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_device *device, 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_device *device, 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, &device->vdisk->part0); 43 part_stat_inc(cpu, &device->vdisk->part0, ios[rw]); 44 part_stat_add(cpu, &device->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(&device->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_device *device, struct drbd_request *req) 53 { 54 int rw = bio_data_dir(req->master_bio); 55 unsigned long duration = jiffies - req->start_jif; 56 int cpu; 57 cpu = part_stat_lock(); 58 part_stat_add(cpu, &device->vdisk->part0, ticks[rw], duration); 59 part_round_stats(cpu, &device->vdisk->part0); 60 part_dec_in_flight(&device->vdisk->part0, rw); 61 part_stat_unlock(); 62 } 63 64 static struct drbd_request *drbd_req_new(struct drbd_device *device, 65 struct bio *bio_src) 66 { 67 struct drbd_request *req; 68 69 req = mempool_alloc(drbd_request_mempool, GFP_NOIO | __GFP_ZERO); 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->device = device; 76 req->master_bio = bio_src; 77 req->epoch = 0; 78 79 drbd_clear_interval(&req->i); 80 req->i.sector = bio_src->bi_iter.bi_sector; 81 req->i.size = bio_src->bi_iter.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 INIT_LIST_HEAD(&req->req_pending_master_completion); 88 INIT_LIST_HEAD(&req->req_pending_local); 89 90 /* one reference to be put by __drbd_make_request */ 91 atomic_set(&req->completion_ref, 1); 92 /* one kref as long as completion_ref > 0 */ 93 kref_init(&req->kref); 94 return req; 95 } 96 97 static void drbd_remove_request_interval(struct rb_root *root, 98 struct drbd_request *req) 99 { 100 struct drbd_device *device = req->device; 101 struct drbd_interval *i = &req->i; 102 103 drbd_remove_interval(root, i); 104 105 /* Wake up any processes waiting for this request to complete. */ 106 if (i->waiting) 107 wake_up(&device->misc_wait); 108 } 109 110 void drbd_req_destroy(struct kref *kref) 111 { 112 struct drbd_request *req = container_of(kref, struct drbd_request, kref); 113 struct drbd_device *device = req->device; 114 const unsigned s = req->rq_state; 115 116 if ((req->master_bio && !(s & RQ_POSTPONED)) || 117 atomic_read(&req->completion_ref) || 118 (s & RQ_LOCAL_PENDING) || 119 ((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) { 120 drbd_err(device, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n", 121 s, atomic_read(&req->completion_ref)); 122 return; 123 } 124 125 /* If called from mod_rq_state (expected normal case) or 126 * drbd_send_and_submit (the less likely normal path), this holds the 127 * req_lock, and req->tl_requests will typicaly be on ->transfer_log, 128 * though it may be still empty (never added to the transfer log). 129 * 130 * If called from do_retry(), we do NOT hold the req_lock, but we are 131 * still allowed to unconditionally list_del(&req->tl_requests), 132 * because it will be on a local on-stack list only. */ 133 list_del_init(&req->tl_requests); 134 135 /* finally remove the request from the conflict detection 136 * respective block_id verification interval tree. */ 137 if (!drbd_interval_empty(&req->i)) { 138 struct rb_root *root; 139 140 if (s & RQ_WRITE) 141 root = &device->write_requests; 142 else 143 root = &device->read_requests; 144 drbd_remove_request_interval(root, req); 145 } else if (s & (RQ_NET_MASK & ~RQ_NET_DONE) && req->i.size != 0) 146 drbd_err(device, "drbd_req_destroy: Logic BUG: interval empty, but: rq_state=0x%x, sect=%llu, size=%u\n", 147 s, (unsigned long long)req->i.sector, req->i.size); 148 149 /* if it was a write, we may have to set the corresponding 150 * bit(s) out-of-sync first. If it had a local part, we need to 151 * release the reference to the activity log. */ 152 if (s & RQ_WRITE) { 153 /* Set out-of-sync unless both OK flags are set 154 * (local only or remote failed). 155 * Other places where we set out-of-sync: 156 * READ with local io-error */ 157 158 /* There is a special case: 159 * we may notice late that IO was suspended, 160 * and postpone, or schedule for retry, a write, 161 * before it even was submitted or sent. 162 * In that case we do not want to touch the bitmap at all. 163 */ 164 if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) { 165 if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK)) 166 drbd_set_out_of_sync(device, req->i.sector, req->i.size); 167 168 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS)) 169 drbd_set_in_sync(device, req->i.sector, req->i.size); 170 } 171 172 /* one might be tempted to move the drbd_al_complete_io 173 * to the local io completion callback drbd_request_endio. 174 * but, if this was a mirror write, we may only 175 * drbd_al_complete_io after this is RQ_NET_DONE, 176 * otherwise the extent could be dropped from the al 177 * before it has actually been written on the peer. 178 * if we crash before our peer knows about the request, 179 * but after the extent has been dropped from the al, 180 * we would forget to resync the corresponding extent. 181 */ 182 if (s & RQ_IN_ACT_LOG) { 183 if (get_ldev_if_state(device, D_FAILED)) { 184 drbd_al_complete_io(device, &req->i); 185 put_ldev(device); 186 } else if (__ratelimit(&drbd_ratelimit_state)) { 187 drbd_warn(device, "Should have called drbd_al_complete_io(, %llu, %u), " 188 "but my Disk seems to have failed :(\n", 189 (unsigned long long) req->i.sector, req->i.size); 190 } 191 } 192 } 193 194 mempool_free(req, drbd_request_mempool); 195 } 196 197 static void wake_all_senders(struct drbd_connection *connection) 198 { 199 wake_up(&connection->sender_work.q_wait); 200 } 201 202 /* must hold resource->req_lock */ 203 void start_new_tl_epoch(struct drbd_connection *connection) 204 { 205 /* no point closing an epoch, if it is empty, anyways. */ 206 if (connection->current_tle_writes == 0) 207 return; 208 209 connection->current_tle_writes = 0; 210 atomic_inc(&connection->current_tle_nr); 211 wake_all_senders(connection); 212 } 213 214 void complete_master_bio(struct drbd_device *device, 215 struct bio_and_error *m) 216 { 217 bio_endio(m->bio, m->error); 218 dec_ap_bio(device); 219 } 220 221 222 /* Helper for __req_mod(). 223 * Set m->bio to the master bio, if it is fit to be completed, 224 * or leave it alone (it is initialized to NULL in __req_mod), 225 * if it has already been completed, or cannot be completed yet. 226 * If m->bio is set, the error status to be returned is placed in m->error. 227 */ 228 static 229 void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m) 230 { 231 const unsigned s = req->rq_state; 232 struct drbd_device *device = req->device; 233 int rw; 234 int error, ok; 235 236 /* we must not complete the master bio, while it is 237 * still being processed by _drbd_send_zc_bio (drbd_send_dblock) 238 * not yet acknowledged by the peer 239 * not yet completed by the local io subsystem 240 * these flags may get cleared in any order by 241 * the worker, 242 * the receiver, 243 * the bio_endio completion callbacks. 