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 generic_start_io_acct(bio_data_dir(req->master_bio), req->i.size >> 9, 40 &device->vdisk->part0); 41 } 42 43 /* Update disk stats when completing request upwards */ 44 static void _drbd_end_io_acct(struct drbd_device *device, struct drbd_request *req) 45 { 46 generic_end_io_acct(bio_data_dir(req->master_bio), 47 &device->vdisk->part0, req->start_jif); 48 } 49 50 static struct drbd_request *drbd_req_new(struct drbd_device *device, struct bio *bio_src) 51 { 52 struct drbd_request *req; 53 54 req = mempool_alloc(drbd_request_mempool, GFP_NOIO); 55 if (!req) 56 return NULL; 57 memset(req, 0, sizeof(*req)); 58 59 drbd_req_make_private_bio(req, bio_src); 60 req->rq_state = (bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0) 61 | (bio_op(bio_src) == REQ_OP_WRITE_SAME ? RQ_WSAME : 0) 62 | (bio_op(bio_src) == REQ_OP_WRITE_ZEROES ? RQ_UNMAP : 0) 63 | (bio_op(bio_src) == REQ_OP_DISCARD ? RQ_UNMAP : 0); 64 req->device = device; 65 req->master_bio = bio_src; 66 req->epoch = 0; 67 68 drbd_clear_interval(&req->i); 69 req->i.sector = bio_src->bi_iter.bi_sector; 70 req->i.size = bio_src->bi_iter.bi_size; 71 req->i.local = true; 72 req->i.waiting = false; 73 74 INIT_LIST_HEAD(&req->tl_requests); 75 INIT_LIST_HEAD(&req->w.list); 76 INIT_LIST_HEAD(&req->req_pending_master_completion); 77 INIT_LIST_HEAD(&req->req_pending_local); 78 79 /* one reference to be put by __drbd_make_request */ 80 atomic_set(&req->completion_ref, 1); 81 /* one kref as long as completion_ref > 0 */ 82 kref_init(&req->kref); 83 return req; 84 } 85 86 static void drbd_remove_request_interval(struct rb_root *root, 87 struct drbd_request *req) 88 { 89 struct drbd_device *device = req->device; 90 struct drbd_interval *i = &req->i; 91 92 drbd_remove_interval(root, i); 93 94 /* Wake up any processes waiting for this request to complete. */ 95 if (i->waiting) 96 wake_up(&device->misc_wait); 97 } 98 99 void drbd_req_destroy(struct kref *kref) 100 { 101 struct drbd_request *req = container_of(kref, struct drbd_request, kref); 102 struct drbd_device *device = req->device; 103 const unsigned s = req->rq_state; 104 105 if ((req->master_bio && !(s & RQ_POSTPONED)) || 106 atomic_read(&req->completion_ref) || 107 (s & RQ_LOCAL_PENDING) || 108 ((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) { 109 drbd_err(device, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n", 110 s, atomic_read(&req->completion_ref)); 111 return; 112 } 113 114 /* If called from mod_rq_state (expected normal case) or 115 * drbd_send_and_submit (the less likely normal path), this holds the 116 * req_lock, and req->tl_requests will typicaly be on ->transfer_log, 117 * though it may be still empty (never added to the transfer log). 118 * 119 * If called from do_retry(), we do NOT hold the req_lock, but we are 120 * still allowed to unconditionally list_del(&req->tl_requests), 121 * because it will be on a local on-stack list only. */ 122 list_del_init(&req->tl_requests); 123 124 /* finally remove the request from the conflict detection 125 * respective block_id verification interval tree. */ 126 if (!drbd_interval_empty(&req->i)) { 127 struct rb_root *root; 128 129 if (s & RQ_WRITE) 130 root = &device->write_requests; 131 else 132 root = &device->read_requests; 133 drbd_remove_request_interval(root, req); 134 } else if (s & (RQ_NET_MASK & ~RQ_NET_DONE) && req->i.size != 0) 135 drbd_err(device, "drbd_req_destroy: Logic BUG: interval empty, but: rq_state=0x%x, sect=%llu, size=%u\n", 136 s, (unsigned long long)req->i.sector, req->i.size); 137 138 /* if it was a write, we may have to set the corresponding 139 * bit(s) out-of-sync first. If it had a local part, we need to 140 * release the reference to the activity log. */ 141 if (s & RQ_WRITE) { 142 /* Set out-of-sync unless both OK flags are set 143 * (local only or remote failed). 144 * Other places where we set out-of-sync: 145 * READ with local io-error */ 146 147 /* There is a special case: 148 * we may notice late that IO was suspended, 149 * and postpone, or schedule for retry, a write, 150 * before it even was submitted or sent. 151 * In that case we do not want to touch the bitmap at all. 152 */ 153 if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) { 154 if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK)) 155 drbd_set_out_of_sync(device, req->i.sector, req->i.size); 156 157 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS)) 158 drbd_set_in_sync(device, req->i.sector, req->i.size); 159 } 160 161 /* one might be tempted to move the drbd_al_complete_io 162 * to the local io completion callback drbd_request_endio. 163 * but, if this was a mirror write, we may only 164 * drbd_al_complete_io after this is RQ_NET_DONE, 165 * otherwise the extent could be dropped from the al 166 * before it has actually been written on the peer. 167 * if we crash before our peer knows about the request, 168 * but after the extent has been dropped from the al, 169 * we would forget to resync the corresponding extent. 170 */ 171 if (s & RQ_IN_ACT_LOG) { 172 if (get_ldev_if_state(device, D_FAILED)) { 173 drbd_al_complete_io(device, &req->i); 174 put_ldev(device); 175 } else if (__ratelimit(&drbd_ratelimit_state)) { 176 drbd_warn(device, "Should have called drbd_al_complete_io(, %llu, %u), " 177 "but my Disk seems to have failed :(\n", 178 (unsigned long long) req->i.sector, req->i.size); 179 } 180 } 181 } 182 183 mempool_free(req, drbd_request_mempool); 184 } 185 186 static void wake_all_senders(struct drbd_connection *connection) 187 { 188 wake_up(&connection->sender_work.q_wait); 189 } 190 191 /* must hold resource->req_lock */ 192 void start_new_tl_epoch(struct drbd_connection *connection) 193 { 194 /* no point closing an epoch, if it is empty, anyways. */ 195 if (connection->current_tle_writes == 0) 196 return; 197 198 connection->current_tle_writes = 0; 199 atomic_inc(&connection->current_tle_nr); 200 wake_all_senders(connection); 201 } 202 203 void complete_master_bio(struct drbd_device *device, 204 struct bio_and_error *m) 205 { 206 m->bio->bi_error = m->error; 207 bio_endio(m->bio); 208 dec_ap_bio(device); 209 } 210 211 212 /* Helper for __req_mod(). 213 * Set m->bio to the master bio, if it is fit to be completed, 214 * or leave it alone (it is initialized to NULL in __req_mod), 215 * if it has already been completed, or cannot be completed yet. 216 * If m->bio is set, the error status to be returned is placed in m->error. 217 */ 218 static 219 void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m) 220 { 221 const unsigned s = req->rq_state; 222 struct drbd_device *device = req->device; 223 int error, ok; 224 225 /* we must not complete the master bio, while it is 226 * still being processed by _drbd_send_zc_bio (drbd_send_dblock) 227 * not yet acknowledged by the peer 228 * not yet completed by the local io subsystem 229 * these flags may get cleared in any order by 230 * the worker, 231 * the receiver, 232 * the bio_endio completion callbacks. 233 */ 234 if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) || 235 (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) || 236 (s & RQ_COMPLETION_SUSP)) { 237 drbd_err(device, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s); 238 return; 239 } 240 241 if (!req->master_bio) { 242 drbd_err(device, "drbd_req_complete: Logic BUG, master_bio == NULL!