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_alloc_clone(device->ldev->backing_bdev, bio_src, 34 GFP_NOIO, &drbd_io_bio_set); 35 req->private_bio->bi_private = req; 36 req->private_bio->bi_end_io = drbd_request_endio; 37 38 req->rq_state = (bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0) 39 | (bio_op(bio_src) == REQ_OP_WRITE_ZEROES ? RQ_ZEROES : 0) 40 | (bio_op(bio_src) == REQ_OP_DISCARD ? RQ_UNMAP : 0); 41 req->device = device; 42 req->master_bio = bio_src; 43 req->epoch = 0; 44 45 drbd_clear_interval(&req->i); 46 req->i.sector = bio_src->bi_iter.bi_sector; 47 req->i.size = bio_src->bi_iter.bi_size; 48 req->i.local = true; 49 req->i.waiting = false; 50 51 INIT_LIST_HEAD(&req->tl_requests); 52 INIT_LIST_HEAD(&req->w.list); 53 INIT_LIST_HEAD(&req->req_pending_master_completion); 54 INIT_LIST_HEAD(&req->req_pending_local); 55 56 /* one reference to be put by __drbd_make_request */ 57 atomic_set(&req->completion_ref, 1); 58 /* one kref as long as completion_ref > 0 */ 59 kref_init(&req->kref); 60 return req; 61 } 62 63 static void drbd_remove_request_interval(struct rb_root *root, 64 struct drbd_request *req) 65 { 66 struct drbd_device *device = req->device; 67 struct drbd_interval *i = &req->i; 68 69 drbd_remove_interval(root, i); 70 71 /* Wake up any processes waiting for this request to complete. */ 72 if (i->waiting) 73 wake_up(&device->misc_wait); 74 } 75 76 void drbd_req_destroy(struct kref *kref) 77 { 78 struct drbd_request *req = container_of(kref, struct drbd_request, kref); 79 struct drbd_device *device = req->device; 80 const unsigned s = req->rq_state; 81 82 if ((req->master_bio && !(s & RQ_POSTPONED)) || 83 atomic_read(&req->completion_ref) || 84 (s & RQ_LOCAL_PENDING) || 85 ((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) { 86 drbd_err(device, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n", 87 s, atomic_read(&req->completion_ref)); 88 return; 89 } 90 91 /* If called from mod_rq_state (expected normal case) or 92 * drbd_send_and_submit (the less likely normal path), this holds the 93 * req_lock, and req->tl_requests will typicaly be on ->transfer_log, 94 * though it may be still empty (never added to the transfer log). 95 * 96 * If called from do_retry(), we do NOT hold the req_lock, but we are 97 * still allowed to unconditionally list_del(&req->tl_requests), 98 * because it will be on a local on-stack list only. */ 99 list_del_init(&req->tl_requests); 100 101 /* finally remove the request from the conflict detection 102 * respective block_id verification interval tree. */ 103 if (!drbd_interval_empty(&req->i)) { 104 struct rb_root *root; 105 106 if (s & RQ_WRITE) 107 root = &device->write_requests; 108 else 109 root = &device->read_requests; 110 drbd_remove_request_interval(root, req); 111 } else if (s & (RQ_NET_MASK & ~RQ_NET_DONE) && req->i.size != 0) 112 drbd_err(device, "drbd_req_destroy: Logic BUG: interval empty, but: rq_state=0x%x, sect=%llu, size=%u\n", 113 s, (unsigned long long)req->i.sector, req->i.size); 114 115 /* if it was a write, we may have to set the corresponding 116 * bit(s) out-of-sync first. If it had a local part, we need to 117 * release the reference to the activity log. */ 118 if (s & RQ_WRITE) { 119 /* Set out-of-sync unless both OK flags are set 120 * (local only or remote failed). 121 * Other places where we set out-of-sync: 122 * READ with local io-error */ 123 124 /* There is a special case: 125 * we may notice late that IO was suspended, 126 * and postpone, or schedule for retry, a write, 127 * before it even was submitted or sent. 128 * In that case we do not want to touch the bitmap at all. 129 */ 130 if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) { 131 if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK)) 132 drbd_set_out_of_sync(device, req->i.sector, req->i.size); 133 134 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS)) 135 drbd_set_in_sync(device, req->i.sector, req->i.size); 136 } 137 138 /* one might be tempted to move the drbd_al_complete_io 139 * to the local io completion callback drbd_request_endio. 140 * but, if this was a mirror write, we may only 141 * drbd_al_complete_io after this is RQ_NET_DONE, 142 * otherwise the extent could be dropped from the al 143 * before it has actually been written on the peer. 144 * if we crash before our peer knows about the request, 145 * but after the extent has been dropped from the al, 146 * we would forget to resync the corresponding extent. 147 */ 148 if (s & RQ_IN_ACT_LOG) { 149 if (get_ldev_if_state(device, D_FAILED)) { 150 drbd_al_complete_io(device, &req->i); 151 put_ldev(device); 152 } else if (__ratelimit(&drbd_ratelimit_state)) { 153 drbd_warn(device, "Should have called drbd_al_complete_io(, %llu, %u), " 154 "but my Disk seems to have failed :(\n", 155 (unsigned long long) req->i.sector, req->i.size); 156 } 157 } 158 } 159 160 mempool_free(req, &drbd_request_mempool); 161 } 162 163 static void wake_all_senders(struct drbd_connection *connection) 164 { 165 wake_up(&connection->sender_work.q_wait); 166 } 167 168 /* must hold resource->req_lock */ 169 void start_new_tl_epoch(struct drbd_connection *connection) 170 { 171 /* no point closing an epoch, if it is empty, anyways. */ 172 if (connection->current_tle_writes == 0) 173 return; 174 175 connection->current_tle_writes = 0; 176 atomic_inc(&connection->current_tle_nr); 177 wake_all_senders(connection); 178 } 179 180 void complete_master_bio(struct drbd_device *device, 181 struct bio_and_error *m) 182 { 183 if (unlikely(m->error)) 184 m->bio->bi_status = errno_to_blk_status(m->error); 185 bio_endio(m->bio); 186 dec_ap_bio(device); 187 } 188 189 190 /* Helper for __req_mod(). 191 * Set m->bio to the master bio, if it is fit to be completed, 192 * or leave it alone (it is initialized to NULL in __req_mod), 193 * if it has already been completed, or cannot be completed yet. 194 * If m->bio is set, the error status to be returned is placed in m->error. 195 */ 196 static 197 void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m) 198 { 199 const unsigned s = req->rq_state; 200 struct drbd_device *device = req->device; 201 int error, ok; 202 203 /* we must not complete the master bio, while it is 204 * still being processed by _drbd_send_zc_bio (drbd_send_dblock) 205 * not yet acknowledged by the peer 206 * not yet completed by the local io subsystem 207 * these flags may get cleared in any order by 208 * the worker, 209 * the receiver, 210 * the bio_endio completion callbacks. 211 */ 212 if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) || 213 (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) || 214 (s & RQ_COMPLETION_SUSP)) { 215 drbd_err(device, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s); 216 return; 217 } 218 219 if (!req->master_bio) { 220 drbd_err(device, "drbd_req_complete: Logic BUG, master_bio == NULL!\n"); 221 return; 222 } 223 224 /* 225 * figure out whether to report success or failure. 226 * 227 * report success when at least one of the operations succeeded. 228 * or, to put the other way, 229 * only report failure, when both operations failed. 230 * 231 * what to do about the failures is handled elsewhere. 232 * what we need to do here is just: complete the master_bio. 233 * 234 * local completion error, if any, has been stored as ERR_PTR 235 * in private_bio within drbd_request_endio. 236 */ 237 ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK); 238 error = PTR_ERR(req->private_bio); 239 240 /* Before we can signal completion to the upper layers, 241 * we may need to close the current transfer log epoch. 