244 */ 245 if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) || 246 (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) || 247 (s & RQ_COMPLETION_SUSP)) { 248 drbd_err(device, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s); 249 return; 250 } 251 252 if (!req->master_bio) { 253 drbd_err(device, "drbd_req_complete: Logic BUG, master_bio == NULL!\n"); 254 return; 255 } 256 257 rw = bio_rw(req->master_bio); 258 259 /* 260 * figure out whether to report success or failure. 261 * 262 * report success when at least one of the operations succeeded. 263 * or, to put the other way, 264 * only report failure, when both operations failed. 265 * 266 * what to do about the failures is handled elsewhere. 267 * what we need to do here is just: complete the master_bio. 268 * 269 * local completion error, if any, has been stored as ERR_PTR 270 * in private_bio within drbd_request_endio. 271 */ 272 ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK); 273 error = PTR_ERR(req->private_bio); 274 275 /* Before we can signal completion to the upper layers, 276 * we may need to close the current transfer log epoch. 277 * We are within the request lock, so we can simply compare 278 * the request epoch number with the current transfer log 279 * epoch number. If they match, increase the current_tle_nr, 280 * and reset the transfer log epoch write_cnt. 281 */ 282 if (rw == WRITE && 283 req->epoch == atomic_read(&first_peer_device(device)->connection->current_tle_nr)) 284 start_new_tl_epoch(first_peer_device(device)->connection); 285 286 /* Update disk stats */ 287 _drbd_end_io_acct(device, req); 288 289 /* If READ failed, 290 * have it be pushed back to the retry work queue, 291 * so it will re-enter __drbd_make_request(), 292 * and be re-assigned to a suitable local or remote path, 293 * or failed if we do not have access to good data anymore. 294 * 295 * Unless it was failed early by __drbd_make_request(), 296 * because no path was available, in which case 297 * it was not even added to the transfer_log. 298 * 299 * READA may fail, and will not be retried. 300 * 301 * WRITE should have used all available paths already. 302 */ 303 if (!ok && rw == READ && !list_empty(&req->tl_requests)) 304 req->rq_state |= RQ_POSTPONED; 305 306 if (!(req->rq_state & RQ_POSTPONED)) { 307 m->error = ok ? 0 : (error ?: -EIO); 308 m->bio = req->master_bio; 309 req->master_bio = NULL; 310 /* We leave it in the tree, to be able to verify later 311 * write-acks in protocol != C during resync. 312 * But we mark it as "complete", so it won't be counted as 313 * conflict in a multi-primary setup. */ 314 req->i.completed = true; 315 } 316 317 if (req->i.waiting) 318 wake_up(&device->misc_wait); 319 320 /* Either we are about to complete to upper layers, 321 * or we will restart this request. 322 * In either case, the request object will be destroyed soon, 323 * so better remove it from all lists. */ 324 list_del_init(&req->req_pending_master_completion); 325 } 326 327 /* still holds resource->req_lock */ 328 static int drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put) 329 { 330 struct drbd_device *device = req->device; 331 D_ASSERT(device, m || (req->rq_state & RQ_POSTPONED)); 332 333 if (!atomic_sub_and_test(put, &req->completion_ref)) 334 return 0; 335 336 drbd_req_complete(req, m); 337 338 if (req->rq_state & RQ_POSTPONED) { 339 /* don't destroy the req object just yet, 340 * but queue it for retry */ 341 drbd_restart_request(req); 342 return 0; 343 } 344 345 return 1; 346 } 347 348 static void set_if_null_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req) 349 { 350 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; 351 if (!connection) 352 return; 353 if (connection->req_next == NULL) 354 connection->req_next = req; 355 } 356 357 static void advance_conn_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req) 358 { 359 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; 360 if (!connection) 361 return; 362 if (connection->req_next != req) 363 return; 364 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) { 365 const unsigned s = req->rq_state; 366 if (s & RQ_NET_QUEUED) 367 break; 368 } 369 if (&req->tl_requests == &connection->transfer_log) 370 req = NULL; 371 connection->req_next = req; 372 } 373 374 static void set_if_null_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req) 375 { 376 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; 377 if (!connection) 378 return; 379 if (connection->req_ack_pending == NULL) 380 connection->req_ack_pending = req; 381 } 382 383 static void advance_conn_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req) 384 { 385 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; 386 if (!connection) 387 return; 388 if (connection->req_ack_pending != req) 389 return; 390 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) { 391 const unsigned s = req->rq_state; 392 if ((s & RQ_NET_SENT) && (s & RQ_NET_PENDING)) 393 break; 394 } 395 if (&req->tl_requests == &connection->transfer_log) 396 req = NULL; 397 connection->req_ack_pending = req; 398 } 399 400 static void set_if_null_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req) 401 { 402 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; 403 if (!connection) 404 return; 405 if (connection->req_not_net_done == NULL) 406 connection->req_not_net_done = req; 407 } 408 409 static void advance_conn_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req) 410 { 411 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; 412 if (!connection) 413 return; 414 if (connection->req_not_net_done != req) 415 return; 416 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) { 417 const unsigned s = req->rq_state; 418 if ((s & RQ_NET_SENT) && !(s & RQ_NET_DONE)) 419 break; 420 } 421 if (&req->tl_requests == &connection->transfer_log) 422 req = NULL; 423 connection->req_not_net_done = req; 424 } 425 426 /* I'd like this to be the only place that manipulates 427 * req->completion_ref and req->kref. */ 428 static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m, 429 int clear, int set) 430 { 431 struct drbd_device *device = req->device; 432 struct drbd_peer_device *peer_device = first_peer_device(device); 433 unsigned s = req->rq_state; 434 int c_put = 0; 435 int k_put = 0; 436 437 if (drbd_suspended(device) && !