File Index
Files and directories
Sign in
FILE

scsi_lib.c

File preview with syntax highlighting.

linux-6.19.14/drivers/scsi/scsi_lib.c C
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) 1999 Eric Youngdale
   4 * Copyright (C) 2014 Christoph Hellwig
   5 *
   6 *  SCSI queueing library.
   7 *      Initial versions: Eric Youngdale (eric@andante.org).
   8 *                        Based upon conversations with large numbers
   9 *                        of people at Linux Expo.
  10 */
  11
  12#include <linux/bio.h>
  13#include <linux/bitops.h>
  14#include <linux/blkdev.h>
  15#include <linux/completion.h>
  16#include <linux/kernel.h>
  17#include <linux/export.h>
  18#include <linux/init.h>
  19#include <linux/pci.h>
  20#include <linux/delay.h>
  21#include <linux/hardirq.h>
  22#include <linux/scatterlist.h>
  23#include <linux/blk-mq.h>
  24#include <linux/blk-integrity.h>
  25#include <linux/ratelimit.h>
  26#include <linux/unaligned.h>
  27
  28#include <scsi/scsi.h>
  29#include <scsi/scsi_cmnd.h>
  30#include <scsi/scsi_dbg.h>
  31#include <scsi/scsi_device.h>
  32#include <scsi/scsi_driver.h>
  33#include <scsi/scsi_eh.h>
  34#include <scsi/scsi_host.h>
  35#include <scsi/scsi_transport.h> /* scsi_init_limits() */
  36#include <scsi/scsi_dh.h>
  37
  38#include <trace/events/scsi.h>
  39
  40#include "scsi_debugfs.h"
  41#include "scsi_priv.h"
  42#include "scsi_logging.h"
  43
  44/*
  45 * Size of integrity metadata is usually small, 1 inline sg should
  46 * cover normal cases.
  47 */
  48#ifdef CONFIG_ARCH_NO_SG_CHAIN
  49#define  SCSI_INLINE_PROT_SG_CNT  0
  50#define  SCSI_INLINE_SG_CNT  0
  51#else
  52#define  SCSI_INLINE_PROT_SG_CNT  1
  53#define  SCSI_INLINE_SG_CNT  2
  54#endif
  55
  56static struct kmem_cache *scsi_sense_cache;
  57static DEFINE_MUTEX(scsi_sense_cache_mutex);
  58
  59static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd);
  60
  61int scsi_init_sense_cache(struct Scsi_Host *shost)
  62{
  63	int ret = 0;
  64
  65	mutex_lock(&scsi_sense_cache_mutex);
  66	if (!scsi_sense_cache) {
  67		scsi_sense_cache =
  68			kmem_cache_create_usercopy("scsi_sense_cache",
  69				SCSI_SENSE_BUFFERSIZE, 0, SLAB_HWCACHE_ALIGN,
  70				0, SCSI_SENSE_BUFFERSIZE, NULL);
  71		if (!scsi_sense_cache)
  72			ret = -ENOMEM;
  73	}
  74	mutex_unlock(&scsi_sense_cache_mutex);
  75	return ret;
  76}
  77
  78static void
  79scsi_set_blocked(struct scsi_cmnd *cmd, int reason)
  80{
  81	struct Scsi_Host *host = cmd->device->host;
  82	struct scsi_device *device = cmd->device;
  83	struct scsi_target *starget = scsi_target(device);
  84
  85	/*
  86	 * Set the appropriate busy bit for the device/host.
  87	 *
  88	 * If the host/device isn't busy, assume that something actually
  89	 * completed, and that we should be able to queue a command now.
  90	 *
  91	 * Note that the prior mid-layer assumption that any host could
  92	 * always queue at least one command is now broken.  The mid-layer
  93	 * will implement a user specifiable stall (see
  94	 * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
  95	 * if a command is requeued with no other commands outstanding
  96	 * either for the device or for the host.
  97	 */
  98	switch (reason) {
  99	case SCSI_MLQUEUE_HOST_BUSY:
 100		atomic_set(&host->host_blocked, host->max_host_blocked);
 101		break;
 102	case SCSI_MLQUEUE_DEVICE_BUSY:
 103	case SCSI_MLQUEUE_EH_RETRY:
 104		atomic_set(&device->device_blocked,
 105			   device->max_device_blocked);
 106		break;
 107	case SCSI_MLQUEUE_TARGET_BUSY:
 108		atomic_set(&starget->target_blocked,
 109			   starget->max_target_blocked);
 110		break;
 111	}
 112}
 113
 114static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd, unsigned long msecs)
 115{
 116	struct request *rq = scsi_cmd_to_rq(cmd);
 117
 118	if (rq->rq_flags & RQF_DONTPREP) {
 119		rq->rq_flags &= ~RQF_DONTPREP;
 120		scsi_mq_uninit_cmd(cmd);
 121	} else {
 122		WARN_ON_ONCE(true);
 123	}
 124
 125	blk_mq_requeue_request(rq, false);
 126	if (!scsi_host_in_recovery(cmd->device->host))
 127		blk_mq_delay_kick_requeue_list(rq->q, msecs);
 128}
 129
 130/**
 131 * __scsi_queue_insert - private queue insertion
 132 * @cmd: The SCSI command being requeued
 133 * @reason:  The reason for the requeue
 134 * @unbusy: Whether the queue should be unbusied
 135 *
 136 * This is a private queue insertion.  The public interface
 137 * scsi_queue_insert() always assumes the queue should be unbusied
 138 * because it's always called before the completion.  This function is
 139 * for a requeue after completion, which should only occur in this
 140 * file.
 141 */
 142static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, bool unbusy)
 143{
 144	struct scsi_device *device = cmd->device;
 145
 146	SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
 147		"Inserting command %p into mlqueue\n", cmd));
 148
 149	scsi_set_blocked(cmd, reason);
 150
 151	/*
 152	 * Decrement the counters, since these commands are no longer
 153	 * active on the host/device.
 154	 */
 155	if (unbusy)
 156		scsi_device_unbusy(device, cmd);
 157
 158	/*
 159	 * Requeue this command.  It will go before all other commands
 160	 * that are already in the queue. Schedule requeue work under
 161	 * lock such that the kblockd_schedule_work() call happens
 162	 * before blk_mq_destroy_queue() finishes.
 163	 */
 164	cmd->result = 0;
 165
 166	blk_mq_requeue_request(scsi_cmd_to_rq(cmd),
 167			       !scsi_host_in_recovery(cmd->device->host));
 168}
 169
 170/**
 171 * scsi_queue_insert - Reinsert a command in the queue.
 172 * @cmd:    command that we are adding to queue.
 173 * @reason: why we are inserting command to queue.
 174 *
 175 * We do this for one of two cases. Either the host is busy and it cannot accept
 176 * any more commands for the time being, or the device returned QUEUE_FULL and
 177 * can accept no more commands.
 178 *
 179 * Context: This could be called either from an interrupt context or a normal
 180 * process context.
 181 */
 182void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
 183{
 184	__scsi_queue_insert(cmd, reason, true);
 185}
 186
 187/**
 188 * scsi_failures_reset_retries - reset all failures to zero
 189 * @failures: &struct scsi_failures with specific failure modes set
 190 */
 191void scsi_failures_reset_retries(struct scsi_failures *failures)
 192{
 193	struct scsi_failure *failure;
 194
 195	failures->total_retries = 0;
 196
 197	for (failure = failures->failure_definitions; failure->result;
 198	     failure++)
 199		failure->retries = 0;
 200}
 201EXPORT_SYMBOL_GPL(scsi_failures_reset_retries);
 202
 203/**
 204 * scsi_check_passthrough - Determine if passthrough scsi_cmnd needs a retry.
 205 * @scmd: scsi_cmnd to check.
 206 * @failures: scsi_failures struct that lists failures to check for.
 207 *
 208 * Returns -EAGAIN if the caller should retry else 0.
 209 */
 210static int scsi_check_passthrough(struct scsi_cmnd *scmd,
 211				  struct scsi_failures *failures)
 212{
 213	struct scsi_failure *failure;
 214	struct scsi_sense_hdr sshdr;
 215	enum sam_status status;
 216
 217	if (!scmd->result)
 218		return 0;
 219
 220	if (!failures)
 221		return 0;
 222
 223	for (failure = failures->failure_definitions; failure->result;
 224	     failure++) {
 225		if (failure->result == SCMD_FAILURE_RESULT_ANY)
 226			goto maybe_retry;
 227
 228		if (host_byte(scmd->result) &&
 229		    host_byte(scmd->result) == host_byte(failure->result))
 230			goto maybe_retry;
 231
 232		status = status_byte(scmd->result);
 233		if (!status)
 234			continue;
 235
 236		if (failure->result == SCMD_FAILURE_STAT_ANY &&
 237		    !scsi_status_is_good(scmd->result))
 238			goto maybe_retry;
 239
 240		if (status != status_byte(failure->result))
 241			continue;
 242
 243		if (status_byte(failure->result) != SAM_STAT_CHECK_CONDITION ||
 244		    failure->sense == SCMD_FAILURE_SENSE_ANY)
 245			goto maybe_retry;
 246
 247		if (!scsi_command_normalize_sense(scmd, &sshdr))
 248			return 0;
 249
 250		if (failure->sense != sshdr.sense_key)
 251			continue;
 252
 253		if (failure->asc == SCMD_FAILURE_ASC_ANY)
 254			goto maybe_retry;
 255
 256		if (failure->asc != sshdr.asc)
 257			continue;
 258
 259		if (failure->ascq == SCMD_FAILURE_ASCQ_ANY ||
 260		    failure->ascq == sshdr.ascq)
 261			goto maybe_retry;
 262	}
 263
 264	return 0;
 265
 266maybe_retry:
 267	if (failure->allowed) {
 268		if (failure->allowed == SCMD_FAILURE_NO_LIMIT ||
 269		    ++failure->retries <= failure->allowed)
 270			return -EAGAIN;
 271	} else {
 272		if (failures->total_allowed == SCMD_FAILURE_NO_LIMIT ||
 273		    ++failures->total_retries <= failures->total_allowed)
 274			return -EAGAIN;
 275	}
 276
 277	return 0;
 278}
 279
 280/**
 281 * scsi_execute_cmd - insert request and wait for the result
 282 * @sdev:	scsi_device
 283 * @cmd:	scsi command
 284 * @opf:	block layer request cmd_flags
 285 * @buffer:	data buffer
 286 * @bufflen:	len of buffer
 287 * @timeout:	request timeout in HZ
 288 * @ml_retries:	number of times SCSI midlayer will retry request
 289 * @args:	Optional args. See struct definition for field descriptions
 290 *
 291 * Returns the scsi_cmnd result field if a command was executed, or a negative
 292 * Linux error code if we didn't get that far.