242 * We are within the request lock, so we can simply compare 243 * the request epoch number with the current transfer log 244 * epoch number. If they match, increase the current_tle_nr, 245 * and reset the transfer log epoch write_cnt. 246 */ 247 if (op_is_write(bio_op(req->master_bio)) && 248 req->epoch == atomic_read(&first_peer_device(device)->connection->current_tle_nr)) 249 start_new_tl_epoch(first_peer_device(device)->connection); 250 251 /* Update disk stats */ 252 bio_end_io_acct(req->master_bio, req->start_jif); 253 254 /* If READ failed, 255 * have it be pushed back to the retry work queue, 256 * so it will re-enter __drbd_make_request(), 257 * and be re-assigned to a suitable local or remote path, 258 * or failed if we do not have access to good data anymore. 259 * 260 * Unless it was failed early by __drbd_make_request(), 261 * because no path was available, in which case 262 * it was not even added to the transfer_log. 263 * 264 * read-ahead may fail, and will not be retried. 265 * 266 * WRITE should have used all available paths already. 267 */ 268 if (!ok && 269 bio_op(req->master_bio) == REQ_OP_READ && 270 !(req->master_bio->bi_opf & REQ_RAHEAD) && 271 !list_empty(&req->tl_requests)) 272 req->rq_state |= RQ_POSTPONED; 273 274 if (!(req->rq_state & RQ_POSTPONED)) { 275 m->error = ok ? 0 : (error ?: -EIO); 276 m->bio = req->master_bio; 277 req->master_bio = NULL; 278 /* We leave it in the tree, to be able to verify later 279 * write-acks in protocol != C during resync. 280 * But we mark it as "complete", so it won't be counted as 281 * conflict in a multi-primary setup. */ 282 req->i.completed = true; 283 } 284 285 if (req->i.waiting) 286 wake_up(&device->misc_wait); 287 288 /* Either we are about to complete to upper layers, 289 * or we will restart this request. 290 * In either case, the request object will be destroyed soon, 291 * so better remove it from all lists. */ 292 list_del_init(&req->req_pending_master_completion); 293 } 294 295 /* still holds resource->req_lock */ 296 static void drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put) 297 { 298 struct drbd_device *device = req->device; 299 D_ASSERT(device, m || (req->rq_state & RQ_POSTPONED)); 300 301 if (!put) 302 return; 303 304 if (!atomic_sub_and_test(put, &req->completion_ref)) 305 return; 306 307 drbd_req_complete(req, m); 308 309 /* local completion may still come in later, 310 * we need to keep the req object around. */ 311 if (req->rq_state & RQ_LOCAL_ABORTED) 312 return; 313 314 if (req->rq_state & RQ_POSTPONED) { 315 /* don't destroy the req object just yet, 316 * but queue it for retry */ 317 drbd_restart_request(req); 318 return; 319 } 320 321 kref_put(&req->kref, drbd_req_destroy); 322 } 323 324 static void set_if_null_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req) 325 { 326 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; 327 if (!connection) 328 return; 329 if (connection->req_next == NULL) 330 connection->req_next = req; 331 } 332 333 static void advance_conn_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req) 334 { 335 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; 336 struct drbd_request *iter = req; 337 if (!connection) 338 return; 339 if (connection->req_next != req) 340 return; 341 342 req = NULL; 343 list_for_each_entry_continue(iter, &connection->transfer_log, tl_requests) { 344 const unsigned int s = iter->rq_state; 345 346 if (s & RQ_NET_QUEUED) { 347 req = iter; 348 break; 349 } 350 } 351 connection->req_next = req; 352 } 353 354 static void set_if_null_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req) 355 { 356 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; 357 if (!connection) 358 return; 359 if (connection->req_ack_pending == NULL) 360 connection->req_ack_pending = req; 361 } 362 363 static void advance_conn_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 struct drbd_request *iter = req; 367 if (!connection) 368 return; 369 if (connection->req_ack_pending != req) 370 return; 371 372 req = NULL; 373 list_for_each_entry_continue(iter, &connection->transfer_log, tl_requests) { 374 const unsigned int s = iter->rq_state; 375 376 if ((s & RQ_NET_SENT) && (s & RQ_NET_PENDING)) { 377 req = iter; 378 break; 379 } 380 } 381 connection->req_ack_pending = req; 382 } 383 384 static void set_if_null_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req) 385 { 386 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; 387 if (!connection) 388 return; 389 if (connection->req_not_net_done == NULL) 390 connection->req_not_net_done = req; 391 } 392 393 static void advance_conn_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req) 394 { 395 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; 396 struct drbd_request *iter = req; 397 if (!connection) 398 return; 399 if (connection->req_not_net_done != req) 400 return; 401 402 req = NULL; 403 list_for_each_entry_continue(iter, &connection->transfer_log, tl_requests) { 404 const unsigned int s = iter->rq_state; 405 406 if ((s & RQ_NET_SENT) && !(s & RQ_NET_DONE)) { 407 req = iter; 408 break; 409 } 410 } 411 connection->req_not_net_done = req; 412 } 413 414 /* I'd like this to be the only place that manipulates 415 * req->completion_ref and req->kref. */ 416 static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m, 417 int clear, int set) 418 { 419 struct drbd_device *device = req->device; 420 struct drbd_peer_device *peer_device = first_peer_device(device); 421 unsigned s = req->rq_state; 422 int c_put = 0; 423 424 if (drbd_suspended(device) && !((s | clear) & RQ_COMPLETION_SUSP)) 425 set |= RQ_COMPLETION_SUSP; 426 427 /* apply */ 428 429 req->rq_state &= ~clear; 430 req->rq_state |= set; 431 432 /* no change? */ 433 if (req->rq_state == s) 434 return; 435 436 /* intent: get references */ 437 438 kref_get(&req->kref); 439 440 if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING)) 441 atomic_inc(&req->completion_ref); 442 443 if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) { 444 inc_ap_pending(device); 445 atomic_inc(&req->completion_ref); 446 } 447 448 if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED)) { 449 atomic_inc(&req->completion_ref); 450 set_if_null_req_next(peer_device, req); 451 } 452 453 if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK)) 454 kref_get(&req->kref); /* wait for the DONE */ 455 456 if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT)) { 457 /* potentially already completed in the ack_receiver thread */ 458 if (!(s & RQ_NET_DONE)) { 459 atomic_add(req->i.size >> 9, &device->ap_in_flight); 460 set_if_null_req_not_net_done(peer_device, req); 461 } 462 if (req->rq_state & RQ_NET_PENDING) 463 set_if_null_req_ack_pending(peer_device, req); 464 } 465 466 if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP)) 467 atomic_inc(&req->completion_ref); 468 469 /* progress: put references */ 470 471 if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP)) 472 ++c_put; 473 474 if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) { 475 D_ASSERT(device, req->rq_state & RQ_LOCAL_PENDING); 476 ++c_put; 477 } 478 479 if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) { 480 if (req->rq_state & RQ_LOCAL_ABORTED) 481 kref_put(&req->kref, drbd_req_destroy); 482 else 483 ++c_put; 484 list_del_init(&req->req_pending_local); 485 } 486 487 if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) { 488 dec_ap_pending(device); 489 ++c_put; 490 req->acked_jif = jiffies; 491 advance_conn_req_ack_pending(peer_device, req); 492 } 493 494 if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED)) { 495 ++c_put; 496 advance_conn_req_next(peer_device, req); 497 } 498 499 if (!