((s | clear) & RQ_COMPLETION_SUSP)) 438 set |= RQ_COMPLETION_SUSP; 439 440 /* apply */ 441 442 req->rq_state &= ~clear; 443 req->rq_state |= set; 444 445 /* no change? */ 446 if (req->rq_state == s) 447 return; 448 449 /* intent: get references */ 450 451 if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING)) 452 atomic_inc(&req->completion_ref); 453 454 if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) { 455 inc_ap_pending(device); 456 atomic_inc(&req->completion_ref); 457 } 458 459 if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED)) { 460 atomic_inc(&req->completion_ref); 461 set_if_null_req_next(peer_device, req); 462 } 463 464 if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK)) 465 kref_get(&req->kref); /* wait for the DONE */ 466 467 if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT)) { 468 /* potentially already completed in the asender thread */ 469 if (!(s & RQ_NET_DONE)) { 470 atomic_add(req->i.size >> 9, &device->ap_in_flight); 471 set_if_null_req_not_net_done(peer_device, req); 472 } 473 if (s & RQ_NET_PENDING) 474 set_if_null_req_ack_pending(peer_device, req); 475 } 476 477 if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP)) 478 atomic_inc(&req->completion_ref); 479 480 /* progress: put references */ 481 482 if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP)) 483 ++c_put; 484 485 if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) { 486 D_ASSERT(device, req->rq_state & RQ_LOCAL_PENDING); 487 /* local completion may still come in later, 488 * we need to keep the req object around. */ 489 kref_get(&req->kref); 490 ++c_put; 491 } 492 493 if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) { 494 if (req->rq_state & RQ_LOCAL_ABORTED) 495 ++k_put; 496 else 497 ++c_put; 498 list_del_init(&req->req_pending_local); 499 } 500 501 if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) { 502 dec_ap_pending(device); 503 ++c_put; 504 req->acked_jif = jiffies; 505 advance_conn_req_ack_pending(peer_device, req); 506 } 507 508 if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED)) { 509 ++c_put; 510 advance_conn_req_next(peer_device, req); 511 } 512 513 if (!(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) { 514 if (s & RQ_NET_SENT) 515 atomic_sub(req->i.size >> 9, &device->ap_in_flight); 516 if (s & RQ_EXP_BARR_ACK) 517 ++k_put; 518 req->net_done_jif = jiffies; 519 520 /* in ahead/behind mode, or just in case, 521 * before we finally destroy this request, 522 * the caching pointers must not reference it anymore */ 523 advance_conn_req_next(peer_device, req); 524 advance_conn_req_ack_pending(peer_device, req); 525 advance_conn_req_not_net_done(peer_device, req); 526 } 527 528 /* potentially complete and destroy */ 529 530 if (k_put || c_put) { 531 /* Completion does it's own kref_put. If we are going to 532 * kref_sub below, we need req to be still around then. */ 533 int at_least = k_put + !!c_put; 534 int refcount = atomic_read(&req->kref.refcount); 535 if (refcount < at_least) 536 drbd_err(device, 537 "mod_rq_state: Logic BUG: %x -> %x: refcount = %d, should be >= %d\n", 538 s, req->rq_state, refcount, at_least); 539 } 540 541 /* If we made progress, retry conflicting peer requests, if any. */ 542 if (req->i.waiting) 543 wake_up(&device->misc_wait); 544 545 if (c_put) 546 k_put += drbd_req_put_completion_ref(req, m, c_put); 547 if (k_put) 548 kref_sub(&req->kref, k_put, drbd_req_destroy); 549 } 550 551 static void drbd_report_io_error(struct drbd_device *device, struct drbd_request *req) 552 { 553 char b[BDEVNAME_SIZE]; 554 555 if (!__ratelimit(&drbd_ratelimit_state)) 556 return; 557 558 drbd_warn(device, "local %s IO error sector %llu+%u on %s\n", 559 (req->rq_state & RQ_WRITE) ? "WRITE" : "READ", 560 (unsigned long long)req->i.sector, 561 req->i.size >> 9, 562 bdevname(device->ldev->backing_bdev, b)); 563 } 564 565 /* Helper for HANDED_OVER_TO_NETWORK. 566 * Is this a protocol A write (neither WRITE_ACK nor RECEIVE_ACK expected)? 567 * Is it also still "PENDING"? 568 * --> If so, clear PENDING and set NET_OK below. 569 * If it is a protocol A write, but not RQ_PENDING anymore, neg-ack was faster 570 * (and we must not set RQ_NET_OK) */ 571 static inline bool is_pending_write_protocol_A(struct drbd_request *req) 572 { 573 return (req->rq_state & 574 (RQ_WRITE|RQ_NET_PENDING|RQ_EXP_WRITE_ACK|RQ_EXP_RECEIVE_ACK)) 575 == (RQ_WRITE|RQ_NET_PENDING); 576 } 577 578 /* obviously this could be coded as many single functions 579 * instead of one huge switch, 580 * or by putting the code directly in the respective locations 581 * (as it has been before). 582 * 583 * but having it this way 584 * enforces that it is all in this one place, where it is easier to audit, 585 * it makes it obvious that whatever "event" "happens" to a request should 586 * happen "atomically" within the req_lock, 587 * and it enforces that we have to think in a very structured manner 588 * about the "events" that may happen to a request during its life time ... 589 */ 590 int __req_mod(struct drbd_request *req, enum drbd_req_event what, 591 struct bio_and_error *m) 592 { 593 struct drbd_device *const device = req->device; 594 struct drbd_peer_device *const peer_device = first_peer_device(device); 595 struct drbd_connection *const connection = peer_device ? peer_device->connection : NULL; 596 struct net_conf *nc; 597 int p, rv = 0; 598 599 if (m) 600 m->bio = NULL; 601 602 switch (what) { 603 default: 604 drbd_err(device, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__); 605 break; 606 607 /* does not happen... 608 * initialization done in drbd_req_new 609 case CREATED: 610 break; 611 */ 612 613 case TO_BE_SENT: /* via network */ 614 /* reached via __drbd_make_request 615 * and from w_read_retry_remote */ 616 D_ASSERT(device, !(req->rq_state & RQ_NET_MASK)); 617 rcu_read_lock(); 618 nc = rcu_dereference(connection->net_conf); 619 p = nc->wire_protocol; 620 rcu_read_unlock(); 621 req->rq_state |= 622 p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK : 623 p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0; 624 mod_rq_state(req, m, 0, RQ_NET_PENDING); 625 break; 626 627 case TO_BE_SUBMITTED: /* locally */ 628 /* reached via __drbd_make_request */ 629 D_ASSERT(device, !(req->rq_state & RQ_LOCAL_MASK)); 630 mod_rq_state(req, m, 0, RQ_LOCAL_PENDING); 631 break; 632 633 case COMPLETED_OK: 634 if (req->rq_state & RQ_WRITE) 635 device->writ_cnt += req->i.size >> 9; 636 else 637 device->read_cnt += req->i.size >> 9; 638 639 mod_rq_state(req, m, RQ_LOCAL_PENDING, 640 RQ_LOCAL_COMPLETED|RQ_LOCAL_OK); 641 break; 642 643 case ABORT_DISK_IO: 644 mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED); 645 break; 646 647 case WRITE_COMPLETED_WITH_ERROR: 648 drbd_report_io_error(device, req); 649 __drbd_chk_io_error(device, DRBD_WRITE_ERROR); 650 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED); 651 break; 652 653 case READ_COMPLETED_WITH_ERROR: 654 drbd_set_out_of_sync(device, req->i.sector, req->i.size); 655 drbd_report_io_error(device, req); 656 __drbd_chk_io_error(device, DRBD_READ_ERROR); 657 /* fall through. */ 658 case READ_AHEAD_COMPLETED_WITH_ERROR: 659 /* it is legal to fail READA, no __drbd_chk_io_error in that case. */ 660 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED); 661 break; 662 663 case DISCARD_COMPLETED_NOTSUPP: 664 case DISCARD_COMPLETED_WITH_ERROR: 665 /* I'd rather not detach from local disk just because it 666 * failed a REQ_DISCARD. */ 667 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED); 668 break; 669 670 case QUEUE_FOR_NET_READ: 671 /* READ or READA, and 672 * no local disk, 673 * or target area marked as invalid, 674 * or just got an io-error. */ 675 /* from __drbd_make_request 676 * or from bio_endio during read io-error recovery */ 677 678 /* So we can verify the handle in the answer packet. 