 293 */
 294int scsi_execute_cmd(struct scsi_device *sdev, const unsigned char *cmd,
 295		     blk_opf_t opf, void *buffer, unsigned int bufflen,
 296		     int timeout, int ml_retries,
 297		     const struct scsi_exec_args *args)
 298{
 299	static const struct scsi_exec_args default_args;
 300	struct request *req;
 301	struct scsi_cmnd *scmd;
 302	int ret;
 303
 304	if (!args)
 305		args = &default_args;
 306	else if (WARN_ON_ONCE(args->sense &&
 307			      args->sense_len != SCSI_SENSE_BUFFERSIZE))
 308		return -EINVAL;
 309
 310retry:
 311	req = scsi_alloc_request(sdev->request_queue, opf, args->req_flags);
 312	if (IS_ERR(req))
 313		return PTR_ERR(req);
 314
 315	if (bufflen) {
 316		ret = blk_rq_map_kern(req, buffer, bufflen, GFP_NOIO);
 317		if (ret)
 318			goto out;
 319	}
 320	scmd = blk_mq_rq_to_pdu(req);
 321	scmd->cmd_len = COMMAND_SIZE(cmd[0]);
 322	memcpy(scmd->cmnd, cmd, scmd->cmd_len);
 323	scmd->allowed = ml_retries;
 324	scmd->flags |= args->scmd_flags;
 325	req->timeout = timeout;
 326	req->rq_flags |= RQF_QUIET;
 327
 328	/*
 329	 * head injection *required* here otherwise quiesce won't work
 330	 */
 331	blk_execute_rq(req, true);
 332
 333	if (scsi_check_passthrough(scmd, args->failures) == -EAGAIN) {
 334		blk_mq_free_request(req);
 335		goto retry;
 336	}
 337
 338	/*
 339	 * Some devices (USB mass-storage in particular) may transfer
 340	 * garbage data together with a residue indicating that the data
 341	 * is invalid.  Prevent the garbage from being misinterpreted
 342	 * and prevent security leaks by zeroing out the excess data.
 343	 */
 344	if (unlikely(scmd->resid_len > 0 && scmd->resid_len <= bufflen))
 345		memset(buffer + bufflen - scmd->resid_len, 0, scmd->resid_len);
 346
 347	if (args->resid)
 348		*args->resid = scmd->resid_len;
 349	if (args->sense)
 350		memcpy(args->sense, scmd->sense_buffer, SCSI_SENSE_BUFFERSIZE);
 351	if (args->sshdr)
 352		scsi_normalize_sense(scmd->sense_buffer, scmd->sense_len,
 353				     args->sshdr);
 354
 355	ret = scmd->result;
 356 out:
 357	blk_mq_free_request(req);
 358
 359	return ret;
 360}
 361EXPORT_SYMBOL(scsi_execute_cmd);
 362
 363/*
 364 * Wake up the error handler if necessary. Avoid as follows that the error
 365 * handler is not woken up if host in-flight requests number ==
 366 * shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
 367 * with an RCU read lock in this function to ensure that this function in
 368 * its entirety either finishes before scsi_eh_scmd_add() increases the
 369 * host_failed counter or that it notices the shost state change made by
 370 * scsi_eh_scmd_add().
 371 */
 372static void scsi_dec_host_busy(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
 373{
 374	unsigned long flags;
 375
 376	rcu_read_lock();
 377	__clear_bit(SCMD_STATE_INFLIGHT, &cmd->state);
 378	if (unlikely(scsi_host_in_recovery(shost))) {
 379		/*
 380		 * Ensure the clear of SCMD_STATE_INFLIGHT is visible to
 381		 * other CPUs before counting busy requests. Otherwise,
 382		 * reordering can cause CPUs to race and miss an eh wakeup
 383		 * when no CPU sees all busy requests as done or timed out.
 384		 */
 385		smp_mb();
 386
 387		unsigned int busy = scsi_host_busy(shost);
 388
 389		spin_lock_irqsave(shost->host_lock, flags);
 390		if (shost->host_failed || shost->host_eh_scheduled)
 391			scsi_eh_wakeup(shost, busy);
 392		spin_unlock_irqrestore(shost->host_lock, flags);
 393	}
 394	rcu_read_unlock();
 395}
 396
 397void scsi_device_unbusy(struct scsi_device *sdev, struct scsi_cmnd *cmd)
 398{
 399	struct Scsi_Host *shost = sdev->host;
 400	struct scsi_target *starget = scsi_target(sdev);
 401
 402	scsi_dec_host_busy(shost, cmd);
 403
 404	if (starget->can_queue > 0)
 405		atomic_dec(&starget->target_busy);
 406
 407	if (sdev->budget_map.map)
 408		sbitmap_put(&sdev->budget_map, cmd->budget_token);
 409	cmd->budget_token = -1;
 410}
 411
 412/*
 413 * Kick the queue of SCSI device @sdev if @sdev != current_sdev. Called with
 414 * interrupts disabled.
 415 */
 416static void scsi_kick_sdev_queue(struct scsi_device *sdev, void *data)
 417{
 418	struct scsi_device *current_sdev = data;
 419
 420	if (sdev != current_sdev)
 421		blk_mq_run_hw_queues(sdev->request_queue, true);
 422}
 423
 424/*
 425 * Called for single_lun devices on IO completion. Clear starget_sdev_user,
 426 * and call blk_run_queue for all the scsi_devices on the target -
 427 * including current_sdev first.
 428 *
 429 * Called with *no* scsi locks held.
 430 */
 431static void scsi_single_lun_run(struct scsi_device *current_sdev)
 432{
 433	struct Scsi_Host *shost = current_sdev->host;
 434	struct scsi_target *starget = scsi_target(current_sdev);
 435	unsigned long flags;
 436
 437	spin_lock_irqsave(shost->host_lock, flags);
 438	starget->starget_sdev_user = NULL;
 439	spin_unlock_irqrestore(shost->host_lock, flags);
 440
 441	/*
 442	 * Call blk_run_queue for all LUNs on the target, starting with
 443	 * current_sdev. We race with others (to set starget_sdev_user),
 444	 * but in most cases, we will be first. Ideally, each LU on the
 445	 * target would get some limited time or requests on the target.
 446	 */
 447	blk_mq_run_hw_queues(current_sdev->request_queue,
 448			     shost->queuecommand_may_block);
 449
 450	spin_lock_irqsave(shost->host_lock, flags);
 451	if (!starget->starget_sdev_user)
 452		__starget_for_each_device(starget, current_sdev,
 453					  scsi_kick_sdev_queue);
 454	spin_unlock_irqrestore(shost->host_lock, flags);
 455}
 456
 457static inline bool scsi_device_is_busy(struct scsi_device *sdev)
 458{
 459	if (scsi_device_busy(sdev) >= sdev->queue_depth)
 460		return true;
 461	if (atomic_read(&sdev->device_blocked) > 0)
 462		return true;
 463	return false;
 464}
 465
 466static inline bool scsi_target_is_busy(struct scsi_target *starget)
 467{
 468	if (starget->can_queue > 0) {
 469		if (atomic_read(&starget->target_busy) >= starget->can_queue)
 470			return true;
 471		if (atomic_read(&starget->target_blocked) > 0)
 472			return true;
 473	}
 474	return false;
 475}
 476
 477static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
 478{
 479	if (atomic_read(&shost->host_blocked) > 0)
 480		return true;
 481	if (shost->host_self_blocked)
 482		return true;
 483	return false;
 484}
 485
 486static void scsi_starved_list_run(struct Scsi_Host *shost)
 487{
 488	LIST_HEAD(starved_list);
 489	struct scsi_device *sdev;
 490	unsigned long flags;
 491
 492	spin_lock_irqsave(shost->host_lock, flags);
 493	list_splice_init(&shost->starved_list, &starved_list);
 494
 495	while (!list_empty(&starved_list)) {
 496		struct request_queue *slq;
 497
 498		/*
 499		 * As long as shost is accepting commands and we have
 500		 * starved queues, call blk_run_queue. scsi_request_fn
 501		 * drops the queue_lock and can add us back to the
 502		 * starved_list.
 503		 *
 504		 * host_lock protects the starved_list and starved_entry.
 505		 * scsi_request_fn must get the host_lock before checking
 506		 * or modifying starved_list or starved_entry.
 507		 */
 508		if (scsi_host_is_busy(shost))
 509			break;
 510
 511		sdev = list_entry(starved_list.next,
 512				  struct scsi_device, starved_entry);
 513		list_del_init(&sdev->starved_entry);
 514		if (scsi_target_is_busy(scsi_target(sdev))) {
 515			list_move_tail(&sdev->starved_entry,
 516				       &shost->starved_list);
 517			continue;
 518		}
 519
 520		/*
 521		 * Once we drop the host lock, a racing scsi_remove_device()
 522		 * call may remove the sdev from the starved list and destroy
 523		 * it and the queue.  Mitigate by taking a reference to the
 524		 * queue and never touching the sdev again after we drop the
 525		 * host lock.  Note: if __scsi_remove_device() invokes
 526		 * blk_mq_destroy_queue() before the queue is run from this
 527		 * function then blk_run_queue() will return immediately since
 528		 * blk_mq_destroy_queue() marks the queue with QUEUE_FLAG_DYING.
 529		 */
 530		slq = sdev->request_queue;
 531		if (!blk_get_queue(slq))
 532			continue;
 533		spin_unlock_irqrestore(shost->host_lock, flags);
 534
 535		blk_mq_run_hw_queues(slq, false);
 536		blk_put_queue(slq);
 537
 538		spin_lock_irqsave(shost->host_lock, flags);
 539	}
 540	/* put any unprocessed entries back */
 541	list_splice(&starved_list, &shost->starved_list);
 542	spin_unlock_irqrestore(shost->host_lock, flags);
 543}
 544
 545/**
 546 * scsi_run_queue - Select a proper request queue to serve next.
 547 * @q:  last request's queue
 548 *
 549 * The previous command was completely finished, start a new one if possible.
 550 */
 551static void scsi_run_queue(struct request_queue *q)
 552{
 553	struct scsi_device *sdev = q->queuedata;
 554
 555	if (scsi_target(sdev)->single_lun)
 556		scsi_single_lun_run(sdev);
 557	if (!list_empty(&sdev->host->starved_list))
 558		scsi_starved_list_run(sdev->host);
 559
 560	/* Note: blk_mq_kick_requeue_list() runs the queue asynchronously. */
 561	blk_mq_kick_requeue_list(q);
 562}
 563
 564void scsi_requeue_run_queue(struct work_struct *work)
 565{
 566	struct scsi_device *sdev;
 567	struct request_queue *q;
 568
 569	sdev = container_of(work, struct scsi_device, requeue_work);
 570	q = sdev->request_queue;
 571	scsi_run_queue(q);
 572}
 573
 574void scsi_run_host_queues(struct Scsi_Host *shost)
 575{
 576	struct scsi_device *sdev;
 577
 578	shost_for_each_device(sdev, shost)
 579		scsi_run_queue(sdev->request_queue);
 580}
 581
 582static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
 583{
 584	if (!blk_rq_is_passthrough(scsi_cmd_to_rq(cmd))) {
 585		struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
 586
 587		if (drv->uninit_command)
 588			drv->uninit_command(cmd);
 589	}
 590}
 591
 592void scsi_free_sgtables(struct scsi_cmnd *cmd)
 593{
 594	if (cmd->sdb.table.nents)
 595		sg_free_table_chained(&cmd->sdb.table,
 596				SCSI_INLINE_SG_CNT);
 597	if (scsi_prot_sg_count(cmd))
 598		sg_free_table_chained(&cmd->prot_sdb->table,
 599				SCSI_INLINE_PROT_SG_CNT);
 600}
 601EXPORT_SYMBOL_GPL(scsi_free_sgtables);
 602
 603static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
 604{
 605	scsi_free_sgtables(cmd);
 606	scsi_uninit_cmd(cmd);
 607}
 608
 609static void scsi_run_queue_async(struct scsi_device *sdev)
 610{
 611	if (scsi_host_in_recovery(sdev->host))
 612		return;
 613
 614	if (scsi_target(sdev)->single_lun ||
 615	    !list_empty(&sdev->host->starved_list)) {
 616		kblockd_schedule_work(&sdev->requeue_work);
 617	} else {
 618		/*
 619		 * smp_mb() present in sbitmap_queue_clear() or implied in
 620		 * .end_io is for ordering writing .device_busy in
 621		 * scsi_device_unbusy() and reading sdev->restarts.
