1 /* 2 * Copyright (C) 1999 Eric Youngdale 3 * Copyright (C) 2014 Christoph Hellwig 4 * 5 * SCSI queueing library. 6 * Initial versions: Eric Youngdale (eric@andante.org). 7 * Based upon conversations with large numbers 8 * of people at Linux Expo. 9 */ 10 11 #include <linux/bio.h> 12 #include <linux/bitops.h> 13 #include <linux/blkdev.h> 14 #include <linux/completion.h> 15 #include <linux/kernel.h> 16 #include <linux/export.h> 17 #include <linux/init.h> 18 #include <linux/pci.h> 19 #include <linux/delay.h> 20 #include <linux/hardirq.h> 21 #include <linux/scatterlist.h> 22 #include <linux/blk-mq.h> 23 #include <linux/ratelimit.h> 24 #include <asm/unaligned.h> 25 26 #include <scsi/scsi.h> 27 #include <scsi/scsi_cmnd.h> 28 #include <scsi/scsi_dbg.h> 29 #include <scsi/scsi_device.h> 30 #include <scsi/scsi_driver.h> 31 #include <scsi/scsi_eh.h> 32 #include <scsi/scsi_host.h> 33 #include <scsi/scsi_dh.h> 34 35 #include <trace/events/scsi.h> 36 37 #include "scsi_priv.h" 38 #include "scsi_logging.h" 39 40 41 struct kmem_cache *scsi_sdb_cache; 42 43 /* 44 * When to reinvoke queueing after a resource shortage. It's 3 msecs to 45 * not change behaviour from the previous unplug mechanism, experimentation 46 * may prove this needs changing. 47 */ 48 #define SCSI_QUEUE_DELAY 3 49 50 static void 51 scsi_set_blocked(struct scsi_cmnd *cmd, int reason) 52 { 53 struct Scsi_Host *host = cmd->device->host; 54 struct scsi_device *device = cmd->device; 55 struct scsi_target *starget = scsi_target(device); 56 57 /* 58 * Set the appropriate busy bit for the device/host. 59 * 60 * If the host/device isn't busy, assume that something actually 61 * completed, and that we should be able to queue a command now. 62 * 63 * Note that the prior mid-layer assumption that any host could 64 * always queue at least one command is now broken. The mid-layer 65 * will implement a user specifiable stall (see 66 * scsi_host.max_host_blocked and scsi_device.max_device_blocked) 67 * if a command is requeued with no other commands outstanding 68 * either for the device or for the host. 69 */ 70 switch (reason) { 71 case SCSI_MLQUEUE_HOST_BUSY: 72 atomic_set(&host->host_blocked, host->max_host_blocked); 73 break; 74 case SCSI_MLQUEUE_DEVICE_BUSY: 75 case SCSI_MLQUEUE_EH_RETRY: 76 atomic_set(&device->device_blocked, 77 device->max_device_blocked); 78 break; 79 case SCSI_MLQUEUE_TARGET_BUSY: 80 atomic_set(&starget->target_blocked, 81 starget->max_target_blocked); 82 break; 83 } 84 } 85 86 static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd) 87 { 88 struct scsi_device *sdev = cmd->device; 89 90 blk_mq_requeue_request(cmd->request, true); 91 put_device(&sdev->sdev_gendev); 92 } 93 94 /** 95 * __scsi_queue_insert - private queue insertion 96 * @cmd: The SCSI command being requeued 97 * @reason: The reason for the requeue 98 * @unbusy: Whether the queue should be unbusied 99 * 100 * This is a private queue insertion. The public interface 101 * scsi_queue_insert() always assumes the queue should be unbusied 102 * because it's always called before the completion. This function is 103 * for a requeue after completion, which should only occur in this 104 * file. 105 */ 106 static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, int unbusy) 107 { 108 struct scsi_device *device = cmd->device; 109 struct request_queue *q = device->request_queue; 110 unsigned long flags; 111 112 SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd, 113 "Inserting command %p into mlqueue\n", cmd)); 114 115 scsi_set_blocked(cmd, reason); 116 117 /* 118 * Decrement the counters, since these commands are no longer 119 * active on the host/device. 120 */ 121 if (unbusy) 122 scsi_device_unbusy(device); 123 124 /* 125 * Requeue this command. It will go before all other commands 126 * that are already in the queue. Schedule requeue work under 127 * lock such that the kblockd_schedule_work() call happens 128 * before blk_cleanup_queue() finishes. 129 */ 130 cmd->result = 0; 131 if (q->mq_ops) { 132 scsi_mq_requeue_cmd(cmd); 133 return; 134 } 135 spin_lock_irqsave(q->queue_lock, flags); 136 blk_requeue_request(q, cmd->request); 137 kblockd_schedule_work(&device->requeue_work); 138 spin_unlock_irqrestore(q->queue_lock, flags); 139 } 140 141 /* 142 * Function: scsi_queue_insert() 143 * 144 * Purpose: Insert a command in the midlevel queue. 145 * 146 * Arguments: cmd - command that we are adding to queue. 147 * reason - why we are inserting command to queue. 148 * 149 * Lock status: Assumed that lock is not held upon entry. 150 * 151 * Returns: Nothing. 152 * 153 * Notes: We do this for one of two cases. Either the host is busy 154 * and it cannot accept any more commands for the time being, 155 * or the device returned QUEUE_FULL and can accept no more 156 * commands. 157 * Notes: This could be called either from an interrupt context or a 158 * normal process context. 159 */ 160 void scsi_queue_insert(struct scsi_cmnd *cmd, int reason) 161 { 162 __scsi_queue_insert(cmd, reason, 1); 163 } 164 165 static int __scsi_execute(struct scsi_device *sdev, const unsigned char *cmd, 166 int data_direction, void *buffer, unsigned bufflen, 167 unsigned char *sense, int timeout, int retries, u64 flags, 168 req_flags_t rq_flags, int *resid) 169 { 170 struct request *req; 171 int write = (data_direction == DMA_TO_DEVICE); 172 int ret = DRIVER_ERROR << 24; 173 174 req = blk_get_request(sdev->request_queue, write, __GFP_RECLAIM); 175 if (IS_ERR(req)) 176 return ret; 177 blk_rq_set_block_pc(req); 178 179 if (bufflen && blk_rq_map_kern(sdev->request_queue, req, 180 buffer, bufflen, __GFP_RECLAIM)) 181 goto out; 182 183 req->cmd_len = COMMAND_SIZE(cmd[0]); 184 memcpy(req->cmd, cmd, req->cmd_len); 185 req->sense = sense; 186 req->sense_len = 0; 187 req->retries = retries; 188 req->timeout = timeout; 189 req->cmd_flags |= flags; 190 req->rq_flags |= rq_flags | RQF_QUIET | RQF_PREEMPT; 191 192 /* 193 * head injection *required* here otherwise quiesce won't work 194 */ 195 blk_execute_rq(req->q, NULL, req, 1); 196 197 /* 198 * Some devices (USB mass-storage in particular) may transfer 199 * garbage data together with a residue indicating that the data 200 * is invalid. Prevent the garbage from being misinterpreted 201 * and prevent security leaks by zeroing out the excess data. 202 */ 203 if (unlikely(req->resid_len > 0 && req->resid_len <= bufflen)) 204 memset(buffer + (bufflen - req->resid_len), 0, req->resid_len); 205 206 if (resid) 207 *resid = req->resid_len; 208 ret = req->errors; 209 out: 210 blk_put_request(req); 211 212 return ret; 213 } 214 215 /** 216 * scsi_execute - insert request and wait for the result 217 * @sdev: scsi device 218 * @cmd: scsi command 219 * @data_direction: data direction 220 * @buffer: data buffer 221 * @bufflen: len of buffer 222 * @sense: optional sense buffer 223 * @timeout: request timeout in seconds 224 * @retries: number of times to retry request 225 * @flags: or into request flags; 226 * @resid: optional residual length 227 * 228 * returns the req->errors value which is the scsi_cmnd result 229 * field. 230 */ 231 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd, 232 int data_direction, void *buffer, unsigned bufflen, 233 unsigned char *sense, int timeout, int retries, u64 flags, 234 int *resid) 235 { 236 return __scsi_execute(sdev, cmd, data_direction, buffer, bufflen, sense, 237 timeout, retries, flags, 0, resid); 238 } 239 EXPORT_SYMBOL(scsi_execute); 240 241 int scsi_execute_req_flags(struct scsi_device *sdev, const unsigned char *cmd, 242 int data_direction, void *buffer, unsigned bufflen, 243 struct scsi_sense_hdr *sshdr, int timeout, int retries, 244 int *resid, u64 flags, req_flags_t rq_flags) 245 { 246 char *sense = NULL; 247 int result; 248 249 if (sshdr) { 250 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO); 251 if (!sense) 252 return DRIVER_ERROR << 24; 253 } 254 result = __scsi_execute(sdev, cmd, data_direction, buffer, bufflen, 255 sense, timeout, retries, flags, rq_flags, resid); 256 if (sshdr) 257 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr); 258 259 kfree(sense); 260 return result; 261 } 262 EXPORT_SYMBOL(scsi_execute_req_flags); 263 264 /* 265 * Function: scsi_init_cmd_errh() 266 * 267 * Purpose: Initialize cmd fields related to error handling. 268 * 269 * Arguments: cmd - command that is ready to be queued. 270 * 271 * Notes: This function has the job of initializing a number of 272 * fields related to error handling. Typically this will 273 * be called once for each command, as required. 274 */ 275 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd) 276 { 277 cmd->serial_number = 0; 278 scsi_set_resid(cmd, 0); 279 memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE); 280 if (cmd->cmd_len == 0) 281 cmd->cmd_len = scsi_command_size(cmd->cmnd); 282 } 283 284 void scsi_device_unbusy(struct scsi_device *sdev) 285 { 286 struct Scsi_Host *shost = sdev->host; 287 struct scsi_target *starget = scsi_target(sdev); 288 unsigned long flags; 289 290 atomic_dec(&shost->host_busy); 291 if (starget->can_queue > 0) 292 atomic_dec(&starget->target_busy); 293 294 if (unlikely(scsi_host_in_recovery(shost) && 295 (shost->host_failed || shost->host_eh_scheduled))) { 296 spin_lock_irqsave(shost->host_lock, flags); 297 scsi_eh_wakeup(shost); 298 spin_unlock_irqrestore(shost->host_lock, flags); 299 } 300 301 atomic_dec(&sdev->device_busy); 302 } 303 304 static void scsi_kick_queue(struct request_queue *q) 305 { 306 if (q->mq_ops) 307 blk_mq_start_hw_queues(q); 308 else 309 blk_run_queue(q); 310 } 311 312 /* 313 * Called for single_lun devices on IO completion. Clear starget_sdev_user, 314 * and call blk_run_queue for all the scsi_devices on the target - 315 * including current_sdev first. 316 * 317 * Called with *no* scsi locks held. 318 */ 319 static void scsi_single_lun_run(struct scsi_device *current_sdev) 320 { 321 struct Scsi_Host *shost = current_sdev->host; 322 struct scsi_device *sdev, *tmp; 323 struct scsi_target *starget = scsi_target(current_sdev); 324 unsigned long flags; 325 326 spin_lock_irqsave(shost->host_lock, flags); 327 starget->starget_sdev_user = NULL; 328 spin_unlock_irqrestore(shost->host_lock, flags); 329 330 /* 331 * Call blk_run_queue for all LUNs on the target, starting with 332 * current_sdev. We race with others (to set starget_sdev_user), 333 * but in most cases, we will be first. Ideally, each LU on the 334 * target would get some limited time or requests on the target. 335 */ 336 scsi_kick_queue(current_sdev->request_queue); 337 338 spin_lock_irqsave(shost->host_lock, flags); 339 if (starget->starget_sdev_user) 340 goto out; 341 list_for_each_entry_safe(sdev, tmp, &starget->devices, 342 same_target_siblings) { 343 if (sdev == current_sdev) 344 continue; 345 if (scsi_device_get(sdev)) 346 continue; 347 348 spin_unlock_irqrestore(shost->host_lock, flags); 349 scsi_kick_queue(sdev->request_queue); 350 spin_lock_irqsave(shost->host_lock, flags); 351 352 scsi_device_put(sdev); 353 } 354 out: 355 spin_unlock_irqrestore(shost->host_lock, flags); 356 } 357 358 static inline bool scsi_device_is_busy(struct scsi_device *sdev) 359 { 360 if (atomic_read(&sdev->device_busy) >= sdev->queue_depth) 361 return true; 362 if (atomic_read(&sdev->device_blocked) > 0) 363 return true; 364 return false; 365 } 366 367 static inline bool scsi_target_is_busy(struct scsi_target *starget) 368 { 369 if (starget->can_queue > 0) { 370 if (atomic_read(&starget->target_busy) >= starget->can_queue) 371 return true; 372 if (atomic_read(&starget->target_blocked) > 0) 373 return true; 374 } 375 return false; 376 } 377 378 static inline bool scsi_host_is_busy(struct Scsi_Host *shost) 379 { 380 if (shost->can_queue > 0 && 381 atomic_read(&shost->host_busy) >= shost->can_queue) 382 return true; 383 if (atomic_read(&shost->host_blocked) > 0) 384 return true; 385 if (shost->host_self_blocked) 386 return true; 387 return false; 388 } 389 390 static void scsi_starved_list_run(struct Scsi_Host *shost) 391 { 392 LIST_HEAD(starved_list); 393 struct scsi_device *sdev; 394 unsigned long flags; 395 396 spin_lock_irqsave(shost->host_lock, flags); 397 list_splice_init(&shost->starved_list, &starved_list); 398 399 while (!list_empty(&starved_list)) { 400 struct request_queue *slq; 401 402 /* 403 * As long as shost is accepting commands and we have 404 * starved queues, call blk_run_queue. scsi_request_fn 405 * drops the queue_lock and can add us back to the 406 * starved_list. 407 * 408 * host_lock protects the starved_list and starved_entry. 