\n"); 243 return; 244 } 245 246 /* 247 * figure out whether to report success or failure. 248 * 249 * report success when at least one of the operations succeeded. 250 * or, to put the other way, 251 * only report failure, when both operations failed. 252 * 253 * what to do about the failures is handled elsewhere. 254 * what we need to do here is just: complete the master_bio. 255 * 256 * local completion error, if any, has been stored as ERR_PTR 257 * in private_bio within drbd_request_endio. 258 */ 259 ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK); 260 error = PTR_ERR(req->private_bio); 261 262 /* Before we can signal completion to the upper layers, 263 * we may need to close the current transfer log epoch. 264 * We are within the request lock, so we can simply compare 265 * the request epoch number with the current transfer log 266 * epoch number. If they match, increase the current_tle_nr, 267 * and reset the transfer log epoch write_cnt. 268 */ 269 if (op_is_write(bio_op(req->master_bio)) && 270 req->epoch == atomic_read(&first_peer_device(device)->connection->current_tle_nr)) 271 start_new_tl_epoch(first_peer_device(device)->connection); 272 273 /* Update disk stats */ 274 _drbd_end_io_acct(device, req); 275 276 /* If READ failed, 277 * have it be pushed back to the retry work queue, 278 * so it will re-enter __drbd_make_request(), 279 * and be re-assigned to a suitable local or remote path, 280 * or failed if we do not have access to good data anymore. 281 * 282 * Unless it was failed early by __drbd_make_request(), 283 * because no path was available, in which case 284 * it was not even added to the transfer_log. 285 * 286 * read-ahead may fail, and will not be retried. 287 * 288 * WRITE should have used all available paths already. 289 */ 290 if (!ok && 291 bio_op(req->master_bio) == REQ_OP_READ && 292 !(req->master_bio->bi_opf & REQ_RAHEAD) && 293 !list_empty(&req->tl_requests)) 294 req->rq_state |= RQ_POSTPONED; 295 296 if (!(req->rq_state & RQ_POSTPONED)) { 297 m->error = ok ? 0 : (error ?: -EIO); 298 m->bio = req->master_bio; 299 req->master_bio = NULL; 300 /* We leave it in the tree, to be able to verify later 301 * write-acks in protocol != C during resync. 302 * But we mark it as "complete", so it won't be counted as 303 * conflict in a multi-primary setup. */ 304 req->i.completed = true; 305 } 306 307 if (req->i.waiting) 308 wake_up(&device->misc_wait); 309 310 /* Either we are about to complete to upper layers, 311 * or we will restart this request. 312 * In either case, the request object will be destroyed soon, 313 * so better remove it from all lists. */ 314 list_del_init(&req->req_pending_master_completion); 315 } 316 317 /* still holds resource->req_lock */ 318 static int drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put) 319 { 320 struct drbd_device *device = req->device; 321 D_ASSERT(device, m || (req->rq_state & RQ_POSTPONED)); 322 323 if (!atomic_sub_and_test(put, &req->completion_ref)) 324 return 0; 325 326 drbd_req_complete(req, m); 327 328 if (req->rq_state & RQ_POSTPONED) { 329 /* don't destroy the req object just yet, 330 * but queue it for retry */ 331 drbd_restart_request(req); 332 return 0; 333 } 334 335 return 1; 336 } 337 338 static void set_if_null_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req) 339 { 340 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; 341 if (!connection) 342 return; 343 if (connection->req_next == NULL) 344 connection->req_next = req; 345 } 346 347 static void advance_conn_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req) 348 { 349 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; 350 if (!connection) 351 return; 352 if (connection->req_next != req) 353 return; 354 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) { 355 const unsigned s = req->rq_state; 356 if (s & RQ_NET_QUEUED) 357 break; 358 } 359 if (&req->tl_requests == &connection->transfer_log) 360 req = NULL; 361 connection->req_next = req; 362 } 363 364 static void set_if_null_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req) 365 { 366 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; 367 if (!connection) 368 return; 369 if (connection->req_ack_pending == NULL) 370 connection->req_ack_pending = req; 371 } 372 373 static void advance_conn_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req) 374 { 375 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; 376 if (!connection) 377 return; 378 if (connection->req_ack_pending != req) 379 return; 380 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) { 381 const unsigned s = req->rq_state; 382 if ((s & RQ_NET_SENT) && (s & RQ_NET_PENDING)) 383 break; 384 } 385 if (&req->tl_requests == &connection->transfer_log) 386 req = NULL; 387 connection->req_ack_pending = req; 388 } 389 390 static void set_if_null_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req) 391 { 392 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; 393 if (!connection) 394 return; 395 if (connection->req_not_net_done == NULL) 396 connection->req_not_net_done = req; 397 } 398 399 static void advance_conn_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req) 400 { 401 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; 402 if (!connection) 403 return; 404 if (connection->req_not_net_done != req) 405 return; 406 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) { 407 const unsigned s = req->rq_state; 408 if ((s & RQ_NET_SENT) && !(s & RQ_NET_DONE)) 409 break; 410 } 411 if (&req->tl_requests == &connection->transfer_log) 412 req = NULL; 413 connection->req_not_net_done = req; 414 } 415 416 /* I'd like this to be the only place that manipulates 417 * req->completion_ref and req->kref. */ 418 static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m, 419 int clear, int set) 420 { 421 struct drbd_device *device = req->device; 422 struct drbd_peer_device *peer_device = first_peer_device(device); 423 unsigned s = req->rq_state; 424 int c_put = 0; 425 426 if (drbd_suspended(device) && !((s | clear) & RQ_COMPLETION_SUSP)) 427 set |= RQ_COMPLETION_SUSP; 428 429 /* apply */ 430 431 req->rq_state &= ~clear; 432 req->rq_state |= set; 433 434 /* no change? */ 435 if (req->rq_state == s) 436 return; 437 438 /* intent: get references */ 439 440 kref_get(&req->kref); 441 442 if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING)) 443 atomic_inc(&req->completion_ref); 444 445 if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) { 446 inc_ap_pending(device); 447 atomic_inc(&req->completion_ref); 448 } 449 450 if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED)) { 451 atomic_inc(&req->completion_ref); 452 set_if_null_req_next(peer_device, req); 453 } 454 455 if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK)) 456 kref_get(&req->kref); /* wait for the DONE */ 457 458 if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT)) { 459 /* potentially already completed in the ack_receiver thread */ 460 if (!