 622		 */
 623		int old = atomic_read(&sdev->restarts);
 624
 625		/*
 626		 * ->restarts has to be kept as non-zero if new budget
 627		 *  contention occurs.
 628		 *
 629		 *  No need to run queue when either another re-run
 630		 *  queue wins in updating ->restarts or a new budget
 631		 *  contention occurs.
 632		 */
 633		if (old && atomic_cmpxchg(&sdev->restarts, old, 0) == old)
 634			blk_mq_run_hw_queues(sdev->request_queue, true);
 635	}
 636}
 637
 638/* Returns false when no more bytes to process, true if there are more */
 639static bool scsi_end_request(struct request *req, blk_status_t error,
 640		unsigned int bytes)
 641{
 642	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
 643	struct scsi_device *sdev = cmd->device;
 644	struct request_queue *q = sdev->request_queue;
 645
 646	if (blk_update_request(req, error, bytes))
 647		return true;
 648
 649	if (q->limits.features & BLK_FEAT_ADD_RANDOM)
 650		add_disk_randomness(req->q->disk);
 651
 652	WARN_ON_ONCE(!blk_rq_is_passthrough(req) &&
 653		     !(cmd->flags & SCMD_INITIALIZED));
 654	cmd->flags = 0;
 655
 656	/*
 657	 * Calling rcu_barrier() is not necessary here because the
 658	 * SCSI error handler guarantees that the function called by
 659	 * call_rcu() has been called before scsi_end_request() is
 660	 * called.
 661	 */
 662	destroy_rcu_head(&cmd->rcu);
 663
 664	/*
 665	 * In the MQ case the command gets freed by __blk_mq_end_request,
 666	 * so we have to do all cleanup that depends on it earlier.
 667	 *
 668	 * We also can't kick the queues from irq context, so we
 669	 * will have to defer it to a workqueue.
 670	 */
 671	scsi_mq_uninit_cmd(cmd);
 672
 673	/*
 674	 * queue is still alive, so grab the ref for preventing it
 675	 * from being cleaned up during running queue.
 676	 */
 677	percpu_ref_get(&q->q_usage_counter);
 678
 679	__blk_mq_end_request(req, error);
 680
 681	scsi_run_queue_async(sdev);
 682
 683	percpu_ref_put(&q->q_usage_counter);
 684	return false;
 685}
 686
 687/**
 688 * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
 689 * @result:	scsi error code
 690 *
 691 * Translate a SCSI result code into a blk_status_t value.
 692 */
 693static blk_status_t scsi_result_to_blk_status(int result)
 694{
 695	/*
 696	 * Check the scsi-ml byte first in case we converted a host or status
 697	 * byte.
 698	 */
 699	switch (scsi_ml_byte(result)) {
 700	case SCSIML_STAT_OK:
 701		break;
 702	case SCSIML_STAT_RESV_CONFLICT:
 703		return BLK_STS_RESV_CONFLICT;
 704	case SCSIML_STAT_NOSPC:
 705		return BLK_STS_NOSPC;
 706	case SCSIML_STAT_MED_ERROR:
 707		return BLK_STS_MEDIUM;
 708	case SCSIML_STAT_TGT_FAILURE:
 709		return BLK_STS_TARGET;
 710	case SCSIML_STAT_DL_TIMEOUT:
 711		return BLK_STS_DURATION_LIMIT;
 712	}
 713
 714	switch (host_byte(result)) {
 715	case DID_OK:
 716		if (scsi_status_is_good(result))
 717			return BLK_STS_OK;
 718		return BLK_STS_IOERR;
 719	case DID_TRANSPORT_FAILFAST:
 720	case DID_TRANSPORT_MARGINAL:
 721		return BLK_STS_TRANSPORT;
 722	default:
 723		return BLK_STS_IOERR;
 724	}
 725}
 726
 727/**
 728 * scsi_rq_err_bytes - determine number of bytes till the next failure boundary
 729 * @rq: request to examine
 730 *
 731 * Description:
 732 *     A request could be merge of IOs which require different failure
 733 *     handling.  This function determines the number of bytes which
 734 *     can be failed from the beginning of the request without
 735 *     crossing into area which need to be retried further.
 736 *
 737 * Return:
 738 *     The number of bytes to fail.
 739 */
 740static unsigned int scsi_rq_err_bytes(const struct request *rq)
 741{
 742	blk_opf_t ff = rq->cmd_flags & REQ_FAILFAST_MASK;
 743	unsigned int bytes = 0;
 744	struct bio *bio;
 745
 746	if (!(rq->rq_flags & RQF_MIXED_MERGE))
 747		return blk_rq_bytes(rq);
 748
 749	/*
 750	 * Currently the only 'mixing' which can happen is between
 751	 * different fastfail types.  We can safely fail portions
 752	 * which have all the failfast bits that the first one has -
 753	 * the ones which are at least as eager to fail as the first
 754	 * one.
 755	 */
 756	for (bio = rq->bio; bio; bio = bio->bi_next) {
 757		if ((bio->bi_opf & ff) != ff)
 758			break;
 759		bytes += bio->bi_iter.bi_size;
 760	}
 761
 762	/* this could lead to infinite loop */
 763	BUG_ON(blk_rq_bytes(rq) && !bytes);
 764	return bytes;
 765}
 766
 767static bool scsi_cmd_runtime_exceeced(struct scsi_cmnd *cmd)
 768{
 769	struct request *req = scsi_cmd_to_rq(cmd);
 770	unsigned long wait_for;
 771
 772	if (cmd->allowed == SCSI_CMD_RETRIES_NO_LIMIT)
 773		return false;
 774
 775	wait_for = (cmd->allowed + 1) * req->timeout;
 776	if (time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
 777		scmd_printk(KERN_ERR, cmd, "timing out command, waited %lus\n",
 778			    wait_for/HZ);
 779		return true;
 780	}
 781	return false;
 782}
 783
 784/*
 785 * When ALUA transition state is returned, reprep the cmd to
 786 * use the ALUA handler's transition timeout. Delay the reprep
 787 * 1 sec to avoid aggressive retries of the target in that
 788 * state.
 789 */
 790#define ALUA_TRANSITION_REPREP_DELAY	1000
 791
 792/* Helper for scsi_io_completion() when special action required. */
 793static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
 794{
 795	struct request *req = scsi_cmd_to_rq(cmd);
 796	int level = 0;
 797	enum {ACTION_FAIL, ACTION_REPREP, ACTION_DELAYED_REPREP,
 798	      ACTION_RETRY, ACTION_DELAYED_RETRY} action;
 799	struct scsi_sense_hdr sshdr;
 800	bool sense_valid;
 801	bool sense_current = true;      /* false implies "deferred sense" */
 802	blk_status_t blk_stat;
 803
 804	sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
 805	if (sense_valid)
 806		sense_current = !scsi_sense_is_deferred(&sshdr);
 807
 808	blk_stat = scsi_result_to_blk_status(result);
 809
 810	if (host_byte(result) == DID_RESET) {
 811		/* Third party bus reset or reset for error recovery
 812		 * reasons.  Just retry the command and see what
 813		 * happens.
 814		 */
 815		action = ACTION_RETRY;
 816	} else if (sense_valid && sense_current) {
 817		switch (sshdr.sense_key) {
 818		case UNIT_ATTENTION:
 819			if (cmd->device->removable) {
 820				/* Detected disc change.  Set a bit
 821				 * and quietly refuse further access.
 822				 */
 823				cmd->device->changed = 1;
 824				action = ACTION_FAIL;
 825			} else {
 826				/* Must have been a power glitch, or a
 827				 * bus reset.  Could not have been a
 828				 * media change, so we just retry the
 829				 * command and see what happens.
 830				 */
 831				action = ACTION_RETRY;
 832			}
 833			break;
 834		case ILLEGAL_REQUEST:
 835			/* If we had an ILLEGAL REQUEST returned, then
 836			 * we may have performed an unsupported
 837			 * command.  The only thing this should be
 838			 * would be a ten byte read where only a six
 839			 * byte read was supported.  Also, on a system
 840			 * where READ CAPACITY failed, we may have
 841			 * read past the end of the disk.
 842			 */
 843			if ((cmd->device->use_10_for_rw &&
 844			    sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
 845			    (cmd->cmnd[0] == READ_10 ||
 846			     cmd->cmnd[0] == WRITE_10)) {
 847				/* This will issue a new 6-byte command. */
 848				cmd->device->use_10_for_rw = 0;
 849				action = ACTION_REPREP;
 850			} else if (sshdr.asc == 0x10) /* DIX */ {
 851				action = ACTION_FAIL;
 852				blk_stat = BLK_STS_PROTECTION;
 853			/* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
 854			} else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
 855				action = ACTION_FAIL;
 856				blk_stat = BLK_STS_TARGET;
 857			} else
 858				action = ACTION_FAIL;
 859			break;
 860		case ABORTED_COMMAND:
 861			action = ACTION_FAIL;
 862			if (sshdr.asc == 0x10) /* DIF */
 863				blk_stat = BLK_STS_PROTECTION;
 864			break;
 865		case NOT_READY:
 866			/* If the device is in the process of becoming
 867			 * ready, or has a temporary blockage, retry.
 868			 */
 869			if (sshdr.asc == 0x04) {
 870				switch (sshdr.ascq) {
 871				case 0x01: /* becoming ready */
 872				case 0x04: /* format in progress */
 873				case 0x05: /* rebuild in progress */
 874				case 0x06: /* recalculation in progress */
 875				case 0x07: /* operation in progress */
 876				case 0x08: /* Long write in progress */
 877				case 0x09: /* self test in progress */
 878				case 0x11: /* notify (enable spinup) required */
 879				case 0x14: /* space allocation in progress */
 880				case 0x1a: /* start stop unit in progress */
 881				case 0x1b: /* sanitize in progress */
 882				case 0x1d: /* configuration in progress */
 883					action = ACTION_DELAYED_RETRY;
 884					break;
 885				case 0x0a: /* ALUA state transition */
 886					action = ACTION_DELAYED_REPREP;
 887					break;
 888				/*
 889				 * Depopulation might take many hours,
 890				 * thus it is not worthwhile to retry.