(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) { 500 if (s & RQ_NET_SENT) 501 atomic_sub(req->i.size >> 9, &device->ap_in_flight); 502 if (s & RQ_EXP_BARR_ACK) 503 kref_put(&req->kref, drbd_req_destroy); 504 req->net_done_jif = jiffies; 505 506 /* in ahead/behind mode, or just in case, 507 * before we finally destroy this request, 508 * the caching pointers must not reference it anymore */ 509 advance_conn_req_next(peer_device, req); 510 advance_conn_req_ack_pending(peer_device, req); 511 advance_conn_req_not_net_done(peer_device, req); 512 } 513 514 /* potentially complete and destroy */ 515 516 /* If we made progress, retry conflicting peer requests, if any. */ 517 if (req->i.waiting) 518 wake_up(&device->misc_wait); 519 520 drbd_req_put_completion_ref(req, m, c_put); 521 kref_put(&req->kref, drbd_req_destroy); 522 } 523 524 static void drbd_report_io_error(struct drbd_device *device, struct drbd_request *req) 525 { 526 if (!__ratelimit(&drbd_ratelimit_state)) 527 return; 528 529 drbd_warn(device, "local %s IO error sector %llu+%u on %pg\n", 530 (req->rq_state & RQ_WRITE) ? "WRITE" : "READ", 531 (unsigned long long)req->i.sector, 532 req->i.size >> 9, 533 device->ldev->backing_bdev); 534 } 535 536 /* Helper for HANDED_OVER_TO_NETWORK. 537 * Is this a protocol A write (neither WRITE_ACK nor RECEIVE_ACK expected)? 538 * Is it also still "PENDING"? 539 * --> If so, clear PENDING and set NET_OK below. 540 * If it is a protocol A write, but not RQ_PENDING anymore, neg-ack was faster 541 * (and we must not set RQ_NET_OK) */ 542 static inline bool is_pending_write_protocol_A(struct drbd_request *req) 543 { 544 return (req->rq_state & 545 (RQ_WRITE|RQ_NET_PENDING|RQ_EXP_WRITE_ACK|RQ_EXP_RECEIVE_ACK)) 546 == (RQ_WRITE|RQ_NET_PENDING); 547 } 548 549 /* obviously this could be coded as many single functions 550 * instead of one huge switch, 551 * or by putting the code directly in the respective locations 552 * (as it has been before). 553 * 554 * but having it this way 555 * enforces that it is all in this one place, where it is easier to audit, 556 * it makes it obvious that whatever "event" "happens" to a request should 557 * happen "atomically" within the req_lock, 558 * and it enforces that we have to think in a very structured manner 559 * about the "events" that may happen to a request during its life time ... 560 */ 561 int __req_mod(struct drbd_request *req, enum drbd_req_event what, 562 struct bio_and_error *m) 563 { 564 struct drbd_device *const device = req->device; 565 struct drbd_peer_device *const peer_device = first_peer_device(device); 566 struct drbd_connection *const connection = peer_device ? peer_device->connection : NULL; 567 struct net_conf *nc; 568 int p, rv = 0; 569 570 if (m) 571 m->bio = NULL; 572 573 switch (what) { 574 default: 575 drbd_err(device, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__); 576 break; 577 578 /* does not happen... 579 * initialization done in drbd_req_new 580 case CREATED: 581 break; 582 */ 583 584 case TO_BE_SENT: /* via network */ 585 /* reached via __drbd_make_request 586 * and from w_read_retry_remote */ 587 D_ASSERT(device, !(req->rq_state & RQ_NET_MASK)); 588 rcu_read_lock(); 589 nc = rcu_dereference(connection->net_conf); 590 p = nc->wire_protocol; 591 rcu_read_unlock(); 592 req->rq_state |= 593 p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK : 594 p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0; 595 mod_rq_state(req, m, 0, RQ_NET_PENDING); 596 break; 597 598 case TO_BE_SUBMITTED: /* locally */ 599 /* reached via __drbd_make_request */ 600 D_ASSERT(device, !(req->rq_state & RQ_LOCAL_MASK)); 601 mod_rq_state(req, m, 0, RQ_LOCAL_PENDING); 602 break; 603 604 case COMPLETED_OK: 605 if (req->rq_state & RQ_WRITE) 606 device->writ_cnt += req->i.size >> 9; 607 else 608 device->read_cnt += req->i.size >> 9; 609 610 mod_rq_state(req, m, RQ_LOCAL_PENDING, 611 RQ_LOCAL_COMPLETED|RQ_LOCAL_OK); 612 break; 613 614 case ABORT_DISK_IO: 615 mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED); 616 break; 617 618 case WRITE_COMPLETED_WITH_ERROR: 619 drbd_report_io_error(device, req); 620 __drbd_chk_io_error(device, DRBD_WRITE_ERROR); 621 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED); 622 break; 623 624 case READ_COMPLETED_WITH_ERROR: 625 drbd_set_out_of_sync(device, req->i.sector, req->i.size); 626 drbd_report_io_error(device, req); 627 __drbd_chk_io_error(device, DRBD_READ_ERROR); 628 fallthrough; 629 case READ_AHEAD_COMPLETED_WITH_ERROR: 630 /* it is legal to fail read-ahead, no __drbd_chk_io_error in that case. */ 631 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED); 632 break; 633 634 case DISCARD_COMPLETED_NOTSUPP: 635 case DISCARD_COMPLETED_WITH_ERROR: 636 /* I'd rather not detach from local disk just because it 637 * failed a REQ_OP_DISCARD. */ 638 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED); 639 break; 640 641 case QUEUE_FOR_NET_READ: 642 /* READ, and 643 * no local disk, 644 * or target area marked as invalid, 645 * or just got an io-error. */ 646 /* from __drbd_make_request 647 * or from bio_endio during read io-error recovery */ 648 649 /* So we can verify the handle in the answer packet. 650 * Corresponding drbd_remove_request_interval is in 651 * drbd_req_complete() */ 652 D_ASSERT(device, drbd_interval_empty(&req->i)); 653 drbd_insert_interval(&device->read_requests, &req->i); 654 655 set_bit(UNPLUG_REMOTE, &device->flags); 656 657 D_ASSERT(device, req->rq_state & RQ_NET_PENDING); 658 D_ASSERT(device, (req->rq_state & RQ_LOCAL_MASK) == 0); 659 mod_rq_state(req, m, 0, RQ_NET_QUEUED); 660 req->w.cb = w_send_read_req; 661 drbd_queue_work(&connection->sender_work, 662 &req->w); 663 break; 664 665 case QUEUE_FOR_NET_WRITE: 666 /* assert something? */ 667 /* from __drbd_make_request only */ 668 669 /* Corresponding drbd_remove_request_interval is in 670 * drbd_req_complete() */ 671 D_ASSERT(device, drbd_interval_empty(&req->i)); 672 drbd_insert_interval(&device->write_requests, &req->i); 673 674 /* NOTE 675 * In case the req ended up on the transfer log before being 676 * queued on the worker, it could lead to this request being 677 * missed during cleanup after connection loss. 678 * So we have to do both operations here, 679 * within the same lock that protects the transfer log. 680 * 681 * _req_add_to_epoch(req); this has to be after the 682 * _maybe_start_new_epoch(req); which happened in 683 * __drbd_make_request, because we now may set the bit 684 * again ourselves to close the current epoch. 