(s & RQ_NET_DONE)) { 461 atomic_add(req->i.size >> 9, &device->ap_in_flight); 462 set_if_null_req_not_net_done(peer_device, req); 463 } 464 if (req->rq_state & RQ_NET_PENDING) 465 set_if_null_req_ack_pending(peer_device, req); 466 } 467 468 if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP)) 469 atomic_inc(&req->completion_ref); 470 471 /* progress: put references */ 472 473 if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP)) 474 ++c_put; 475 476 if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) { 477 D_ASSERT(device, req->rq_state & RQ_LOCAL_PENDING); 478 ++c_put; 479 } 480 481 if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) { 482 if (req->rq_state & RQ_LOCAL_ABORTED) 483 kref_put(&req->kref, drbd_req_destroy); 484 else 485 ++c_put; 486 list_del_init(&req->req_pending_local); 487 } 488 489 if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) { 490 dec_ap_pending(device); 491 ++c_put; 492 req->acked_jif = jiffies; 493 advance_conn_req_ack_pending(peer_device, req); 494 } 495 496 if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED)) { 497 ++c_put; 498 advance_conn_req_next(peer_device, req); 499 } 500 501 if (!(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) { 502 if (s & RQ_NET_SENT) 503 atomic_sub(req->i.size >> 9, &device->ap_in_flight); 504 if (s & RQ_EXP_BARR_ACK) 505 kref_put(&req->kref, drbd_req_destroy); 506 req->net_done_jif = jiffies; 507 508 /* in ahead/behind mode, or just in case, 509 * before we finally destroy this request, 510 * the caching pointers must not reference it anymore */ 511 advance_conn_req_next(peer_device, req); 512 advance_conn_req_ack_pending(peer_device, req); 513 advance_conn_req_not_net_done(peer_device, req); 514 } 515 516 /* potentially complete and destroy */ 517 518 /* If we made progress, retry conflicting peer requests, if any. */ 519 if (req->i.waiting) 520 wake_up(&device->misc_wait); 521 522 if (c_put) { 523 if (drbd_req_put_completion_ref(req, m, c_put)) 524 kref_put(&req->kref, drbd_req_destroy); 525 } else { 526 kref_put(&req->kref, drbd_req_destroy); 527 } 528 } 529 530 static void drbd_report_io_error(struct drbd_device *device, struct drbd_request *req) 531 { 532 char b[BDEVNAME_SIZE]; 533 534 if (!__ratelimit(&drbd_ratelimit_state)) 535 return; 536 537 drbd_warn(device, "local %s IO error sector %llu+%u on %s\n", 538 (req->rq_state & RQ_WRITE) ? "WRITE" : "READ", 539 (unsigned long long)req->i.sector, 540 req->i.size >> 9, 541 bdevname(device->ldev->backing_bdev, b)); 542 } 543 544 /* Helper for HANDED_OVER_TO_NETWORK. 545 * Is this a protocol A write (neither WRITE_ACK nor RECEIVE_ACK expected)? 546 * Is it also still "PENDING"? 547 * --> If so, clear PENDING and set NET_OK below. 548 * If it is a protocol A write, but not RQ_PENDING anymore, neg-ack was faster 549 * (and we must not set RQ_NET_OK) */ 550 static inline bool is_pending_write_protocol_A(struct drbd_request *req) 551 { 552 return (req->rq_state & 553 (RQ_WRITE|RQ_NET_PENDING|RQ_EXP_WRITE_ACK|RQ_EXP_RECEIVE_ACK)) 554 == (RQ_WRITE|RQ_NET_PENDING); 555 } 556 557 /* obviously this could be coded as many single functions 558 * instead of one huge switch, 559 * or by putting the code directly in the respective locations 560 * (as it has been before). 561 * 562 * but having it this way 563 * enforces that it is all in this one place, where it is easier to audit, 564 * it makes it obvious that whatever "event" "happens" to a request should 565 * happen "atomically" within the req_lock, 566 * and it enforces that we have to think in a very structured manner 567 * about the "events" that may happen to a request during its life time ... 568 */ 569 int __req_mod(struct drbd_request *req, enum drbd_req_event what, 570 struct bio_and_error *m) 571 { 572 struct drbd_device *const device = req->device; 573 struct drbd_peer_device *const peer_device = first_peer_device(device); 574 struct drbd_connection *const connection = peer_device ? peer_device->connection : NULL; 575 struct net_conf *nc; 576 int p, rv = 0; 577 578 if (m) 579 m->bio = NULL; 580 581 switch (what) { 582 default: 583 drbd_err(device, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__); 584 break; 585 586 /* does not happen... 587 * initialization done in drbd_req_new 588 case CREATED: 589 break; 590 */ 591 592 case TO_BE_SENT: /* via network */ 593 /* reached via __drbd_make_request 594 * and from w_read_retry_remote */ 595 D_ASSERT(device, !(req->rq_state & RQ_NET_MASK)); 596 rcu_read_lock(); 597 nc = rcu_dereference(connection->net_conf); 598 p = nc->wire_protocol; 599 rcu_read_unlock(); 600 req->rq_state |= 601 p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK : 602 p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0; 603 mod_rq_state(req, m, 0, RQ_NET_PENDING); 604 break; 605 606 case TO_BE_SUBMITTED: /* locally */ 607 /* reached via __drbd_make_request */ 608 D_ASSERT(device, !(req->rq_state & RQ_LOCAL_MASK)); 609 mod_rq_state(req, m, 0, RQ_LOCAL_PENDING); 610 break; 611 612 case COMPLETED_OK: 613 if (req->rq_state & RQ_WRITE) 614 device->writ_cnt += req->i.size >> 9; 615 else 616 device->read_cnt += req->i.size >> 9; 617 618 mod_rq_state(req, m, RQ_LOCAL_PENDING, 619 RQ_LOCAL_COMPLETED|RQ_LOCAL_OK); 620 break; 621 622 case ABORT_DISK_IO: 623 mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED); 624 break; 625 626 case WRITE_COMPLETED_WITH_ERROR: 627 drbd_report_io_error(device, req); 628 __drbd_chk_io_error(device, DRBD_WRITE_ERROR); 629 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED); 630 break; 631 632 case READ_COMPLETED_WITH_ERROR: 633 drbd_set_out_of_sync(device, req->i.sector, req->i.size); 634 drbd_report_io_error(device, req); 635 __drbd_chk_io_error(device, DRBD_READ_ERROR); 636 /* fall through. */ 637 case READ_AHEAD_COMPLETED_WITH_ERROR: 638 /* it is legal to fail read-ahead, no __drbd_chk_io_error in that case. */ 639 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED); 640 break; 641 642 case DISCARD_COMPLETED_NOTSUPP: 643 case DISCARD_COMPLETED_WITH_ERROR: 644 /* I'd rather not detach from local disk just because it 645 * failed a REQ_DISCARD. */ 646 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED); 647 break; 648 649 case QUEUE_FOR_NET_READ: 650 /* READ, and 651 * no local disk, 652 * or target area marked as invalid, 653 * or just got an io-error. */ 654 /* from __drbd_make_request 655 * or from bio_endio during read io-error recovery */ 656 657 /* So we can verify the handle in the answer packet. 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->read_requests, &req->i); 662 663 set_bit(UNPLUG_REMOTE, &device->flags); 664 665 D_ASSERT(device, req->rq_state & RQ_NET_PENDING); 666 D_ASSERT(device, (req->rq_state & RQ_LOCAL_MASK) == 0); 667 mod_rq_state(req, m, 0, RQ_NET_QUEUED); 668 req->w.cb = w_send_read_req; 669 drbd_queue_work(&connection->sender_work, 670 &req->w); 671 break; 672 673 case QUEUE_FOR_NET_WRITE: 674 /* assert something? */ 675 /* from __drbd_make_request only */ 676 677 /* Corresponding drbd_remove_request_interval is in 678 * drbd_req_complete() */ 679 D_ASSERT(device, drbd_interval_empty(&req->i)); 680 drbd_insert_interval(&device->write_requests, &req->i); 681 682 /* NOTE 683 * In case the req ended up on the transfer log before being 684 * queued on the worker, it could lead to this request being 685 * missed during cleanup after connection loss. 686 * So we have to do both operations here, 687 * within the same lock that protects the transfer log. 688 * 689 * _req_add_to_epoch(req); this has to be after the 690 * _maybe_start_new_epoch(req); which happened in 691 * __drbd_make_request, because we now may set the bit 692 * again ourselves to close the current epoch. 