409 * scsi_request_fn must get the host_lock before checking 410 * or modifying starved_list or starved_entry. 411 */ 412 if (scsi_host_is_busy(shost)) 413 break; 414 415 sdev = list_entry(starved_list.next, 416 struct scsi_device, starved_entry); 417 list_del_init(&sdev->starved_entry); 418 if (scsi_target_is_busy(scsi_target(sdev))) { 419 list_move_tail(&sdev->starved_entry, 420 &shost->starved_list); 421 continue; 422 } 423 424 /* 425 * Once we drop the host lock, a racing scsi_remove_device() 426 * call may remove the sdev from the starved list and destroy 427 * it and the queue. Mitigate by taking a reference to the 428 * queue and never touching the sdev again after we drop the 429 * host lock. Note: if __scsi_remove_device() invokes 430 * blk_cleanup_queue() before the queue is run from this 431 * function then blk_run_queue() will return immediately since 432 * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING. 433 */ 434 slq = sdev->request_queue; 435 if (!blk_get_queue(slq)) 436 continue; 437 spin_unlock_irqrestore(shost->host_lock, flags); 438 439 scsi_kick_queue(slq); 440 blk_put_queue(slq); 441 442 spin_lock_irqsave(shost->host_lock, flags); 443 } 444 /* put any unprocessed entries back */ 445 list_splice(&starved_list, &shost->starved_list); 446 spin_unlock_irqrestore(shost->host_lock, flags); 447 } 448 449 /* 450 * Function: scsi_run_queue() 451 * 452 * Purpose: Select a proper request queue to serve next 453 * 454 * Arguments: q - last request's queue 455 * 456 * Returns: Nothing 457 * 458 * Notes: The previous command was completely finished, start 459 * a new one if possible. 460 */ 461 static void scsi_run_queue(struct request_queue *q) 462 { 463 struct scsi_device *sdev = q->queuedata; 464 465 if (scsi_target(sdev)->single_lun) 466 scsi_single_lun_run(sdev); 467 if (!list_empty(&sdev->host->starved_list)) 468 scsi_starved_list_run(sdev->host); 469 470 if (q->mq_ops) 471 blk_mq_start_stopped_hw_queues(q, false); 472 else 473 blk_run_queue(q); 474 } 475 476 void scsi_requeue_run_queue(struct work_struct *work) 477 { 478 struct scsi_device *sdev; 479 struct request_queue *q; 480 481 sdev = container_of(work, struct scsi_device, requeue_work); 482 q = sdev->request_queue; 483 scsi_run_queue(q); 484 } 485 486 /* 487 * Function: scsi_requeue_command() 488 * 489 * Purpose: Handle post-processing of completed commands. 490 * 491 * Arguments: q - queue to operate on 492 * cmd - command that may need to be requeued. 493 * 494 * Returns: Nothing 495 * 496 * Notes: After command completion, there may be blocks left 497 * over which weren't finished by the previous command 498 * this can be for a number of reasons - the main one is 499 * I/O errors in the middle of the request, in which case 500 * we need to request the blocks that come after the bad 501 * sector. 502 * Notes: Upon return, cmd is a stale pointer. 503 */ 504 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd) 505 { 506 struct scsi_device *sdev = cmd->device; 507 struct request *req = cmd->request; 508 unsigned long flags; 509 510 spin_lock_irqsave(q->queue_lock, flags); 511 blk_unprep_request(req); 512 req->special = NULL; 513 scsi_put_command(cmd); 514 blk_requeue_request(q, req); 515 spin_unlock_irqrestore(q->queue_lock, flags); 516 517 scsi_run_queue(q); 518 519 put_device(&sdev->sdev_gendev); 520 } 521 522 void scsi_run_host_queues(struct Scsi_Host *shost) 523 { 524 struct scsi_device *sdev; 525 526 shost_for_each_device(sdev, shost) 527 scsi_run_queue(sdev->request_queue); 528 } 529 530 static void scsi_uninit_cmd(struct scsi_cmnd *cmd) 531 { 532 if (cmd->request->cmd_type == REQ_TYPE_FS) { 533 struct scsi_driver *drv = scsi_cmd_to_driver(cmd); 534 535 if (drv->uninit_command) 536 drv->uninit_command(cmd); 537 } 538 } 539 540 static void scsi_mq_free_sgtables(struct scsi_cmnd *cmd) 541 { 542 struct scsi_data_buffer *sdb; 543 544 if (cmd->sdb.table.nents) 545 sg_free_table_chained(&cmd->sdb.table, true); 546 if (cmd->request->next_rq) { 547 sdb = cmd->request->next_rq->special; 548 if (sdb) 549 sg_free_table_chained(&sdb->table, true); 550 } 551 if (scsi_prot_sg_count(cmd)) 552 sg_free_table_chained(&cmd->prot_sdb->table, true); 553 } 554 555 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd) 556 { 557 struct scsi_device *sdev = cmd->device; 558 struct Scsi_Host *shost = sdev->host; 559 unsigned long flags; 560 561 scsi_mq_free_sgtables(cmd); 562 scsi_uninit_cmd(cmd); 563 564 if (shost->use_cmd_list) { 565 BUG_ON(list_empty(&cmd->list)); 566 spin_lock_irqsave(&sdev->list_lock, flags); 567 list_del_init(&cmd->list); 568 spin_unlock_irqrestore(&sdev->list_lock, flags); 569 } 570 } 571 572 /* 573 * Function: scsi_release_buffers() 574 * 575 * Purpose: Free resources allocate for a scsi_command. 576 * 577 * Arguments: cmd - command that we are bailing. 578 * 579 * Lock status: Assumed that no lock is held upon entry. 580 * 581 * Returns: Nothing 582 * 583 * Notes: In the event that an upper level driver rejects a 584 * command, we must release resources allocated during 585 * the __init_io() function. Primarily this would involve 586 * the scatter-gather table. 587 */ 588 static void scsi_release_buffers(struct scsi_cmnd *cmd) 589 { 590 if (cmd->sdb.table.nents) 591 sg_free_table_chained(&cmd->sdb.table, false); 592 593 memset(&cmd->sdb, 0, sizeof(cmd->sdb)); 594 595 if (scsi_prot_sg_count(cmd)) 596 sg_free_table_chained(&cmd->prot_sdb->table, false); 597 } 598 599 static void scsi_release_bidi_buffers(struct scsi_cmnd *cmd) 600 { 601 struct scsi_data_buffer *bidi_sdb = cmd->request->next_rq->special; 602 603 sg_free_table_chained(&bidi_sdb->table, false); 604 kmem_cache_free(scsi_sdb_cache, bidi_sdb); 605 cmd->request->next_rq->special = NULL; 606 } 607 608 static bool scsi_end_request(struct request *req, int error, 609 unsigned int bytes, unsigned int bidi_bytes) 610 { 611 struct scsi_cmnd *cmd = req->special; 612 struct scsi_device *sdev = cmd->device; 613 struct request_queue *q = sdev->request_queue; 614 615 if (blk_update_request(req, error, bytes)) 616 return true; 617 618 /* Bidi request must be completed as a whole */ 619 if (unlikely(bidi_bytes) && 620 blk_update_request(req->next_rq, error, bidi_bytes)) 621 return true; 622 623 if (blk_queue_add_random(q)) 624 add_disk_randomness(req->rq_disk); 625 626 if (req->mq_ctx) { 627 /* 628 * In the MQ case the command gets freed by __blk_mq_end_request, 629 * so we have to do all cleanup that depends on it earlier. 630 * 631 * We also can't kick the queues from irq context, so we 632 * will have to defer it to a workqueue. 633 */ 634 scsi_mq_uninit_cmd(cmd); 635 636 __blk_mq_end_request(req, error); 637 638 if (scsi_target(sdev)->single_lun || 639 !list_empty(&sdev->host->starved_list)) 640 kblockd_schedule_work(&sdev->requeue_work); 641 else 642 blk_mq_start_stopped_hw_queues(q, true); 643 } else { 644 unsigned long flags; 645 646 if (bidi_bytes) 647 scsi_release_bidi_buffers(cmd); 648 649 spin_lock_irqsave(q->queue_lock, flags); 650 blk_finish_request(req, error); 651 spin_unlock_irqrestore(q->queue_lock, flags); 652 653 scsi_release_buffers(cmd); 654 655 scsi_put_command(cmd); 656 scsi_run_queue(q); 657 } 658 659 put_device(&sdev->sdev_gendev); 660 return false; 661 } 662 663 /** 664 * __scsi_error_from_host_byte - translate SCSI error code into errno 665 * @cmd: SCSI command (unused) 666 * @result: scsi error code 667 * 668 * Translate SCSI error code into standard UNIX errno. 669 * Return values: 670 * -ENOLINK temporary transport failure 671 * -EREMOTEIO permanent target failure, do not retry 672 * -EBADE permanent nexus failure, retry on other path 673 * -ENOSPC No write space available 674 * -ENODATA Medium error 675 * -EIO unspecified I/O error 676 */ 677 static int __scsi_error_from_host_byte(struct scsi_cmnd *cmd, int result) 678 { 679 int error = 0; 680 681 switch(host_byte(result)) { 682 case DID_TRANSPORT_FAILFAST: 683 error = -ENOLINK; 684 break; 685 case DID_TARGET_FAILURE: 686 set_host_byte(cmd, DID_OK); 687 error = -EREMOTEIO; 688 break; 689 case DID_NEXUS_FAILURE: 690 set_host_byte(cmd, DID_OK); 691 error = -EBADE; 692 break; 693 case DID_ALLOC_FAILURE: 694 set_host_byte(cmd, DID_OK); 695 error = -ENOSPC; 696 break; 697 case DID_MEDIUM_ERROR: 698 set_host_byte(cmd, DID_OK); 699 error = -ENODATA; 700 break; 701 default: 702 error = -EIO; 703 break; 704 } 705 706 return error; 707 } 708 709 /* 710 * Function: scsi_io_completion() 711 * 712 * Purpose: Completion processing for block device I/O requests. 713 * 714 * Arguments: cmd - command that is finished. 715 * 716 * Lock status: Assumed that no lock is held upon entry. 717 * 718 * Returns: Nothing 719 * 720 * Notes: We will finish off the specified number of sectors. If we 721 * are done, the command block will be released and the queue 722 * function will be goosed. If we are not done then we have to 723 * figure out what to do next: 724 * 725 * a) We can call scsi_requeue_command(). The request 726 * will be unprepared and put back on the queue. Then 727 * a new command will be created for it. This should 728 * be used if we made forward progress, or if we want 729 * to switch from READ(10) to READ(6) for example. 730 * 731 * b) We can call __scsi_queue_insert(). The request will 732 * be put back on the queue and retried using the same 733 * command as before, possibly after a delay. 734 * 735 * c) We can call scsi_end_request() with -EIO to fail 736 * the remainder of the request. 737 */ 738 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes) 739 { 740 int result = cmd->result; 741 struct request_queue *q = cmd->device->request_queue; 742 struct request *req = cmd->request; 743 int error = 0; 744 struct scsi_sense_hdr sshdr; 745 bool sense_valid = false; 746 int sense_deferred = 0, level = 0; 747 enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY, 748 ACTION_DELAYED_RETRY} action; 749 unsigned long wait_for = (cmd->allowed + 1) * req->timeout; 750 751 if (result) { 752 sense_valid = scsi_command_normalize_sense(cmd, &sshdr); 753 if (sense_valid) 754 sense_deferred = scsi_sense_is_deferred(&sshdr); 755 } 756 757 if (req->cmd_type == REQ_TYPE_BLOCK_PC) { /* SG_IO ioctl from block level */ 758 if (result) { 759 if (sense_valid && req->sense) { 760 /* 761 * SG_IO wants current and deferred errors 762 */ 763 int len = 8 + cmd->sense_buffer[7]; 764 765 if (len > SCSI_SENSE_BUFFERSIZE) 766 len = SCSI_SENSE_BUFFERSIZE; 767 memcpy(req->sense, cmd->sense_buffer, len); 768 req->sense_len = len; 769 } 770 if (!sense_deferred) 771 error = __scsi_error_from_host_byte(cmd, result); 772 } 773 /* 774 * __scsi_error_from_host_byte may have reset the host_byte 775 */ 776 req->errors = cmd->result; 777 778 req->resid_len = scsi_get_resid(cmd); 779 780 if (scsi_bidi_cmnd(cmd)) { 781 /* 782 * Bidi commands Must be complete as a whole, 783 * both sides at once. 784 */ 785 req->next_rq->resid_len = scsi_in(cmd)->resid; 786 if (scsi_end_request(req, 0, blk_rq_bytes(req), 787 blk_rq_bytes(req->next_rq))) 788 BUG(); 789 return; 790 } 791 } else if (blk_rq_bytes(req) == 0 && result && !sense_deferred) { 792 /* 793 * Certain non BLOCK_PC requests are commands that don't 794 * actually transfer anything (FLUSH), so cannot use 795 * good_bytes != blk_rq_bytes(req) as the signal for an error. 