc && device->state.conn >= C_WF_REPORT_PARAMS) {
1701                ko_count = nc->ko_count;
1702                timeout = nc->timeout;
1703        }
1704
1705        if (get_ldev(device)) { /* implicit state.disk >= D_INCONSISTENT */
1706                dt = rcu_dereference(device->ldev->disk_conf)->disk_timeout * HZ / 10;
1707                put_ldev(device);
1708        }
1709        rcu_read_unlock();
1710
1711
1712        ent = timeout * HZ/10 * ko_count;
1713        et = min_not_zero(dt, ent);
1714
1715        if (!et)
1716                return; /* Recurring timer stopped */
1717
1718        now = jiffies;
1719        nt = now + et;
1720
1721        spin_lock_irq(&device->resource->req_lock);
1722        req_read = list_first_entry_or_null(&device->pending_completion[0], struct drbd_request, req_pending_local);
1723        req_write = list_first_entry_or_null(&device->pending_completion[1], struct drbd_request, req_pending_local);
1724
1725        /* maybe the oldest request waiting for the peer is in fact still
1726         * blocking in tcp sendmsg.  That's ok, though, that's handled via the
1727         * socket send timeout, requesting a ping, and bumping ko-count in
1728         * we_should_drop_the_connection().
1729         */
1730
1731        /* check the oldest request we did successfully sent,
1732         * but which is still waiting for an ACK. */
1733        req_peer = connection->req_ack_pending;
1734
1735        /* if we don't have such request (e.g. protocoll A)
1736         * check the oldest requests which is still waiting on its epoch
1737         * closing barrier ack. */
1738        if (!req_peer)
1739                req_peer = connection->req_not_net_done;
1740
1741        /* evaluate the oldest peer request only in one timer! */
1742        if (req_peer && req_peer->device != device)
1743                req_peer = NULL;
1744
1745        /* do we have something to evaluate? */
1746        if (req_peer == NULL && req_write == NULL && req_read == NULL)
1747                goto out;
1748
1749        oldest_submit_jif =
1750                (req_write && req_read)
1751                ? ( time_before(req_write->pre_submit_jif, req_read->pre_submit_jif)
1752                  ? req_write->pre_submit_jif : req_read->pre_submit_jif )
1753                : req_write ? req_write->pre_submit_jif
1754                : req_read ? req_read->pre_submit_jif : now;
1755
1756        if (ent && req_peer && net_timeout_reached(req_peer, connection, now, ent, ko_count, timeout))
1757                _conn_request_state(connection, NS(conn, C_TIMEOUT), CS_VERBOSE | CS_HARD);
1758
1759        if (dt && oldest_submit_jif != now &&
1760                 time_after(now, oldest_submit_jif + dt) &&
1761                !time_in_range(now, device->last_reattach_jif, device->last_reattach_jif + dt)) {
1762                drbd_warn(device, "Local backing device failed to meet the disk-timeout\n");
1763                __drbd_chk_io_error(device, DRBD_FORCE_DETACH);
1764        }
1765
1766        /* Reschedule timer for the nearest not already expired timeout.
1767         * Fallback to now + min(effective network timeout, disk timeout). */
1768        ent = (ent && req_peer && time_before(now, req_peer->pre_send_jif + ent))
1769                ? req_peer->pre_send_jif + ent : now + et;
1770        dt = (dt && oldest_submit_jif != now && time_before(now, oldest_submit_jif + dt))
1771                ? oldest_submit_jif + dt : now + et;
1772        nt = time_before(ent, dt) ? ent : dt;
1773out:
1774        spin_unlock_irq(&device->resource->req_lock);
1775        mod_timer(&device->request_timer, nt);
1776}
1777