679 * Corresponding drbd_remove_request_interval is in 680 * drbd_req_complete() */ 681 D_ASSERT(device, drbd_interval_empty(&req->i)); 682 drbd_insert_interval(&device->read_requests, &req->i); 683 684 set_bit(UNPLUG_REMOTE, &device->flags); 685 686 D_ASSERT(device, req->rq_state & RQ_NET_PENDING); 687 D_ASSERT(device, (req->rq_state & RQ_LOCAL_MASK) == 0); 688 mod_rq_state(req, m, 0, RQ_NET_QUEUED); 689 req->w.cb = w_send_read_req; 690 drbd_queue_work(&connection->sender_work, 691 &req->w); 692 break; 693 694 case QUEUE_FOR_NET_WRITE: 695 /* assert something? */ 696 /* from __drbd_make_request only */ 697 698 /* Corresponding drbd_remove_request_interval is in 699 * drbd_req_complete() */ 700 D_ASSERT(device, drbd_interval_empty(&req->i)); 701 drbd_insert_interval(&device->write_requests, &req->i); 702 703 /* NOTE 704 * In case the req ended up on the transfer log before being 705 * queued on the worker, it could lead to this request being 706 * missed during cleanup after connection loss. 707 * So we have to do both operations here, 708 * within the same lock that protects the transfer log. 709 * 710 * _req_add_to_epoch(req); this has to be after the 711 * _maybe_start_new_epoch(req); which happened in 712 * __drbd_make_request, because we now may set the bit 713 * again ourselves to close the current epoch. 714 * 715 * Add req to the (now) current epoch (barrier). */ 716 717 /* otherwise we may lose an unplug, which may cause some remote 718 * io-scheduler timeout to expire, increasing maximum latency, 719 * hurting performance. */ 720 set_bit(UNPLUG_REMOTE, &device->flags); 721 722 /* queue work item to send data */ 723 D_ASSERT(device, req->rq_state & RQ_NET_PENDING); 724 mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK); 725 req->w.cb = w_send_dblock; 726 drbd_queue_work(&connection->sender_work, 727 &req->w); 728 729 /* close the epoch, in case it outgrew the limit */ 730 rcu_read_lock(); 731 nc = rcu_dereference(connection->net_conf); 732 p = nc->max_epoch_size; 733 rcu_read_unlock(); 734 if (connection->current_tle_writes >= p) 735 start_new_tl_epoch(connection); 736 737 break; 738 739 case QUEUE_FOR_SEND_OOS: 740 mod_rq_state(req, m, 0, RQ_NET_QUEUED); 741 req->w.cb = w_send_out_of_sync; 742 drbd_queue_work(&connection->sender_work, 743 &req->w); 744 break; 745 746 case READ_RETRY_REMOTE_CANCELED: 747 case SEND_CANCELED: 748 case SEND_FAILED: 749 /* real cleanup will be done from tl_clear. just update flags 750 * so it is no longer marked as on the worker queue */ 751 mod_rq_state(req, m, RQ_NET_QUEUED, 0); 752 break; 753 754 case HANDED_OVER_TO_NETWORK: 755 /* assert something? */ 756 if (is_pending_write_protocol_A(req)) 757 /* this is what is dangerous about protocol A: 758 * pretend it was successfully written on the peer. */ 759 mod_rq_state(req, m, RQ_NET_QUEUED|RQ_NET_PENDING, 760 RQ_NET_SENT|RQ_NET_OK); 761 else 762 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT); 763 /* It is still not yet RQ_NET_DONE until the 764 * corresponding epoch barrier got acked as well, 765 * so we know what to dirty on connection loss. */ 766 break; 767 768 case OOS_HANDED_TO_NETWORK: 769 /* Was not set PENDING, no longer QUEUED, so is now DONE 770 * as far as this connection is concerned. */ 771 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE); 772 break; 773 774 case CONNECTION_LOST_WHILE_PENDING: 775 /* transfer log cleanup after connection loss */ 776 mod_rq_state(req, m, 777 RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP, 778 RQ_NET_DONE); 779 break; 780 781 case CONFLICT_RESOLVED: 782 /* for superseded conflicting writes of multiple primaries, 783 * there is no need to keep anything in the tl, potential 784 * node crashes are covered by the activity log. 785 * 786 * If this request had been marked as RQ_POSTPONED before, 787 * it will actually not be completed, but "restarted", 788 * resubmitted from the retry worker context. */ 789 D_ASSERT(device, req->rq_state & RQ_NET_PENDING); 790 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK); 791 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK); 792 break; 793 794 case WRITE_ACKED_BY_PEER_AND_SIS: 795 req->rq_state |= RQ_NET_SIS; 796 case WRITE_ACKED_BY_PEER: 797 /* Normal operation protocol C: successfully written on peer. 798 * During resync, even in protocol != C, 799 * we requested an explicit write ack anyways. 800 * Which means we cannot even assert anything here. 801 * Nothing more to do here. 802 * We want to keep the tl in place for all protocols, to cater 803 * for volatile write-back caches on lower level devices. */ 804 goto ack_common; 805 case RECV_ACKED_BY_PEER: 806 D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK); 807 /* protocol B; pretends to be successfully written on peer. 808 * see also notes above in HANDED_OVER_TO_NETWORK about 809 * protocol != C */ 810 ack_common: 811 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK); 812 break; 813 814 case POSTPONE_WRITE: 815 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK); 816 /* If this node has already detected the write conflict, the 817 * worker will be waiting on misc_wait. Wake it up once this 818 * request has completed locally. 819 */ 820 D_ASSERT(device, req->rq_state & RQ_NET_PENDING); 821 req->rq_state |= RQ_POSTPONED; 822 if (req->i.waiting) 823 wake_up(&device->misc_wait); 824 /* Do not clear RQ_NET_PENDING. This request will make further 825 * progress via restart_conflicting_writes() or 826 * fail_postponed_requests(). Hopefully. */ 827 break; 828 829 case NEG_ACKED: 830 mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0); 831 break; 832 833 case FAIL_FROZEN_DISK_IO: 834 if (!(req->rq_state & RQ_LOCAL_COMPLETED)) 835 break; 836 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0); 837 break; 838 839 case RESTART_FROZEN_DISK_IO: 840 if (!(req->rq_state & RQ_LOCAL_COMPLETED)) 841 break; 842 843 mod_rq_state(req, m, 844 RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED, 845 RQ_LOCAL_PENDING); 846 847 rv = MR_READ; 848 if (bio_data_dir(req->master_bio) == WRITE) 849 rv = MR_WRITE; 850 851 get_ldev(device); /* always succeeds in this call path */ 852 req->w.cb = w_restart_disk_io; 853 drbd_queue_work(&connection->sender_work, 854 &req->w); 855 break; 856 857 case RESEND: 858 /* Simply complete (local only) READs. */ 859 if (!