 891				 */
 892				case 0x24: /* depopulation in progress */
 893				case 0x25: /* depopulation restore in progress */
 894					fallthrough;
 895				default:
 896					action = ACTION_FAIL;
 897					break;
 898				}
 899			} else
 900				action = ACTION_FAIL;
 901			break;
 902		case VOLUME_OVERFLOW:
 903			/* See SSC3rXX or current. */
 904			action = ACTION_FAIL;
 905			break;
 906		case DATA_PROTECT:
 907			action = ACTION_FAIL;
 908			if ((sshdr.asc == 0x0C && sshdr.ascq == 0x12) ||
 909			    (sshdr.asc == 0x55 &&
 910			     (sshdr.ascq == 0x0E || sshdr.ascq == 0x0F))) {
 911				/* Insufficient zone resources */
 912				blk_stat = BLK_STS_ZONE_OPEN_RESOURCE;
 913			}
 914			break;
 915		case COMPLETED:
 916			fallthrough;
 917		default:
 918			action = ACTION_FAIL;
 919			break;
 920		}
 921	} else
 922		action = ACTION_FAIL;
 923
 924	if (action != ACTION_FAIL && scsi_cmd_runtime_exceeced(cmd))
 925		action = ACTION_FAIL;
 926
 927	switch (action) {
 928	case ACTION_FAIL:
 929		/* Give up and fail the remainder of the request */
 930		if (!(req->rq_flags & RQF_QUIET)) {
 931			static DEFINE_RATELIMIT_STATE(_rs,
 932					DEFAULT_RATELIMIT_INTERVAL,
 933					DEFAULT_RATELIMIT_BURST);
 934
 935			if (unlikely(scsi_logging_level))
 936				level =
 937				     SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
 938						    SCSI_LOG_MLCOMPLETE_BITS);
 939
 940			/*
 941			 * if logging is enabled the failure will be printed
 942			 * in scsi_log_completion(), so avoid duplicate messages
 943			 */
 944			if (!level && __ratelimit(&_rs)) {
 945				scsi_print_result(cmd, NULL, FAILED);
 946				if (sense_valid)
 947					scsi_print_sense(cmd);
 948				scsi_print_command(cmd);
 949			}
 950		}
 951		if (!scsi_end_request(req, blk_stat, scsi_rq_err_bytes(req)))
 952			return;
 953		fallthrough;
 954	case ACTION_REPREP:
 955		scsi_mq_requeue_cmd(cmd, 0);
 956		break;
 957	case ACTION_DELAYED_REPREP:
 958		scsi_mq_requeue_cmd(cmd, ALUA_TRANSITION_REPREP_DELAY);
 959		break;
 960	case ACTION_RETRY:
 961		/* Retry the same command immediately */
 962		__scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
 963		break;
 964	case ACTION_DELAYED_RETRY:
 965		/* Retry the same command after a delay */
 966		__scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
 967		break;
 968	}
 969}
 970
 971/*
 972 * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
 973 * new result that may suppress further error checking. Also modifies
 974 * *blk_statp in some cases.
 975 */
 976static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
 977					blk_status_t *blk_statp)
 978{
 979	bool sense_valid;
 980	bool sense_current = true;	/* false implies "deferred sense" */
 981	struct request *req = scsi_cmd_to_rq(cmd);
 982	struct scsi_sense_hdr sshdr;
 983
 984	sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
 985	if (sense_valid)
 986		sense_current = !scsi_sense_is_deferred(&sshdr);
 987
 988	if (blk_rq_is_passthrough(req)) {
 989		if (sense_valid) {
 990			/*
 991			 * SG_IO wants current and deferred errors
 992			 */
 993			cmd->sense_len = min(8 + cmd->sense_buffer[7],
 994					     SCSI_SENSE_BUFFERSIZE);
 995		}
 996		if (sense_current)
 997			*blk_statp = scsi_result_to_blk_status(result);
 998	} else if (blk_rq_bytes(req) == 0 && sense_current) {
 999		/*
1000		 * Flush commands do not transfers any data, and thus cannot use
1001		 * good_bytes != blk_rq_bytes(req) as the signal for an error.
1002		 * This sets *blk_statp explicitly for the problem case.
1003		 */
1004		*blk_statp = scsi_result_to_blk_status(result);
1005	}
1006	/*
1007	 * Recovered errors need reporting, but they're always treated as
1008	 * success, so fiddle the result code here.  For passthrough requests
1009	 * we already took a copy of the original into sreq->result which
1010	 * is what gets returned to the user
1011	 */
1012	if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
1013		bool do_print = true;
1014		/*
1015		 * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
1016		 * skip print since caller wants ATA registers. Only occurs
1017		 * on SCSI ATA PASS_THROUGH commands when CK_COND=1
1018		 */
1019		if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
1020			do_print = false;
1021		else if (req->rq_flags & RQF_QUIET)
1022			do_print = false;
1023		if (do_print)
1024			scsi_print_sense(cmd);
1025		result = 0;
1026		/* for passthrough, *blk_statp may be set */
1027		*blk_statp = BLK_STS_OK;
1028	}
1029	/*
1030	 * Another corner case: the SCSI status byte is non-zero but 'good'.
1031	 * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
1032	 * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
1033	 * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
1034	 * intermediate statuses (both obsolete in SAM-4) as good.
1035	 */
1036	if ((result & 0xff) && scsi_status_is_good(result)) {
1037		result = 0;
1038		*blk_statp = BLK_STS_OK;
1039	}
1040	return result;
1041}
1042
1043/**
1044 * scsi_io_completion - Completion processing for SCSI commands.
1045 * @cmd:	command that is finished.
1046 * @good_bytes:	number of processed bytes.
1047 *
1048 * We will finish off the specified number of sectors. If we are done, the
1049 * command block will be released and the queue function will be goosed. If we
1050 * are not done then we have to figure out what to do next:
1051 *
1052 *   a) We can call scsi_mq_requeue_cmd().  The request will be
1053 *	unprepared and put back on the queue.  Then a new command will
1054 *	be created for it.  This should be used if we made forward
1055 *	progress, or if we want to switch from READ(10) to READ(6) for
1056 *	example.
1057 *
1058 *   b) We can call scsi_io_completion_action().  The request will be
1059 *	put back on the queue and retried using the same command as
1060 *	before, possibly after a delay.
1061 *
1062 *   c) We can call scsi_end_request() with blk_stat other than
1063 *	BLK_STS_OK, to fail the remainder of the request.
1064 */
1065void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
1066{
1067	int result = cmd->result;
1068	struct request *req = scsi_cmd_to_rq(cmd);
1069	blk_status_t blk_stat = BLK_STS_OK;
1070
1071	if (unlikely(result))	/* a nz result may or may not be an error */
1072		result = scsi_io_completion_nz_result(cmd, result, &blk_stat);
1073
1074	/*
1075	 * Next deal with any sectors which we were able to correctly
1076	 * handle.
1077	 */
1078	SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
1079		"%u sectors total, %d bytes done.\n",
1080		blk_rq_sectors(req), good_bytes));
1081
1082	/*
1083	 * Failed, zero length commands always need to drop down
1084	 * to retry code. Fast path should return in this block.
1085	 */
1086	if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
1087		if (likely(!scsi_end_request(req, blk_stat, good_bytes)))
1088			return; /* no bytes remaining */
1089	}
1090
1091	/* Kill remainder if no retries. */
1092	if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
1093		if (scsi_end_request(req, blk_stat, blk_rq_bytes(req)))
1094			WARN_ONCE(true,
1095			    "Bytes remaining after failed, no-retry command");
1096		return;
1097	}
1098
1099	/*
1100	 * If there had been no error, but we have leftover bytes in the
1101	 * request just queue the command up again.
1102	 */
1103	if (likely(result == 0))
1104		scsi_mq_requeue_cmd(cmd, 0);
1105	else
1106		scsi_io_completion_action(cmd, result);
1107}
1108
1109static inline bool scsi_cmd_needs_dma_drain(struct scsi_device *sdev,
1110		struct request *rq)
1111{
1112	return sdev->dma_drain_len && blk_rq_is_passthrough(rq) &&
1113	       !op_is_write(req_op(rq)) &&
1114	       sdev->host->hostt->dma_need_drain(rq);
1115}
1116
1117/**
1118 * scsi_alloc_sgtables - Allocate and initialize data and integrity scatterlists
1119 * @cmd: SCSI command data structure to initialize.
1120 *
1121 * Initializes @cmd->sdb and also @cmd->prot_sdb if data integrity is enabled
1122 * for @cmd.
1123 *
1124 * Returns:
1125 * * BLK_STS_OK       - on success
1126 * * BLK_STS_RESOURCE - if the failure is retryable
1127 * * BLK_STS_IOERR    - if the failure is fatal
1128 */
1129blk_status_t scsi_alloc_sgtables(struct scsi_cmnd *cmd)
1130{
1131	struct scsi_device *sdev = cmd->device;
1132	struct request *rq = scsi_cmd_to_rq(cmd);
1133	unsigned short nr_segs = blk_rq_nr_phys_segments(rq);
1134	struct scatterlist *last_sg = NULL;
1135	blk_status_t ret;
1136	bool need_drain = scsi_cmd_needs_dma_drain(sdev, rq);
1137	int count;
1138
1139	if (WARN_ON_ONCE(!nr_segs))
1140		return BLK_STS_IOERR;
1141
1142	/*
1143	 * Make sure there is space for the drain.  The driver must adjust
1144	 * max_hw_segments to be prepared for this.
1145	 */
1146	if (need_drain)
1147		nr_segs++;
1148
1149	/*
1150	 * If sg table allocation fails, requeue request later.
1151	 */
1152	if (unlikely(sg_alloc_table_chained(&cmd->sdb.table, nr_segs,
1153			cmd->sdb.table.sgl, SCSI_INLINE_SG_CNT)))
1154		return BLK_STS_RESOURCE;
1155
1156	/*
1157	 * Next, walk the list, and fill in the addresses and sizes of
1158	 * each segment.
1159	 */
1160	count = __blk_rq_map_sg(rq, cmd->sdb.table.sgl, &last_sg);
1161
1162	if (blk_rq_bytes(rq) & rq->q->limits.dma_pad_mask) {
1163		unsigned int pad_len =
1164			(rq->q->limits.dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
1165
1166		last_sg->length += pad_len;
1167		cmd->extra_len += pad_len;
1168	}
1169
1170	if (need_drain) {
1171		sg_unmark_end(last_sg);
1172		last_sg = sg_next(last_sg);
1173		sg_set_buf(last_sg, sdev->dma_drain_buf, sdev->dma_drain_len);
1174		sg_mark_end(last_sg);
1175
1176		cmd->extra_len += sdev->dma_drain_len;
1177		count++;
1178	}
1179
1180	BUG_ON(count > cmd->sdb.table.nents);
1181	cmd->sdb.table.nents = count;
1182	cmd->sdb.length = blk_rq_payload_bytes(rq);
1183
1184	if (blk_integrity_rq(rq)) {
1185		struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1186
1187		if (WARN_ON_ONCE(!prot_sdb)) {
1188			/*
1189			 * This can happen if someone (e.g. multipath)
1190			 * queues a command to a device on an adapter
1191			 * that does not support DIX.
1192			 */
1193			ret = BLK_STS_IOERR;
1194			goto out_free_sgtables;
1195		}
1196
1197		if (sg_alloc_table_chained(&prot_sdb->table,
1198				rq->nr_integrity_segments,
1199				prot_sdb->table.sgl,
1200				SCSI_INLINE_PROT_SG_CNT)) {
1201			ret = BLK_STS_RESOURCE;
1202			goto out_free_sgtables;
1203		}
1204
1205		count = blk_rq_map_integrity_sg(rq, prot_sdb->table.sgl);
1206		cmd->prot_sdb = prot_sdb;
1207		cmd->prot_sdb->table.nents = count;
1208	}
1209
1210	return BLK_STS_OK;
1211out_free_sgtables:
1212	scsi_free_sgtables(cmd);
1213	return ret;
1214}
1215EXPORT_SYMBOL(scsi_alloc_sgtables);
1216
1217/**
1218 * scsi_initialize_rq - initialize struct scsi_cmnd partially
1219 * @rq: Request associated with the SCSI command to be initialized.