685 * 686 * Add req to the (now) current epoch (barrier). */ 687 688 /* otherwise we may lose an unplug, which may cause some remote 689 * io-scheduler timeout to expire, increasing maximum latency, 690 * hurting performance. */ 691 set_bit(UNPLUG_REMOTE, &device->flags); 692 693 /* queue work item to send data */ 694 D_ASSERT(device, req->rq_state & RQ_NET_PENDING); 695 mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK); 696 req->w.cb = w_send_dblock; 697 drbd_queue_work(&connection->sender_work, 698 &req->w); 699 700 /* close the epoch, in case it outgrew the limit */ 701 rcu_read_lock(); 702 nc = rcu_dereference(connection->net_conf); 703 p = nc->max_epoch_size; 704 rcu_read_unlock(); 705 if (connection->current_tle_writes >= p) 706 start_new_tl_epoch(connection); 707 708 break; 709 710 case QUEUE_FOR_SEND_OOS: 711 mod_rq_state(req, m, 0, RQ_NET_QUEUED); 712 req->w.cb = w_send_out_of_sync; 713 drbd_queue_work(&connection->sender_work, 714 &req->w); 715 break; 716 717 case READ_RETRY_REMOTE_CANCELED: 718 case SEND_CANCELED: 719 case SEND_FAILED: 720 /* real cleanup will be done from tl_clear. just update flags 721 * so it is no longer marked as on the worker queue */ 722 mod_rq_state(req, m, RQ_NET_QUEUED, 0); 723 break; 724 725 case HANDED_OVER_TO_NETWORK: 726 /* assert something? */ 727 if (is_pending_write_protocol_A(req)) 728 /* this is what is dangerous about protocol A: 729 * pretend it was successfully written on the peer. */ 730 mod_rq_state(req, m, RQ_NET_QUEUED|RQ_NET_PENDING, 731 RQ_NET_SENT|RQ_NET_OK); 732 else 733 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT); 734 /* It is still not yet RQ_NET_DONE until the 735 * corresponding epoch barrier got acked as well, 736 * so we know what to dirty on connection loss. */ 737 break; 738 739 case OOS_HANDED_TO_NETWORK: 740 /* Was not set PENDING, no longer QUEUED, so is now DONE 741 * as far as this connection is concerned. */ 742 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE); 743 break; 744 745 case CONNECTION_LOST_WHILE_PENDING: 746 /* transfer log cleanup after connection loss */ 747 mod_rq_state(req, m, 748 RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP, 749 RQ_NET_DONE); 750 break; 751 752 case CONFLICT_RESOLVED: 753 /* for superseded conflicting writes of multiple primaries, 754 * there is no need to keep anything in the tl, potential 755 * node crashes are covered by the activity log. 756 * 757 * If this request had been marked as RQ_POSTPONED before, 758 * it will actually not be completed, but "restarted", 759 * resubmitted from the retry worker context. */ 760 D_ASSERT(device, req->rq_state & RQ_NET_PENDING); 761 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK); 762 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK); 763 break; 764 765 case WRITE_ACKED_BY_PEER_AND_SIS: 766 req->rq_state |= RQ_NET_SIS; 767 fallthrough; 768 case WRITE_ACKED_BY_PEER: 769 /* Normal operation protocol C: successfully written on peer. 770 * During resync, even in protocol != C, 771 * we requested an explicit write ack anyways. 772 * Which means we cannot even assert anything here. 773 * Nothing more to do here. 774 * We want to keep the tl in place for all protocols, to cater 775 * for volatile write-back caches on lower level devices. */ 776 goto ack_common; 777 case RECV_ACKED_BY_PEER: 778 D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK); 779 /* protocol B; pretends to be successfully written on peer. 780 * see also notes above in HANDED_OVER_TO_NETWORK about 781 * protocol != C */ 782 ack_common: 783 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK); 784 break; 785 786 case POSTPONE_WRITE: 787 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK); 788 /* If this node has already detected the write conflict, the 789 * worker will be waiting on misc_wait. Wake it up once this 790 * request has completed locally. 791 */ 792 D_ASSERT(device, req->rq_state & RQ_NET_PENDING); 793 req->rq_state |= RQ_POSTPONED; 794 if (req->i.waiting) 795 wake_up(&device->misc_wait); 796 /* Do not clear RQ_NET_PENDING. This request will make further 797 * progress via restart_conflicting_writes() or 798 * fail_postponed_requests(). Hopefully. */ 799 break; 800 801 case NEG_ACKED: 802 mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0); 803 break; 804 805 case FAIL_FROZEN_DISK_IO: 806 if (!(req->rq_state & RQ_LOCAL_COMPLETED)) 807 break; 808 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0); 809 break; 810 811 case RESTART_FROZEN_DISK_IO: 812 if (!(req->rq_state & RQ_LOCAL_COMPLETED)) 813 break; 814 815 mod_rq_state(req, m, 816 RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED, 817 RQ_LOCAL_PENDING); 818 819 rv = MR_READ; 820 if (bio_data_dir(req->master_bio) == WRITE) 821 rv = MR_WRITE; 822 823 get_ldev(device); /* always succeeds in this call path */ 824 req->w.cb = w_restart_disk_io; 825 drbd_queue_work(&connection->sender_work, 826 &req->w); 827 break; 828 829 case RESEND: 830 /* Simply complete (local only) READs. */ 831 if (!(req->rq_state & RQ_WRITE) && !req->w.cb) { 832 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0); 833 break; 834 } 835 836 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK 837 before the connection loss (B&C only); only P_BARRIER_ACK 838 (or the local completion?) was missing when we suspended. 839 Throwing them out of the TL here by pretending we got a BARRIER_ACK. 840 During connection handshake, we ensure that the peer was not rebooted. */ 841 if (!(req->rq_state & RQ_NET_OK)) { 842 /* FIXME could this possibly be a req->dw.cb == w_send_out_of_sync? 843 * in that case we must not set RQ_NET_PENDING. */ 844 845 mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING); 846 if (req->w.cb) { 847 /* w.cb expected to be w_send_dblock, or w_send_read_req */ 848 drbd_queue_work(&connection->sender_work, 849 &req->w); 850 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ; 851 } /* else: FIXME can this happen? */ 852 break; 853 } 854 fallthrough; /* to BARRIER_ACKED */ 855 856 case BARRIER_ACKED: 857 /* barrier ack for READ requests does not make sense */ 858 if (!(req->rq_state & RQ_WRITE)) 859 break; 860 861 if (req->rq_state & RQ_NET_PENDING) { 862 /* barrier came in before all requests were acked. 863 * this is bad, because if the connection is lost now, 864 * we won't be able to clean them up... */ 865 drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n"); 866 } 867 /* Allowed to complete requests, even while suspended. 868 * As this is called for all requests within a matching epoch, 869 * we need to filter, and only set RQ_NET_DONE for those that 870 * have actually been on the wire. */ 871 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 872 (req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0); 873 break; 874 875 case DATA_RECEIVED: 876 D_ASSERT(device, req->rq_state & RQ_NET_PENDING); 877 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE); 878 break; 879 880 case QUEUE_AS_DRBD_BARRIER: 881 start_new_tl_epoch(connection); 882 mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE); 883 break; 884 } 885 886 return rv; 887 } 888 889 /* we may do a local read if: 890 * - we are consistent (of course), 891 * - or we are generally inconsistent, 892 * BUT we are still/already IN SYNC for this area. 893 * since size may be bigger than BM_BLOCK_SIZE, 894 * we may need to check several bits. 895 */ 896 static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size) 897 { 898 unsigned long sbnr, ebnr; 899 sector_t esector, nr_sectors; 900 901 if (device->state.disk == D_UP_TO_DATE) 902 return true; 903 if (device->state.disk != D_INCONSISTENT) 904 return false; 905 esector = sector + (size >> 9) - 1; 906 nr_sectors = get_capacity(device->vdisk); 907 D_ASSERT(device, sector < nr_sectors); 908 D_ASSERT(device, esector < nr_sectors); 909 910 sbnr = BM_SECT_TO_BIT(sector); 911 ebnr = BM_SECT_TO_BIT(esector); 912 913 return drbd_bm_count_bits(device, sbnr, ebnr) == 0; 914 } 915 916 static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector, 917 enum drbd_read_balancing rbm) 918 { 919 int stripe_shift; 920 921 switch (rbm) { 922 case RB_CONGESTED_REMOTE: 923 return false; 924 case RB_LEAST_PENDING: 925 return atomic_read(&device->local_cnt) > 926 atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt); 927 case RB_32K_STRIPING: /* stripe_shift = 15 */ 928 case RB_64K_STRIPING: 929 case RB_128K_STRIPING: 930 case RB_256K_STRIPING: 931 case RB_512K_STRIPING: 932 case RB_1M_STRIPING: /* stripe_shift = 20 */ 933 stripe_shift = (rbm - RB_32K_STRIPING + 15); 934 return (sector >> (stripe_shift - 9)) & 1; 935 case RB_ROUND_ROBIN: 936 return test_and_change_bit(READ_BALANCE_RR, &device->flags); 937 case RB_PREFER_REMOTE: 938 return true; 939 case RB_PREFER_LOCAL: 940 default: 941 return false; 942 } 943 } 944 945 /* 946 * complete_conflicting_writes - wait for any conflicting write requests 947 * 948 * The write_requests tree contains all active write requests which we 949 * currently know about. Wait for any requests to complete which conflict with 950 * the new one. 951 * 952 * Only way out: remove the conflicting intervals from the tree. 953 */ 954 static void complete_conflicting_writes(struct drbd_request *req) 955 { 956 DEFINE_WAIT(wait); 957 struct drbd_device *device = req->device; 958 struct drbd_interval *i; 959 sector_t sector = req->i.sector; 960 int size = req->i.size; 961 962 for (;;) { 963 drbd_for_each_overlap(i, &device->write_requests, sector, size) { 964 /* Ignore, if already completed to upper layers. */ 965 if (i->completed) 966 continue; 967 /* Handle the first found overlap. After the schedule 968 * we have to restart the tree walk. */ 969 break; 970 } 971 if (!i) /* if any */ 972 break; 973 974 /* Indicate to wake up device->misc_wait on progress. */ 975 prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE); 976 i->waiting = true; 977 spin_unlock_irq(&device->resource->req_lock); 978 schedule(); 979 spin_lock_irq(&device->resource->req_lock); 980 } 981 finish_wait(&device->misc_wait, &wait); 982 } 983 984 /* called within req_lock */ 985 static void maybe_pull_ahead(struct drbd_device *device) 986 { 987 struct drbd_connection *connection = first_peer_device(device)->connection; 988 struct net_conf *nc; 989 bool congested = false; 990 enum drbd_on_congestion on_congestion; 991 992 rcu_read_lock(); 993 nc = rcu_dereference(connection->net_conf); 994 on_congestion = nc ? nc->on_congestion : OC_BLOCK; 995 rcu_read_unlock(); 996 if (on_congestion == OC_BLOCK || 997 connection->agreed_pro_version < 96) 998 return; 999 1000 if (on_congestion == OC_PULL_AHEAD && device->state.conn == C_AHEAD) 1001 return; /* nothing to do ... */ 1002 1003 /* If I don't even have good local storage, we can not reasonably try 1004 * to pull ahead of the peer. We also need the local reference to make 1005 * sure device->act_log is there. 1006 */ 1007 if (!get_ldev_if_state(device, D_UP_TO_DATE)) 1008 return; 1009 1010 if (nc->cong_fill && 1011 atomic_read(&device->ap_in_flight) >= nc->cong_fill) { 1012 drbd_info(device, "Congestion-fill threshold reached\n"); 1013 congested = true; 1014 } 1015 1016 if (device->act_log->used >= nc->cong_extents) { 1017 drbd_info(device, "Congestion-extents threshold reached\n"); 1018 congested = true; 1019 } 1020 1021 if (congested) { 1022 /* start a new epoch for non-mirrored writes */ 1023 start_new_tl_epoch(first_peer_device(device)->connection); 1024 1025 if (on_congestion == OC_PULL_AHEAD) 1026 _drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL); 1027 else /*nc->on_congestion == OC_DISCONNECT */ 1028 _drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL); 1029 } 1030 put_ldev(device); 1031 } 1032 1033 /* If this returns false, and req->private_bio is still set, 1034 * this should be submitted locally. 1035 * 1036 * If it returns false, but req->private_bio is not set, 1037 * we do not have access to good data :( 1038 * 1039 * Otherwise, this destroys req->private_bio, if any, 1040 * and returns true. 1041 */ 1042 static bool do_remote_read(struct drbd_request *req) 1043 { 1044 struct drbd_device *device = req->device; 1045 enum drbd_read_balancing rbm; 1046 1047 if (req->private_bio) { 1048 if (!drbd_may_do_local_read(device, 1049 req->i.sector, req->i.size)) { 1050 bio_put(req->private_bio); 1051 req->private_bio = NULL; 1052 put_ldev(device); 1053 } 1054 } 1055 1056 if (device->state.pdsk != D_UP_TO_DATE) 1057 return false; 1058 1059 if (req->private_bio == NULL) 1060 return true; 1061 1062 /* TODO: improve read balancing decisions, take into account drbd 1063 * protocol, pending requests etc. */ 1064 1065 rcu_read_lock(); 1066 rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing; 1067 rcu_read_unlock(); 1068 1069 if (rbm == RB_PREFER_LOCAL && req->private_bio) 1070 return false; /* submit locally */ 1071 1072 if (remote_due_to_read_balancing(device, req->i.sector, rbm)) { 1073 if (req->private_bio) { 1074 bio_put(req->private_bio); 1075 req->private_bio = NULL; 1076 put_ldev(device); 1077 } 1078 return true; 1079 } 1080 1081 return false; 1082 } 1083 1084 bool drbd_should_do_remote(union drbd_dev_state s) 1085 { 1086 return s.pdsk == D_UP_TO_DATE || 1087 (s.pdsk >= D_INCONSISTENT && 1088 s.conn >= C_WF_BITMAP_T && 1089 s.conn < C_AHEAD); 1090 /* Before proto 96 that was >= CONNECTED instead of >= C_WF_BITMAP_T. 1091 That is equivalent since before 96 IO was frozen in the C_WF_BITMAP* 1092 states. */ 1093 } 1094 1095 static bool drbd_should_send_out_of_sync(union drbd_dev_state s) 1096 { 1097 return s.conn == C_AHEAD || s.conn == C_WF_BITMAP_S; 1098 /* pdsk = D_INCONSISTENT as a consequence. Protocol 96 check not necessary 1099 since we enter state C_AHEAD only if proto >= 96 */ 1100 } 1101 1102 /* returns number of connections (== 1, for drbd 8.4) 1103 * expected to actually write this data, 1104 * which does NOT include those that we are L_AHEAD for. */ 1105 static int drbd_process_write_request(struct drbd_request *req) 1106 { 1107 struct drbd_device *device = req->device; 1108 int remote, send_oos; 1109 1110 remote = drbd_should_do_remote(device->state); 1111 send_oos = drbd_should_send_out_of_sync(device->state); 1112 1113 /* Need to replicate writes. Unless it is an empty flush, 1114 * which is better mapped to a DRBD P_BARRIER packet, 1115 * also for drbd wire protocol compatibility reasons. 1116 * If this was a flush, just start a new epoch. 1117 * Unless the current epoch was empty anyways, or we are not currently 1118 * replicating, in which case there is no point. */ 1119 if (unlikely(req->i.size == 0)) { 1120 /* The only size==0 bios we expect are empty flushes. */ 1121 D_ASSERT(device, req->master_bio->bi_opf & REQ_PREFLUSH); 1122 if (remote) 1123 _req_mod(req, QUEUE_AS_DRBD_BARRIER); 1124 return remote; 1125 } 1126 1127 if (!remote && !send_oos) 1128 return 0; 1129 1130 D_ASSERT(device, !(remote && send_oos)); 1131 1132 if (remote) { 1133 _req_mod(req, TO_BE_SENT); 1134 _req_mod(req, QUEUE_FOR_NET_WRITE); 1135 } else if (drbd_set_out_of_sync(device, req->i.sector, req->i.size)) 1136 _req_mod(req, QUEUE_FOR_SEND_OOS); 1137 1138 return remote; 1139 } 1140 1141 static void drbd_process_discard_or_zeroes_req(struct drbd_request *req, int flags) 1142 { 1143 int err = drbd_issue_discard_or_zero_out(req->device, 1144 req->i.sector, req->i.size >> 9, flags); 1145 if (err) 1146 req->private_bio->bi_status = BLK_STS_IOERR; 1147 bio_endio(req->private_bio); 1148 } 1149 1150 static void 1151 drbd_submit_req_private_bio(struct drbd_request *req) 1152 { 1153 struct drbd_device *device = req->device; 1154 struct bio *bio = req->private_bio; 1155 unsigned int type; 1156 1157 if (bio_op(bio) != REQ_OP_READ) 1158 type = DRBD_FAULT_DT_WR; 1159 else if (bio->bi_opf & REQ_RAHEAD) 1160 type = DRBD_FAULT_DT_RA; 1161 else 1162 type = DRBD_FAULT_DT_RD; 1163 1164 /* State may have changed since we grabbed our reference on the 1165 * ->ldev member. Double check, and short-circuit to endio. 