(req->rq_state & RQ_WRITE) && !req->w.cb) { 860 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0); 861 break; 862 } 863 864 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK 865 before the connection loss (B&C only); only P_BARRIER_ACK 866 (or the local completion?) was missing when we suspended. 867 Throwing them out of the TL here by pretending we got a BARRIER_ACK. 868 During connection handshake, we ensure that the peer was not rebooted. */ 869 if (!(req->rq_state & RQ_NET_OK)) { 870 /* FIXME could this possibly be a req->dw.cb == w_send_out_of_sync? 871 * in that case we must not set RQ_NET_PENDING. */ 872 873 mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING); 874 if (req->w.cb) { 875 /* w.cb expected to be w_send_dblock, or w_send_read_req */ 876 drbd_queue_work(&connection->sender_work, 877 &req->w); 878 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ; 879 } /* else: FIXME can this happen? */ 880 break; 881 } 882 /* else, fall through to BARRIER_ACKED */ 883 884 case BARRIER_ACKED: 885 /* barrier ack for READ requests does not make sense */ 886 if (!(req->rq_state & RQ_WRITE)) 887 break; 888 889 if (req->rq_state & RQ_NET_PENDING) { 890 /* barrier came in before all requests were acked. 891 * this is bad, because if the connection is lost now, 892 * we won't be able to clean them up... */ 893 drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n"); 894 } 895 /* Allowed to complete requests, even while suspended. 896 * As this is called for all requests within a matching epoch, 897 * we need to filter, and only set RQ_NET_DONE for those that 898 * have actually been on the wire. */ 899 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 900 (req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0); 901 break; 902 903 case DATA_RECEIVED: 904 D_ASSERT(device, req->rq_state & RQ_NET_PENDING); 905 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE); 906 break; 907 908 case QUEUE_AS_DRBD_BARRIER: 909 start_new_tl_epoch(connection); 910 mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE); 911 break; 912 }; 913 914 return rv; 915 } 916 917 /* we may do a local read if: 918 * - we are consistent (of course), 919 * - or we are generally inconsistent, 920 * BUT we are still/already IN SYNC for this area. 921 * since size may be bigger than BM_BLOCK_SIZE, 922 * we may need to check several bits. 923 */ 924 static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size) 925 { 926 unsigned long sbnr, ebnr; 927 sector_t esector, nr_sectors; 928 929 if (device->state.disk == D_UP_TO_DATE) 930 return true; 931 if (device->state.disk != D_INCONSISTENT) 932 return false; 933 esector = sector + (size >> 9) - 1; 934 nr_sectors = drbd_get_capacity(device->this_bdev); 935 D_ASSERT(device, sector < nr_sectors); 936 D_ASSERT(device, esector < nr_sectors); 937 938 sbnr = BM_SECT_TO_BIT(sector); 939 ebnr = BM_SECT_TO_BIT(esector); 940 941 return drbd_bm_count_bits(device, sbnr, ebnr) == 0; 942 } 943 944 static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector, 945 enum drbd_read_balancing rbm) 946 { 947 struct backing_dev_info *bdi; 948 int stripe_shift; 949 950 switch (rbm) { 951 case RB_CONGESTED_REMOTE: 952 bdi = &device->ldev->backing_bdev->bd_disk->queue->backing_dev_info; 953 return bdi_read_congested(bdi); 954 case RB_LEAST_PENDING: 955 return atomic_read(&device->local_cnt) > 956 atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt); 957 case RB_32K_STRIPING: /* stripe_shift = 15 */ 958 case RB_64K_STRIPING: 959 case RB_128K_STRIPING: 960 case RB_256K_STRIPING: 961 case RB_512K_STRIPING: 962 case RB_1M_STRIPING: /* stripe_shift = 20 */ 963 stripe_shift = (rbm - RB_32K_STRIPING + 15); 964 return (sector >> (stripe_shift - 9)) & 1; 965 case RB_ROUND_ROBIN: 966 return test_and_change_bit(READ_BALANCE_RR, &device->flags); 967 case RB_PREFER_REMOTE: 968 return true; 969 case RB_PREFER_LOCAL: 970 default: 971 return false; 972 } 973 } 974 975 /* 976 * complete_conflicting_writes - wait for any conflicting write requests 977 * 978 * The write_requests tree contains all active write requests which we 979 * currently know about. Wait for any requests to complete which conflict with 980 * the new one. 981 * 982 * Only way out: remove the conflicting intervals from the tree. 983 */ 984 static void complete_conflicting_writes(struct drbd_request *req) 985 { 986 DEFINE_WAIT(wait); 987 struct drbd_device *device = req->device; 988 struct drbd_interval *i; 989 sector_t sector = req->i.sector; 990 int size = req->i.size; 991 992 i = drbd_find_overlap(&device->write_requests, sector, size); 993 if (!i) 994 return; 995 996 for (;;) { 997 prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE); 998 i = drbd_find_overlap(&device->write_requests, sector, size); 999 if (!i) 1000 break; 1001 /* Indicate to wake up device->misc_wait on progress. */ 1002 i->waiting = true; 1003 spin_unlock_irq(&device->resource->req_lock); 1004 schedule(); 1005 spin_lock_irq(&device->resource->req_lock); 1006 } 1007 finish_wait(&device->misc_wait, &wait); 1008 } 1009 1010 /* called within req_lock and rcu_read_lock() */ 1011 static void maybe_pull_ahead(struct drbd_device *device) 1012 { 1013 struct drbd_connection *connection = first_peer_device(device)->connection; 1014 struct net_conf *nc; 1015 bool congested = false; 1016 enum drbd_on_congestion on_congestion; 1017 1018 rcu_read_lock(); 1019 nc = rcu_dereference(connection->net_conf); 1020 on_congestion = nc ? nc->on_congestion : OC_BLOCK; 1021 rcu_read_unlock(); 1022 if (on_congestion == OC_BLOCK || 1023 connection->agreed_pro_version < 96) 1024 return; 1025 1026 if (on_congestion == OC_PULL_AHEAD && device->state.conn == C_AHEAD) 1027 return; /* nothing to do ... */ 1028 1029 /* If I don't even have good local storage, we can not reasonably try 1030 * to pull ahead of the peer. We also need the local reference to make 1031 * sure device->act_log is there. 1032 */ 1033 if (!get_ldev_if_state(device, D_UP_TO_DATE)) 1034 return; 1035 1036 if (nc->cong_fill && 1037 atomic_read(&device->ap_in_flight) >= nc->cong_fill) { 1038 drbd_info(device, "Congestion-fill threshold reached\n"); 1039 congested = true; 1040 } 1041 1042 if (device->act_log->used >= nc->cong_extents) { 1043 drbd_info(device, "Congestion-extents threshold reached\n"); 1044 congested = true; 1045 } 1046 1047 if (congested) { 1048 /* start a new epoch for non-mirrored writes */ 1049 start_new_tl_epoch(first_peer_device(device)->connection); 1050 1051 if (on_congestion == OC_PULL_AHEAD) 1052 _drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL); 1053 else /*nc->on_congestion == OC_DISCONNECT */ 1054 _drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL); 1055 } 1056 put_ldev(device); 1057 } 1058 1059 /* If this returns false, and req->private_bio is still set, 1060 * this should be submitted locally. 