1220 *
1221 * This function initializes the members of struct scsi_cmnd that must be
1222 * initialized before request processing starts and that won't be
1223 * reinitialized if a SCSI command is requeued.
1224 */
1225static void scsi_initialize_rq(struct request *rq)
1226{
1227	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1228
1229	memset(cmd->cmnd, 0, sizeof(cmd->cmnd));
1230	cmd->cmd_len = MAX_COMMAND_SIZE;
1231	cmd->sense_len = 0;
1232	init_rcu_head(&cmd->rcu);
1233	cmd->jiffies_at_alloc = jiffies;
1234	cmd->retries = 0;
1235}
1236
1237/**
1238 * scsi_alloc_request - allocate a block request and partially
1239 *                      initialize its &scsi_cmnd
1240 * @q: the device's request queue
1241 * @opf: the request operation code
1242 * @flags: block layer allocation flags
1243 *
1244 * Return: &struct request pointer on success or %NULL on failure
1245 */
1246struct request *scsi_alloc_request(struct request_queue *q, blk_opf_t opf,
1247				   blk_mq_req_flags_t flags)
1248{
1249	struct request *rq;
1250
1251	rq = blk_mq_alloc_request(q, opf, flags);
1252	if (!IS_ERR(rq))
1253		scsi_initialize_rq(rq);
1254	return rq;
1255}
1256EXPORT_SYMBOL_GPL(scsi_alloc_request);
1257
1258/*
1259 * Only called when the request isn't completed by SCSI, and not freed by
1260 * SCSI
1261 */
1262static void scsi_cleanup_rq(struct request *rq)
1263{
1264	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1265
1266	cmd->flags = 0;
1267
1268	if (rq->rq_flags & RQF_DONTPREP) {
1269		scsi_mq_uninit_cmd(cmd);
1270		rq->rq_flags &= ~RQF_DONTPREP;
1271	}
1272}
1273
1274/* Called before a request is prepared. See also scsi_mq_prep_fn(). */
1275void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1276{
1277	struct request *rq = scsi_cmd_to_rq(cmd);
1278
1279	if (!blk_rq_is_passthrough(rq) && !(cmd->flags & SCMD_INITIALIZED)) {
1280		cmd->flags |= SCMD_INITIALIZED;
1281		scsi_initialize_rq(rq);
1282	}
1283
1284	cmd->device = dev;
1285	INIT_LIST_HEAD(&cmd->eh_entry);
1286	INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1287}
1288
1289static blk_status_t scsi_setup_scsi_cmnd(struct scsi_device *sdev,
1290		struct request *req)
1291{
1292	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1293
1294	/*
1295	 * Passthrough requests may transfer data, in which case they must
1296	 * a bio attached to them.  Or they might contain a SCSI command
1297	 * that does not transfer data, in which case they may optionally
1298	 * submit a request without an attached bio.
1299	 */
1300	if (req->bio) {
1301		blk_status_t ret = scsi_alloc_sgtables(cmd);
1302		if (unlikely(ret != BLK_STS_OK))
1303			return ret;
1304	} else {
1305		BUG_ON(blk_rq_bytes(req));
1306
1307		memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1308	}
1309
1310	cmd->transfersize = blk_rq_bytes(req);
1311	return BLK_STS_OK;
1312}
1313
1314static blk_status_t
1315scsi_device_state_check(struct scsi_device *sdev, struct request *req)
1316{
1317	switch (sdev->sdev_state) {
1318	case SDEV_CREATED:
1319		return BLK_STS_OK;
1320	case SDEV_OFFLINE:
1321	case SDEV_TRANSPORT_OFFLINE:
1322		/*
1323		 * If the device is offline we refuse to process any
1324		 * commands.  The device must be brought online
1325		 * before trying any recovery commands.
1326		 */
1327		if (!sdev->offline_already) {
1328			sdev->offline_already = true;
1329			sdev_printk(KERN_ERR, sdev,
1330				    "rejecting I/O to offline device\n");
1331		}
1332		return BLK_STS_IOERR;
1333	case SDEV_DEL:
1334		/*
1335		 * If the device is fully deleted, we refuse to
1336		 * process any commands as well.
1337		 */
1338		sdev_printk(KERN_ERR, sdev,
1339			    "rejecting I/O to dead device\n");
1340		return BLK_STS_IOERR;
1341	case SDEV_BLOCK:
1342	case SDEV_CREATED_BLOCK:
1343		return BLK_STS_RESOURCE;
1344	case SDEV_QUIESCE:
1345		/*
1346		 * If the device is blocked we only accept power management
1347		 * commands.
1348		 */
1349		if (req && WARN_ON_ONCE(!(req->rq_flags & RQF_PM)))
1350			return BLK_STS_RESOURCE;
1351		return BLK_STS_OK;
1352	default:
1353		/*
1354		 * For any other not fully online state we only allow
1355		 * power management commands.
1356		 */
1357		if (req && !(req->rq_flags & RQF_PM))
1358			return BLK_STS_OFFLINE;
1359		return BLK_STS_OK;
1360	}
1361}
1362
1363/*
1364 * scsi_dev_queue_ready: if we can send requests to sdev, assign one token
1365 * and return the token else return -1.
1366 */
1367static inline int scsi_dev_queue_ready(struct request_queue *q,
1368				  struct scsi_device *sdev)
1369{
1370	int token;
1371
1372	if (!sdev->budget_map.map)
1373		return INT_MAX;
1374
1375	token = sbitmap_get(&sdev->budget_map);
1376	if (token < 0)
1377		return -1;
1378
1379	if (!atomic_read(&sdev->device_blocked))
1380		return token;
1381
1382	/*
1383	 * Only unblock if no other commands are pending and
1384	 * if device_blocked has decreased to zero
1385	 */
1386	if (scsi_device_busy(sdev) > 1 ||
1387	    atomic_dec_return(&sdev->device_blocked) > 0) {
1388		sbitmap_put(&sdev->budget_map, token);
1389		return -1;
1390	}
1391
1392	SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1393			 "unblocking device at zero depth\n"));
1394
1395	return token;
1396}
1397
1398/*
1399 * scsi_target_queue_ready: checks if there we can send commands to target
1400 * @sdev: scsi device on starget to check.
1401 */
1402static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1403					   struct scsi_device *sdev)
1404{
1405	struct scsi_target *starget = scsi_target(sdev);
1406	unsigned int busy;
1407
1408	if (starget->single_lun) {
1409		spin_lock_irq(shost->host_lock);
1410		if (starget->starget_sdev_user &&
1411		    starget->starget_sdev_user != sdev) {
1412			spin_unlock_irq(shost->host_lock);
1413			return 0;
1414		}
1415		starget->starget_sdev_user = sdev;
1416		spin_unlock_irq(shost->host_lock);
1417	}
1418
1419	if (starget->can_queue <= 0)
1420		return 1;
1421
1422	busy = atomic_inc_return(&starget->target_busy) - 1;
1423	if (atomic_read(&starget->target_blocked) > 0) {
1424		if (busy)
1425			goto starved;
1426
1427		/*
1428		 * unblock after target_blocked iterates to zero
1429		 */
1430		if (atomic_dec_return(&starget->target_blocked) > 0)
1431			goto out_dec;
1432
1433		SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1434				 "unblocking target at zero depth\n"));
1435	}
1436
1437	if (busy >= starget->can_queue)
1438		goto starved;
1439
1440	return 1;
1441
1442starved:
1443	spin_lock_irq(shost->host_lock);
1444	list_move_tail(&sdev->starved_entry, &shost->starved_list);
1445	spin_unlock_irq(shost->host_lock);
1446out_dec:
1447	if (starget->can_queue > 0)
1448		atomic_dec(&starget->target_busy);
1449	return 0;
1450}
1451
1452/*
1453 * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1454 * return 0. We must end up running the queue again whenever 0 is
1455 * returned, else IO can hang.
1456 */
1457static inline int scsi_host_queue_ready(struct request_queue *q,
1458				   struct Scsi_Host *shost,
1459				   struct scsi_device *sdev,
1460				   struct scsi_cmnd *cmd)
1461{
1462	if (atomic_read(&shost->host_blocked) > 0) {
1463		if (scsi_host_busy(shost) > 0)
1464			goto starved;
1465
1466		/*
1467		 * unblock after host_blocked iterates to zero
1468		 */
1469		if (atomic_dec_return(&shost->host_blocked) > 0)
1470			goto out_dec;
1471
1472		SCSI_LOG_MLQUEUE(3,
1473			shost_printk(KERN_INFO, shost,
1474				     "unblocking host at zero depth\n"));
1475	}
1476
1477	if (shost->host_self_blocked)
1478		goto starved;
1479
1480	/* We're OK to process the command, so we can't be starved */
1481	if (!list_empty(&sdev->starved_entry)) {
1482		spin_lock_irq(shost->host_lock);
1483		if (!list_empty(&sdev->starved_entry))
1484			list_del_init(&sdev->starved_entry);
1485		spin_unlock_irq(shost->host_lock);
1486	}
1487
1488	__set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1489
1490	return 1;
1491
1492starved:
1493	spin_lock_irq(shost->host_lock);
1494	if (list_empty(&sdev->starved_entry))
1495		list_add_tail(&sdev->starved_entry, &shost->starved_list);
1496	spin_unlock_irq(shost->host_lock);
1497out_dec:
1498	scsi_dec_host_busy(shost, cmd);
1499	return 0;
1500}
1501
1502/*
1503 * Busy state exporting function for request stacking drivers.
1504 *
1505 * For efficiency, no lock is taken to check the busy state of
1506 * shost/starget/sdev, since the returned value is not guaranteed and
1507 * may be changed after request stacking drivers call the function,
1508 * regardless of taking lock or not.
1509 *
1510 * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1511 * needs to return 'not busy'. Otherwise, request stacking drivers
1512 * may hold requests forever.
1513 */
1514static bool scsi_mq_lld_busy(struct request_queue *q)
1515{
1516	struct scsi_device *sdev = q->queuedata;
1517	struct Scsi_Host *shost;
1518
1519	if (blk_queue_dying(q))
1520		return false;
1521
1522	shost = sdev->host;
1523
1524	/*
1525	 * Ignore host/starget busy state.
1526	 * Since block layer does not have a concept of fairness across
1527	 * multiple queues, congestion of host/starget needs to be handled
1528	 * in SCSI layer.
1529	 */
1530	if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1531		return true;
1532
1533	return false;
1534}
1535
1536/*
1537 * Block layer request completion callback. May be called from interrupt
1538 * context.