1166 * In case the last activity log transaction failed to get on 1167 * stable storage, and this is a WRITE, we may not even submit 1168 * this bio. */ 1169 if (get_ldev(device)) { 1170 if (drbd_insert_fault(device, type)) 1171 bio_io_error(bio); 1172 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES) 1173 drbd_process_discard_or_zeroes_req(req, EE_ZEROOUT | 1174 ((bio->bi_opf & REQ_NOUNMAP) ? 0 : EE_TRIM)); 1175 else if (bio_op(bio) == REQ_OP_DISCARD) 1176 drbd_process_discard_or_zeroes_req(req, EE_TRIM); 1177 else 1178 submit_bio_noacct(bio); 1179 put_ldev(device); 1180 } else 1181 bio_io_error(bio); 1182 } 1183 1184 static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req) 1185 { 1186 spin_lock_irq(&device->resource->req_lock); 1187 list_add_tail(&req->tl_requests, &device->submit.writes); 1188 list_add_tail(&req->req_pending_master_completion, 1189 &device->pending_master_completion[1 /* WRITE */]); 1190 spin_unlock_irq(&device->resource->req_lock); 1191 queue_work(device->submit.wq, &device->submit.worker); 1192 /* do_submit() may sleep internally on al_wait, too */ 1193 wake_up(&device->al_wait); 1194 } 1195 1196 /* returns the new drbd_request pointer, if the caller is expected to 1197 * drbd_send_and_submit() it (to save latency), or NULL if we queued the 1198 * request on the submitter thread. 1199 * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request. 1200 */ 1201 static struct drbd_request * 1202 drbd_request_prepare(struct drbd_device *device, struct bio *bio) 1203 { 1204 const int rw = bio_data_dir(bio); 1205 struct drbd_request *req; 1206 1207 /* allocate outside of all locks; */ 1208 req = drbd_req_new(device, bio); 1209 if (!req) { 1210 dec_ap_bio(device); 1211 /* only pass the error to the upper layers. 1212 * if user cannot handle io errors, that's not our business. */ 1213 drbd_err(device, "could not kmalloc() req\n"); 1214 bio->bi_status = BLK_STS_RESOURCE; 1215 bio_endio(bio); 1216 return ERR_PTR(-ENOMEM); 1217 } 1218 1219 /* Update disk stats */ 1220 req->start_jif = bio_start_io_acct(req->master_bio); 1221 1222 if (!get_ldev(device)) { 1223 bio_put(req->private_bio); 1224 req->private_bio = NULL; 1225 } 1226 1227 /* process discards always from our submitter thread */ 1228 if (bio_op(bio) == REQ_OP_WRITE_ZEROES || 1229 bio_op(bio) == REQ_OP_DISCARD) 1230 goto queue_for_submitter_thread; 1231 1232 if (rw == WRITE && req->private_bio && req->i.size 1233 && !test_bit(AL_SUSPENDED, &device->flags)) { 1234 if (!drbd_al_begin_io_fastpath(device, &req->i)) 1235 goto queue_for_submitter_thread; 1236 req->rq_state |= RQ_IN_ACT_LOG; 1237 req->in_actlog_jif = jiffies; 1238 } 1239 return req; 1240 1241 queue_for_submitter_thread: 1242 atomic_inc(&device->ap_actlog_cnt); 1243 drbd_queue_write(device, req); 1244 return NULL; 1245 } 1246 1247 /* Require at least one path to current data. 1248 * We don't want to allow writes on C_STANDALONE D_INCONSISTENT: 1249 * We would not allow to read what was written, 1250 * we would not have bumped the data generation uuids, 1251 * we would cause data divergence for all the wrong reasons. 1252 * 1253 * If we don't see at least one D_UP_TO_DATE, we will fail this request, 1254 * which either returns EIO, or, if OND_SUSPEND_IO is set, suspends IO, 1255 * and queues for retry later. 1256 */ 1257 static bool may_do_writes(struct drbd_device *device) 1258 { 1259 const union drbd_dev_state s = device->state; 1260 return s.disk == D_UP_TO_DATE || s.pdsk == D_UP_TO_DATE; 1261 } 1262 1263 struct drbd_plug_cb { 1264 struct blk_plug_cb cb; 1265 struct drbd_request *most_recent_req; 1266 /* do we need more? */ 1267 }; 1268 1269 static void drbd_unplug(struct blk_plug_cb *cb, bool from_schedule) 1270 { 1271 struct drbd_plug_cb *plug = container_of(cb, struct drbd_plug_cb, cb); 1272 struct drbd_resource *resource = plug->cb.data; 1273 struct drbd_request *req = plug->most_recent_req; 1274 1275 kfree(cb); 1276 if (!req) 1277 return; 1278 1279 spin_lock_irq(&resource->req_lock); 1280 /* In case the sender did not process it yet, raise the flag to 1281 * have it followed with P_UNPLUG_REMOTE just after. */ 1282 req->rq_state |= RQ_UNPLUG; 1283 /* but also queue a generic unplug */ 1284 drbd_queue_unplug(req->device); 1285 kref_put(&req->kref, drbd_req_destroy); 1286 spin_unlock_irq(&resource->req_lock); 1287 } 1288 1289 static struct drbd_plug_cb* drbd_check_plugged(struct drbd_resource *resource) 1290 { 1291 /* A lot of text to say 1292 * return (struct drbd_plug_cb*)blk_check_plugged(); */ 1293 struct drbd_plug_cb *plug; 1294 struct blk_plug_cb *cb = blk_check_plugged(drbd_unplug, resource, sizeof(*plug)); 1295 1296 if (cb) 1297 plug = container_of(cb, struct drbd_plug_cb, cb); 1298 else 1299 plug = NULL; 1300 return plug; 1301 } 1302 1303 static void drbd_update_plug(struct drbd_plug_cb *plug, struct drbd_request *req) 1304 { 1305 struct drbd_request *tmp = plug->most_recent_req; 1306 /* Will be sent to some peer. 1307 * Remember to tag it with UNPLUG_REMOTE on unplug */ 1308 kref_get(&req->kref); 1309 plug->most_recent_req = req; 1310 if (tmp) 1311 kref_put(&tmp->kref, drbd_req_destroy); 1312 } 1313 1314 static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req) 1315 { 1316 struct drbd_resource *resource = device->resource; 1317 const int rw = bio_data_dir(req->master_bio); 1318 struct bio_and_error m = { NULL, }; 1319 bool no_remote = false; 1320 bool submit_private_bio = false; 1321 1322 spin_lock_irq(&resource->req_lock); 1323 if (rw == WRITE) { 1324 /* This may temporarily give up the req_lock, 1325 * but will re-aquire it before it returns here. 1326 * Needs to be before the check on drbd_suspended() */ 1327 complete_conflicting_writes(req); 1328 /* no more giving up req_lock from now on! */ 1329 1330 /* check for congestion, and potentially stop sending 1331 * full data updates, but start sending "dirty bits" only. */ 1332 maybe_pull_ahead(device); 1333 } 1334 1335 1336 if (drbd_suspended(device)) { 1337 /* push back and retry: */ 1338 req->rq_state |= RQ_POSTPONED; 1339 if (req->private_bio) { 1340 bio_put(req->private_bio); 1341 req->private_bio = NULL; 1342 put_ldev(device); 1343 } 1344 goto out; 1345 } 1346 1347 /* We fail READ early, if we can not serve it. 1348 * We must do this before req is registered on any lists. 1349 * Otherwise, drbd_req_complete() will queue failed READ for retry. */ 1350 if (rw != WRITE) { 1351 if (!do_remote_read(req) && !req->private_bio) 1352 goto nodata; 1353 } 1354 1355 /* which transfer log epoch does this belong to? */ 1356 req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr); 1357 1358 /* no point in adding empty flushes to the transfer log, 1359 * they are mapped to drbd barriers already. */ 1360 if (likely(req->i.size!=0)) { 1361 if (rw == WRITE) 1362 first_peer_device(device)->connection->current_tle_writes++; 1363 1364 list_add_tail(&req->tl_requests, &first_peer_device(device)->connection->transfer_log); 1365 } 1366 1367 if (rw == WRITE) { 1368 if (req->private_bio && !may_do_writes(device)) { 1369 bio_put(req->private_bio); 1370 req->private_bio = NULL; 1371 put_ldev(device); 1372 goto nodata; 1373 } 1374 if (!drbd_process_write_request(req)) 1375 no_remote = true; 1376 } else { 1377 /* We either have a private_bio, or we can read from remote. 