1061 * 1062 * If it returns false, but req->private_bio is not set, 1063 * we do not have access to good data :( 1064 * 1065 * Otherwise, this destroys req->private_bio, if any, 1066 * and returns true. 1067 */ 1068 static bool do_remote_read(struct drbd_request *req) 1069 { 1070 struct drbd_device *device = req->device; 1071 enum drbd_read_balancing rbm; 1072 1073 if (req->private_bio) { 1074 if (!drbd_may_do_local_read(device, 1075 req->i.sector, req->i.size)) { 1076 bio_put(req->private_bio); 1077 req->private_bio = NULL; 1078 put_ldev(device); 1079 } 1080 } 1081 1082 if (device->state.pdsk != D_UP_TO_DATE) 1083 return false; 1084 1085 if (req->private_bio == NULL) 1086 return true; 1087 1088 /* TODO: improve read balancing decisions, take into account drbd 1089 * protocol, pending requests etc. */ 1090 1091 rcu_read_lock(); 1092 rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing; 1093 rcu_read_unlock(); 1094 1095 if (rbm == RB_PREFER_LOCAL && req->private_bio) 1096 return false; /* submit locally */ 1097 1098 if (remote_due_to_read_balancing(device, req->i.sector, rbm)) { 1099 if (req->private_bio) { 1100 bio_put(req->private_bio); 1101 req->private_bio = NULL; 1102 put_ldev(device); 1103 } 1104 return true; 1105 } 1106 1107 return false; 1108 } 1109 1110 /* returns number of connections (== 1, for drbd 8.4) 1111 * expected to actually write this data, 1112 * which does NOT include those that we are L_AHEAD for. */ 1113 static int drbd_process_write_request(struct drbd_request *req) 1114 { 1115 struct drbd_device *device = req->device; 1116 int remote, send_oos; 1117 1118 remote = drbd_should_do_remote(device->state); 1119 send_oos = drbd_should_send_out_of_sync(device->state); 1120 1121 /* Need to replicate writes. Unless it is an empty flush, 1122 * which is better mapped to a DRBD P_BARRIER packet, 1123 * also for drbd wire protocol compatibility reasons. 1124 * If this was a flush, just start a new epoch. 1125 * Unless the current epoch was empty anyways, or we are not currently 1126 * replicating, in which case there is no point. */ 1127 if (unlikely(req->i.size == 0)) { 1128 /* The only size==0 bios we expect are empty flushes. */ 1129 D_ASSERT(device, req->master_bio->bi_rw & REQ_FLUSH); 1130 if (remote) 1131 _req_mod(req, QUEUE_AS_DRBD_BARRIER); 1132 return remote; 1133 } 1134 1135 if (!remote && !send_oos) 1136 return 0; 1137 1138 D_ASSERT(device, !(remote && send_oos)); 1139 1140 if (remote) { 1141 _req_mod(req, TO_BE_SENT); 1142 _req_mod(req, QUEUE_FOR_NET_WRITE); 1143 } else if (drbd_set_out_of_sync(device, req->i.sector, req->i.size)) 1144 _req_mod(req, QUEUE_FOR_SEND_OOS); 1145 1146 return remote; 1147 } 1148 1149 static void 1150 drbd_submit_req_private_bio(struct drbd_request *req) 1151 { 1152 struct drbd_device *device = req->device; 1153 struct bio *bio = req->private_bio; 1154 const int rw = bio_rw(bio); 1155 1156 bio->bi_bdev = device->ldev->backing_bdev; 1157 1158 /* State may have changed since we grabbed our reference on the 1159 * ->ldev member. Double check, and short-circuit to endio. 1160 * In case the last activity log transaction failed to get on 1161 * stable storage, and this is a WRITE, we may not even submit 1162 * this bio. */ 1163 if (get_ldev(device)) { 1164 req->pre_submit_jif = jiffies; 1165 if (drbd_insert_fault(device, 1166 rw == WRITE ? DRBD_FAULT_DT_WR 1167 : rw == READ ? DRBD_FAULT_DT_RD 1168 : DRBD_FAULT_DT_RA)) 1169 bio_endio(bio, -EIO); 1170 else 1171 generic_make_request(bio); 1172 put_ldev(device); 1173 } else 1174 bio_endio(bio, -EIO); 1175 } 1176 1177 static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req) 1178 { 1179 spin_lock_irq(&device->resource->req_lock); 1180 list_add_tail(&req->tl_requests, &device->submit.writes); 1181 list_add_tail(&req->req_pending_master_completion, 1182 &device->pending_master_completion[1 /* WRITE */]); 1183 spin_unlock_irq(&device->resource->req_lock); 1184 queue_work(device->submit.wq, &device->submit.worker); 1185 /* do_submit() may sleep internally on al_wait, too */ 1186 wake_up(&device->al_wait); 1187 } 1188 1189 /* returns the new drbd_request pointer, if the caller is expected to 1190 * drbd_send_and_submit() it (to save latency), or NULL if we queued the 1191 * request on the submitter thread. 1192 * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request. 1193 */ 1194 static struct drbd_request * 1195 drbd_request_prepare(struct drbd_device *device, struct bio *bio, unsigned long start_jif) 1196 { 1197 const int rw = bio_data_dir(bio); 1198 struct drbd_request *req; 1199 1200 /* allocate outside of all locks; */ 1201 req = drbd_req_new(device, bio); 1202 if (!req) { 1203 dec_ap_bio(device); 1204 /* only pass the error to the upper layers. 1205 * if user cannot handle io errors, that's not our business. */ 1206 drbd_err(device, "could not kmalloc() req\n"); 1207 bio_endio(bio, -ENOMEM); 1208 return ERR_PTR(-ENOMEM); 1209 } 1210 req->start_jif = start_jif; 1211 1212 if (!get_ldev(device)) { 1213 bio_put(req->private_bio); 1214 req->private_bio = NULL; 1215 } 1216 1217 /* Update disk stats */ 1218 _drbd_start_io_acct(device, req); 1219 1220 if (rw == WRITE && req->private_bio && req->i.size 1221 && !test_bit(AL_SUSPENDED, &device->flags)) { 1222 if (!drbd_al_begin_io_fastpath(device, &req->i)) { 1223 atomic_inc(&device->ap_actlog_cnt); 1224 drbd_queue_write(device, req); 1225 return NULL; 1226 } 1227 req->rq_state |= RQ_IN_ACT_LOG; 1228 req->in_actlog_jif = jiffies; 1229 } 1230 1231 return req; 1232 } 1233 1234 static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req) 1235 { 1236 struct drbd_resource *resource = device->resource; 1237 const int rw = bio_rw(req->master_bio); 1238 struct bio_and_error m = { NULL, }; 1239 bool no_remote = false; 1240 bool submit_private_bio = false; 1241 1242 spin_lock_irq(&resource->req_lock); 1243 if (rw == WRITE) { 1244 /* This may temporarily give up the req_lock, 1245 * but will re-aquire it before it returns here. 1246 * Needs to be before the check on drbd_suspended() */ 1247 complete_conflicting_writes(req); 1248 /* no more giving up req_lock from now on! */ 1249 1250 /* check for congestion, and potentially stop sending 1251 * full data updates, but start sending "dirty bits" only. */ 1252 maybe_pull_ahead(device); 1253 } 1254 1255 1256 if (drbd_suspended(device)) { 1257 /* push back and retry: */ 1258 req->rq_state |= RQ_POSTPONED; 1259 if (req->private_bio) { 1260 bio_put(req->private_bio); 1261 req->private_bio = NULL; 1262 put_ldev(device); 1263 } 1264 goto out; 1265 } 1266 1267 /* We fail READ/READA early, if we can not serve it. 1268 * We must do this before req is registered on any lists. 1269 * Otherwise, drbd_req_complete() will queue failed READ for retry. */ 1270 if (rw != WRITE) { 1271 if (!do_remote_read(req) && !req->private_bio) 1272 goto nodata; 1273 } 1274 1275 /* which transfer log epoch does this belong to? */ 1276 req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr); 1277 1278 /* no point in adding empty flushes to the transfer log, 1279 * they are mapped to drbd barriers already. */ 1280 if (likely(req->i.size!