796 * This sets the error explicitly for the problem case. 797 */ 798 error = __scsi_error_from_host_byte(cmd, result); 799 } 800 801 /* no bidi support for !REQ_TYPE_BLOCK_PC yet */ 802 BUG_ON(blk_bidi_rq(req)); 803 804 /* 805 * Next deal with any sectors which we were able to correctly 806 * handle. 807 */ 808 SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd, 809 "%u sectors total, %d bytes done.\n", 810 blk_rq_sectors(req), good_bytes)); 811 812 /* 813 * Recovered errors need reporting, but they're always treated 814 * as success, so fiddle the result code here. For BLOCK_PC 815 * we already took a copy of the original into rq->errors which 816 * is what gets returned to the user 817 */ 818 if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) { 819 /* if ATA PASS-THROUGH INFORMATION AVAILABLE skip 820 * print since caller wants ATA registers. Only occurs on 821 * SCSI ATA PASS_THROUGH commands when CK_COND=1 822 */ 823 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d)) 824 ; 825 else if (!(req->rq_flags & RQF_QUIET)) 826 scsi_print_sense(cmd); 827 result = 0; 828 /* BLOCK_PC may have set error */ 829 error = 0; 830 } 831 832 /* 833 * special case: failed zero length commands always need to 834 * drop down into the retry code. Otherwise, if we finished 835 * all bytes in the request we are done now. 836 */ 837 if (!(blk_rq_bytes(req) == 0 && error) && 838 !scsi_end_request(req, error, good_bytes, 0)) 839 return; 840 841 /* 842 * Kill remainder if no retrys. 843 */ 844 if (error && scsi_noretry_cmd(cmd)) { 845 if (scsi_end_request(req, error, blk_rq_bytes(req), 0)) 846 BUG(); 847 return; 848 } 849 850 /* 851 * If there had been no error, but we have leftover bytes in the 852 * requeues just queue the command up again. 853 */ 854 if (result == 0) 855 goto requeue; 856 857 error = __scsi_error_from_host_byte(cmd, result); 858 859 if (host_byte(result) == DID_RESET) { 860 /* Third party bus reset or reset for error recovery 861 * reasons. Just retry the command and see what 862 * happens. 863 */ 864 action = ACTION_RETRY; 865 } else if (sense_valid && !sense_deferred) { 866 switch (sshdr.sense_key) { 867 case UNIT_ATTENTION: 868 if (cmd->device->removable) { 869 /* Detected disc change. Set a bit 870 * and quietly refuse further access. 871 */ 872 cmd->device->changed = 1; 873 action = ACTION_FAIL; 874 } else { 875 /* Must have been a power glitch, or a 876 * bus reset. Could not have been a 877 * media change, so we just retry the 878 * command and see what happens. 879 */ 880 action = ACTION_RETRY; 881 } 882 break; 883 case ILLEGAL_REQUEST: 884 /* If we had an ILLEGAL REQUEST returned, then 885 * we may have performed an unsupported 886 * command. The only thing this should be 887 * would be a ten byte read where only a six 888 * byte read was supported. Also, on a system 889 * where READ CAPACITY failed, we may have 890 * read past the end of the disk. 891 */ 892 if ((cmd->device->use_10_for_rw && 893 sshdr.asc == 0x20 && sshdr.ascq == 0x00) && 894 (cmd->cmnd[0] == READ_10 || 895 cmd->cmnd[0] == WRITE_10)) { 896 /* This will issue a new 6-byte command. */ 897 cmd->device->use_10_for_rw = 0; 898 action = ACTION_REPREP; 899 } else if (sshdr.asc == 0x10) /* DIX */ { 900 action = ACTION_FAIL; 901 error = -EILSEQ; 902 /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */ 903 } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) { 904 action = ACTION_FAIL; 905 error = -EREMOTEIO; 906 } else 907 action = ACTION_FAIL; 908 break; 909 case ABORTED_COMMAND: 910 action = ACTION_FAIL; 911 if (sshdr.asc == 0x10) /* DIF */ 912 error = -EILSEQ; 913 break; 914 case NOT_READY: 915 /* If the device is in the process of becoming 916 * ready, or has a temporary blockage, retry. 917 */ 918 if (sshdr.asc == 0x04) { 919 switch (sshdr.ascq) { 920 case 0x01: /* becoming ready */ 921 case 0x04: /* format in progress */ 922 case 0x05: /* rebuild in progress */ 923 case 0x06: /* recalculation in progress */ 924 case 0x07: /* operation in progress */ 925 case 0x08: /* Long write in progress */ 926 case 0x09: /* self test in progress */ 927 case 0x14: /* space allocation in progress */ 928 action = ACTION_DELAYED_RETRY; 929 break; 930 default: 931 action = ACTION_FAIL; 932 break; 933 } 934 } else 935 action = ACTION_FAIL; 936 break; 937 case VOLUME_OVERFLOW: 938 /* See SSC3rXX or current. */ 939 action = ACTION_FAIL; 940 break; 941 default: 942 action = ACTION_FAIL; 943 break; 944 } 945 } else 946 action = ACTION_FAIL; 947 948 if (action != ACTION_FAIL && 949 time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) 950 action = ACTION_FAIL; 951 952 switch (action) { 953 case ACTION_FAIL: 954 /* Give up and fail the remainder of the request */ 955 if (!(req->rq_flags & RQF_QUIET)) { 956 static DEFINE_RATELIMIT_STATE(_rs, 957 DEFAULT_RATELIMIT_INTERVAL, 958 DEFAULT_RATELIMIT_BURST); 959 960 if (unlikely(scsi_logging_level)) 961 level = SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT, 962 SCSI_LOG_MLCOMPLETE_BITS); 963 964 /* 965 * if logging is enabled the failure will be printed 966 * in scsi_log_completion(), so avoid duplicate messages 967 */ 968 if (!level && __ratelimit(&_rs)) { 969 scsi_print_result(cmd, NULL, FAILED); 970 if (driver_byte(result) & DRIVER_SENSE) 971 scsi_print_sense(cmd); 972 scsi_print_command(cmd); 973 } 974 } 975 if (!scsi_end_request(req, error, blk_rq_err_bytes(req), 0)) 976 return; 977 /*FALLTHRU*/ 978 case ACTION_REPREP: 979 requeue: 980 /* Unprep the request and put it back at the head of the queue. 981 * A new command will be prepared and issued. 982 */ 983 if (q->mq_ops) { 984 cmd->request->rq_flags &= ~RQF_DONTPREP; 985 scsi_mq_uninit_cmd(cmd); 986 scsi_mq_requeue_cmd(cmd); 987 } else { 988 scsi_release_buffers(cmd); 989 scsi_requeue_command(q, cmd); 990 } 991 break; 992 case ACTION_RETRY: 993 /* Retry the same command immediately */ 994 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, 0); 995 break; 996 case ACTION_DELAYED_RETRY: 997 /* Retry the same command after a delay */ 998 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, 0); 999 break; 1000 } 1001 } 1002 1003 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb) 1004 { 1005 int count; 1006 1007 /* 1008 * If sg table allocation fails, requeue request later. 1009 */ 1010 if (unlikely(sg_alloc_table_chained(&sdb->table, 1011 blk_rq_nr_phys_segments(req), sdb->table.sgl))) 1012 return BLKPREP_DEFER; 1013 1014 /* 1015 * Next, walk the list, and fill in the addresses and sizes of 1016 * each segment. 1017 */ 1018 count = blk_rq_map_sg(req->q, req, sdb->table.sgl); 1019 BUG_ON(count > sdb->table.nents); 1020 sdb->table.nents = count; 1021 sdb->length = blk_rq_bytes(req); 1022 return BLKPREP_OK; 1023 } 1024 1025 /* 1026 * Function: scsi_init_io() 1027 * 1028 * Purpose: SCSI I/O initialize function. 1029 * 1030 * Arguments: cmd - Command descriptor we wish to initialize 1031 * 1032 * Returns: 0 on success 1033 * BLKPREP_DEFER if the failure is retryable 1034 * BLKPREP_KILL if the failure is fatal 1035 */ 1036 int scsi_init_io(struct scsi_cmnd *cmd) 1037 { 1038 struct scsi_device *sdev = cmd->device; 1039 struct request *rq = cmd->request; 1040 bool is_mq = (rq->mq_ctx != NULL); 1041 int error; 1042 1043 BUG_ON(!blk_rq_nr_phys_segments(rq)); 1044 1045 error = scsi_init_sgtable(rq, &cmd->sdb); 1046 if (error) 1047 goto err_exit; 1048 1049 if (blk_bidi_rq(rq)) { 1050 if (!rq->q->mq_ops) { 1051 struct scsi_data_buffer *bidi_sdb = 1052 kmem_cache_zalloc(scsi_sdb_cache, GFP_ATOMIC); 1053 if (!bidi_sdb) { 1054 error = BLKPREP_DEFER; 1055 goto err_exit; 1056 } 1057 1058 rq->next_rq->special = bidi_sdb; 1059 } 1060 1061 error = scsi_init_sgtable(rq->next_rq, rq->next_rq->special); 1062 if (error) 1063 goto err_exit; 1064 } 1065 1066 if (blk_integrity_rq(rq)) { 1067 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb; 1068 int ivecs, count; 1069 1070 if (prot_sdb == NULL) { 1071 /* 1072 * This can happen if someone (e.g. multipath) 1073 * queues a command to a device on an adapter 1074 * that does not support DIX. 1075 */ 1076 WARN_ON_ONCE(1); 1077 error = BLKPREP_KILL; 1078 goto err_exit; 1079 } 1080 1081 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio); 1082 1083 if (sg_alloc_table_chained(&prot_sdb->table, ivecs, 1084 prot_sdb->table.sgl)) { 1085 error = BLKPREP_DEFER; 1086 goto err_exit; 1087 } 1088 1089 count = blk_rq_map_integrity_sg(rq->q, rq->bio, 1090 prot_sdb->table.sgl); 1091 BUG_ON(unlikely(count > ivecs)); 1092 BUG_ON(unlikely(count > queue_max_integrity_segments(rq->q))); 1093 1094 cmd->prot_sdb = prot_sdb; 1095 cmd->prot_sdb->table.nents = count; 1096 } 1097 1098 return BLKPREP_OK; 1099 err_exit: 1100 if (is_mq) { 1101 scsi_mq_free_sgtables(cmd); 1102 } else { 1103 scsi_release_buffers(cmd); 1104 cmd->request->special = NULL; 1105 scsi_put_command(cmd); 1106 put_device(&sdev->sdev_gendev); 1107 } 1108 return error; 1109 } 1110 EXPORT_SYMBOL(scsi_init_io); 1111 1112 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev, 1113 struct request *req) 1114 { 1115 struct scsi_cmnd *cmd; 1116 1117 if (!req->special) { 1118 /* Bail if we can't get a reference to the device */ 1119 if (!get_device(&sdev->sdev_gendev)) 1120 return NULL; 1121 1122 cmd = scsi_get_command(sdev, GFP_ATOMIC); 1123 if (unlikely(!cmd)) { 1124 put_device(&sdev->sdev_gendev); 1125 return NULL; 1126 } 1127 req->special = cmd; 1128 } else { 1129 cmd = req->special; 1130 } 1131 1132 /* pull a tag out of the request if we have one */ 1133 cmd->tag = req->tag; 1134 cmd->request = req; 1135 1136 cmd->cmnd = req->cmd; 1137 cmd->prot_op = SCSI_PROT_NORMAL; 1138 1139 return cmd; 1140 } 1141 1142 static int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req) 1143 { 1144 struct scsi_cmnd *cmd = req->special; 1145 1146 /* 1147 * BLOCK_PC requests may transfer data, in which case they must 1148 * a bio attached to them. Or they might contain a SCSI command 1149 * that does not transfer data, in which case they may optionally 1150 * submit a request without an attached bio. 1151 */ 1152 if (req->bio) { 1153 int ret = scsi_init_io(cmd); 1154 if (unlikely(ret)) 1155 return ret; 1156 } else { 1157 BUG_ON(blk_rq_bytes(req)); 1158 1159 memset(&cmd->sdb, 0, sizeof(cmd->sdb)); 1160 } 1161 1162 cmd->cmd_len = req->cmd_len; 1163 cmd->transfersize = blk_rq_bytes(req); 1164 cmd->allowed = req->retries; 1165 return BLKPREP_OK; 1166 } 1167 1168 /* 1169 * Setup a REQ_TYPE_FS command. These are simple request from filesystems 1170 * that still need to be translated to SCSI CDBs from the ULD. 1171 */ 1172 static int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req) 1173 { 1174 struct scsi_cmnd *cmd = req->special; 1175 1176 if (unlikely(sdev->handler && sdev->handler->prep_fn)) { 1177 int ret = sdev->handler->prep_fn(sdev, req); 1178 if (ret != BLKPREP_OK) 1179 return ret; 1180 } 1181 1182 memset(cmd->cmnd, 0, BLK_MAX_CDB); 1183 return scsi_cmd_to_driver(cmd)->init_command(cmd); 1184 } 1185 1186 static int scsi_setup_cmnd(struct scsi_device *sdev, struct request *req) 1187 { 1188 struct scsi_cmnd *cmd = req->special; 1189 1190 if (!blk_rq_bytes(req)) 1191 cmd->sc_data_direction = DMA_NONE; 1192 else if (rq_data_dir(req) == WRITE) 1193 cmd->sc_data_direction = DMA_TO_DEVICE; 1194 else 1195 cmd->sc_data_direction = DMA_FROM_DEVICE; 1196 1197 switch (req->cmd_type) { 1198 case REQ_TYPE_FS: 1199 return scsi_setup_fs_cmnd(sdev, req); 1200 case REQ_TYPE_BLOCK_PC: 1201 return scsi_setup_blk_pc_cmnd(sdev, req); 1202 default: 1203 return BLKPREP_KILL; 1204 } 1205 } 1206 1207 static int 1208 scsi_prep_state_check(struct scsi_device *sdev, struct request *req) 1209 { 1210 int ret = BLKPREP_OK; 1211 1212 /* 1213 * If the device is not in running state we will reject some 1214 * or all commands. 