1378 * Otherwise we had done the goto nodata above. */ 1379 if (req->private_bio == NULL) { 1380 _req_mod(req, TO_BE_SENT); 1381 _req_mod(req, QUEUE_FOR_NET_READ); 1382 } else 1383 no_remote = true; 1384 } 1385 1386 if (no_remote == false) { 1387 struct drbd_plug_cb *plug = drbd_check_plugged(resource); 1388 if (plug) 1389 drbd_update_plug(plug, req); 1390 } 1391 1392 /* If it took the fast path in drbd_request_prepare, add it here. 1393 * The slow path has added it already. */ 1394 if (list_empty(&req->req_pending_master_completion)) 1395 list_add_tail(&req->req_pending_master_completion, 1396 &device->pending_master_completion[rw == WRITE]); 1397 if (req->private_bio) { 1398 /* needs to be marked within the same spinlock */ 1399 req->pre_submit_jif = jiffies; 1400 list_add_tail(&req->req_pending_local, 1401 &device->pending_completion[rw == WRITE]); 1402 _req_mod(req, TO_BE_SUBMITTED); 1403 /* but we need to give up the spinlock to submit */ 1404 submit_private_bio = true; 1405 } else if (no_remote) { 1406 nodata: 1407 if (__ratelimit(&drbd_ratelimit_state)) 1408 drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n", 1409 (unsigned long long)req->i.sector, req->i.size >> 9); 1410 /* A write may have been queued for send_oos, however. 1411 * So we can not simply free it, we must go through drbd_req_put_completion_ref() */ 1412 } 1413 1414 out: 1415 drbd_req_put_completion_ref(req, &m, 1); 1416 spin_unlock_irq(&resource->req_lock); 1417 1418 /* Even though above is a kref_put(), this is safe. 1419 * As long as we still need to submit our private bio, 1420 * we hold a completion ref, and the request cannot disappear. 1421 * If however this request did not even have a private bio to submit 1422 * (e.g. remote read), req may already be invalid now. 1423 * That's why we cannot check on req->private_bio. */ 1424 if (submit_private_bio) 1425 drbd_submit_req_private_bio(req); 1426 if (m.bio) 1427 complete_master_bio(device, &m); 1428 } 1429 1430 void __drbd_make_request(struct drbd_device *device, struct bio *bio) 1431 { 1432 struct drbd_request *req = drbd_request_prepare(device, bio); 1433 if (IS_ERR_OR_NULL(req)) 1434 return; 1435 drbd_send_and_submit(device, req); 1436 } 1437 1438 static void submit_fast_path(struct drbd_device *device, struct list_head *incoming) 1439 { 1440 struct blk_plug plug; 1441 struct drbd_request *req, *tmp; 1442 1443 blk_start_plug(&plug); 1444 list_for_each_entry_safe(req, tmp, incoming, tl_requests) { 1445 const int rw = bio_data_dir(req->master_bio); 1446 1447 if (rw == WRITE /* rw != WRITE should not even end up here! */ 1448 && req->private_bio && req->i.size 1449 && !test_bit(AL_SUSPENDED, &device->flags)) { 1450 if (!drbd_al_begin_io_fastpath(device, &req->i)) 1451 continue; 1452 1453 req->rq_state |= RQ_IN_ACT_LOG; 1454 req->in_actlog_jif = jiffies; 1455 atomic_dec(&device->ap_actlog_cnt); 1456 } 1457 1458 list_del_init(&req->tl_requests); 1459 drbd_send_and_submit(device, req); 1460 } 1461 blk_finish_plug(&plug); 1462 } 1463 1464 static bool prepare_al_transaction_nonblock(struct drbd_device *device, 1465 struct list_head *incoming, 1466 struct list_head *pending, 1467 struct list_head *later) 1468 { 1469 struct drbd_request *req; 1470 int wake = 0; 1471 int err; 1472 1473 spin_lock_irq(&device->al_lock); 1474 while ((req = list_first_entry_or_null(incoming, struct drbd_request, tl_requests))) { 1475 err = drbd_al_begin_io_nonblock(device, &req->i); 1476 if (err == -ENOBUFS) 1477 break; 1478 if (err == -EBUSY) 1479 wake = 1; 1480 if (err) 1481 list_move_tail(&req->tl_requests, later); 1482 else 1483 list_move_tail(&req->tl_requests, pending); 1484 } 1485 spin_unlock_irq(&device->al_lock); 1486 if (wake) 1487 wake_up(&device->al_wait); 1488 return !list_empty(pending); 1489 } 1490 1491 static void send_and_submit_pending(struct drbd_device *device, struct list_head *pending) 1492 { 1493 struct blk_plug plug; 1494 struct drbd_request *req; 1495 1496 blk_start_plug(&plug); 1497 while ((req = list_first_entry_or_null(pending, struct drbd_request, tl_requests))) { 1498 req->rq_state |= RQ_IN_ACT_LOG; 1499 req->in_actlog_jif = jiffies; 1500 atomic_dec(&device->ap_actlog_cnt); 1501 list_del_init(&req->tl_requests); 1502 drbd_send_and_submit(device, req); 1503 } 1504 blk_finish_plug(&plug); 1505 } 1506 1507 void do_submit(struct work_struct *ws) 1508 { 1509 struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker); 1510 LIST_HEAD(incoming); /* from drbd_make_request() */ 1511 LIST_HEAD(pending); /* to be submitted after next AL-transaction commit */ 1512 LIST_HEAD(busy); /* blocked by resync requests */ 1513 1514 /* grab new incoming requests */ 1515 spin_lock_irq(&device->resource->req_lock); 1516 list_splice_tail_init(&device->submit.writes, &incoming); 1517 spin_unlock_irq(&device->resource->req_lock); 1518 1519 for (;;) { 1520 DEFINE_WAIT(wait); 1521 1522 /* move used-to-be-busy back to front of incoming */ 1523 list_splice_init(&busy, &incoming); 1524 submit_fast_path(device, &incoming); 1525 if (list_empty(&incoming)) 1526 break; 1527 1528 for (;;) { 1529 prepare_to_wait(&device->al_wait, &wait, TASK_UNINTERRUPTIBLE); 1530 1531 list_splice_init(&busy, &incoming); 1532 prepare_al_transaction_nonblock(device, &incoming, &pending, &busy); 1533 if (!list_empty(&pending)) 1534 break; 1535 1536 schedule(); 1537 1538 /* If all currently "hot" activity log extents are kept busy by 1539 * incoming requests, we still must not totally starve new 1540 * requests to "cold" extents. 1541 * Something left on &incoming means there had not been 1542 * enough update slots available, and the activity log 1543 * has been marked as "starving". 1544 * 1545 * Try again now, without looking for new requests, 1546 * effectively blocking all new requests until we made 1547 * at least _some_ progress with what we currently have. 1548 */ 1549 if (!list_empty(&incoming)) 1550 continue; 1551 1552 /* Nothing moved to pending, but nothing left 1553 * on incoming: all moved to busy! 1554 * Grab new and iterate. */ 1555 spin_lock_irq(&device->resource->req_lock); 1556 list_splice_tail_init(&device->submit.writes, &incoming); 1557 spin_unlock_irq(&device->resource->req_lock); 1558 } 1559 finish_wait(&device->al_wait, &wait); 1560 1561 /* If the transaction was full, before all incoming requests 1562 * had been processed, skip ahead to commit, and iterate 1563 * without splicing in more incoming requests from upper layers. 1564 * 1565 * Else, if all incoming have been processed, 1566 * they have become either "pending" (to be submitted after 1567 * next transaction commit) or "busy" (blocked by resync). 1568 * 1569 * Maybe more was queued, while we prepared the transaction? 1570 * Try to stuff those into this transaction as well. 1571 * Be strictly non-blocking here, 1572 * we already have something to commit. 1573 * 1574 * Commit if we don't make any more progres. 