=0)) { 1281 if (rw == WRITE) 1282 first_peer_device(device)->connection->current_tle_writes++; 1283 1284 list_add_tail(&req->tl_requests, &first_peer_device(device)->connection->transfer_log); 1285 } 1286 1287 if (rw == WRITE) { 1288 if (!drbd_process_write_request(req)) 1289 no_remote = true; 1290 } else { 1291 /* We either have a private_bio, or we can read from remote. 1292 * Otherwise we had done the goto nodata above. */ 1293 if (req->private_bio == NULL) { 1294 _req_mod(req, TO_BE_SENT); 1295 _req_mod(req, QUEUE_FOR_NET_READ); 1296 } else 1297 no_remote = true; 1298 } 1299 1300 /* If it took the fast path in drbd_request_prepare, add it here. 1301 * The slow path has added it already. */ 1302 if (list_empty(&req->req_pending_master_completion)) 1303 list_add_tail(&req->req_pending_master_completion, 1304 &device->pending_master_completion[rw == WRITE]); 1305 if (req->private_bio) { 1306 /* needs to be marked within the same spinlock */ 1307 list_add_tail(&req->req_pending_local, 1308 &device->pending_completion[rw == WRITE]); 1309 _req_mod(req, TO_BE_SUBMITTED); 1310 /* but we need to give up the spinlock to submit */ 1311 submit_private_bio = true; 1312 } else if (no_remote) { 1313 nodata: 1314 if (__ratelimit(&drbd_ratelimit_state)) 1315 drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n", 1316 (unsigned long long)req->i.sector, req->i.size >> 9); 1317 /* A write may have been queued for send_oos, however. 1318 * So we can not simply free it, we must go through drbd_req_put_completion_ref() */ 1319 } 1320 1321 out: 1322 if (drbd_req_put_completion_ref(req, &m, 1)) 1323 kref_put(&req->kref, drbd_req_destroy); 1324 spin_unlock_irq(&resource->req_lock); 1325 1326 /* Even though above is a kref_put(), this is safe. 1327 * As long as we still need to submit our private bio, 1328 * we hold a completion ref, and the request cannot disappear. 1329 * If however this request did not even have a private bio to submit 1330 * (e.g. remote read), req may already be invalid now. 1331 * That's why we cannot check on req->private_bio. */ 1332 if (submit_private_bio) 1333 drbd_submit_req_private_bio(req); 1334 if (m.bio) 1335 complete_master_bio(device, &m); 1336 } 1337 1338 void __drbd_make_request(struct drbd_device *device, struct bio *bio, unsigned long start_jif) 1339 { 1340 struct drbd_request *req = drbd_request_prepare(device, bio, start_jif); 1341 if (IS_ERR_OR_NULL(req)) 1342 return; 1343 drbd_send_and_submit(device, req); 1344 } 1345 1346 static void submit_fast_path(struct drbd_device *device, struct list_head *incoming) 1347 { 1348 struct drbd_request *req, *tmp; 1349 list_for_each_entry_safe(req, tmp, incoming, tl_requests) { 1350 const int rw = bio_data_dir(req->master_bio); 1351 1352 if (rw == WRITE /* rw != WRITE should not even end up here! */ 1353 && req->private_bio && req->i.size 1354 && !test_bit(AL_SUSPENDED, &device->flags)) { 1355 if (!drbd_al_begin_io_fastpath(device, &req->i)) 1356 continue; 1357 1358 req->rq_state |= RQ_IN_ACT_LOG; 1359 req->in_actlog_jif = jiffies; 1360 atomic_dec(&device->ap_actlog_cnt); 1361 } 1362 1363 list_del_init(&req->tl_requests); 1364 drbd_send_and_submit(device, req); 1365 } 1366 } 1367 1368 static bool prepare_al_transaction_nonblock(struct drbd_device *device, 1369 struct list_head *incoming, 1370 struct list_head *pending, 1371 struct list_head *later) 1372 { 1373 struct drbd_request *req, *tmp; 1374 int wake = 0; 1375 int err; 1376 1377 spin_lock_irq(&device->al_lock); 1378 list_for_each_entry_safe(req, tmp, incoming, tl_requests) { 1379 err = drbd_al_begin_io_nonblock(device, &req->i); 1380 if (err == -ENOBUFS) 1381 break; 1382 if (err == -EBUSY) 1383 wake = 1; 1384 if (err) 1385 list_move_tail(&req->tl_requests, later); 1386 else 1387 list_move_tail(&req->tl_requests, pending); 1388 } 1389 spin_unlock_irq(&device->al_lock); 1390 if (wake) 1391 wake_up(&device->al_wait); 1392 return !list_empty(pending); 1393 } 1394 1395 void send_and_submit_pending(struct drbd_device *device, struct list_head *pending) 1396 { 1397 struct drbd_request *req, *tmp; 1398 1399 list_for_each_entry_safe(req, tmp, pending, tl_requests) { 1400 req->rq_state |= RQ_IN_ACT_LOG; 1401 req->in_actlog_jif = jiffies; 1402 atomic_dec(&device->ap_actlog_cnt); 1403 list_del_init(&req->tl_requests); 1404 drbd_send_and_submit(device, req); 1405 } 1406 } 1407 1408 void do_submit(struct work_struct *ws) 1409 { 1410 struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker); 1411 LIST_HEAD(incoming); /* from drbd_make_request() */ 1412 LIST_HEAD(pending); /* to be submitted after next AL-transaction commit */ 1413 LIST_HEAD(busy); /* blocked by resync requests */ 1414 1415 /* grab new incoming requests */ 1416 spin_lock_irq(&device->resource->req_lock); 1417 list_splice_tail_init(&device->submit.writes, &incoming); 1418 spin_unlock_irq(&device->resource->req_lock); 1419 1420 for (;;) { 1421 DEFINE_WAIT(wait); 1422 1423 /* move used-to-be-busy back to front of incoming */ 1424 list_splice_init(&busy, &incoming); 1425 submit_fast_path(device, &incoming); 1426 if (list_empty(&incoming)) 1427 break; 1428 1429 for (;;) { 1430 prepare_to_wait(&device->al_wait, &wait, TASK_UNINTERRUPTIBLE); 1431 1432 list_splice_init(&busy, &incoming); 1433 prepare_al_transaction_nonblock(device, &incoming, &pending, &busy); 1434 if (!list_empty(&pending)) 1435 break; 1436 1437 schedule(); 1438 1439 /* If all currently "hot" activity log extents are kept busy by 1440 * incoming requests, we still must not totally starve new 1441 * requests to "cold" extents. 1442 * Something left on &incoming means there had not been 1443 * enough update slots available, and the activity log 1444 * has been marked as "starving". 1445 * 1446 * Try again now, without looking for new requests, 1447 * effectively blocking all new requests until we made 1448 * at least _some_ progress with what we currently have. 1449 */ 1450 if (!list_empty(&incoming)) 1451 continue; 1452 1453 /* Nothing moved to pending, but nothing left 1454 * on incoming: all moved to busy! 1455 * Grab new and iterate. */ 1456 spin_lock_irq(&device->resource->req_lock); 1457 list_splice_tail_init(&device->submit.writes, &incoming); 1458 spin_unlock_irq(&device->resource->req_lock); 1459 } 1460 finish_wait(&device->al_wait, &wait); 1461 1462 /* If the transaction was full, before all incoming requests 1463 * had been processed, skip ahead to commit, and iterate 1464 * without splicing in more incoming requests from upper layers. 1465 * 1466 * Else, if all incoming have been processed, 1467 * they have become either "pending" (to be submitted after 1468 * next transaction commit) or "busy" (blocked by resync). 