1215 */ 1216 if (unlikely(sdev->sdev_state != SDEV_RUNNING)) { 1217 switch (sdev->sdev_state) { 1218 case SDEV_OFFLINE: 1219 case SDEV_TRANSPORT_OFFLINE: 1220 /* 1221 * If the device is offline we refuse to process any 1222 * commands. The device must be brought online 1223 * before trying any recovery commands. 1224 */ 1225 sdev_printk(KERN_ERR, sdev, 1226 "rejecting I/O to offline device\n"); 1227 ret = BLKPREP_KILL; 1228 break; 1229 case SDEV_DEL: 1230 /* 1231 * If the device is fully deleted, we refuse to 1232 * process any commands as well. 1233 */ 1234 sdev_printk(KERN_ERR, sdev, 1235 "rejecting I/O to dead device\n"); 1236 ret = BLKPREP_KILL; 1237 break; 1238 case SDEV_BLOCK: 1239 case SDEV_CREATED_BLOCK: 1240 ret = BLKPREP_DEFER; 1241 break; 1242 case SDEV_QUIESCE: 1243 /* 1244 * If the devices is blocked we defer normal commands. 1245 */ 1246 if (!(req->rq_flags & RQF_PREEMPT)) 1247 ret = BLKPREP_DEFER; 1248 break; 1249 default: 1250 /* 1251 * For any other not fully online state we only allow 1252 * special commands. In particular any user initiated 1253 * command is not allowed. 1254 */ 1255 if (!(req->rq_flags & RQF_PREEMPT)) 1256 ret = BLKPREP_KILL; 1257 break; 1258 } 1259 } 1260 return ret; 1261 } 1262 1263 static int 1264 scsi_prep_return(struct request_queue *q, struct request *req, int ret) 1265 { 1266 struct scsi_device *sdev = q->queuedata; 1267 1268 switch (ret) { 1269 case BLKPREP_KILL: 1270 case BLKPREP_INVALID: 1271 req->errors = DID_NO_CONNECT << 16; 1272 /* release the command and kill it */ 1273 if (req->special) { 1274 struct scsi_cmnd *cmd = req->special; 1275 scsi_release_buffers(cmd); 1276 scsi_put_command(cmd); 1277 put_device(&sdev->sdev_gendev); 1278 req->special = NULL; 1279 } 1280 break; 1281 case BLKPREP_DEFER: 1282 /* 1283 * If we defer, the blk_peek_request() returns NULL, but the 1284 * queue must be restarted, so we schedule a callback to happen 1285 * shortly. 1286 */ 1287 if (atomic_read(&sdev->device_busy) == 0) 1288 blk_delay_queue(q, SCSI_QUEUE_DELAY); 1289 break; 1290 default: 1291 req->rq_flags |= RQF_DONTPREP; 1292 } 1293 1294 return ret; 1295 } 1296 1297 static int scsi_prep_fn(struct request_queue *q, struct request *req) 1298 { 1299 struct scsi_device *sdev = q->queuedata; 1300 struct scsi_cmnd *cmd; 1301 int ret; 1302 1303 ret = scsi_prep_state_check(sdev, req); 1304 if (ret != BLKPREP_OK) 1305 goto out; 1306 1307 cmd = scsi_get_cmd_from_req(sdev, req); 1308 if (unlikely(!cmd)) { 1309 ret = BLKPREP_DEFER; 1310 goto out; 1311 } 1312 1313 ret = scsi_setup_cmnd(sdev, req); 1314 out: 1315 return scsi_prep_return(q, req, ret); 1316 } 1317 1318 static void scsi_unprep_fn(struct request_queue *q, struct request *req) 1319 { 1320 scsi_uninit_cmd(req->special); 1321 } 1322 1323 /* 1324 * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else 1325 * return 0. 1326 * 1327 * Called with the queue_lock held. 1328 */ 1329 static inline int scsi_dev_queue_ready(struct request_queue *q, 1330 struct scsi_device *sdev) 1331 { 1332 unsigned int busy; 1333 1334 busy = atomic_inc_return(&sdev->device_busy) - 1; 1335 if (atomic_read(&sdev->device_blocked)) { 1336 if (busy) 1337 goto out_dec; 1338 1339 /* 1340 * unblock after device_blocked iterates to zero 1341 */ 1342 if (atomic_dec_return(&sdev->device_blocked) > 0) { 1343 /* 1344 * For the MQ case we take care of this in the caller. 1345 */ 1346 if (!q->mq_ops) 1347 blk_delay_queue(q, SCSI_QUEUE_DELAY); 1348 goto out_dec; 1349 } 1350 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev, 1351 "unblocking device at zero depth\n")); 1352 } 1353 1354 if (busy >= sdev->queue_depth) 1355 goto out_dec; 1356 1357 return 1; 1358 out_dec: 1359 atomic_dec(&sdev->device_busy); 1360 return 0; 1361 } 1362 1363 /* 1364 * scsi_target_queue_ready: checks if there we can send commands to target 1365 * @sdev: scsi device on starget to check. 1366 */ 1367 static inline int scsi_target_queue_ready(struct Scsi_Host *shost, 1368 struct scsi_device *sdev) 1369 { 1370 struct scsi_target *starget = scsi_target(sdev); 1371 unsigned int busy; 1372 1373 if (starget->single_lun) { 1374 spin_lock_irq(shost->host_lock); 1375 if (starget->starget_sdev_user && 1376 starget->starget_sdev_user != sdev) { 1377 spin_unlock_irq(shost->host_lock); 1378 return 0; 1379 } 1380 starget->starget_sdev_user = sdev; 1381 spin_unlock_irq(shost->host_lock); 1382 } 1383 1384 if (starget->can_queue <= 0) 1385 return 1; 1386 1387 busy = atomic_inc_return(&starget->target_busy) - 1; 1388 if (atomic_read(&starget->target_blocked) > 0) { 1389 if (busy) 1390 goto starved; 1391 1392 /* 1393 * unblock after target_blocked iterates to zero 1394 */ 1395 if (atomic_dec_return(&starget->target_blocked) > 0) 1396 goto out_dec; 1397 1398 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget, 1399 "unblocking target at zero depth\n")); 1400 } 1401 1402 if (busy >= starget->can_queue) 1403 goto starved; 1404 1405 return 1; 1406 1407 starved: 1408 spin_lock_irq(shost->host_lock); 1409 list_move_tail(&sdev->starved_entry, &shost->starved_list); 1410 spin_unlock_irq(shost->host_lock); 1411 out_dec: 1412 if (starget->can_queue > 0) 1413 atomic_dec(&starget->target_busy); 1414 return 0; 1415 } 1416 1417 /* 1418 * scsi_host_queue_ready: if we can send requests to shost, return 1 else 1419 * return 0. We must end up running the queue again whenever 0 is 1420 * returned, else IO can hang. 1421 */ 1422 static inline int scsi_host_queue_ready(struct request_queue *q, 1423 struct Scsi_Host *shost, 1424 struct scsi_device *sdev) 1425 { 1426 unsigned int busy; 1427 1428 if (scsi_host_in_recovery(shost)) 1429 return 0; 1430 1431 busy = atomic_inc_return(&shost->host_busy) - 1; 1432 if (atomic_read(&shost->host_blocked) > 0) { 1433 if (busy) 1434 goto starved; 1435 1436 /* 1437 * unblock after host_blocked iterates to zero 1438 */ 1439 if (atomic_dec_return(&shost->host_blocked) > 0) 1440 goto out_dec; 1441 1442 SCSI_LOG_MLQUEUE(3, 1443 shost_printk(KERN_INFO, shost, 1444 "unblocking host at zero depth\n")); 1445 } 1446 1447 if (shost->can_queue > 0 && busy >= shost->can_queue) 1448 goto starved; 1449 if (shost->host_self_blocked) 1450 goto starved; 1451 1452 /* We're OK to process the command, so we can't be starved */ 1453 if (!list_empty(&sdev->starved_entry)) { 1454 spin_lock_irq(shost->host_lock); 1455 if (!list_empty(&sdev->starved_entry)) 1456 list_del_init(&sdev->starved_entry); 1457 spin_unlock_irq(shost->host_lock); 1458 } 1459 1460 return 1; 1461 1462 starved: 1463 spin_lock_irq(shost->host_lock); 1464 if (list_empty(&sdev->starved_entry)) 1465 list_add_tail(&sdev->starved_entry, &shost->starved_list); 1466 spin_unlock_irq(shost->host_lock); 1467 out_dec: 1468 atomic_dec(&shost->host_busy); 1469 return 0; 1470 } 1471 1472 /* 1473 * Busy state exporting function for request stacking drivers. 1474 * 1475 * For efficiency, no lock is taken to check the busy state of 1476 * shost/starget/sdev, since the returned value is not guaranteed and 1477 * may be changed after request stacking drivers call the function, 1478 * regardless of taking lock or not. 1479 * 1480 * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi 1481 * needs to return 'not busy'. Otherwise, request stacking drivers 1482 * may hold requests forever. 1483 */ 1484 static int scsi_lld_busy(struct request_queue *q) 1485 { 1486 struct scsi_device *sdev = q->queuedata; 1487 struct Scsi_Host *shost; 1488 1489 if (blk_queue_dying(q)) 1490 return 0; 1491 1492 shost = sdev->host; 1493 1494 /* 1495 * Ignore host/starget busy state. 1496 * Since block layer does not have a concept of fairness across 1497 * multiple queues, congestion of host/starget needs to be handled 1498 * in SCSI layer. 1499 */ 1500 if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev)) 1501 return 1; 1502 1503 return 0; 1504 } 1505 1506 /* 1507 * Kill a request for a dead device 1508 */ 1509 static void scsi_kill_request(struct request *req, struct request_queue *q) 1510 { 1511 struct scsi_cmnd *cmd = req->special; 1512 struct scsi_device *sdev; 1513 struct scsi_target *starget; 1514 struct Scsi_Host *shost; 1515 1516 blk_start_request(req); 1517 1518 scmd_printk(KERN_INFO, cmd, "killing request\n"); 1519 1520 sdev = cmd->device; 1521 starget = scsi_target(sdev); 1522 shost = sdev->host; 1523 scsi_init_cmd_errh(cmd); 1524 cmd->result = DID_NO_CONNECT << 16; 1525 atomic_inc(&cmd->device->iorequest_cnt); 1526 1527 /* 1528 * SCSI request completion path will do scsi_device_unbusy(), 1529 * bump busy counts. To bump the counters, we need to dance 1530 * with the locks as normal issue path does. 1531 */ 1532 atomic_inc(&sdev->device_busy); 1533 atomic_inc(&shost->host_busy); 1534 if (starget->can_queue > 0) 1535 atomic_inc(&starget->target_busy); 1536 1537 blk_complete_request(req); 1538 } 1539 1540 static void scsi_softirq_done(struct request *rq) 1541 { 1542 struct scsi_cmnd *cmd = rq->special; 1543 unsigned long wait_for = (cmd->allowed + 1) * rq->timeout; 1544 int disposition; 1545 1546 INIT_LIST_HEAD(&cmd->eh_entry); 1547 1548 atomic_inc(&cmd->device->iodone_cnt); 1549 if (cmd->result) 1550 atomic_inc(&cmd->device->ioerr_cnt); 1551 1552 disposition = scsi_decide_disposition(cmd); 1553 if (disposition != SUCCESS && 1554 time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) { 1555 sdev_printk(KERN_ERR, cmd->device, 1556 "timing out command, waited %lus\n", 1557 wait_for/HZ); 1558 disposition = SUCCESS; 1559 } 1560 1561 scsi_log_completion(cmd, disposition); 1562 1563 switch (disposition) { 1564 case SUCCESS: 1565 scsi_finish_command(cmd); 1566 break; 1567 case NEEDS_RETRY: 1568 scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY); 1569 break; 1570 case ADD_TO_MLQUEUE: 1571 scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY); 1572 break; 1573 default: 1574 if (!scsi_eh_scmd_add(cmd, 0)) 1575 scsi_finish_command(cmd); 1576 } 1577 } 1578 1579 /** 1580 * scsi_dispatch_command - Dispatch a command to the low-level driver. 1581 * @cmd: command block we are dispatching. 1582 * 1583 * Return: nonzero return request was rejected and device's queue needs to be 1584 * plugged. 1585 */ 1586 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd) 1587 { 1588 struct Scsi_Host *host = cmd->device->host; 1589 int rtn = 0; 1590 1591 atomic_inc(&cmd->device->iorequest_cnt); 1592 1593 /* check if the device is still usable */ 1594 if (unlikely(cmd->device->sdev_state == SDEV_DEL)) { 1595 /* in SDEV_DEL we error all commands. DID_NO_CONNECT 1596 * returns an immediate error upwards, and signals 1597 * that the device is no longer present */ 1598 cmd->result = DID_NO_CONNECT << 16; 1599 goto done; 1600 } 1601 1602 /* Check to see if the scsi lld made this device blocked. */ 1603 if (unlikely(scsi_device_blocked(cmd->device))) { 1604 /* 1605 * in blocked state, the command is just put back on 1606 * the device queue. The suspend state has already 1607 * blocked the queue so future requests should not 1608 * occur until the device transitions out of the 1609 * suspend state. 1610 */ 1611 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd, 1612 "queuecommand : device blocked\n")); 1613 return SCSI_MLQUEUE_DEVICE_BUSY; 1614 } 1615 1616 /* Store the LUN value in cmnd, if needed. */ 1617 if (cmd->device->lun_in_cdb) 1618 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) | 1619 (cmd->device->lun << 5 & 0xe0); 1620 1621 scsi_log_send(cmd); 1622 1623 /* 1624 * Before we queue this command, check if the command 1625 * length exceeds what the host adapter can handle. 