693 * 694 * Add req to the (now) current epoch (barrier). */ 695 696 /* otherwise we may lose an unplug, which may cause some remote 697 * io-scheduler timeout to expire, increasing maximum latency, 698 * hurting performance. */ 699 set_bit(UNPLUG_REMOTE, &device->flags); 700 701 /* queue work item to send data */ 702 D_ASSERT(device, req->rq_state & RQ_NET_PENDING); 703 mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK); 704 req->w.cb = w_send_dblock; 705 drbd_queue_work(&connection->sender_work, 706 &req->w); 707 708 /* close the epoch, in case it outgrew the limit */ 709 rcu_read_lock(); 710 nc = rcu_dereference(connection->net_conf); 711 p = nc->max_epoch_size; 712 rcu_read_unlock(); 713 if (connection->current_tle_writes >= p) 714 start_new_tl_epoch(connection); 715 716 break; 717 718 case QUEUE_FOR_SEND_OOS: 719 mod_rq_state(req, m, 0, RQ_NET_QUEUED); 720 req->w.cb = w_send_out_of_sync; 721 drbd_queue_work(&connection->sender_work, 722 &req->w); 723 break; 724 725 case READ_RETRY_REMOTE_CANCELED: 726 case SEND_CANCELED: 727 case SEND_FAILED: 728 /* real cleanup will be done from tl_clear. just update flags 729 * so it is no longer marked as on the worker queue */ 730 mod_rq_state(req, m, RQ_NET_QUEUED, 0); 731 break; 732 733 case HANDED_OVER_TO_NETWORK: 734 /* assert something? */ 735 if (is_pending_write_protocol_A(req)) 736 /* this is what is dangerous about protocol A: 737 * pretend it was successfully written on the peer. */ 738 mod_rq_state(req, m, RQ_NET_QUEUED|RQ_NET_PENDING, 739 RQ_NET_SENT|RQ_NET_OK); 740 else 741 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT); 742 /* It is still not yet RQ_NET_DONE until the 743 * corresponding epoch barrier got acked as well, 744 * so we know what to dirty on connection loss. */ 745 break; 746 747 case OOS_HANDED_TO_NETWORK: 748 /* Was not set PENDING, no longer QUEUED, so is now DONE 749 * as far as this connection is concerned. */ 750 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE); 751 break; 752 753 case CONNECTION_LOST_WHILE_PENDING: 754 /* transfer log cleanup after connection loss */ 755 mod_rq_state(req, m, 756 RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP, 757 RQ_NET_DONE); 758 break; 759 760 case CONFLICT_RESOLVED: 761 /* for superseded conflicting writes of multiple primaries, 762 * there is no need to keep anything in the tl, potential 763 * node crashes are covered by the activity log. 764 * 765 * If this request had been marked as RQ_POSTPONED before, 766 * it will actually not be completed, but "restarted", 767 * resubmitted from the retry worker context. */ 768 D_ASSERT(device, req->rq_state & RQ_NET_PENDING); 769 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK); 770 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK); 771 break; 772 773 case WRITE_ACKED_BY_PEER_AND_SIS: 774 req->rq_state |= RQ_NET_SIS; 775 case WRITE_ACKED_BY_PEER: 776 /* Normal operation protocol C: successfully written on peer. 777 * During resync, even in protocol != C, 778 * we requested an explicit write ack anyways. 779 * Which means we cannot even assert anything here. 780 * Nothing more to do here. 781 * We want to keep the tl in place for all protocols, to cater 782 * for volatile write-back caches on lower level devices. */ 783 goto ack_common; 784 case RECV_ACKED_BY_PEER: 785 D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK); 786 /* protocol B; pretends to be successfully written on peer. 787 * see also notes above in HANDED_OVER_TO_NETWORK about 788 * protocol != C */ 789 ack_common: 790 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK); 791 break; 792 793 case POSTPONE_WRITE: 794 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK); 795 /* If this node has already detected the write conflict, the 796 * worker will be waiting on misc_wait. Wake it up once this 797 * request has completed locally. 798 */ 799 D_ASSERT(device, req->rq_state & RQ_NET_PENDING); 800 req->rq_state |= RQ_POSTPONED; 801 if (req->i.waiting) 802 wake_up(&device->misc_wait); 803 /* Do not clear RQ_NET_PENDING. This request will make further 804 * progress via restart_conflicting_writes() or 805 * fail_postponed_requests(). Hopefully. */ 806 break; 807 808 case NEG_ACKED: 809 mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0); 810 break; 811 812 case FAIL_FROZEN_DISK_IO: 813 if (!(req->rq_state & RQ_LOCAL_COMPLETED)) 814 break; 815 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0); 816 break; 817 818 case RESTART_FROZEN_DISK_IO: 819 if (!(req->rq_state & RQ_LOCAL_COMPLETED)) 820 break; 821 822 mod_rq_state(req, m, 823 RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED, 824 RQ_LOCAL_PENDING); 825 826 rv = MR_READ; 827 if (bio_data_dir(req->master_bio) == WRITE) 828 rv = MR_WRITE; 829 830 get_ldev(device); /* always succeeds in this call path */ 831 req->w.cb = w_restart_disk_io; 832 drbd_queue_work(&connection->sender_work, 833 &req->w); 834 break; 835 836 case RESEND: 837 /* Simply complete (local only) READs. */ 838 if (!(req->rq_state & RQ_WRITE) && !req->w.cb) { 839 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0); 840 break; 841 } 842 843 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK 844 before the connection loss (B&C only); only P_BARRIER_ACK 845 (or the local completion?) was missing when we suspended. 846 Throwing them out of the TL here by pretending we got a BARRIER_ACK. 847 During connection handshake, we ensure that the peer was not rebooted. */ 848 if (!(req->rq_state & RQ_NET_OK)) { 849 /* FIXME could this possibly be a req->dw.cb == w_send_out_of_sync? 850 * in that case we must not set RQ_NET_PENDING. */ 851 852 mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING); 853 if (req->w.cb) { 854 /* w.cb expected to be w_send_dblock, or w_send_read_req */ 855 drbd_queue_work(&connection->sender_work, 856 &req->w); 857 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ; 858 } /* else: FIXME can this happen? */ 859 break; 860 } 861 /* else, fall through to BARRIER_ACKED */ 862 863 case BARRIER_ACKED: 864 /* barrier ack for READ requests does not make sense */ 865 if (!(req->rq_state & RQ_WRITE)) 866 break; 867 868 if (req->rq_state & RQ_NET_PENDING) { 869 /* barrier came in before all requests were acked. 870 * this is bad, because if the connection is lost now, 871 * we won't be able to clean them up... */ 872 drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n"); 873 } 874 /* Allowed to complete requests, even while suspended. 