1539 */
1540static void scsi_complete(struct request *rq)
1541{
1542	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1543	enum scsi_disposition disposition;
1544
1545	if (blk_mq_is_reserved_rq(rq)) {
1546		/* Only pass-through requests are supported in this code path. */
1547		WARN_ON_ONCE(!blk_rq_is_passthrough(scsi_cmd_to_rq(cmd)));
1548		scsi_mq_uninit_cmd(cmd);
1549		__blk_mq_end_request(rq, scsi_result_to_blk_status(cmd->result));
1550		return;
1551	}
1552
1553	INIT_LIST_HEAD(&cmd->eh_entry);
1554
1555	atomic_inc(&cmd->device->iodone_cnt);
1556	if (cmd->result)
1557		atomic_inc(&cmd->device->ioerr_cnt);
1558
1559	disposition = scsi_decide_disposition(cmd);
1560	if (disposition != SUCCESS && scsi_cmd_runtime_exceeced(cmd))
1561		disposition = SUCCESS;
1562
1563	scsi_log_completion(cmd, disposition);
1564
1565	switch (disposition) {
1566	case SUCCESS:
1567		scsi_finish_command(cmd);
1568		break;
1569	case NEEDS_RETRY:
1570		scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1571		break;
1572	case ADD_TO_MLQUEUE:
1573		scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1574		break;
1575	default:
1576		scsi_eh_scmd_add(cmd);
1577		break;
1578	}
1579}
1580
1581/**
1582 * scsi_dispatch_cmd - Dispatch a command to the low-level driver.
1583 * @cmd: command block we are dispatching.
1584 *
1585 * Return: nonzero return request was rejected and device's queue needs to be
1586 * plugged.
1587 */
1588static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1589{
1590	struct Scsi_Host *host = cmd->device->host;
1591	int rtn = 0;
1592
1593	atomic_inc(&cmd->device->iorequest_cnt);
1594
1595	/* check if the device is still usable */
1596	if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1597		/* in SDEV_DEL we error all commands. DID_NO_CONNECT
1598		 * returns an immediate error upwards, and signals
1599		 * that the device is no longer present */
1600		cmd->result = DID_NO_CONNECT << 16;
1601		goto done;
1602	}
1603
1604	/* Check to see if the scsi lld made this device blocked. */
1605	if (unlikely(scsi_device_blocked(cmd->device))) {
1606		/*
1607		 * in blocked state, the command is just put back on
1608		 * the device queue.  The suspend state has already
1609		 * blocked the queue so future requests should not
1610		 * occur until the device transitions out of the
1611		 * suspend state.
1612		 */
1613		SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1614			"queuecommand : device blocked\n"));
1615		atomic_dec(&cmd->device->iorequest_cnt);
1616		return SCSI_MLQUEUE_DEVICE_BUSY;
1617	}
1618
1619	/* Store the LUN value in cmnd, if needed. */
1620	if (cmd->device->lun_in_cdb)
1621		cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1622			       (cmd->device->lun << 5 & 0xe0);
1623
1624	scsi_log_send(cmd);
1625
1626	/*
1627	 * Before we queue this command, check if the command
1628	 * length exceeds what the host adapter can handle.
1629	 */
1630	if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1631		SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1632			       "queuecommand : command too long. "
1633			       "cdb_size=%d host->max_cmd_len=%d\n",
1634			       cmd->cmd_len, cmd->device->host->max_cmd_len));
1635		cmd->result = (DID_ABORT << 16);
1636		goto done;
1637	}
1638
1639	if (unlikely(host->shost_state == SHOST_DEL)) {
1640		cmd->result = (DID_NO_CONNECT << 16);
1641		goto done;
1642
1643	}
1644
1645	trace_scsi_dispatch_cmd_start(cmd);
1646	rtn = host->hostt->queuecommand(host, cmd);
1647	if (rtn) {
1648		atomic_dec(&cmd->device->iorequest_cnt);
1649		trace_scsi_dispatch_cmd_error(cmd, rtn);
1650		if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1651		    rtn != SCSI_MLQUEUE_TARGET_BUSY)
1652			rtn = SCSI_MLQUEUE_HOST_BUSY;
1653
1654		SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1655			"queuecommand : request rejected\n"));
1656	}
1657
1658	return rtn;
1659 done:
1660	scsi_done(cmd);
1661	return 0;
1662}
1663
1664/* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1665static unsigned int scsi_mq_inline_sgl_size(struct Scsi_Host *shost)
1666{
1667	return min_t(unsigned int, shost->sg_tablesize, SCSI_INLINE_SG_CNT) *
1668		sizeof(struct scatterlist);
1669}
1670
1671static blk_status_t scsi_prepare_cmd(struct request *req)
1672{
1673	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1674	struct scsi_device *sdev = req->q->queuedata;
1675	struct Scsi_Host *shost = sdev->host;
1676	bool in_flight = test_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1677	struct scatterlist *sg;
1678
1679	scsi_init_command(sdev, cmd);
1680
1681	cmd->eh_eflags = 0;
1682	cmd->prot_type = 0;
1683	cmd->prot_flags = 0;
1684	cmd->submitter = 0;
1685	memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1686	cmd->underflow = 0;
1687	cmd->transfersize = 0;
1688	cmd->host_scribble = NULL;
1689	cmd->result = 0;
1690	cmd->extra_len = 0;
1691	cmd->state = 0;
1692	if (in_flight)
1693		__set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1694
1695	cmd->prot_op = SCSI_PROT_NORMAL;
1696	if (blk_rq_bytes(req))
1697		cmd->sc_data_direction = rq_dma_dir(req);
1698	else
1699		cmd->sc_data_direction = DMA_NONE;
1700
1701	sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1702	cmd->sdb.table.sgl = sg;
1703
1704	if (scsi_host_get_prot(shost)) {
1705		memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1706
1707		cmd->prot_sdb->table.sgl =
1708			(struct scatterlist *)(cmd->prot_sdb + 1);
1709	}
1710
1711	/*
1712	 * Special handling for passthrough commands, which don't go to the ULP
1713	 * at all:
1714	 */
1715	if (blk_rq_is_passthrough(req))
1716		return scsi_setup_scsi_cmnd(sdev, req);
1717
1718	if (sdev->handler && sdev->handler->prep_fn) {
1719		blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1720
1721		if (ret != BLK_STS_OK)
1722			return ret;
1723	}
1724
1725	/* Usually overridden by the ULP */
1726	cmd->allowed = 0;
1727	memset(cmd->cmnd, 0, sizeof(cmd->cmnd));
1728	return scsi_cmd_to_driver(cmd)->init_command(cmd);
1729}
1730
1731static void scsi_done_internal(struct scsi_cmnd *cmd, bool complete_directly)
1732{
1733	struct request *req = scsi_cmd_to_rq(cmd);
1734
1735	switch (cmd->submitter) {
1736	case SUBMITTED_BY_BLOCK_LAYER:
1737		break;
1738	case SUBMITTED_BY_SCSI_ERROR_HANDLER:
1739		return scsi_eh_done(cmd);
1740	case SUBMITTED_BY_SCSI_RESET_IOCTL:
1741		return;
1742	}
1743
1744	if (unlikely(blk_should_fake_timeout(scsi_cmd_to_rq(cmd)->q)))
1745		return;
1746	if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1747		return;
1748	trace_scsi_dispatch_cmd_done(cmd);
1749
1750	if (complete_directly)
1751		blk_mq_complete_request_direct(req, scsi_complete);
1752	else
1753		blk_mq_complete_request(req);
1754}
1755
1756void scsi_done(struct scsi_cmnd *cmd)
1757{
1758	scsi_done_internal(cmd, false);
1759}
1760EXPORT_SYMBOL(scsi_done);
1761
1762void scsi_done_direct(struct scsi_cmnd *cmd)
1763{
1764	scsi_done_internal(cmd, true);
1765}
1766EXPORT_SYMBOL(scsi_done_direct);
1767
1768static void scsi_mq_put_budget(struct request_queue *q, int budget_token)
1769{
1770	struct scsi_device *sdev = q->queuedata;
1771
1772	if (sdev->budget_map.map)
1773		sbitmap_put(&sdev->budget_map, budget_token);
1774}
1775
1776/*
1777 * When to reinvoke queueing after a resource shortage. It's 3 msecs to
1778 * not change behaviour from the previous unplug mechanism, experimentation
1779 * may prove this needs changing.
1780 */
1781#define SCSI_QUEUE_DELAY 3
1782
1783static int scsi_mq_get_budget(struct request_queue *q)
1784{
1785	struct scsi_device *sdev = q->queuedata;
1786	int token = scsi_dev_queue_ready(q, sdev);
1787
1788	if (token >= 0)
1789		return token;
1790
1791	atomic_inc(&sdev->restarts);
1792
1793	/*
1794	 * Orders atomic_inc(&sdev->restarts) and atomic_read(&sdev->device_busy).
1795	 * .restarts must be incremented before .device_busy is read because the
1796	 * code in scsi_run_queue_async() depends on the order of these operations.
1797	 */
1798	smp_mb__after_atomic();
1799
1800	/*
1801	 * If all in-flight requests originated from this LUN are completed
1802	 * before reading .device_busy, sdev->device_busy will be observed as
1803	 * zero, then blk_mq_delay_run_hw_queues() will dispatch this request
1804	 * soon. Otherwise, completion of one of these requests will observe
1805	 * the .restarts flag, and the request queue will be run for handling
1806	 * this request, see scsi_end_request().
1807	 */
1808	if (unlikely(scsi_device_busy(sdev) == 0 &&
1809				!scsi_device_blocked(sdev)))
1810		blk_mq_delay_run_hw_queues(sdev->request_queue, SCSI_QUEUE_DELAY);
1811	return -1;
1812}
1813
1814static void scsi_mq_set_rq_budget_token(struct request *req, int token)
1815{
1816	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1817
1818	cmd->budget_token = token;
1819}
1820
1821static int scsi_mq_get_rq_budget_token(struct request *req)
1822{
1823	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1824
1825	return cmd->budget_token;
1826}
1827
1828static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1829			 const struct blk_mq_queue_data *bd)
1830{
1831	struct request *req = bd->rq;
1832	struct request_queue *q = req->q;
1833	struct scsi_device *sdev = q->queuedata;
1834	struct Scsi_Host *shost = sdev->host;
1835	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1836	blk_status_t ret;
1837	int reason;
1838
1839	WARN_ON_ONCE(cmd->budget_token < 0);
1840
1841	/*
1842	 * Bypass the SCSI device, SCSI target and SCSI host checks for
1843	 * reserved commands.
1844	 */
1845	if (!blk_mq_is_reserved_rq(req)) {
1846		/*
1847		 * If the device is not in running state we will reject some or
1848		 * all commands.
1849		 */
1850		if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1851			ret = scsi_device_state_check(sdev, req);
1852			if (ret != BLK_STS_OK)
1853				goto out_put_budget;
1854		}
1855
1856		ret = BLK_STS_RESOURCE;
1857		if (!scsi_target_queue_ready(shost, sdev))
1858			goto out_put_budget;
1859		if (unlikely(scsi_host_in_recovery(shost))) {
1860			if (cmd->flags & SCMD_FAIL_IF_RECOVERING)
1861				ret = BLK_STS_OFFLINE;
1862			goto out_dec_target_busy;
1863		}
1864		if (!scsi_host_queue_ready(q, shost, sdev, cmd))
1865			goto out_dec_target_busy;
1866	}
1867
1868	/*
1869	 * Only clear the driver-private command data if the LLD does not supply
1870	 * a function to initialize that data.