1626 */ 1627 if (cmd->cmd_len > cmd->device->host->max_cmd_len) { 1628 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd, 1629 "queuecommand : command too long. " 1630 "cdb_size=%d host->max_cmd_len=%d\n", 1631 cmd->cmd_len, cmd->device->host->max_cmd_len)); 1632 cmd->result = (DID_ABORT << 16); 1633 goto done; 1634 } 1635 1636 if (unlikely(host->shost_state == SHOST_DEL)) { 1637 cmd->result = (DID_NO_CONNECT << 16); 1638 goto done; 1639 1640 } 1641 1642 trace_scsi_dispatch_cmd_start(cmd); 1643 rtn = host->hostt->queuecommand(host, cmd); 1644 if (rtn) { 1645 trace_scsi_dispatch_cmd_error(cmd, rtn); 1646 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY && 1647 rtn != SCSI_MLQUEUE_TARGET_BUSY) 1648 rtn = SCSI_MLQUEUE_HOST_BUSY; 1649 1650 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd, 1651 "queuecommand : request rejected\n")); 1652 } 1653 1654 return rtn; 1655 done: 1656 cmd->scsi_done(cmd); 1657 return 0; 1658 } 1659 1660 /** 1661 * scsi_done - Invoke completion on finished SCSI command. 1662 * @cmd: The SCSI Command for which a low-level device driver (LLDD) gives 1663 * ownership back to SCSI Core -- i.e. the LLDD has finished with it. 1664 * 1665 * Description: This function is the mid-level's (SCSI Core) interrupt routine, 1666 * which regains ownership of the SCSI command (de facto) from a LLDD, and 1667 * calls blk_complete_request() for further processing. 1668 * 1669 * This function is interrupt context safe. 1670 */ 1671 static void scsi_done(struct scsi_cmnd *cmd) 1672 { 1673 trace_scsi_dispatch_cmd_done(cmd); 1674 blk_complete_request(cmd->request); 1675 } 1676 1677 /* 1678 * Function: scsi_request_fn() 1679 * 1680 * Purpose: Main strategy routine for SCSI. 1681 * 1682 * Arguments: q - Pointer to actual queue. 1683 * 1684 * Returns: Nothing 1685 * 1686 * Lock status: IO request lock assumed to be held when called. 1687 */ 1688 static void scsi_request_fn(struct request_queue *q) 1689 __releases(q->queue_lock) 1690 __acquires(q->queue_lock) 1691 { 1692 struct scsi_device *sdev = q->queuedata; 1693 struct Scsi_Host *shost; 1694 struct scsi_cmnd *cmd; 1695 struct request *req; 1696 1697 /* 1698 * To start with, we keep looping until the queue is empty, or until 1699 * the host is no longer able to accept any more requests. 1700 */ 1701 shost = sdev->host; 1702 for (;;) { 1703 int rtn; 1704 /* 1705 * get next queueable request. We do this early to make sure 1706 * that the request is fully prepared even if we cannot 1707 * accept it. 1708 */ 1709 req = blk_peek_request(q); 1710 if (!req) 1711 break; 1712 1713 if (unlikely(!scsi_device_online(sdev))) { 1714 sdev_printk(KERN_ERR, sdev, 1715 "rejecting I/O to offline device\n"); 1716 scsi_kill_request(req, q); 1717 continue; 1718 } 1719 1720 if (!scsi_dev_queue_ready(q, sdev)) 1721 break; 1722 1723 /* 1724 * Remove the request from the request list. 1725 */ 1726 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req))) 1727 blk_start_request(req); 1728 1729 spin_unlock_irq(q->queue_lock); 1730 cmd = req->special; 1731 if (unlikely(cmd == NULL)) { 1732 printk(KERN_CRIT "impossible request in %s.\n" 1733 "please mail a stack trace to " 1734 "linux-scsi@vger.kernel.org\n", 1735 __func__); 1736 blk_dump_rq_flags(req, "foo"); 1737 BUG(); 1738 } 1739 1740 /* 1741 * We hit this when the driver is using a host wide 1742 * tag map. For device level tag maps the queue_depth check 1743 * in the device ready fn would prevent us from trying 1744 * to allocate a tag. Since the map is a shared host resource 1745 * we add the dev to the starved list so it eventually gets 1746 * a run when a tag is freed. 1747 */ 1748 if (blk_queue_tagged(q) && !(req->rq_flags & RQF_QUEUED)) { 1749 spin_lock_irq(shost->host_lock); 1750 if (list_empty(&sdev->starved_entry)) 1751 list_add_tail(&sdev->starved_entry, 1752 &shost->starved_list); 1753 spin_unlock_irq(shost->host_lock); 1754 goto not_ready; 1755 } 1756 1757 if (!scsi_target_queue_ready(shost, sdev)) 1758 goto not_ready; 1759 1760 if (!scsi_host_queue_ready(q, shost, sdev)) 1761 goto host_not_ready; 1762 1763 if (sdev->simple_tags) 1764 cmd->flags |= SCMD_TAGGED; 1765 else 1766 cmd->flags &= ~SCMD_TAGGED; 1767 1768 /* 1769 * Finally, initialize any error handling parameters, and set up 1770 * the timers for timeouts. 1771 */ 1772 scsi_init_cmd_errh(cmd); 1773 1774 /* 1775 * Dispatch the command to the low-level driver. 1776 */ 1777 cmd->scsi_done = scsi_done; 1778 rtn = scsi_dispatch_cmd(cmd); 1779 if (rtn) { 1780 scsi_queue_insert(cmd, rtn); 1781 spin_lock_irq(q->queue_lock); 1782 goto out_delay; 1783 } 1784 spin_lock_irq(q->queue_lock); 1785 } 1786 1787 return; 1788 1789 host_not_ready: 1790 if (scsi_target(sdev)->can_queue > 0) 1791 atomic_dec(&scsi_target(sdev)->target_busy); 1792 not_ready: 1793 /* 1794 * lock q, handle tag, requeue req, and decrement device_busy. We 1795 * must return with queue_lock held. 1796 * 1797 * Decrementing device_busy without checking it is OK, as all such 1798 * cases (host limits or settings) should run the queue at some 1799 * later time. 1800 */ 1801 spin_lock_irq(q->queue_lock); 1802 blk_requeue_request(q, req); 1803 atomic_dec(&sdev->device_busy); 1804 out_delay: 1805 if (!atomic_read(&sdev->device_busy) && !scsi_device_blocked(sdev)) 1806 blk_delay_queue(q, SCSI_QUEUE_DELAY); 1807 } 1808 1809 static inline int prep_to_mq(int ret) 1810 { 1811 switch (ret) { 1812 case BLKPREP_OK: 1813 return BLK_MQ_RQ_QUEUE_OK; 1814 case BLKPREP_DEFER: 1815 return BLK_MQ_RQ_QUEUE_BUSY; 1816 default: 1817 return BLK_MQ_RQ_QUEUE_ERROR; 1818 } 1819 } 1820 1821 static int scsi_mq_prep_fn(struct request *req) 1822 { 1823 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req); 1824 struct scsi_device *sdev = req->q->queuedata; 1825 struct Scsi_Host *shost = sdev->host; 1826 unsigned char *sense_buf = cmd->sense_buffer; 1827 struct scatterlist *sg; 1828 1829 memset(cmd, 0, sizeof(struct scsi_cmnd)); 1830 1831 req->special = cmd; 1832 1833 cmd->request = req; 1834 cmd->device = sdev; 1835 cmd->sense_buffer = sense_buf; 1836 1837 cmd->tag = req->tag; 1838 1839 cmd->cmnd = req->cmd; 1840 cmd->prot_op = SCSI_PROT_NORMAL; 1841 1842 INIT_LIST_HEAD(&cmd->list); 1843 INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler); 1844 cmd->jiffies_at_alloc = jiffies; 1845 1846 if (shost->use_cmd_list) { 1847 spin_lock_irq(&sdev->list_lock); 1848 list_add_tail(&cmd->list, &sdev->cmd_list); 1849 spin_unlock_irq(&sdev->list_lock); 1850 } 1851 1852 sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size; 1853 cmd->sdb.table.sgl = sg; 1854 1855 if (scsi_host_get_prot(shost)) { 1856 cmd->prot_sdb = (void *)sg + 1857 min_t(unsigned int, 1858 shost->sg_tablesize, SG_CHUNK_SIZE) * 1859 sizeof(struct scatterlist); 1860 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer)); 1861 1862 cmd->prot_sdb->table.sgl = 1863 (struct scatterlist *)(cmd->prot_sdb + 1); 1864 } 1865 1866 if (blk_bidi_rq(req)) { 1867 struct request *next_rq = req->next_rq; 1868 struct scsi_data_buffer *bidi_sdb = blk_mq_rq_to_pdu(next_rq); 1869 1870 memset(bidi_sdb, 0, sizeof(struct scsi_data_buffer)); 1871 bidi_sdb->table.sgl = 1872 (struct scatterlist *)(bidi_sdb + 1); 1873 1874 next_rq->special = bidi_sdb; 1875 } 1876 1877 blk_mq_start_request(req); 1878 1879 return scsi_setup_cmnd(sdev, req); 1880 } 1881 1882 static void scsi_mq_done(struct scsi_cmnd *cmd) 1883 { 1884 trace_scsi_dispatch_cmd_done(cmd); 1885 blk_mq_complete_request(cmd->request, cmd->request->errors); 1886 } 1887 1888 static int scsi_queue_rq(struct blk_mq_hw_ctx *hctx, 1889 const struct blk_mq_queue_data *bd) 1890 { 1891 struct request *req = bd->rq; 1892 struct request_queue *q = req->q; 1893 struct scsi_device *sdev = q->queuedata; 1894 struct Scsi_Host *shost = sdev->host; 1895 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req); 1896 int ret; 1897 int reason; 1898 1899 ret = prep_to_mq(scsi_prep_state_check(sdev, req)); 1900 if (ret != BLK_MQ_RQ_QUEUE_OK) 1901 goto out; 1902 1903 ret = BLK_MQ_RQ_QUEUE_BUSY; 1904 if (!get_device(&sdev->sdev_gendev)) 1905 goto out; 1906 1907 if (!scsi_dev_queue_ready(q, sdev)) 1908 goto out_put_device; 1909 if (!scsi_target_queue_ready(shost, sdev)) 1910 goto out_dec_device_busy; 1911 if (!scsi_host_queue_ready(q, shost, sdev)) 1912 goto out_dec_target_busy; 1913 1914 1915 if (!(req->rq_flags & RQF_DONTPREP)) { 1916 ret = prep_to_mq(scsi_mq_prep_fn(req)); 1917 if (ret != BLK_MQ_RQ_QUEUE_OK) 1918 goto out_dec_host_busy; 1919 req->rq_flags |= RQF_DONTPREP; 1920 } else { 1921 blk_mq_start_request(req); 1922 } 1923 1924 if (sdev->simple_tags) 1925 cmd->flags |= SCMD_TAGGED; 1926 else 1927 cmd->flags &= ~SCMD_TAGGED; 1928 1929 scsi_init_cmd_errh(cmd); 1930 cmd->scsi_done = scsi_mq_done; 1931 1932 reason = scsi_dispatch_cmd(cmd); 1933 if (reason) { 1934 scsi_set_blocked(cmd, reason); 1935 ret = BLK_MQ_RQ_QUEUE_BUSY; 1936 goto out_dec_host_busy; 1937 } 1938 1939 return BLK_MQ_RQ_QUEUE_OK; 1940 1941 out_dec_host_busy: 1942 atomic_dec(&shost->host_busy); 1943 out_dec_target_busy: 1944 if (scsi_target(sdev)->can_queue > 0) 1945 atomic_dec(&scsi_target(sdev)->target_busy); 1946 out_dec_device_busy: 1947 atomic_dec(&sdev->device_busy); 1948 out_put_device: 1949 put_device(&sdev->sdev_gendev); 1950 out: 1951 switch (ret) { 1952 case BLK_MQ_RQ_QUEUE_BUSY: 1953 if (atomic_read(&sdev->device_busy) == 0 && 1954 !scsi_device_blocked(sdev)) 1955 blk_mq_delay_queue(hctx, SCSI_QUEUE_DELAY); 1956 break; 1957 case BLK_MQ_RQ_QUEUE_ERROR: 1958 /* 1959 * Make sure to release all allocated ressources when 1960 * we hit an error, as we will never see this command 1961 * again. 1962 */ 1963 if (req->rq_flags & RQF_DONTPREP) 1964 scsi_mq_uninit_cmd(cmd); 1965 break; 1966 default: 1967 break; 1968 } 1969 return ret; 1970 } 1971 1972 static enum blk_eh_timer_return scsi_timeout(struct request *req, 1973 bool reserved) 1974 { 1975 if (reserved) 1976 return BLK_EH_RESET_TIMER; 1977 return scsi_times_out(req); 1978 } 1979 1980 static int scsi_init_request(void *data, struct request *rq, 1981 unsigned int hctx_idx, unsigned int request_idx, 1982 unsigned int numa_node) 1983 { 1984 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq); 1985 1986 cmd->sense_buffer = kzalloc_node(SCSI_SENSE_BUFFERSIZE, GFP_KERNEL, 1987 numa_node); 1988 if (!cmd->sense_buffer) 1989 return -ENOMEM; 1990 return 0; 1991 } 1992 1993 static void scsi_exit_request(void *data, struct request *rq, 1994 unsigned int hctx_idx, unsigned int request_idx) 1995 { 1996 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq); 1997 1998 kfree(cmd->sense_buffer); 1999 } 2000 2001 static int scsi_map_queues(struct blk_mq_tag_set *set) 2002 { 2003 struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set); 2004 2005 if (shost->hostt->map_queues) 2006 return shost->hostt->map_queues(shost); 2007 return blk_mq_map_queues(set); 2008 } 2009 2010 static u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost) 2011 { 2012 struct device *host_dev; 2013 u64 bounce_limit = 0xffffffff; 2014 2015 if (shost->unchecked_isa_dma) 2016 return BLK_BOUNCE_ISA; 2017 /* 2018 * Platforms with virtual-DMA translation 2019 * hardware have no practical limit. 2020 */ 2021 if (!PCI_DMA_BUS_IS_PHYS) 2022 return BLK_BOUNCE_ANY; 2023 2024 host_dev = scsi_get_device(shost); 2025 if (host_dev && host_dev->dma_mask) 2026 bounce_limit = (u64)dma_max_pfn(host_dev) << PAGE_SHIFT; 2027 2028 return bounce_limit; 2029 } 2030 2031 static void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q) 2032 { 2033 struct device *dev = shost->dma_dev; 2034 2035 /* 2036 * this limit is imposed by hardware restrictions 2037 */ 2038 blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize, 2039 SG_MAX_SEGMENTS)); 2040 2041 if (scsi_host_prot_dma(shost)) { 2042 shost->sg_prot_tablesize = 2043 min_not_zero(shost->sg_prot_tablesize, 2044 (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS); 2045 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize); 2046 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize); 2047 } 2048 2049 blk_queue_max_hw_sectors(q, shost->max_sectors); 2050 blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost)); 2051 blk_queue_segment_boundary(q, shost->dma_boundary); 2052 dma_set_seg_boundary(dev, shost->dma_boundary); 2053 2054 blk_queue_max_segment_size(q, dma_get_max_seg_size(dev)); 2055 2056 if (!shost->use_clustering) 2057 q->limits.cluster = 0; 2058 2059 /* 2060 * set a reasonable default alignment on word boundaries: the 2061 * host and device may alter it using 2062 * blk_queue_update_dma_alignment() later. 