875 * As this is called for all requests within a matching epoch, 876 * we need to filter, and only set RQ_NET_DONE for those that 877 * have actually been on the wire. */ 878 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 879 (req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0); 880 break; 881 882 case DATA_RECEIVED: 883 D_ASSERT(device, req->rq_state & RQ_NET_PENDING); 884 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE); 885 break; 886 887 case QUEUE_AS_DRBD_BARRIER: 888 start_new_tl_epoch(connection); 889 mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE); 890 break; 891 }; 892 893 return rv; 894 } 895 896 /* we may do a local read if: 897 * - we are consistent (of course), 898 * - or we are generally inconsistent, 899 * BUT we are still/already IN SYNC for this area. 900 * since size may be bigger than BM_BLOCK_SIZE, 901 * we may need to check several bits. 902 */ 903 static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size) 904 { 905 unsigned long sbnr, ebnr; 906 sector_t esector, nr_sectors; 907 908 if (device->state.disk == D_UP_TO_DATE) 909 return true; 910 if (device->state.disk != D_INCONSISTENT) 911 return false; 912 esector = sector + (size >> 9) - 1; 913 nr_sectors = drbd_get_capacity(device->this_bdev); 914 D_ASSERT(device, sector < nr_sectors); 915 D_ASSERT(device, esector < nr_sectors); 916 917 sbnr = BM_SECT_TO_BIT(sector); 918 ebnr = BM_SECT_TO_BIT(esector); 919 920 return drbd_bm_count_bits(device, sbnr, ebnr) == 0; 921 } 922 923 static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector, 924 enum drbd_read_balancing rbm) 925 { 926 struct backing_dev_info *bdi; 927 int stripe_shift; 928 929 switch (rbm) { 930 case RB_CONGESTED_REMOTE: 931 bdi = device->ldev->backing_bdev->bd_disk->queue->backing_dev_info; 932 return bdi_read_congested(bdi); 933 case RB_LEAST_PENDING: 934 return atomic_read(&device->local_cnt) > 935 atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt); 936 case RB_32K_STRIPING: /* stripe_shift = 15 */ 937 case RB_64K_STRIPING: 938 case RB_128K_STRIPING: 939 case RB_256K_STRIPING: 940 case RB_512K_STRIPING: 941 case RB_1M_STRIPING: /* stripe_shift = 20 */ 942 stripe_shift = (rbm - RB_32K_STRIPING + 15); 943 return (sector >> (stripe_shift - 9)) & 1; 944 case RB_ROUND_ROBIN: 945 return test_and_change_bit(READ_BALANCE_RR, &device->flags); 946 case RB_PREFER_REMOTE: 947 return true; 948 case RB_PREFER_LOCAL: 949 default: 950 return false; 951 } 952 } 953 954 /* 955 * complete_conflicting_writes - wait for any conflicting write requests 956 * 957 * The write_requests tree contains all active write requests which we 958 * currently know about. Wait for any requests to complete which conflict with 959 * the new one. 960 * 961 * Only way out: remove the conflicting intervals from the tree. 962 */ 963 static void complete_conflicting_writes(struct drbd_request *req) 964 { 965 DEFINE_WAIT(wait); 966 struct drbd_device *device = req->device; 967 struct drbd_interval *i; 968 sector_t sector = req->i.sector; 969 int size = req->i.size; 970 971 for (;;) { 972 drbd_for_each_overlap(i, &device->write_requests, sector, size) { 973 /* Ignore, if already completed to upper layers. */ 974 if (i->completed) 975 continue; 976 /* Handle the first found overlap. After the schedule 977 * we have to restart the tree walk. */ 978 break; 979 } 980 if (!i) /* if any */ 981 break; 982 983 /* Indicate to wake up device->misc_wait on progress. */ 984 prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE); 985 i->waiting = true; 986 spin_unlock_irq(&device->resource->req_lock); 987 schedule(); 988 spin_lock_irq(&device->resource->req_lock); 989 } 990 finish_wait(&device->misc_wait, &wait); 991 } 992 993 /* called within req_lock */ 994 static void maybe_pull_ahead(struct drbd_device *device) 995 { 996 struct drbd_connection *connection = first_peer_device(device)->connection; 997 struct net_conf *nc; 998 bool congested = false; 999 enum drbd_on_congestion on_congestion; 1000 1001 rcu_read_lock(); 1002 nc = rcu_dereference(connection->net_conf); 1003 on_congestion = nc ? nc->on_congestion : OC_BLOCK; 1004 rcu_read_unlock(); 1005 if (on_congestion == OC_BLOCK || 1006 connection->agreed_pro_version < 96) 1007 return; 1008 1009 if (on_congestion == OC_PULL_AHEAD && device->state.conn == C_AHEAD) 1010 return; /* nothing to do ... */ 1011 1012 /* If I don't even have good local storage, we can not reasonably try 1013 * to pull ahead of the peer. We also need the local reference to make 1014 * sure device->act_log is there. 1015 */ 1016 if (!get_ldev_if_state(device, D_UP_TO_DATE)) 1017 return; 1018 1019 if (nc->cong_fill && 1020 atomic_read(&device->ap_in_flight) >= nc->cong_fill) { 1021 drbd_info(device, "Congestion-fill threshold reached\n"); 1022 congested = true; 1023 } 1024 1025 if (device->act_log->used >= nc->cong_extents) { 1026 drbd_info(device, "Congestion-extents threshold reached\n"); 1027 congested = true; 1028 } 1029 1030 if (congested) { 1031 /* start a new epoch for non-mirrored writes */ 1032 start_new_tl_epoch(first_peer_device(device)->connection); 1033 1034 if (on_congestion == OC_PULL_AHEAD) 1035 _drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL); 1036 else /*nc->on_congestion == OC_DISCONNECT */ 1037 _drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL); 1038 } 1039 put_ldev(device); 1040 } 1041 1042 /* If this returns false, and req->private_bio is still set, 1043 * this should be submitted locally. 1044 * 1045 * If it returns false, but req->private_bio is not set, 1046 * we do not have access to good data :( 1047 * 1048 * Otherwise, this destroys req->private_bio, if any, 1049 * and returns true. 1050 */ 1051 static bool do_remote_read(struct drbd_request *req) 1052 { 1053 struct drbd_device *device = req->device; 1054 enum drbd_read_balancing rbm; 1055 1056 if (req->private_bio) { 1057 if (!drbd_may_do_local_read(device, 1058 req->i.sector, req->i.size)) { 1059 bio_put(req->private_bio); 1060 req->private_bio = NULL; 1061 put_ldev(device); 1062 } 1063 } 1064 1065 if (device->state.pdsk != D_UP_TO_DATE) 1066 return false; 1067 1068 if (req->private_bio == NULL) 1069 return true; 1070 1071 /* TODO: improve read balancing decisions, take into account drbd 1072 * protocol, pending requests etc. */ 1073 1074 rcu_read_lock(); 1075 rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing; 1076 rcu_read_unlock(); 1077 1078 if (rbm == RB_PREFER_LOCAL && req->private_bio) 1079 return false; /* submit locally */ 1080 1081 if (remote_due_to_read_balancing(device, req->i.sector, rbm)) { 1082 if (req->private_bio) { 1083 bio_put(req->private_bio); 1084 req->private_bio = NULL; 1085 put_ldev(device); 1086 } 1087 return true; 1088 } 1089 1090 return false; 1091 } 1092 1093 bool drbd_should_do_remote(union drbd_dev_state s) 1094 { 1095 return s.pdsk == D_UP_TO_DATE || 1096 (s.pdsk >= D_INCONSISTENT && 1097 s.conn >= C_WF_BITMAP_T && 1098 s.conn < C_AHEAD); 1099 /* Before proto 96 that was >= CONNECTED instead of >= C_WF_BITMAP_T. 1100 That is equivalent since before 96 IO was frozen in the C_WF_BITMAP* 1101 states. */ 1102 } 1103 1104 static bool drbd_should_send_out_of_sync(union drbd_dev_state s) 1105 { 1106 return s.conn == C_AHEAD || s.conn == C_WF_BITMAP_S; 1107 /* pdsk = D_INCONSISTENT as a consequence. Protocol 96 check not necessary 1108 since we enter state C_AHEAD only if proto >= 96 */ 1109 } 1110 1111 /* returns number of connections (== 1, for drbd 8.4) 1112 * expected to actually write this data, 1113 * which does NOT include those that we are L_AHEAD for. */ 1114 static int drbd_process_write_request(struct drbd_request *req) 1115 { 1116 struct drbd_device *device = req->device; 1117 int remote, send_oos; 1118 1119 remote = drbd_should_do_remote(device->state); 1120 send_oos = drbd_should_send_out_of_sync(device->state); 1121 1122 /* Need to replicate writes. Unless it is an empty flush, 1123 * which is better mapped to a DRBD P_BARRIER packet, 1124 * also for drbd wire protocol compatibility reasons. 1125 * If this was a flush, just start a new epoch. 