1575 */ 1576 1577 while (list_empty(&incoming)) { 1578 LIST_HEAD(more_pending); 1579 LIST_HEAD(more_incoming); 1580 bool made_progress; 1581 1582 /* It is ok to look outside the lock, 1583 * it's only an optimization anyways */ 1584 if (list_empty(&device->submit.writes)) 1585 break; 1586 1587 spin_lock_irq(&device->resource->req_lock); 1588 list_splice_tail_init(&device->submit.writes, &more_incoming); 1589 spin_unlock_irq(&device->resource->req_lock); 1590 1591 if (list_empty(&more_incoming)) 1592 break; 1593 1594 made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending, &busy); 1595 1596 list_splice_tail_init(&more_pending, &pending); 1597 list_splice_tail_init(&more_incoming, &incoming); 1598 if (!made_progress) 1599 break; 1600 } 1601 1602 drbd_al_begin_io_commit(device); 1603 send_and_submit_pending(device, &pending); 1604 } 1605 } 1606 1607 void drbd_submit_bio(struct bio *bio) 1608 { 1609 struct drbd_device *device = bio->bi_bdev->bd_disk->private_data; 1610 1611 bio = bio_split_to_limits(bio); 1612 1613 /* 1614 * what we "blindly" assume: 1615 */ 1616 D_ASSERT(device, IS_ALIGNED(bio->bi_iter.bi_size, 512)); 1617 1618 inc_ap_bio(device); 1619 __drbd_make_request(device, bio); 1620 } 1621 1622 static bool net_timeout_reached(struct drbd_request *net_req, 1623 struct drbd_connection *connection, 1624 unsigned long now, unsigned long ent, 1625 unsigned int ko_count, unsigned int timeout) 1626 { 1627 struct drbd_device *device = net_req->device; 1628 1629 if (!time_after(now, net_req->pre_send_jif + ent)) 1630 return false; 1631 1632 if (time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent)) 1633 return false; 1634 1635 if (net_req->rq_state & RQ_NET_PENDING) { 1636 drbd_warn(device, "Remote failed to finish a request within %ums > ko-count (%u) * timeout (%u * 0.1s)\n", 1637 jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout); 1638 return true; 1639 } 1640 1641 /* We received an ACK already (or are using protocol A), 1642 * but are waiting for the epoch closing barrier ack. 1643 * Check if we sent the barrier already. We should not blame the peer 1644 * for being unresponsive, if we did not even ask it yet. */ 1645 if (net_req->epoch == connection->send.current_epoch_nr) { 1646 drbd_warn(device, 1647 "We did not send a P_BARRIER for %ums > ko-count (%u) * timeout (%u * 0.1s); drbd kernel thread blocked?\n", 1648 jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout); 1649 return false; 1650 } 1651 1652 /* Worst case: we may have been blocked for whatever reason, then 1653 * suddenly are able to send a lot of requests (and epoch separating 1654 * barriers) in quick succession. 1655 * The timestamp of the net_req may be much too old and not correspond 1656 * to the sending time of the relevant unack'ed barrier packet, so 1657 * would trigger a spurious timeout. The latest barrier packet may 1658 * have a too recent timestamp to trigger the timeout, potentially miss 1659 * a timeout. Right now we don't have a place to conveniently store 1660 * these timestamps. 1661 * But in this particular situation, the application requests are still 1662 * completed to upper layers, DRBD should still "feel" responsive. 1663 * No need yet to kill this connection, it may still recover. 1664 * If not, eventually we will have queued enough into the network for 1665 * us to block. From that point of view, the timestamp of the last sent 1666 * barrier packet is relevant enough. 1667 */ 1668 if (time_after(now, connection->send.last_sent_barrier_jif + ent)) { 1669 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", 1670 connection->send.last_sent_barrier_jif, now, 1671 jiffies_to_msecs(now - connection->send.last_sent_barrier_jif), ko_count, timeout); 1672 return true; 1673 } 1674 return false; 1675 } 1676 1677 /* A request is considered timed out, if 1678 * - we have some effective timeout from the configuration, 1679 * with some state restrictions applied, 1680 * - the oldest request is waiting for a response from the network 1681 * resp. the local disk, 1682 * - the oldest request is in fact older than the effective timeout, 1683 * - the connection was established (resp. disk was attached) 1684 * for longer than the timeout already. 1685 * Note that for 32bit jiffies and very stable connections/disks, 1686 * we may have a wrap around, which is catched by 1687 * !time_in_range(now, last_..._jif, last_..._jif + timeout). 1688 * 1689 * Side effect: once per 32bit wrap-around interval, which means every 1690 * ~198 days with 250 HZ, we have a window where the timeout would need 1691 * to expire twice (worst case) to become effective. Good enough. 1692 */ 1693 1694 void request_timer_fn(struct timer_list *t) 1695 { 1696 struct drbd_device *device = from_timer(device, t, request_timer); 1697 struct drbd_connection *connection = first_peer_device(device)->connection; 1698 struct drbd_request *req_read, *req_write, *req_peer; /* oldest request */ 1699 struct net_conf *nc; 1700 unsigned long oldest_submit_jif; 1701 unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */ 1702 unsigned long now; 1703 unsigned int ko_count = 0, timeout = 0; 1704 1705 rcu_read_lock(); 1706 nc = rcu_dereference(connection->net_conf); 1707 if (nc && device->state.conn >= C_WF_REPORT_PARAMS) { 1708 ko_count = nc->ko_count; 1709 timeout = nc->timeout; 1710 } 1711 1712 if (get_ldev(device)) { /* implicit state.disk >= D_INCONSISTENT */ 1713 dt = rcu_dereference(device->ldev->disk_conf)->disk_timeout * HZ / 10; 1714 put_ldev(device); 1715 } 1716 rcu_read_unlock(); 1717 1718 1719 ent = timeout * HZ/10 * ko_count; 1720 et = min_not_zero(dt, ent); 1721 1722 if (!et) 1723 return; /* Recurring timer stopped */ 1724 1725 now = jiffies; 1726 nt = now + et; 1727 1728 spin_lock_irq(&device->resource->req_lock); 1729 req_read = list_first_entry_or_null(&device->pending_completion[0], struct drbd_request, req_pending_local); 1730 req_write = list_first_entry_or_null(&device->pending_completion[1], struct drbd_request, req_pending_local); 1731 1732 /* maybe the oldest request waiting for the peer is in fact still 1733 * blocking in tcp sendmsg. That's ok, though, that's handled via the 1734 * socket send timeout, requesting a ping, and bumping ko-count in 1735 * we_should_drop_the_connection(). 1736 */ 1737 1738 /* check the oldest request we did successfully sent, 1739 * but which is still waiting for an ACK. */ 1740 req_peer = connection->req_ack_pending; 1741 1742 /* if we don't have such request (e.g. protocoll A) 1743 * check the oldest requests which is still waiting on its epoch 1744 * closing barrier ack. */ 1745 if (!req_peer) 1746 req_peer = connection->req_not_net_done; 1747 1748 /* evaluate the oldest peer request only in one timer! */ 1749 if (req_peer && req_peer->device != device) 1750 req_peer = NULL; 1751 1752 /* do we have something to evaluate? */ 1753 if (req_peer == NULL && req_write == NULL && req_read == NULL) 1754 goto out; 1755 1756 oldest_submit_jif = 1757 (req_write && req_read) 1758 ? ( time_before(req_write->pre_submit_jif, req_read->pre_submit_jif) 1759 ? req_write->pre_submit_jif : req_read->pre_submit_jif ) 1760 : req_write ? req_write->pre_submit_jif 1761 : req_read ? req_read->pre_submit_jif : now; 1762 1763 if (ent && req_peer && net_timeout_reached(req_peer, connection, now, ent, ko_count, timeout)) 1764 _conn_request_state(connection, NS(conn, C_TIMEOUT), CS_VERBOSE | CS_HARD); 1765 1766 if (dt && oldest_submit_jif != now && 1767 time_after(now, oldest_submit_jif + dt) && 1768 !time_in_range(now, device->last_reattach_jif, device->last_reattach_jif + dt)) { 1769 drbd_warn(device, "Local backing device failed to meet the disk-timeout\n"); 1770 __drbd_chk_io_error(device, DRBD_FORCE_DETACH); 1771 } 1772 1773 /* Reschedule timer for the nearest not already expired timeout. 1774 * Fallback to now + min(effective network timeout, disk timeout). */ 1775 ent = (ent && req_peer && time_before(now, req_peer->pre_send_jif + ent)) 1776 ? req_peer->pre_send_jif + ent : now + et; 1777 dt = (dt && oldest_submit_jif != now && time_before(now, oldest_submit_jif + dt)) 1778 ? oldest_submit_jif + dt : now + et; 1779 nt = time_before(ent, dt) ? ent : dt; 1780 out: 1781 spin_unlock_irq(&device->resource->req_lock); 1782 mod_timer(&device->request_timer, nt); 1783 } 1784