2063 */ 2064 blk_queue_dma_alignment(q, 0x03); 2065 } 2066 2067 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost, 2068 request_fn_proc *request_fn) 2069 { 2070 struct request_queue *q; 2071 2072 q = blk_init_queue(request_fn, NULL); 2073 if (!q) 2074 return NULL; 2075 __scsi_init_queue(shost, q); 2076 return q; 2077 } 2078 EXPORT_SYMBOL(__scsi_alloc_queue); 2079 2080 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev) 2081 { 2082 struct request_queue *q; 2083 2084 q = __scsi_alloc_queue(sdev->host, scsi_request_fn); 2085 if (!q) 2086 return NULL; 2087 2088 blk_queue_prep_rq(q, scsi_prep_fn); 2089 blk_queue_unprep_rq(q, scsi_unprep_fn); 2090 blk_queue_softirq_done(q, scsi_softirq_done); 2091 blk_queue_rq_timed_out(q, scsi_times_out); 2092 blk_queue_lld_busy(q, scsi_lld_busy); 2093 return q; 2094 } 2095 2096 static struct blk_mq_ops scsi_mq_ops = { 2097 .queue_rq = scsi_queue_rq, 2098 .complete = scsi_softirq_done, 2099 .timeout = scsi_timeout, 2100 .init_request = scsi_init_request, 2101 .exit_request = scsi_exit_request, 2102 .map_queues = scsi_map_queues, 2103 }; 2104 2105 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev) 2106 { 2107 sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set); 2108 if (IS_ERR(sdev->request_queue)) 2109 return NULL; 2110 2111 sdev->request_queue->queuedata = sdev; 2112 __scsi_init_queue(sdev->host, sdev->request_queue); 2113 return sdev->request_queue; 2114 } 2115 2116 int scsi_mq_setup_tags(struct Scsi_Host *shost) 2117 { 2118 unsigned int cmd_size, sgl_size, tbl_size; 2119 2120 tbl_size = shost->sg_tablesize; 2121 if (tbl_size > SG_CHUNK_SIZE) 2122 tbl_size = SG_CHUNK_SIZE; 2123 sgl_size = tbl_size * sizeof(struct scatterlist); 2124 cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size; 2125 if (scsi_host_get_prot(shost)) 2126 cmd_size += sizeof(struct scsi_data_buffer) + sgl_size; 2127 2128 memset(&shost->tag_set, 0, sizeof(shost->tag_set)); 2129 shost->tag_set.ops = &scsi_mq_ops; 2130 shost->tag_set.nr_hw_queues = shost->nr_hw_queues ? : 1; 2131 shost->tag_set.queue_depth = shost->can_queue; 2132 shost->tag_set.cmd_size = cmd_size; 2133 shost->tag_set.numa_node = NUMA_NO_NODE; 2134 shost->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE; 2135 shost->tag_set.flags |= 2136 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy); 2137 shost->tag_set.driver_data = shost; 2138 2139 return blk_mq_alloc_tag_set(&shost->tag_set); 2140 } 2141 2142 void scsi_mq_destroy_tags(struct Scsi_Host *shost) 2143 { 2144 blk_mq_free_tag_set(&shost->tag_set); 2145 } 2146 2147 /* 2148 * Function: scsi_block_requests() 2149 * 2150 * Purpose: Utility function used by low-level drivers to prevent further 2151 * commands from being queued to the device. 2152 * 2153 * Arguments: shost - Host in question 2154 * 2155 * Returns: Nothing 2156 * 2157 * Lock status: No locks are assumed held. 2158 * 2159 * Notes: There is no timer nor any other means by which the requests 2160 * get unblocked other than the low-level driver calling 2161 * scsi_unblock_requests(). 2162 */ 2163 void scsi_block_requests(struct Scsi_Host *shost) 2164 { 2165 shost->host_self_blocked = 1; 2166 } 2167 EXPORT_SYMBOL(scsi_block_requests); 2168 2169 /* 2170 * Function: scsi_unblock_requests() 2171 * 2172 * Purpose: Utility function used by low-level drivers to allow further 2173 * commands from being queued to the device. 2174 * 2175 * Arguments: shost - Host in question 2176 * 2177 * Returns: Nothing 2178 * 2179 * Lock status: No locks are assumed held. 2180 * 2181 * Notes: There is no timer nor any other means by which the requests 2182 * get unblocked other than the low-level driver calling 2183 * scsi_unblock_requests(). 2184 * 2185 * This is done as an API function so that changes to the 2186 * internals of the scsi mid-layer won't require wholesale 2187 * changes to drivers that use this feature. 2188 */ 2189 void scsi_unblock_requests(struct Scsi_Host *shost) 2190 { 2191 shost->host_self_blocked = 0; 2192 scsi_run_host_queues(shost); 2193 } 2194 EXPORT_SYMBOL(scsi_unblock_requests); 2195 2196 int __init scsi_init_queue(void) 2197 { 2198 scsi_sdb_cache = kmem_cache_create("scsi_data_buffer", 2199 sizeof(struct scsi_data_buffer), 2200 0, 0, NULL); 2201 if (!scsi_sdb_cache) { 2202 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n"); 2203 return -ENOMEM; 2204 } 2205 2206 return 0; 2207 } 2208 2209 void scsi_exit_queue(void) 2210 { 2211 kmem_cache_destroy(scsi_sdb_cache); 2212 } 2213 2214 /** 2215 * scsi_mode_select - issue a mode select 2216 * @sdev: SCSI device to be queried 2217 * @pf: Page format bit (1 == standard, 0 == vendor specific) 2218 * @sp: Save page bit (0 == don't save, 1 == save) 2219 * @modepage: mode page being requested 2220 * @buffer: request buffer (may not be smaller than eight bytes) 2221 * @len: length of request buffer. 2222 * @timeout: command timeout 2223 * @retries: number of retries before failing 2224 * @data: returns a structure abstracting the mode header data 2225 * @sshdr: place to put sense data (or NULL if no sense to be collected). 2226 * must be SCSI_SENSE_BUFFERSIZE big. 2227 * 2228 * Returns zero if successful; negative error number or scsi 2229 * status on error 2230 * 2231 */ 2232 int 2233 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage, 2234 unsigned char *buffer, int len, int timeout, int retries, 2235 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr) 2236 { 2237 unsigned char cmd[10]; 2238 unsigned char *real_buffer; 2239 int ret; 2240 2241 memset(cmd, 0, sizeof(cmd)); 2242 cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0); 2243 2244 if (sdev->use_10_for_ms) { 2245 if (len > 65535) 2246 return -EINVAL; 2247 real_buffer = kmalloc(8 + len, GFP_KERNEL); 2248 if (!real_buffer) 2249 return -ENOMEM; 2250 memcpy(real_buffer + 8, buffer, len); 2251 len += 8; 2252 real_buffer[0] = 0; 2253 real_buffer[1] = 0; 2254 real_buffer[2] = data->medium_type; 2255 real_buffer[3] = data->device_specific; 2256 real_buffer[4] = data->longlba ? 0x01 : 0; 2257 real_buffer[5] = 0; 2258 real_buffer[6] = data->block_descriptor_length >> 8; 2259 real_buffer[7] = data->block_descriptor_length; 2260 2261 cmd[0] = MODE_SELECT_10; 2262 cmd[7] = len >> 8; 2263 cmd[8] = len; 2264 } else { 2265 if (len > 255 || data->block_descriptor_length > 255 || 2266 data->longlba) 2267 return -EINVAL; 2268 2269 real_buffer = kmalloc(4 + len, GFP_KERNEL); 2270 if (!real_buffer) 2271 return -ENOMEM; 2272 memcpy(real_buffer + 4, buffer, len); 2273 len += 4; 2274 real_buffer[0] = 0; 2275 real_buffer[1] = data->medium_type; 2276 real_buffer[2] = data->device_specific; 2277 real_buffer[3] = data->block_descriptor_length; 2278 2279 2280 cmd[0] = MODE_SELECT; 2281 cmd[4] = len; 2282 } 2283 2284 ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len, 2285 sshdr, timeout, retries, NULL); 2286 kfree(real_buffer); 2287 return ret; 2288 } 2289 EXPORT_SYMBOL_GPL(scsi_mode_select); 2290 2291 /** 2292 * scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary. 2293 * @sdev: SCSI device to be queried 2294 * @dbd: set if mode sense will allow block descriptors to be returned 2295 * @modepage: mode page being requested 2296 * @buffer: request buffer (may not be smaller than eight bytes) 2297 * @len: length of request buffer. 2298 * @timeout: command timeout 2299 * @retries: number of retries before failing 2300 * @data: returns a structure abstracting the mode header data 2301 * @sshdr: place to put sense data (or NULL if no sense to be collected). 2302 * must be SCSI_SENSE_BUFFERSIZE big. 2303 * 2304 * Returns zero if unsuccessful, or the header offset (either 4 2305 * or 8 depending on whether a six or ten byte command was 2306 * issued) if successful. 2307 */ 2308 int 2309 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage, 2310 unsigned char *buffer, int len, int timeout, int retries, 2311 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr) 2312 { 2313 unsigned char cmd[12]; 2314 int use_10_for_ms; 2315 int header_length; 2316 int result, retry_count = retries; 2317 struct scsi_sense_hdr my_sshdr; 2318 2319 memset(data, 0, sizeof(*data)); 2320 memset(&cmd[0], 0, 12); 2321 cmd[1] = dbd & 0x18; /* allows DBD and LLBA bits */ 2322 cmd[2] = modepage; 2323 2324 /* caller might not be interested in sense, but we need it */ 2325 if (!sshdr) 2326 sshdr = &my_sshdr; 2327 2328 retry: 2329 use_10_for_ms = sdev->use_10_for_ms; 2330 2331 if (use_10_for_ms) { 2332 if (len < 8) 2333 len = 8; 2334 2335 cmd[0] = MODE_SENSE_10; 2336 cmd[8] = len; 2337 header_length = 8; 2338 } else { 2339 if (len < 4) 2340 len = 4; 2341 2342 cmd[0] = MODE_SENSE; 2343 cmd[4] = len; 2344 header_length = 4; 2345 } 2346 2347 memset(buffer, 0, len); 2348 2349 result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len, 2350 sshdr, timeout, retries, NULL); 2351 2352 /* This code looks awful: what it's doing is making sure an 2353 * ILLEGAL REQUEST sense return identifies the actual command 2354 * byte as the problem. MODE_SENSE commands can return 2355 * ILLEGAL REQUEST if the code page isn't supported */ 2356 2357 if (use_10_for_ms && !scsi_status_is_good(result) && 2358 (driver_byte(result) & DRIVER_SENSE)) { 2359 if (scsi_sense_valid(sshdr)) { 2360 if ((sshdr->sense_key == ILLEGAL_REQUEST) && 2361 (sshdr->asc == 0x20) && (sshdr->ascq == 0)) { 2362 /* 2363 * Invalid command operation code 2364 */ 2365 sdev->use_10_for_ms = 0; 2366 goto retry; 2367 } 2368 } 2369 } 2370 2371 if(scsi_status_is_good(result)) { 2372 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b && 2373 (modepage == 6 || modepage == 8))) { 2374 /* Initio breakage? */ 2375 header_length = 0; 2376 data->length = 13; 2377 data->medium_type = 0; 2378 data->device_specific = 0; 2379 data->longlba = 0; 2380 data->block_descriptor_length = 0; 2381 } else if(use_10_for_ms) { 2382 data->length = buffer[0]*256 + buffer[1] + 2; 2383 data->medium_type = buffer[2]; 2384 data->device_specific = buffer[3]; 2385 data->longlba = buffer[4] & 0x01; 2386 data->block_descriptor_length = buffer[6]*256 2387 + buffer[7]; 2388 } else { 2389 data->length = buffer[0] + 1; 2390 data->medium_type = buffer[1]; 2391 data->device_specific = buffer[2]; 2392 data->block_descriptor_length = buffer[3]; 2393 } 2394 data->header_length = header_length; 2395 } else if ((status_byte(result) == CHECK_CONDITION) && 2396 scsi_sense_valid(sshdr) && 2397 sshdr->sense_key == UNIT_ATTENTION && retry_count) { 2398 retry_count--; 2399 goto retry; 2400 } 2401 2402 return result; 2403 } 2404 EXPORT_SYMBOL(scsi_mode_sense); 2405 2406 /** 2407 * scsi_test_unit_ready - test if unit is ready 2408 * @sdev: scsi device to change the state of. 2409 * @timeout: command timeout 2410 * @retries: number of retries before failing 2411 * @sshdr_external: Optional pointer to struct scsi_sense_hdr for 2412 * returning sense. Make sure that this is cleared before passing 2413 * in. 2414 * 2415 * Returns zero if unsuccessful or an error if TUR failed. For 2416 * removable media, UNIT_ATTENTION sets ->changed flag. 2417 **/ 2418 int 2419 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries, 2420 struct scsi_sense_hdr *sshdr_external) 2421 { 2422 char cmd[] = { 2423 TEST_UNIT_READY, 0, 0, 0, 0, 0, 2424 }; 2425 struct scsi_sense_hdr *sshdr; 2426 int result; 2427 2428 if (!sshdr_external) 2429 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL); 2430 else 2431 sshdr = sshdr_external; 2432 2433 /* try to eat the UNIT_ATTENTION if there are enough retries */ 2434 do { 2435 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr, 2436 timeout, retries, NULL); 2437 if (sdev->removable && scsi_sense_valid(sshdr) && 2438 sshdr->sense_key == UNIT_ATTENTION) 2439 sdev->changed = 1; 2440 } while (scsi_sense_valid(sshdr) && 2441 sshdr->sense_key == UNIT_ATTENTION && --retries); 2442 2443 if (!sshdr_external) 2444 kfree(sshdr); 2445 return result; 2446 } 2447 EXPORT_SYMBOL(scsi_test_unit_ready); 2448 2449 /** 2450 * scsi_device_set_state - Take the given device through the device state model. 