1126 * Unless the current epoch was empty anyways, or we are not currently 1127 * replicating, in which case there is no point. */ 1128 if (unlikely(req->i.size == 0)) { 1129 /* The only size==0 bios we expect are empty flushes. */ 1130 D_ASSERT(device, req->master_bio->bi_opf & REQ_PREFLUSH); 1131 if (remote) 1132 _req_mod(req, QUEUE_AS_DRBD_BARRIER); 1133 return remote; 1134 } 1135 1136 if (!remote && !send_oos) 1137 return 0; 1138 1139 D_ASSERT(device, !(remote && send_oos)); 1140 1141 if (remote) { 1142 _req_mod(req, TO_BE_SENT); 1143 _req_mod(req, QUEUE_FOR_NET_WRITE); 1144 } else if (drbd_set_out_of_sync(device, req->i.sector, req->i.size)) 1145 _req_mod(req, QUEUE_FOR_SEND_OOS); 1146 1147 return remote; 1148 } 1149 1150 static void drbd_process_discard_req(struct drbd_request *req) 1151 { 1152 struct block_device *bdev = req->device->ldev->backing_bdev; 1153 1154 if (blkdev_issue_zeroout(bdev, req->i.sector, req->i.size >> 9, 1155 GFP_NOIO, 0)) 1156 req->private_bio->bi_error = -EIO; 1157 bio_endio(req->private_bio); 1158 } 1159 1160 static void 1161 drbd_submit_req_private_bio(struct drbd_request *req) 1162 { 1163 struct drbd_device *device = req->device; 1164 struct bio *bio = req->private_bio; 1165 unsigned int type; 1166 1167 if (bio_op(bio) != REQ_OP_READ) 1168 type = DRBD_FAULT_DT_WR; 1169 else if (bio->bi_opf & REQ_RAHEAD) 1170 type = DRBD_FAULT_DT_RA; 1171 else 1172 type = DRBD_FAULT_DT_RD; 1173 1174 bio->bi_bdev = device->ldev->backing_bdev; 1175 1176 /* State may have changed since we grabbed our reference on the 1177 * ->ldev member. Double check, and short-circuit to endio. 1178 * In case the last activity log transaction failed to get on 1179 * stable storage, and this is a WRITE, we may not even submit 1180 * this bio. */ 1181 if (get_ldev(device)) { 1182 if (drbd_insert_fault(device, type)) 1183 bio_io_error(bio); 1184 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES || 1185 bio_op(bio) == REQ_OP_DISCARD) 1186 drbd_process_discard_req(req); 1187 else 1188 generic_make_request(bio); 1189 put_ldev(device); 1190 } else 1191 bio_io_error(bio); 1192 } 1193 1194 static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req) 1195 { 1196 spin_lock_irq(&device->resource->req_lock); 1197 list_add_tail(&req->tl_requests, &device->submit.writes); 1198 list_add_tail(&req->req_pending_master_completion, 1199 &device->pending_master_completion[1 /* WRITE */]); 1200 spin_unlock_irq(&device->resource->req_lock); 1201 queue_work(device->submit.wq, &device->submit.worker); 1202 /* do_submit() may sleep internally on al_wait, too */ 1203 wake_up(&device->al_wait); 1204 } 1205 1206 /* returns the new drbd_request pointer, if the caller is expected to 1207 * drbd_send_and_submit() it (to save latency), or NULL if we queued the 1208 * request on the submitter thread. 1209 * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request. 1210 */ 1211 static struct drbd_request * 1212 drbd_request_prepare(struct drbd_device *device, struct bio *bio, unsigned long start_jif) 1213 { 1214 const int rw = bio_data_dir(bio); 1215 struct drbd_request *req; 1216 1217 /* allocate outside of all locks; */ 1218 req = drbd_req_new(device, bio); 1219 if (!req) { 1220 dec_ap_bio(device); 1221 /* only pass the error to the upper layers. 1222 * if user cannot handle io errors, that's not our business. */ 1223 drbd_err(device, "could not kmalloc() req\n"); 1224 bio->bi_error = -ENOMEM; 1225 bio_endio(bio); 1226 return ERR_PTR(-ENOMEM); 1227 } 1228 req->start_jif = start_jif; 1229 1230 if (!get_ldev(device)) { 1231 bio_put(req->private_bio); 1232 req->private_bio = NULL; 1233 } 1234 1235 /* Update disk stats */ 1236 _drbd_start_io_acct(device, req); 1237 1238 /* process discards always from our submitter thread */ 1239 if ((bio_op(bio) & REQ_OP_WRITE_ZEROES) || 1240 (bio_op(bio) & REQ_OP_DISCARD)) 1241 goto queue_for_submitter_thread; 1242 1243 if (rw == WRITE && req->private_bio && req->i.size 1244 && !test_bit(AL_SUSPENDED, &device->flags)) { 1245 if (!drbd_al_begin_io_fastpath(device, &req->i)) 1246 goto queue_for_submitter_thread; 1247 req->rq_state |= RQ_IN_ACT_LOG; 1248 req->in_actlog_jif = jiffies; 1249 } 1250 return req; 1251 1252 queue_for_submitter_thread: 1253 atomic_inc(&device->ap_actlog_cnt); 1254 drbd_queue_write(device, req); 1255 return NULL; 1256 } 1257 1258 /* Require at least one path to current data. 1259 * We don't want to allow writes on C_STANDALONE D_INCONSISTENT: 1260 * We would not allow to read what was written, 1261 * we would not have bumped the data generation uuids, 1262 * we would cause data divergence for all the wrong reasons. 1263 * 1264 * If we don't see at least one D_UP_TO_DATE, we will fail this request, 1265 * which either returns EIO, or, if OND_SUSPEND_IO is set, suspends IO, 1266 * and queues for retry later. 1267 */ 1268 static bool may_do_writes(struct drbd_device *device) 1269 { 1270 const union drbd_dev_state s = device->state; 1271 return s.disk == D_UP_TO_DATE || s.pdsk == D_UP_TO_DATE; 1272 } 1273 1274 static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req) 1275 { 1276 struct drbd_resource *resource = device->resource; 1277 const int rw = bio_data_dir(req->master_bio); 1278 struct bio_and_error m = { NULL, }; 1279 bool no_remote = false; 1280 bool submit_private_bio = false; 1281 1282 spin_lock_irq(&resource->req_lock); 1283 if (rw == WRITE) { 1284 /* This may temporarily give up the req_lock, 1285 * but will re-aquire it before it returns here. 1286 * Needs to be before the check on drbd_suspended() */ 1287 complete_conflicting_writes(req); 1288 /* no more giving up req_lock from now on! */ 1289 1290 /* check for congestion, and potentially stop sending 1291 * full data updates, but start sending "dirty bits" only. */ 1292 maybe_pull_ahead(device); 1293 } 1294 1295 1296 if (drbd_suspended(device)) { 1297 /* push back and retry: */ 1298 req->rq_state |= RQ_POSTPONED; 1299 if (req->private_bio) { 1300 bio_put(req->private_bio); 1301 req->private_bio = NULL; 1302 put_ldev(device); 1303 } 1304 goto out; 1305 } 1306 1307 /* We fail READ early, if we can not serve it. 1308 * We must do this before req is registered on any lists. 1309 * Otherwise, drbd_req_complete() will queue failed READ for retry. */ 1310 if (rw != WRITE) { 1311 if (!do_remote_read(req) && !req->private_bio) 1312 goto nodata; 1313 } 1314 1315 /* which transfer log epoch does this belong to? */ 1316 req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr); 1317 1318 /* no point in adding empty flushes to the transfer log, 1319 * they are mapped to drbd barriers already. */ 1320 if (likely(req->i.size!=0)) { 1321 if (rw == WRITE) 1322 first_peer_device(device)->connection->current_tle_writes++; 1323 1324 list_add_tail(&req->tl_requests, &first_peer_device(device)->connection->transfer_log); 1325 } 1326 1327 if (rw == WRITE) { 1328 if (req->private_bio && !may_do_writes(device)) { 1329 bio_put(req->private_bio); 1330 req->private_bio = NULL; 1331 put_ldev(device); 1332 goto nodata; 1333 } 1334 if (!drbd_process_write_request(req)) 1335 no_remote = true; 1336 } else { 1337 /* We either have a private_bio, or we can read from remote. 