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