2451 * @sdev: scsi device to change the state of. 2452 * @state: state to change to. 2453 * 2454 * Returns zero if unsuccessful or an error if the requested 2455 * transition is illegal. 2456 */ 2457 int 2458 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state) 2459 { 2460 enum scsi_device_state oldstate = sdev->sdev_state; 2461 2462 if (state == oldstate) 2463 return 0; 2464 2465 switch (state) { 2466 case SDEV_CREATED: 2467 switch (oldstate) { 2468 case SDEV_CREATED_BLOCK: 2469 break; 2470 default: 2471 goto illegal; 2472 } 2473 break; 2474 2475 case SDEV_RUNNING: 2476 switch (oldstate) { 2477 case SDEV_CREATED: 2478 case SDEV_OFFLINE: 2479 case SDEV_TRANSPORT_OFFLINE: 2480 case SDEV_QUIESCE: 2481 case SDEV_BLOCK: 2482 break; 2483 default: 2484 goto illegal; 2485 } 2486 break; 2487 2488 case SDEV_QUIESCE: 2489 switch (oldstate) { 2490 case SDEV_RUNNING: 2491 case SDEV_OFFLINE: 2492 case SDEV_TRANSPORT_OFFLINE: 2493 break; 2494 default: 2495 goto illegal; 2496 } 2497 break; 2498 2499 case SDEV_OFFLINE: 2500 case SDEV_TRANSPORT_OFFLINE: 2501 switch (oldstate) { 2502 case SDEV_CREATED: 2503 case SDEV_RUNNING: 2504 case SDEV_QUIESCE: 2505 case SDEV_BLOCK: 2506 break; 2507 default: 2508 goto illegal; 2509 } 2510 break; 2511 2512 case SDEV_BLOCK: 2513 switch (oldstate) { 2514 case SDEV_RUNNING: 2515 case SDEV_CREATED_BLOCK: 2516 break; 2517 default: 2518 goto illegal; 2519 } 2520 break; 2521 2522 case SDEV_CREATED_BLOCK: 2523 switch (oldstate) { 2524 case SDEV_CREATED: 2525 break; 2526 default: 2527 goto illegal; 2528 } 2529 break; 2530 2531 case SDEV_CANCEL: 2532 switch (oldstate) { 2533 case SDEV_CREATED: 2534 case SDEV_RUNNING: 2535 case SDEV_QUIESCE: 2536 case SDEV_OFFLINE: 2537 case SDEV_TRANSPORT_OFFLINE: 2538 case SDEV_BLOCK: 2539 break; 2540 default: 2541 goto illegal; 2542 } 2543 break; 2544 2545 case SDEV_DEL: 2546 switch (oldstate) { 2547 case SDEV_CREATED: 2548 case SDEV_RUNNING: 2549 case SDEV_OFFLINE: 2550 case SDEV_TRANSPORT_OFFLINE: 2551 case SDEV_CANCEL: 2552 case SDEV_CREATED_BLOCK: 2553 break; 2554 default: 2555 goto illegal; 2556 } 2557 break; 2558 2559 } 2560 sdev->sdev_state = state; 2561 return 0; 2562 2563 illegal: 2564 SCSI_LOG_ERROR_RECOVERY(1, 2565 sdev_printk(KERN_ERR, sdev, 2566 "Illegal state transition %s->%s", 2567 scsi_device_state_name(oldstate), 2568 scsi_device_state_name(state)) 2569 ); 2570 return -EINVAL; 2571 } 2572 EXPORT_SYMBOL(scsi_device_set_state); 2573 2574 /** 2575 * sdev_evt_emit - emit a single SCSI device uevent 2576 * @sdev: associated SCSI device 2577 * @evt: event to emit 2578 * 2579 * Send a single uevent (scsi_event) to the associated scsi_device. 2580 */ 2581 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt) 2582 { 2583 int idx = 0; 2584 char *envp[3]; 2585 2586 switch (evt->evt_type) { 2587 case SDEV_EVT_MEDIA_CHANGE: 2588 envp[idx++] = "SDEV_MEDIA_CHANGE=1"; 2589 break; 2590 case SDEV_EVT_INQUIRY_CHANGE_REPORTED: 2591 scsi_rescan_device(&sdev->sdev_gendev); 2592 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED"; 2593 break; 2594 case SDEV_EVT_CAPACITY_CHANGE_REPORTED: 2595 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED"; 2596 break; 2597 case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED: 2598 envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED"; 2599 break; 2600 case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED: 2601 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED"; 2602 break; 2603 case SDEV_EVT_LUN_CHANGE_REPORTED: 2604 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED"; 2605 break; 2606 case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED: 2607 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED"; 2608 break; 2609 default: 2610 /* do nothing */ 2611 break; 2612 } 2613 2614 envp[idx++] = NULL; 2615 2616 kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp); 2617 } 2618 2619 /** 2620 * sdev_evt_thread - send a uevent for each scsi event 2621 * @work: work struct for scsi_device 2622 * 2623 * Dispatch queued events to their associated scsi_device kobjects 2624 * as uevents. 2625 */ 2626 void scsi_evt_thread(struct work_struct *work) 2627 { 2628 struct scsi_device *sdev; 2629 enum scsi_device_event evt_type; 2630 LIST_HEAD(event_list); 2631 2632 sdev = container_of(work, struct scsi_device, event_work); 2633 2634 for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++) 2635 if (test_and_clear_bit(evt_type, sdev->pending_events)) 2636 sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL); 2637 2638 while (1) { 2639 struct scsi_event *evt; 2640 struct list_head *this, *tmp; 2641 unsigned long flags; 2642 2643 spin_lock_irqsave(&sdev->list_lock, flags); 2644 list_splice_init(&sdev->event_list, &event_list); 2645 spin_unlock_irqrestore(&sdev->list_lock, flags); 2646 2647 if (list_empty(&event_list)) 2648 break; 2649 2650 list_for_each_safe(this, tmp, &event_list) { 2651 evt = list_entry(this, struct scsi_event, node); 2652 list_del(&evt->node); 2653 scsi_evt_emit(sdev, evt); 2654 kfree(evt); 2655 } 2656 } 2657 } 2658 2659 /** 2660 * sdev_evt_send - send asserted event to uevent thread 2661 * @sdev: scsi_device event occurred on 2662 * @evt: event to send 2663 * 2664 * Assert scsi device event asynchronously. 2665 */ 2666 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt) 2667 { 2668 unsigned long flags; 2669 2670 #if 0 2671 /* FIXME: currently this check eliminates all media change events 2672 * for polled devices. Need to update to discriminate between AN 2673 * and polled events */ 2674 if (!test_bit(evt->evt_type, sdev->supported_events)) { 2675 kfree(evt); 2676 return; 2677 } 2678 #endif 2679 2680 spin_lock_irqsave(&sdev->list_lock, flags); 2681 list_add_tail(&evt->node, &sdev->event_list); 2682 schedule_work(&sdev->event_work); 2683 spin_unlock_irqrestore(&sdev->list_lock, flags); 2684 } 2685 EXPORT_SYMBOL_GPL(sdev_evt_send); 2686 2687 /** 2688 * sdev_evt_alloc - allocate a new scsi event 2689 * @evt_type: type of event to allocate 2690 * @gfpflags: GFP flags for allocation 2691 * 2692 * Allocates and returns a new scsi_event. 2693 */ 2694 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type, 2695 gfp_t gfpflags) 2696 { 2697 struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags); 2698 if (!evt) 2699 return NULL; 2700 2701 evt->evt_type = evt_type; 2702 INIT_LIST_HEAD(&evt->node); 2703 2704 /* evt_type-specific initialization, if any */ 2705 switch (evt_type) { 2706 case SDEV_EVT_MEDIA_CHANGE: 2707 case SDEV_EVT_INQUIRY_CHANGE_REPORTED: 2708 case SDEV_EVT_CAPACITY_CHANGE_REPORTED: 2709 case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED: 2710 case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED: 2711 case SDEV_EVT_LUN_CHANGE_REPORTED: 2712 case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED: 2713 default: 2714 /* do nothing */ 2715 break; 2716 } 2717 2718 return evt; 2719 } 2720 EXPORT_SYMBOL_GPL(sdev_evt_alloc); 2721 2722 /** 2723 * sdev_evt_send_simple - send asserted event to uevent thread 2724 * @sdev: scsi_device event occurred on 2725 * @evt_type: type of event to send 2726 * @gfpflags: GFP flags for allocation 2727 * 2728 * Assert scsi device event asynchronously, given an event type. 2729 */ 2730 void sdev_evt_send_simple(struct scsi_device *sdev, 2731 enum scsi_device_event evt_type, gfp_t gfpflags) 2732 { 2733 struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags); 2734 if (!evt) { 2735 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n", 2736 evt_type); 2737 return; 2738 } 2739 2740 sdev_evt_send(sdev, evt); 2741 } 2742 EXPORT_SYMBOL_GPL(sdev_evt_send_simple); 2743 2744 /** 2745 * scsi_request_fn_active() - number of kernel threads inside scsi_request_fn() 2746 * @sdev: SCSI device to count the number of scsi_request_fn() callers for. 2747 */ 2748 static int scsi_request_fn_active(struct scsi_device *sdev) 2749 { 2750 struct request_queue *q = sdev->request_queue; 2751 int request_fn_active; 2752 2753 WARN_ON_ONCE(sdev->host->use_blk_mq); 2754 2755 spin_lock_irq(q->queue_lock); 2756 request_fn_active = q->request_fn_active; 2757 spin_unlock_irq(q->queue_lock); 2758 2759 return request_fn_active; 2760 } 2761 2762 /** 2763 * scsi_wait_for_queuecommand() - wait for ongoing queuecommand() calls 2764 * @sdev: SCSI device pointer. 2765 * 2766 * Wait until the ongoing shost->hostt->queuecommand() calls that are 2767 * invoked from scsi_request_fn() have finished. 2768 */ 2769 static void scsi_wait_for_queuecommand(struct scsi_device *sdev) 2770 { 2771 WARN_ON_ONCE(sdev->host->use_blk_mq); 2772 2773 while (scsi_request_fn_active(sdev)) 2774 msleep(20); 2775 } 2776 2777 /** 2778 * scsi_device_quiesce - Block user issued commands. 2779 * @sdev: scsi device to quiesce. 2780 * 2781 * This works by trying to transition to the SDEV_QUIESCE state 2782 * (which must be a legal transition). When the device is in this 2783 * state, only special requests will be accepted, all others will 2784 * be deferred. Since special requests may also be requeued requests, 2785 * a successful return doesn't guarantee the device will be 2786 * totally quiescent. 2787 * 2788 * Must be called with user context, may sleep. 2789 * 2790 * Returns zero if unsuccessful or an error if not. 2791 */ 2792 int 2793 scsi_device_quiesce(struct scsi_device *sdev) 2794 { 2795 int err = scsi_device_set_state(sdev, SDEV_QUIESCE); 2796 if (err) 2797 return err; 2798 2799 scsi_run_queue(sdev->request_queue); 2800 while (atomic_read(&sdev->device_busy)) { 2801 msleep_interruptible(200); 2802 scsi_run_queue(sdev->request_queue); 2803 } 2804 return 0; 2805 } 2806 EXPORT_SYMBOL(scsi_device_quiesce); 2807 2808 /** 2809 * scsi_device_resume - Restart user issued commands to a quiesced device. 2810 * @sdev: scsi device to resume. 2811 * 2812 * Moves the device from quiesced back to running and restarts the 2813 * queues. 2814 * 2815 * Must be called with user context, may sleep. 2816 */ 2817 void scsi_device_resume(struct scsi_device *sdev) 2818 { 2819 /* check if the device state was mutated prior to resume, and if 2820 * so assume the state is being managed elsewhere (for example 2821 * device deleted during suspend) 2822 */ 2823 if (sdev->sdev_state != SDEV_QUIESCE || 2824 scsi_device_set_state(sdev, SDEV_RUNNING)) 2825 return; 2826 scsi_run_queue(sdev->request_queue); 2827 } 2828 EXPORT_SYMBOL(scsi_device_resume); 2829 2830 static void 2831 device_quiesce_fn(struct scsi_device *sdev, void *data) 2832 { 2833 scsi_device_quiesce(sdev); 2834 } 2835 2836 void 2837 scsi_target_quiesce(struct scsi_target *starget) 2838 { 2839 starget_for_each_device(starget, NULL, device_quiesce_fn); 2840 } 2841 EXPORT_SYMBOL(scsi_target_quiesce); 2842 2843 static void 2844 device_resume_fn(struct scsi_device *sdev, void *data) 2845 { 2846 scsi_device_resume(sdev); 2847 } 2848 2849 void 2850 scsi_target_resume(struct scsi_target *starget) 2851 { 2852 starget_for_each_device(starget, NULL, device_resume_fn); 2853 } 2854 EXPORT_SYMBOL(scsi_target_resume); 2855 2856 /** 2857 * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state 2858 * @sdev: device to block 2859 * 2860 * Block request made by scsi lld's to temporarily stop all 2861 * scsi commands on the specified device. May sleep. 