1469 * 1470 * Maybe more was queued, while we prepared the transaction? 1471 * Try to stuff those into this transaction as well. 1472 * Be strictly non-blocking here, 1473 * we already have something to commit. 1474 * 1475 * Commit if we don't make any more progres. 1476 */ 1477 1478 while (list_empty(&incoming)) { 1479 LIST_HEAD(more_pending); 1480 LIST_HEAD(more_incoming); 1481 bool made_progress; 1482 1483 /* It is ok to look outside the lock, 1484 * it's only an optimization anyways */ 1485 if (list_empty(&device->submit.writes)) 1486 break; 1487 1488 spin_lock_irq(&device->resource->req_lock); 1489 list_splice_tail_init(&device->submit.writes, &more_incoming); 1490 spin_unlock_irq(&device->resource->req_lock); 1491 1492 if (list_empty(&more_incoming)) 1493 break; 1494 1495 made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending, &busy); 1496 1497 list_splice_tail_init(&more_pending, &pending); 1498 list_splice_tail_init(&more_incoming, &incoming); 1499 if (!made_progress) 1500 break; 1501 } 1502 1503 drbd_al_begin_io_commit(device); 1504 send_and_submit_pending(device, &pending); 1505 } 1506 } 1507 1508 void drbd_make_request(struct request_queue *q, struct bio *bio) 1509 { 1510 struct drbd_device *device = (struct drbd_device *) q->queuedata; 1511 unsigned long start_jif; 1512 1513 start_jif = jiffies; 1514 1515 /* 1516 * what we "blindly" assume: 1517 */ 1518 D_ASSERT(device, IS_ALIGNED(bio->bi_iter.bi_size, 512)); 1519 1520 inc_ap_bio(device); 1521 __drbd_make_request(device, bio, start_jif); 1522 } 1523 1524 /* This is called by bio_add_page(). 1525 * 1526 * q->max_hw_sectors and other global limits are already enforced there. 1527 * 1528 * We need to call down to our lower level device, 1529 * in case it has special restrictions. 1530 * 1531 * We also may need to enforce configured max-bio-bvecs limits. 1532 * 1533 * As long as the BIO is empty we have to allow at least one bvec, 1534 * regardless of size and offset, so no need to ask lower levels. 1535 */ 1536 int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec) 1537 { 1538 struct drbd_device *device = (struct drbd_device *) q->queuedata; 1539 unsigned int bio_size = bvm->bi_size; 1540 int limit = DRBD_MAX_BIO_SIZE; 1541 int backing_limit; 1542 1543 if (bio_size && get_ldev(device)) { 1544 unsigned int max_hw_sectors = queue_max_hw_sectors(q); 1545 struct request_queue * const b = 1546 device->ldev->backing_bdev->bd_disk->queue; 1547 if (b->merge_bvec_fn) { 1548 backing_limit = b->merge_bvec_fn(b, bvm, bvec); 1549 limit = min(limit, backing_limit); 1550 } 1551 put_ldev(device); 1552 if ((limit >> 9) > max_hw_sectors) 1553 limit = max_hw_sectors << 9; 1554 } 1555 return limit; 1556 } 1557 1558 void request_timer_fn(unsigned long data) 1559 { 1560 struct drbd_device *device = (struct drbd_device *) data; 1561 struct drbd_connection *connection = first_peer_device(device)->connection; 1562 struct drbd_request *req_read, *req_write, *req_peer; /* oldest request */ 1563 struct net_conf *nc; 1564 unsigned long oldest_submit_jif; 1565 unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */ 1566 unsigned long now; 1567 1568 rcu_read_lock(); 1569 nc = rcu_dereference(connection->net_conf); 1570 if (nc && device->state.conn >= C_WF_REPORT_PARAMS) 1571 ent = nc->timeout * HZ/10 * nc->ko_count; 1572 1573 if (get_ldev(device)) { /* implicit state.disk >= D_INCONSISTENT */ 1574 dt = rcu_dereference(device->ldev->disk_conf)->disk_timeout * HZ / 10; 1575 put_ldev(device); 1576 } 1577 rcu_read_unlock(); 1578 1579 et = min_not_zero(dt, ent); 1580 1581 if (!et) 1582 return; /* Recurring timer stopped */ 1583 1584 now = jiffies; 1585 nt = now + et; 1586 1587 spin_lock_irq(&device->resource->req_lock); 1588 req_read = list_first_entry_or_null(&device->pending_completion[0], struct drbd_request, req_pending_local); 1589 req_write = list_first_entry_or_null(&device->pending_completion[1], struct drbd_request, req_pending_local); 1590 req_peer = connection->req_not_net_done; 1591 /* maybe the oldest request waiting for the peer is in fact still 1592 * blocking in tcp sendmsg */ 1593 if (!req_peer && connection->req_next && connection->req_next->pre_send_jif) 1594 req_peer = connection->req_next; 1595 1596 /* evaluate the oldest peer request only in one timer! */ 1597 if (req_peer && req_peer->device != device) 1598 req_peer = NULL; 1599 1600 /* do we have something to evaluate? */ 1601 if (req_peer == NULL && req_write == NULL && req_read == NULL) 1602 goto out; 1603 1604 oldest_submit_jif = 1605 (req_write && req_read) 1606 ? ( time_before(req_write->pre_submit_jif, req_read->pre_submit_jif) 1607 ? req_write->pre_submit_jif : req_read->pre_submit_jif ) 1608 : req_write ? req_write->pre_submit_jif 1609 : req_read ? req_read->pre_submit_jif : now; 1610 1611 /* The request is considered timed out, if 1612 * - we have some effective timeout from the configuration, 1613 * with above state restrictions applied, 1614 * - the oldest request is waiting for a response from the network 1615 * resp. the local disk, 1616 * - the oldest request is in fact older than the effective timeout, 1617 * - the connection was established (resp. disk was attached) 1618 * for longer than the timeout already. 1619 * Note that for 32bit jiffies and very stable connections/disks, 1620 * we may have a wrap around, which is catched by 1621 * !time_in_range(now, last_..._jif, last_..._jif + timeout). 1622 * 1623 * Side effect: once per 32bit wrap-around interval, which means every 1624 * ~198 days with 250 HZ, we have a window where the timeout would need 1625 * to expire twice (worst case) to become effective. Good enough. 1626 */ 1627 if (ent && req_peer && 1628 time_after(now, req_peer->pre_send_jif + ent) && 1629 !time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent)) { 1630 drbd_warn(device, "Remote failed to finish a request within ko-count * timeout\n"); 1631 _drbd_set_state(_NS(device, conn, C_TIMEOUT), CS_VERBOSE | CS_HARD, NULL); 1632 } 1633 if (dt && oldest_submit_jif != now && 1634 time_after(now, oldest_submit_jif + dt) && 1635 !time_in_range(now, device->last_reattach_jif, device->last_reattach_jif + dt)) { 1636 drbd_warn(device, "Local backing device failed to meet the disk-timeout\n"); 1637 __drbd_chk_io_error(device, DRBD_FORCE_DETACH); 1638 } 1639 1640 /* Reschedule timer for the nearest not already expired timeout. 1641 * Fallback to now + min(effective network timeout, disk timeout). */ 1642 ent = (ent && req_peer && time_before(now, req_peer->pre_send_jif + ent)) 1643 ? req_peer->pre_send_jif + ent : now + et; 1644 dt = (dt && oldest_submit_jif != now && time_before(now, oldest_submit_jif + dt)) 1645 ? oldest_submit_jif + dt : now + et; 1646 nt = time_before(ent, dt) ? ent : dt; 1647 out: 1648 spin_unlock_irq(&device->resource->req_lock); 1649 mod_timer(&device->request_timer, nt); 1650 } 1651