1871	 */
1872	if (shost->hostt->cmd_size && !shost->hostt->init_cmd_priv)
1873		memset(scsi_cmd_priv(cmd), 0, shost->hostt->cmd_size);
1874
1875	if (!(req->rq_flags & RQF_DONTPREP)) {
1876		ret = scsi_prepare_cmd(req);
1877		if (ret != BLK_STS_OK)
1878			goto out_dec_host_busy;
1879		req->rq_flags |= RQF_DONTPREP;
1880	} else {
1881		clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1882	}
1883
1884	cmd->flags &= SCMD_PRESERVED_FLAGS;
1885	if (sdev->simple_tags)
1886		cmd->flags |= SCMD_TAGGED;
1887	if (bd->last)
1888		cmd->flags |= SCMD_LAST;
1889
1890	scsi_set_resid(cmd, 0);
1891	memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
1892	cmd->submitter = SUBMITTED_BY_BLOCK_LAYER;
1893
1894	blk_mq_start_request(req);
1895	if (blk_mq_is_reserved_rq(req)) {
1896		reason = shost->hostt->queue_reserved_command(shost, cmd);
1897		if (reason) {
1898			ret = BLK_STS_RESOURCE;
1899			goto out_put_budget;
1900		}
1901		return BLK_STS_OK;
1902	}
1903	reason = scsi_dispatch_cmd(cmd);
1904	if (reason) {
1905		scsi_set_blocked(cmd, reason);
1906		ret = BLK_STS_RESOURCE;
1907		goto out_dec_host_busy;
1908	}
1909
1910	return BLK_STS_OK;
1911
1912out_dec_host_busy:
1913	scsi_dec_host_busy(shost, cmd);
1914out_dec_target_busy:
1915	if (scsi_target(sdev)->can_queue > 0)
1916		atomic_dec(&scsi_target(sdev)->target_busy);
1917out_put_budget:
1918	scsi_mq_put_budget(q, cmd->budget_token);
1919	cmd->budget_token = -1;
1920	switch (ret) {
1921	case BLK_STS_OK:
1922		break;
1923	case BLK_STS_RESOURCE:
1924		if (scsi_device_blocked(sdev))
1925			ret = BLK_STS_DEV_RESOURCE;
1926		break;
1927	case BLK_STS_AGAIN:
1928		cmd->result = DID_BUS_BUSY << 16;
1929		if (req->rq_flags & RQF_DONTPREP)
1930			scsi_mq_uninit_cmd(cmd);
1931		break;
1932	default:
1933		if (unlikely(!scsi_device_online(sdev)))
1934			cmd->result = DID_NO_CONNECT << 16;
1935		else
1936			cmd->result = DID_ERROR << 16;
1937		/*
1938		 * Make sure to release all allocated resources when
1939		 * we hit an error, as we will never see this command
1940		 * again.
1941		 */
1942		if (req->rq_flags & RQF_DONTPREP)
1943			scsi_mq_uninit_cmd(cmd);
1944		scsi_run_queue_async(sdev);
1945		break;
1946	}
1947	return ret;
1948}
1949
1950static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1951				unsigned int hctx_idx, unsigned int numa_node)
1952{
1953	struct Scsi_Host *shost = set->driver_data;
1954	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1955	struct scatterlist *sg;
1956	int ret = 0;
1957
1958	cmd->sense_buffer =
1959		kmem_cache_alloc_node(scsi_sense_cache, GFP_KERNEL, numa_node);
1960	if (!cmd->sense_buffer)
1961		return -ENOMEM;
1962
1963	if (scsi_host_get_prot(shost)) {
1964		sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1965			shost->hostt->cmd_size;
1966		cmd->prot_sdb = (void *)sg + scsi_mq_inline_sgl_size(shost);
1967	}
1968
1969	if (shost->hostt->init_cmd_priv) {
1970		ret = shost->hostt->init_cmd_priv(shost, cmd);
1971		if (ret < 0)
1972			kmem_cache_free(scsi_sense_cache, cmd->sense_buffer);
1973	}
1974
1975	return ret;
1976}
1977
1978static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1979				 unsigned int hctx_idx)
1980{
1981	struct Scsi_Host *shost = set->driver_data;
1982	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1983
1984	if (shost->hostt->exit_cmd_priv)
1985		shost->hostt->exit_cmd_priv(shost, cmd);
1986	kmem_cache_free(scsi_sense_cache, cmd->sense_buffer);
1987}
1988
1989
1990static int scsi_mq_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
1991{
1992	struct Scsi_Host *shost = hctx->driver_data;
1993
1994	if (shost->hostt->mq_poll)
1995		return shost->hostt->mq_poll(shost, hctx->queue_num);
1996
1997	return 0;
1998}
1999
2000static int scsi_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
2001			  unsigned int hctx_idx)
2002{
2003	struct Scsi_Host *shost = data;
2004
2005	hctx->driver_data = shost;
2006	return 0;
2007}
2008
2009static void scsi_map_queues(struct blk_mq_tag_set *set)
2010{
2011	struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
2012
2013	if (shost->hostt->map_queues)
2014		return shost->hostt->map_queues(shost);
2015	blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2016}
2017
2018void scsi_init_limits(struct Scsi_Host *shost, struct queue_limits *lim)
2019{
2020	struct device *dev = shost->dma_dev;
2021
2022	memset(lim, 0, sizeof(*lim));
2023	lim->max_segments =
2024		min_t(unsigned short, shost->sg_tablesize, SG_MAX_SEGMENTS);
2025
2026	if (scsi_host_prot_dma(shost)) {
2027		shost->sg_prot_tablesize =
2028			min_not_zero(shost->sg_prot_tablesize,
2029				     (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
2030		BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
2031		lim->max_integrity_segments = shost->sg_prot_tablesize;
2032	}
2033
2034	lim->max_hw_sectors = shost->max_sectors;
2035	lim->seg_boundary_mask = shost->dma_boundary;
2036	lim->max_segment_size = shost->max_segment_size;
2037	lim->virt_boundary_mask = shost->virt_boundary_mask;
2038	lim->dma_alignment = max_t(unsigned int,
2039		shost->dma_alignment, dma_get_cache_alignment() - 1);
2040
2041	/*
2042	 * Propagate the DMA formation properties to the dma-mapping layer as
2043	 * a courtesy service to the LLDDs.  This needs to check that the buses
2044	 * actually support the DMA API first, though.
2045	 */
2046	if (dev->dma_parms) {
2047		dma_set_seg_boundary(dev, shost->dma_boundary);
2048		dma_set_max_seg_size(dev, shost->max_segment_size);
2049	}
2050}
2051EXPORT_SYMBOL_GPL(scsi_init_limits);
2052
2053static const struct blk_mq_ops scsi_mq_ops_no_commit = {
2054	.get_budget	= scsi_mq_get_budget,
2055	.put_budget	= scsi_mq_put_budget,
2056	.queue_rq	= scsi_queue_rq,
2057	.complete	= scsi_complete,
2058	.timeout	= scsi_timeout,
2059#ifdef CONFIG_BLK_DEBUG_FS
2060	.show_rq	= scsi_show_rq,
2061#endif
2062	.init_request	= scsi_mq_init_request,
2063	.exit_request	= scsi_mq_exit_request,
2064	.cleanup_rq	= scsi_cleanup_rq,
2065	.busy		= scsi_mq_lld_busy,
2066	.map_queues	= scsi_map_queues,
2067	.init_hctx	= scsi_init_hctx,
2068	.poll		= scsi_mq_poll,
2069	.set_rq_budget_token = scsi_mq_set_rq_budget_token,
2070	.get_rq_budget_token = scsi_mq_get_rq_budget_token,
2071};
2072
2073
2074static void scsi_commit_rqs(struct blk_mq_hw_ctx *hctx)
2075{
2076	struct Scsi_Host *shost = hctx->driver_data;
2077
2078	shost->hostt->commit_rqs(shost, hctx->queue_num);
2079}
2080
2081static const struct blk_mq_ops scsi_mq_ops = {
2082	.get_budget	= scsi_mq_get_budget,
2083	.put_budget	= scsi_mq_put_budget,
2084	.queue_rq	= scsi_queue_rq,
2085	.commit_rqs	= scsi_commit_rqs,
2086	.complete	= scsi_complete,
2087	.timeout	= scsi_timeout,
2088#ifdef CONFIG_BLK_DEBUG_FS
2089	.show_rq	= scsi_show_rq,
2090#endif
2091	.init_request	= scsi_mq_init_request,
2092	.exit_request	= scsi_mq_exit_request,
2093	.cleanup_rq	= scsi_cleanup_rq,
2094	.busy		= scsi_mq_lld_busy,
2095	.map_queues	= scsi_map_queues,
2096	.init_hctx	= scsi_init_hctx,
2097	.poll		= scsi_mq_poll,
2098	.set_rq_budget_token = scsi_mq_set_rq_budget_token,
2099	.get_rq_budget_token = scsi_mq_get_rq_budget_token,
2100};
2101
2102int scsi_mq_setup_tags(struct Scsi_Host *shost)
2103{
2104	unsigned int cmd_size, sgl_size;
2105	struct blk_mq_tag_set *tag_set = &shost->tag_set;
2106
2107	sgl_size = max_t(unsigned int, sizeof(struct scatterlist),
2108				scsi_mq_inline_sgl_size(shost));
2109	cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
2110	if (scsi_host_get_prot(shost))
2111		cmd_size += sizeof(struct scsi_data_buffer) +
2112			sizeof(struct scatterlist) * SCSI_INLINE_PROT_SG_CNT;
2113
2114	memset(tag_set, 0, sizeof(*tag_set));
2115	if (shost->hostt->commit_rqs)
2116		tag_set->ops = &scsi_mq_ops;
2117	else
2118		tag_set->ops = &scsi_mq_ops_no_commit;
2119	tag_set->nr_hw_queues = shost->nr_hw_queues ? : 1;
2120	tag_set->nr_maps = shost->nr_maps ? : 1;
2121	tag_set->queue_depth = shost->can_queue + shost->nr_reserved_cmds;
2122	tag_set->reserved_tags = shost->nr_reserved_cmds;
2123	tag_set->cmd_size = cmd_size;
2124	tag_set->numa_node = dev_to_node(shost->dma_dev);
2125	if (shost->hostt->tag_alloc_policy_rr)
2126		tag_set->flags |= BLK_MQ_F_TAG_RR;
2127	if (shost->queuecommand_may_block)
2128		tag_set->flags |= BLK_MQ_F_BLOCKING;
2129	tag_set->driver_data = shost;
2130	if (shost->host_tagset)
2131		tag_set->flags |= BLK_MQ_F_TAG_HCTX_SHARED;
2132
2133	return blk_mq_alloc_tag_set(tag_set);
2134}
2135
2136void scsi_mq_free_tags(struct kref *kref)
2137{
2138	struct Scsi_Host *shost = container_of(kref, typeof(*shost),
2139					       tagset_refcnt);
2140
2141	blk_mq_free_tag_set(&shost->tag_set);
2142	complete(&shost->tagset_freed);
2143}
2144
2145/**
2146 * scsi_get_internal_cmd() - Allocate an internal SCSI command.
2147 * @sdev: SCSI device from which to allocate the command
2148 * @data_direction: Data direction for the allocated command
2149 * @flags: request allocation flags, e.g. BLK_MQ_REQ_RESERVED or
2150 *	BLK_MQ_REQ_NOWAIT.
2151 *
2152 * Allocates a SCSI command for internal LLDD use.
