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