2862 * 2863 * Returns zero if successful or error if not 2864 * 2865 * Notes: 2866 * This routine transitions the device to the SDEV_BLOCK state 2867 * (which must be a legal transition). When the device is in this 2868 * state, all commands are deferred until the scsi lld reenables 2869 * the device with scsi_device_unblock or device_block_tmo fires. 2870 * 2871 * To do: avoid that scsi_send_eh_cmnd() calls queuecommand() after 2872 * scsi_internal_device_block() has blocked a SCSI device and also 2873 * remove the rport mutex lock and unlock calls from srp_queuecommand(). 2874 */ 2875 int 2876 scsi_internal_device_block(struct scsi_device *sdev) 2877 { 2878 struct request_queue *q = sdev->request_queue; 2879 unsigned long flags; 2880 int err = 0; 2881 2882 err = scsi_device_set_state(sdev, SDEV_BLOCK); 2883 if (err) { 2884 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK); 2885 2886 if (err) 2887 return err; 2888 } 2889 2890 /* 2891 * The device has transitioned to SDEV_BLOCK. Stop the 2892 * block layer from calling the midlayer with this device's 2893 * request queue. 2894 */ 2895 if (q->mq_ops) { 2896 blk_mq_stop_hw_queues(q); 2897 } else { 2898 spin_lock_irqsave(q->queue_lock, flags); 2899 blk_stop_queue(q); 2900 spin_unlock_irqrestore(q->queue_lock, flags); 2901 scsi_wait_for_queuecommand(sdev); 2902 } 2903 2904 return 0; 2905 } 2906 EXPORT_SYMBOL_GPL(scsi_internal_device_block); 2907 2908 /** 2909 * scsi_internal_device_unblock - resume a device after a block request 2910 * @sdev: device to resume 2911 * @new_state: state to set devices to after unblocking 2912 * 2913 * Called by scsi lld's or the midlayer to restart the device queue 2914 * for the previously suspended scsi device. Called from interrupt or 2915 * normal process context. 2916 * 2917 * Returns zero if successful or error if not. 2918 * 2919 * Notes: 2920 * This routine transitions the device to the SDEV_RUNNING state 2921 * or to one of the offline states (which must be a legal transition) 2922 * allowing the midlayer to goose the queue for this device. 2923 */ 2924 int 2925 scsi_internal_device_unblock(struct scsi_device *sdev, 2926 enum scsi_device_state new_state) 2927 { 2928 struct request_queue *q = sdev->request_queue; 2929 unsigned long flags; 2930 2931 /* 2932 * Try to transition the scsi device to SDEV_RUNNING or one of the 2933 * offlined states and goose the device queue if successful. 2934 */ 2935 if ((sdev->sdev_state == SDEV_BLOCK) || 2936 (sdev->sdev_state == SDEV_TRANSPORT_OFFLINE)) 2937 sdev->sdev_state = new_state; 2938 else if (sdev->sdev_state == SDEV_CREATED_BLOCK) { 2939 if (new_state == SDEV_TRANSPORT_OFFLINE || 2940 new_state == SDEV_OFFLINE) 2941 sdev->sdev_state = new_state; 2942 else 2943 sdev->sdev_state = SDEV_CREATED; 2944 } else if (sdev->sdev_state != SDEV_CANCEL && 2945 sdev->sdev_state != SDEV_OFFLINE) 2946 return -EINVAL; 2947 2948 if (q->mq_ops) { 2949 blk_mq_start_stopped_hw_queues(q, false); 2950 } else { 2951 spin_lock_irqsave(q->queue_lock, flags); 2952 blk_start_queue(q); 2953 spin_unlock_irqrestore(q->queue_lock, flags); 2954 } 2955 2956 return 0; 2957 } 2958 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock); 2959 2960 static void 2961 device_block(struct scsi_device *sdev, void *data) 2962 { 2963 scsi_internal_device_block(sdev); 2964 } 2965 2966 static int 2967 target_block(struct device *dev, void *data) 2968 { 2969 if (scsi_is_target_device(dev)) 2970 starget_for_each_device(to_scsi_target(dev), NULL, 2971 device_block); 2972 return 0; 2973 } 2974 2975 void 2976 scsi_target_block(struct device *dev) 2977 { 2978 if (scsi_is_target_device(dev)) 2979 starget_for_each_device(to_scsi_target(dev), NULL, 2980 device_block); 2981 else 2982 device_for_each_child(dev, NULL, target_block); 2983 } 2984 EXPORT_SYMBOL_GPL(scsi_target_block); 2985 2986 static void 2987 device_unblock(struct scsi_device *sdev, void *data) 2988 { 2989 scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data); 2990 } 2991 2992 static int 2993 target_unblock(struct device *dev, void *data) 2994 { 2995 if (scsi_is_target_device(dev)) 2996 starget_for_each_device(to_scsi_target(dev), data, 2997 device_unblock); 2998 return 0; 2999 } 3000 3001 void 3002 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state) 3003 { 3004 if (scsi_is_target_device(dev)) 3005 starget_for_each_device(to_scsi_target(dev), &new_state, 3006 device_unblock); 3007 else 3008 device_for_each_child(dev, &new_state, target_unblock); 3009 } 3010 EXPORT_SYMBOL_GPL(scsi_target_unblock); 3011 3012 /** 3013 * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt 3014 * @sgl: scatter-gather list 3015 * @sg_count: number of segments in sg 3016 * @offset: offset in bytes into sg, on return offset into the mapped area 3017 * @len: bytes to map, on return number of bytes mapped 3018 * 3019 * Returns virtual address of the start of the mapped page 3020 */ 3021 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count, 3022 size_t *offset, size_t *len) 3023 { 3024 int i; 3025 size_t sg_len = 0, len_complete = 0; 3026 struct scatterlist *sg; 3027 struct page *page; 3028 3029 WARN_ON(!irqs_disabled()); 3030 3031 for_each_sg(sgl, sg, sg_count, i) { 3032 len_complete = sg_len; /* Complete sg-entries */ 3033 sg_len += sg->length; 3034 if (sg_len > *offset) 3035 break; 3036 } 3037 3038 if (unlikely(i == sg_count)) { 3039 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, " 3040 "elements %d\n", 3041 __func__, sg_len, *offset, sg_count); 3042 WARN_ON(1); 3043 return NULL; 3044 } 3045 3046 /* Offset starting from the beginning of first page in this sg-entry */ 3047 *offset = *offset - len_complete + sg->offset; 3048 3049 /* Assumption: contiguous pages can be accessed as "page + i" */ 3050 page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT)); 3051 *offset &= ~PAGE_MASK; 3052 3053 /* Bytes in this sg-entry from *offset to the end of the page */ 3054 sg_len = PAGE_SIZE - *offset; 3055 if (*len > sg_len) 3056 *len = sg_len; 3057 3058 return kmap_atomic(page); 3059 } 3060 EXPORT_SYMBOL(scsi_kmap_atomic_sg); 3061 3062 /** 3063 * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg 3064 * @virt: virtual address to be unmapped 3065 */ 3066 void scsi_kunmap_atomic_sg(void *virt) 3067 { 3068 kunmap_atomic(virt); 3069 } 3070 EXPORT_SYMBOL(scsi_kunmap_atomic_sg); 3071 3072 void sdev_disable_disk_events(struct scsi_device *sdev) 3073 { 3074 atomic_inc(&sdev->disk_events_disable_depth); 3075 } 3076 EXPORT_SYMBOL(sdev_disable_disk_events); 3077 3078 void sdev_enable_disk_events(struct scsi_device *sdev) 3079 { 3080 if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0)) 3081 return; 3082 atomic_dec(&sdev->disk_events_disable_depth); 3083 } 3084 EXPORT_SYMBOL(sdev_enable_disk_events); 3085 3086 /** 3087 * scsi_vpd_lun_id - return a unique device identification 3088 * @sdev: SCSI device 3089 * @id: buffer for the identification 3090 * @id_len: length of the buffer 3091 * 3092 * Copies a unique device identification into @id based 3093 * on the information in the VPD page 0x83 of the device. 3094 * The string will be formatted as a SCSI name string. 3095 * 3096 * Returns the length of the identification or error on failure. 3097 * If the identifier is longer than the supplied buffer the actual 3098 * identifier length is returned and the buffer is not zero-padded. 3099 */ 3100 int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len) 3101 { 3102 u8 cur_id_type = 0xff; 3103 u8 cur_id_size = 0; 3104 unsigned char *d, *cur_id_str; 3105 unsigned char __rcu *vpd_pg83; 3106 int id_size = -EINVAL; 3107 3108 rcu_read_lock(); 3109 vpd_pg83 = rcu_dereference(sdev->vpd_pg83); 3110 if (!vpd_pg83) { 3111 rcu_read_unlock(); 3112 return -ENXIO; 3113 } 3114 3115 /* 3116 * Look for the correct descriptor. 3117 * Order of preference for lun descriptor: 3118 * - SCSI name string 3119 * - NAA IEEE Registered Extended 3120 * - EUI-64 based 16-byte 3121 * - EUI-64 based 12-byte 3122 * - NAA IEEE Registered 3123 * - NAA IEEE Extended 3124 * - T10 Vendor ID 3125 * as longer descriptors reduce the likelyhood 3126 * of identification clashes. 3127 */ 3128 3129 /* The id string must be at least 20 bytes + terminating NULL byte */ 3130 if (id_len < 21) { 3131 rcu_read_unlock(); 3132 return -EINVAL; 3133 } 3134 3135 memset(id, 0, id_len); 3136 d = vpd_pg83 + 4; 3137 while (d < vpd_pg83 + sdev->vpd_pg83_len) { 3138 /* Skip designators not referring to the LUN */ 3139 if ((d[1] & 0x30) != 0x00) 3140 goto next_desig; 3141 3142 switch (d[1] & 0xf) { 3143 case 0x1: 3144 /* T10 Vendor ID */ 3145 if (cur_id_size > d[3]) 3146 break; 3147 /* Prefer anything */ 3148 if (cur_id_type > 0x01 && cur_id_type != 0xff) 3149 break; 3150 cur_id_size = d[3]; 3151 if (cur_id_size + 4 > id_len) 3152 cur_id_size = id_len - 4; 3153 cur_id_str = d + 4; 3154 cur_id_type = d[1] & 0xf; 3155 id_size = snprintf(id, id_len, "t10.%*pE", 3156 cur_id_size, cur_id_str); 3157 break; 3158 case 0x2: 3159 /* EUI-64 */ 3160 if (cur_id_size > d[3]) 3161 break; 3162 /* Prefer NAA IEEE Registered Extended */ 3163 if (cur_id_type == 0x3 && 3164 cur_id_size == d[3]) 3165 break; 3166 cur_id_size = d[3]; 3167 cur_id_str = d + 4; 3168 cur_id_type = d[1] & 0xf; 3169 switch (cur_id_size) { 3170 case 8: 3171 id_size = snprintf(id, id_len, 3172 "eui.%8phN", 3173 cur_id_str); 3174 break; 3175 case 12: 3176 id_size = snprintf(id, id_len, 3177 "eui.%12phN", 3178 cur_id_str); 3179 break; 3180 case 16: 3181 id_size = snprintf(id, id_len, 3182 "eui.%16phN", 3183 cur_id_str); 3184 break; 3185 default: 3186 cur_id_size = 0; 3187 break; 3188 } 3189 break; 3190 case 0x3: 3191 /* NAA */ 3192 if (cur_id_size > d[3]) 3193 break; 3194 cur_id_size = d[3]; 3195 cur_id_str = d + 4; 3196 cur_id_type = d[1] & 0xf; 3197 switch (cur_id_size) { 3198 case 8: 3199 id_size = snprintf(id, id_len, 3200 "naa.%8phN", 3201 cur_id_str); 3202 break; 3203 case 16: 3204 id_size = snprintf(id, id_len, 3205 "naa.%16phN", 3206 cur_id_str); 3207 break; 3208 default: 3209 cur_id_size = 0; 3210 break; 3211 } 3212 break; 3213 case 0x8: 3214 /* SCSI name string */ 3215 if (cur_id_size + 4 > d[3]) 3216 break; 3217 /* Prefer others for truncated descriptor */ 3218 if (cur_id_size && d[3] > id_len) 3219 break; 3220 cur_id_size = id_size = d[3]; 3221 cur_id_str = d + 4; 3222 cur_id_type = d[1] & 0xf; 3223 if (cur_id_size >= id_len) 3224 cur_id_size = id_len - 1; 3225 memcpy(id, cur_id_str, cur_id_size); 3226 /* Decrease priority for truncated descriptor */ 3227 if (cur_id_size != id_size) 3228 cur_id_size = 6; 3229 break; 3230 default: 3231 break; 3232 } 3233 next_desig: 3234 d += d[3] + 4; 3235 } 3236 rcu_read_unlock(); 3237 3238 return id_size; 3239 } 3240 EXPORT_SYMBOL(scsi_vpd_lun_id); 3241 3242 /* 3243 * scsi_vpd_tpg_id - return a target port group identifier 3244 * @sdev: SCSI device 3245 * 3246 * Returns the Target Port Group identifier from the information 3247 * froom VPD page 0x83 of the device. 3248 * 3249 * Returns the identifier or error on failure. 3250 */ 3251 int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id) 3252 { 3253 unsigned char *d; 3254 unsigned char __rcu *vpd_pg83; 3255 int group_id = -EAGAIN, rel_port = -1; 3256 3257 rcu_read_lock(); 3258 vpd_pg83 = rcu_dereference(sdev->vpd_pg83); 3259 if (!vpd_pg83) { 3260 rcu_read_unlock(); 3261 return -ENXIO; 3262 } 3263 3264 d = sdev->vpd_pg83 + 4; 3265 while (d < sdev->vpd_pg83 + sdev->vpd_pg83_len) { 3266 switch (d[1] & 0xf) { 3267 case 0x4: 3268 /* Relative target port */ 3269 rel_port = get_unaligned_be16(&d[6]); 3270 break; 3271 case 0x5: 3272 /* Target port group */ 3273 group_id = get_unaligned_be16(&d[6]); 3274 break; 3275 default: 3276 break; 3277 } 3278 d += d[3] + 4; 3279 } 3280 rcu_read_unlock(); 3281 3282 if (group_id >= 0 && rel_id && rel_port != -1) 3283 *rel_id = rel_port; 3284 3285 return group_id; 3286 } 3287 EXPORT_SYMBOL(scsi_vpd_tpg_id); 3288