1338 * Otherwise we had done the goto nodata above. */ 1339 if (req->private_bio == NULL) { 1340 _req_mod(req, TO_BE_SENT); 1341 _req_mod(req, QUEUE_FOR_NET_READ); 1342 } else 1343 no_remote = true; 1344 } 1345 1346 /* If it took the fast path in drbd_request_prepare, add it here. 1347 * The slow path has added it already. */ 1348 if (list_empty(&req->req_pending_master_completion)) 1349 list_add_tail(&req->req_pending_master_completion, 1350 &device->pending_master_completion[rw == WRITE]); 1351 if (req->private_bio) { 1352 /* needs to be marked within the same spinlock */ 1353 req->pre_submit_jif = jiffies; 1354 list_add_tail(&req->req_pending_local, 1355 &device->pending_completion[rw == WRITE]); 1356 _req_mod(req, TO_BE_SUBMITTED); 1357 /* but we need to give up the spinlock to submit */ 1358 submit_private_bio = true; 1359 } else if (no_remote) { 1360 nodata: 1361 if (__ratelimit(&drbd_ratelimit_state)) 1362 drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n", 1363 (unsigned long long)req->i.sector, req->i.size >> 9); 1364 /* A write may have been queued for send_oos, however. 1365 * So we can not simply free it, we must go through drbd_req_put_completion_ref() */ 1366 } 1367 1368 out: 1369 if (drbd_req_put_completion_ref(req, &m, 1)) 1370 kref_put(&req->kref, drbd_req_destroy); 1371 spin_unlock_irq(&resource->req_lock); 1372 1373 /* Even though above is a kref_put(), this is safe. 1374 * As long as we still need to submit our private bio, 1375 * we hold a completion ref, and the request cannot disappear. 1376 * If however this request did not even have a private bio to submit 1377 * (e.g. remote read), req may already be invalid now. 1378 * That's why we cannot check on req->private_bio. */ 1379 if (submit_private_bio) 1380 drbd_submit_req_private_bio(req); 1381 if (m.bio) 1382 complete_master_bio(device, &m); 1383 } 1384 1385 void __drbd_make_request(struct drbd_device *device, struct bio *bio, unsigned long start_jif) 1386 { 1387 struct drbd_request *req = drbd_request_prepare(device, bio, start_jif); 1388 if (IS_ERR_OR_NULL(req)) 1389 return; 1390 drbd_send_and_submit(device, req); 1391 } 1392 1393 static void submit_fast_path(struct drbd_device *device, struct list_head *incoming) 1394 { 1395 struct drbd_request *req, *tmp; 1396 list_for_each_entry_safe(req, tmp, incoming, tl_requests) { 1397 const int rw = bio_data_dir(req->master_bio); 1398 1399 if (rw == WRITE /* rw != WRITE should not even end up here! */ 1400 && req->private_bio && req->i.size 1401 && !test_bit(AL_SUSPENDED, &device->flags)) { 1402 if (!drbd_al_begin_io_fastpath(device, &req->i)) 1403 continue; 1404 1405 req->rq_state |= RQ_IN_ACT_LOG; 1406 req->in_actlog_jif = jiffies; 1407 atomic_dec(&device->ap_actlog_cnt); 1408 } 1409 1410 list_del_init(&req->tl_requests); 1411 drbd_send_and_submit(device, req); 1412 } 1413 } 1414 1415 static bool prepare_al_transaction_nonblock(struct drbd_device *device, 1416 struct list_head *incoming, 1417 struct list_head *pending, 1418 struct list_head *later) 1419 { 1420 struct drbd_request *req, *tmp; 1421 int wake = 0; 1422 int err; 1423 1424 spin_lock_irq(&device->al_lock); 1425 list_for_each_entry_safe(req, tmp, incoming, tl_requests) { 1426 err = drbd_al_begin_io_nonblock(device, &req->i); 1427 if (err == -ENOBUFS) 1428 break; 1429 if (err == -EBUSY) 1430 wake = 1; 1431 if (err) 1432 list_move_tail(&req->tl_requests, later); 1433 else 1434 list_move_tail(&req->tl_requests, pending); 1435 } 1436 spin_unlock_irq(&device->al_lock); 1437 if (wake) 1438 wake_up(&device->al_wait); 1439 return !list_empty(pending); 1440 } 1441 1442 void send_and_submit_pending(struct drbd_device *device, struct list_head *pending) 1443 { 1444 struct drbd_request *req, *tmp; 1445 1446 list_for_each_entry_safe(req, tmp, pending, tl_requests) { 1447 req->rq_state |= RQ_IN_ACT_LOG; 1448 req->in_actlog_jif = jiffies; 1449 atomic_dec(&device->ap_actlog_cnt); 1450 list_del_init(&req->tl_requests); 1451 drbd_send_and_submit(device, req); 1452 } 1453 } 1454 1455 void do_submit(struct work_struct *ws) 1456 { 1457 struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker); 1458 LIST_HEAD(incoming); /* from drbd_make_request() */ 1459 LIST_HEAD(pending); /* to be submitted after next AL-transaction commit */ 1460 LIST_HEAD(busy); /* blocked by resync requests */ 1461 1462 /* grab new incoming requests */ 1463 spin_lock_irq(&device->resource->req_lock); 1464 list_splice_tail_init(&device->submit.writes, &incoming); 1465 spin_unlock_irq(&device->resource->req_lock); 1466 1467 for (;;) { 1468 DEFINE_WAIT(wait); 1469 1470 /* move used-to-be-busy back to front of incoming */ 1471 list_splice_init(&busy, &incoming); 1472 submit_fast_path(device, &incoming); 1473 if (list_empty(&incoming)) 1474 break; 1475 1476 for (;;) { 1477 prepare_to_wait(&device->al_wait, &wait, TASK_UNINTERRUPTIBLE); 1478 1479 list_splice_init(&busy, &incoming); 1480 prepare_al_transaction_nonblock(device, &incoming, &pending, &busy); 1481 if (!list_empty(&pending)) 1482 break; 1483 1484 schedule(); 1485 1486 /* If all currently "hot" activity log extents are kept busy by 1487 * incoming requests, we still must not totally starve new 1488 * requests to "cold" extents. 1489 * Something left on &incoming means there had not been 1490 * enough update slots available, and the activity log 1491 * has been marked as "starving". 1492 * 1493 * Try again now, without looking for new requests, 1494 * effectively blocking all new requests until we made 1495 * at least _some_ progress with what we currently have. 1496 */ 1497 if (!list_empty(&incoming)) 1498 continue; 1499 1500 /* Nothing moved to pending, but nothing left 1501 * on incoming: all moved to busy! 1502 * Grab new and iterate. */ 1503 spin_lock_irq(&device->resource->req_lock); 1504 list_splice_tail_init(&device->submit.writes, &incoming); 1505 spin_unlock_irq(&device->resource->req_lock); 1506 } 1507 finish_wait(&device->al_wait, &wait); 1508 1509 /* If the transaction was full, before all incoming requests 1510 * had been processed, skip ahead to commit, and iterate 1511 * without splicing in more incoming requests from upper layers. 1512 * 1513 * Else, if all incoming have been processed, 1514 * they have become either "pending" (to be submitted after 1515 * next transaction commit) or "busy" (blocked by resync). 1516 * 1517 * Maybe more was queued, while we prepared the transaction? 1518 * Try to stuff those into this transaction as well. 1519 * Be strictly non-blocking here, 1520 * we already have something to commit. 1521 * 1522 * Commit if we don't make any more progres. 1523 */ 1524 1525 while (list_empty(&incoming)) { 1526 LIST_HEAD(more_pending); 1527 LIST_HEAD(more_incoming); 1528 bool made_progress; 1529 1530 /* It is ok to look outside the lock, 1531 * it's only an optimization anyways */ 1532 if (list_empty(&device->submit.writes)) 1533 break; 1534 1535 spin_lock_irq(&device->resource->req_lock); 1536 list_splice_tail_init(&device->submit.writes, &more_incoming); 1537 spin_unlock_irq(&device->resource->req_lock); 1538 1539 if (list_empty(&more_incoming)) 1540 break; 1541 1542 made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending, &busy); 1543 1544 list_splice_tail_init(&more_pending, &pending); 1545 list_splice_tail_init(&more_incoming, &incoming); 1546 if (!made_progress) 1547 break; 1548 } 1549 1550 drbd_al_begin_io_commit(device); 1551 send_and_submit_pending(device, &pending); 1552 } 1553 } 1554 1555 blk_qc_t drbd_make_request(struct request_queue *q, struct bio *bio) 1556 { 1557 struct drbd_device *device = (struct drbd_device *) q->queuedata; 1558 unsigned long start_jif; 1559 1560 blk_queue_split(q, &bio, q->bio_split); 1561 1562 start_jif = jiffies; 1563 1564 /* 1565 * what we "blindly" assume: 1566 */ 1567 D_ASSERT(device, IS_ALIGNED(bio->bi_iter.bi_size, 512)); 1568 1569 inc_ap_bio(device); 1570 __drbd_make_request(device, bio, start_jif); 1571 return BLK_QC_T_NONE; 1572 } 1573 1574 static bool net_timeout_reached(struct drbd_request *net_req, 1575 struct drbd_connection *connection, 1576 unsigned long now, unsigned long ent, 1577 unsigned int ko_count, unsigned int timeout) 1578 { 1579 struct drbd_device *device = net_req->device; 1580 1581 if (!time_after(now, net_req->pre_send_jif + ent)) 1582 return false; 1583 1584 if (time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent)) 1585 return false; 1586 1587 if (net_req->rq_state & RQ_NET_PENDING) { 1588 drbd_warn(device, "Remote failed to finish a request within %ums > ko-count (%u) * timeout (%u * 0.1s)\n", 1589 jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout); 1590 return true; 1591 } 1592 1593 /* We received an ACK already (or are using protocol A), 1594 * but are waiting for the epoch closing barrier ack. 1595 * Check if we sent the barrier already. We should not blame the peer 1596 * for being unresponsive, if we did not even ask it yet. */ 1597 if (net_req->epoch == connection->send.current_epoch_nr) { 1598 drbd_warn(device, 1599 "We did not send a P_BARRIER for %ums > ko-count (%u) * timeout (%u * 0.1s); drbd kernel thread blocked?\n", 1600 jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout); 1601 return false; 1602 } 1603 1604 /* Worst case: we may have been blocked for whatever reason, then 1605 * suddenly are able to send a lot of requests (and epoch separating 1606 * barriers) in quick succession. 1607 * The timestamp of the net_req may be much too old and not correspond 1608 * to the sending time of the relevant unack'ed barrier packet, so 1609 * would trigger a spurious timeout. The latest barrier packet may 1610 * have a too recent timestamp to trigger the timeout, potentially miss 1611 * a timeout. Right now we don't have a place to conveniently store 1612 * these timestamps. 1613 * But in this particular situation, the application requests are still 1614 * completed to upper layers, DRBD should still "feel" responsive. 1615 * No need yet to kill this connection, it may still recover. 1616 * If not, eventually we will have queued enough into the network for 1617 * us to block. From that point of view, the timestamp of the last sent 1618 * barrier packet is relevant enough. 1619 */ 1620 if (time_after(now, connection->send.last_sent_barrier_jif + ent)) { 1621 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", 1622 connection->send.last_sent_barrier_jif, now, 1623 jiffies_to_msecs(now - connection->send.last_sent_barrier_jif), ko_count, timeout); 1624 return true; 1625 } 1626 return false; 1627 } 1628 1629 /* A request is considered timed out, if 1630 * - we have some effective timeout from the configuration, 1631 * with some state restrictions applied, 1632 * - the oldest request is waiting for a response from the network 1633 * resp. the local disk, 1634 * - the oldest request is in fact older than the effective timeout, 1635 * - the connection was established (resp. disk was attached) 1636 * for longer than the timeout already. 1637 * Note that for 32bit jiffies and very stable connections/disks, 1638 * we may have a wrap around, which is catched by 1639 * !time_in_range(now, last_..._jif, last_..._jif + timeout). 1640 * 1641 * Side effect: once per 32bit wrap-around interval, which means every 1642 * ~198 days with 250 HZ, we have a window where the timeout would need 1643 * to expire twice (worst case) to become effective. Good enough. 1644 */ 1645 1646 void request_timer_fn(unsigned long data) 1647 { 1648 struct drbd_device *device = (struct drbd_device *) data; 1649 struct drbd_connection *connection = first_peer_device(device)->connection; 1650 struct drbd_request *req_read, *req_write, *req_peer; /* oldest request */ 1651 struct net_conf *nc; 1652 unsigned long oldest_submit_jif; 1653 unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */ 1654 unsigned long now; 1655 unsigned int ko_count = 0, timeout = 0; 1656 1657 rcu_read_lock(); 1658 nc = rcu_dereference(connection->net_conf); 1659 if (nc && device->state.conn >= C_WF_REPORT_PARAMS) { 1660 ko_count = nc->ko_count; 1661 timeout = nc->timeout; 1662 } 1663 1664 if (get_ldev(device)) { /* implicit state.disk >= D_INCONSISTENT */ 1665 dt = rcu_dereference(device->ldev->disk_conf)->disk_timeout * HZ / 10; 1666 put_ldev(device); 1667 } 1668 rcu_read_unlock(); 1669 1670 1671 ent = timeout * HZ/10 * ko_count; 1672 et = min_not_zero(dt, ent); 1673 1674 if (!et) 1675 return; /* Recurring timer stopped */ 1676 1677 now = jiffies; 1678 nt = now + et; 1679 1680 spin_lock_irq(&device->resource->req_lock); 1681 req_read = list_first_entry_or_null(&device->pending_completion[0], struct drbd_request, req_pending_local); 1682 req_write = list_first_entry_or_null(&device->pending_completion[1], struct drbd_request, req_pending_local); 1683 1684 /* maybe the oldest request waiting for the peer is in fact still 1685 * blocking in tcp sendmsg. That's ok, though, that's handled via the 1686 * socket send timeout, requesting a ping, and bumping ko-count in 1687 * we_should_drop_the_connection(). 1688 */ 1689 1690 /* check the oldest request we did successfully sent, 1691 * but which is still waiting for an ACK. */ 1692 req_peer = connection->req_ack_pending; 1693 1694 /* if we don't have such request (e.g. protocoll A) 1695 * check the oldest requests which is still waiting on its epoch 1696 * closing barrier ack. */ 1697 if (!req_peer) 1698 req_peer = connection->req_not_net_done; 1699 1700 /* evaluate the oldest peer request only in one timer! */ 1701 if (req_peer && req_peer->device != device) 1702 req_peer = NULL; 1703 1704 /* do we have something to evaluate? */ 1705 if (req_peer == NULL && req_write == NULL && req_read == NULL) 1706 goto out; 1707 1708 oldest_submit_jif = 1709 (req_write && req_read) 1710 ? ( time_before(req_write->pre_submit_jif, req_read->pre_submit_jif) 1711 ? req_write->pre_submit_jif : req_read->pre_submit_jif ) 1712 : req_write ? req_write->pre_submit_jif 1713 : req_read ? req_read->pre_submit_jif : now; 1714 1715 if (ent && req_peer && net_timeout_reached(req_peer, connection, now, ent, ko_count, timeout)) 1716 _conn_request_state(connection, NS(conn, C_TIMEOUT), CS_VERBOSE | CS_HARD); 1717 1718 if (dt && oldest_submit_jif != now && 1719 time_after(now, oldest_submit_jif + dt) && 1720 !time_in_range(now, device->last_reattach_jif, device->last_reattach_jif + dt)) { 1721 drbd_warn(device, "Local backing device failed to meet the disk-timeout\n"); 1722 __drbd_chk_io_error(device, DRBD_FORCE_DETACH); 1723 } 1724 1725 /* Reschedule timer for the nearest not already expired timeout. 1726 * Fallback to now + min(effective network timeout, disk timeout). */ 1727 ent = (ent && req_peer && time_before(now, req_peer->pre_send_jif + ent)) 1728 ? req_peer->pre_send_jif + ent : now + et; 1729 dt = (dt && oldest_submit_jif != now && time_before(now, oldest_submit_jif + dt)) 1730 ? oldest_submit_jif + dt : now + et; 1731 nt = time_before(ent, dt) ? ent : dt; 1732 out: 1733 spin_unlock_irq(&device->resource->req_lock); 1734 mod_timer(&device->request_timer, nt); 1735 } 1736