2153 */
2154struct scsi_cmnd *scsi_get_internal_cmd(struct scsi_device *sdev,
2155					enum dma_data_direction data_direction,
2156					blk_mq_req_flags_t flags)
2157{
2158	enum req_op op = data_direction == DMA_TO_DEVICE ? REQ_OP_DRV_OUT :
2159							   REQ_OP_DRV_IN;
2160	struct scsi_cmnd *scmd;
2161	struct request *rq;
2162
2163	rq = scsi_alloc_request(sdev->request_queue, op, flags);
2164	if (IS_ERR(rq))
2165		return NULL;
2166	scmd = blk_mq_rq_to_pdu(rq);
2167	scmd->device = sdev;
2168
2169	return scmd;
2170}
2171EXPORT_SYMBOL_GPL(scsi_get_internal_cmd);
2172
2173/**
2174 * scsi_put_internal_cmd() - Free an internal SCSI command.
2175 * @scmd: SCSI command to be freed
2176 */
2177void scsi_put_internal_cmd(struct scsi_cmnd *scmd)
2178{
2179	blk_mq_free_request(blk_mq_rq_from_pdu(scmd));
2180}
2181EXPORT_SYMBOL_GPL(scsi_put_internal_cmd);
2182
2183/**
2184 * scsi_device_from_queue - return sdev associated with a request_queue
2185 * @q: The request queue to return the sdev from
2186 *
2187 * Return the sdev associated with a request queue or NULL if the
2188 * request_queue does not reference a SCSI device.
2189 */
2190struct scsi_device *scsi_device_from_queue(struct request_queue *q)
2191{
2192	struct scsi_device *sdev = NULL;
2193
2194	if (q->mq_ops == &scsi_mq_ops_no_commit ||
2195	    q->mq_ops == &scsi_mq_ops)
2196		sdev = q->queuedata;
2197	if (!sdev || !get_device(&sdev->sdev_gendev))
2198		sdev = NULL;
2199
2200	return sdev;
2201}
2202/*
2203 * pktcdvd should have been integrated into the SCSI layers, but for historical
2204 * reasons like the old IDE driver it isn't.  This export allows it to safely
2205 * probe if a given device is a SCSI one and only attach to that.
2206 */
2207#ifdef CONFIG_CDROM_PKTCDVD_MODULE
2208EXPORT_SYMBOL_GPL(scsi_device_from_queue);
2209#endif
2210
2211/**
2212 * scsi_block_requests - Utility function used by low-level drivers to prevent
2213 * further commands from being queued to the device.
2214 * @shost:  host in question
2215 *
2216 * There is no timer nor any other means by which the requests get unblocked
2217 * other than the low-level driver calling scsi_unblock_requests().
2218 */
2219void scsi_block_requests(struct Scsi_Host *shost)
2220{
2221	shost->host_self_blocked = 1;
2222}
2223EXPORT_SYMBOL(scsi_block_requests);
2224
2225/**
2226 * scsi_unblock_requests - Utility function used by low-level drivers to allow
2227 * further commands to be queued to the device.
2228 * @shost:  host in question
2229 *
2230 * There is no timer nor any other means by which the requests get unblocked
2231 * other than the low-level driver calling scsi_unblock_requests(). This is done
2232 * as an API function so that changes to the internals of the scsi mid-layer
2233 * won't require wholesale changes to drivers that use this feature.
2234 */
2235void scsi_unblock_requests(struct Scsi_Host *shost)
2236{
2237	shost->host_self_blocked = 0;
2238	scsi_run_host_queues(shost);
2239}
2240EXPORT_SYMBOL(scsi_unblock_requests);
2241
2242void scsi_exit_queue(void)
2243{
2244	kmem_cache_destroy(scsi_sense_cache);
2245}
2246
2247/**
2248 *	scsi_mode_select - issue a mode select
2249 *	@sdev:	SCSI device to be queried
2250 *	@pf:	Page format bit (1 == standard, 0 == vendor specific)
2251 *	@sp:	Save page bit (0 == don't save, 1 == save)
2252 *	@buffer: request buffer (may not be smaller than eight bytes)
2253 *	@len:	length of request buffer.
2254 *	@timeout: command timeout
2255 *	@retries: number of retries before failing
2256 *	@data: returns a structure abstracting the mode header data
2257 *	@sshdr: place to put sense data (or NULL if no sense to be collected).
2258 *		must be SCSI_SENSE_BUFFERSIZE big.
2259 *
2260 *	Returns zero if successful; negative error number or scsi
2261 *	status on error
2262 *
2263 */
2264int scsi_mode_select(struct scsi_device *sdev, int pf, int sp,
2265		     unsigned char *buffer, int len, int timeout, int retries,
2266		     struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2267{
2268	unsigned char cmd[10];
2269	unsigned char *real_buffer;
2270	const struct scsi_exec_args exec_args = {
2271		.sshdr = sshdr,
2272	};
2273	int ret;
2274
2275	memset(cmd, 0, sizeof(cmd));
2276	cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2277
2278	/*
2279	 * Use MODE SELECT(10) if the device asked for it or if the mode page
2280	 * and the mode select header cannot fit within the maximumm 255 bytes
2281	 * of the MODE SELECT(6) command.
2282	 */
2283	if (sdev->use_10_for_ms ||
2284	    len + 4 > 255 ||
2285	    data->block_descriptor_length > 255) {
2286		if (len > 65535 - 8)
2287			return -EINVAL;
2288		real_buffer = kmalloc(8 + len, GFP_KERNEL);
2289		if (!real_buffer)
2290			return -ENOMEM;
2291		memcpy(real_buffer + 8, buffer, len);
2292		len += 8;
2293		real_buffer[0] = 0;
2294		real_buffer[1] = 0;
2295		real_buffer[2] = data->medium_type;
2296		real_buffer[3] = data->device_specific;
2297		real_buffer[4] = data->longlba ? 0x01 : 0;
2298		real_buffer[5] = 0;
2299		put_unaligned_be16(data->block_descriptor_length,
2300				   &real_buffer[6]);
2301
2302		cmd[0] = MODE_SELECT_10;
2303		put_unaligned_be16(len, &cmd[7]);
2304	} else {
2305		if (data->longlba)
2306			return -EINVAL;
2307
2308		real_buffer = kmalloc(4 + len, GFP_KERNEL);
2309		if (!real_buffer)
2310			return -ENOMEM;
2311		memcpy(real_buffer + 4, buffer, len);
2312		len += 4;
2313		real_buffer[0] = 0;
2314		real_buffer[1] = data->medium_type;
2315		real_buffer[2] = data->device_specific;
2316		real_buffer[3] = data->block_descriptor_length;
2317
2318		cmd[0] = MODE_SELECT;
2319		cmd[4] = len;
2320	}
2321
2322	ret = scsi_execute_cmd(sdev, cmd, REQ_OP_DRV_OUT, real_buffer, len,
2323			       timeout, retries, &exec_args);
2324	kfree(real_buffer);
2325	return ret;
2326}
2327EXPORT_SYMBOL_GPL(scsi_mode_select);
2328
2329/**
2330 *	scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2331 *	@sdev:	SCSI device to be queried
2332 *	@dbd:	set to prevent mode sense from returning block descriptors
2333 *	@modepage: mode page being requested
2334 *	@subpage: sub-page of the mode page being requested
2335 *	@buffer: request buffer (may not be smaller than eight bytes)
2336 *	@len:	length of request buffer.
2337 *	@timeout: command timeout
2338 *	@retries: number of retries before failing
2339 *	@data: returns a structure abstracting the mode header data
2340 *	@sshdr: place to put sense data (or NULL if no sense to be collected).
2341 *		must be SCSI_SENSE_BUFFERSIZE big.
2342 *
2343 *	Returns zero if successful, or a negative error number on failure
2344 */
2345int
2346scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage, int subpage,
2347		  unsigned char *buffer, int len, int timeout, int retries,
2348		  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2349{
2350	unsigned char cmd[12];
2351	int use_10_for_ms;
2352	int header_length;
2353	int result;
2354	struct scsi_sense_hdr my_sshdr;
2355	struct scsi_failure failure_defs[] = {
2356		{
2357			.sense = UNIT_ATTENTION,
2358			.asc = SCMD_FAILURE_ASC_ANY,
2359			.ascq = SCMD_FAILURE_ASCQ_ANY,
2360			.allowed = retries,
2361			.result = SAM_STAT_CHECK_CONDITION,
2362		},
2363		{}
2364	};
2365	struct scsi_failures failures = {
2366		.failure_definitions = failure_defs,
2367	};
2368	const struct scsi_exec_args exec_args = {
2369		/* caller might not be interested in sense, but we need it */
2370		.sshdr = sshdr ? : &my_sshdr,
2371		.failures = &failures,
2372	};
2373
2374	memset(data, 0, sizeof(*data));
2375	memset(&cmd[0], 0, 12);
2376
2377	dbd = sdev->set_dbd_for_ms ? 8 : dbd;
2378	cmd[1] = dbd & 0x18;	/* allows DBD and LLBA bits */
2379	cmd[2] = modepage;
2380	cmd[3] = subpage;
2381
2382	sshdr = exec_args.sshdr;
2383
2384 retry:
2385	use_10_for_ms = sdev->use_10_for_ms || len > 255;
2386
2387	if (use_10_for_ms) {
2388		if (len < 8 || len > 65535)
2389			return -EINVAL;
2390
2391		cmd[0] = MODE_SENSE_10;
2392		put_unaligned_be16(len, &cmd[7]);
2393		header_length = 8;
2394	} else {
2395		if (len < 4)
2396			return -EINVAL;
2397
2398		cmd[0] = MODE_SENSE;
2399		cmd[4] = len;
2400		header_length = 4;
2401	}
2402
2403	memset(buffer, 0, len);
2404
2405	result = scsi_execute_cmd(sdev, cmd, REQ_OP_DRV_IN, buffer, len,
2406				  timeout, retries, &exec_args);
2407	if (result < 0)
2408		return result;
2409
2410	/* This code looks awful: what it's doing is making sure an
2411	 * ILLEGAL REQUEST sense return identifies the actual command
2412	 * byte as the problem.  MODE_SENSE commands can return
2413	 * ILLEGAL REQUEST if the code page isn't supported */
2414
2415	if (!scsi_status_is_good(result)) {
2416		if (scsi_sense_valid(sshdr)) {
2417			if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2418			    (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2419				/*
2420				 * Invalid command operation code: retry using
2421				 * MODE SENSE(6) if this was a MODE SENSE(10)
2422				 * request, except if the request mode page is
2423				 * too large for MODE SENSE single byte
2424				 * allocation length field.
2425				 */
2426				if (use_10_for_ms) {
2427					if (len > 255)
2428						return -EIO;
2429					sdev->use_10_for_ms = 0;
2430					goto retry;
2431				}
2432			}
2433		}
2434		return -EIO;
2435	}
2436	if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2437		     (modepage == 6 || modepage == 8))) {
2438		/* Initio breakage? */
2439		header_length = 0;
2440		data->length = 13;
2441		data->medium_type = 0;
2442		data->device_specific = 0;
2443		data->longlba = 0;
2444		data->block_descriptor_length = 0;
2445	} else if (use_10_for_ms) {
2446		data->length = get_unaligned_be16(&buffer[0]) + 2;
2447		data->medium_type = buffer[2];
2448		data->device_specific = buffer[3];
2449		data->longlba = buffer[4] & 0x01;
2450		data->block_descriptor_length = get_unaligned_be16(&buffer[6]);
2451	} else {
2452		data->length = buffer[0] + 1;
2453		data->medium_type = buffer[1];
2454		data->device_specific = buffer[2];
2455		data->block_descriptor_length = buffer[3];
2456	}
2457	data->header_length = header_length;
2458
2459	return 0;
2460}
2461EXPORT_SYMBOL(scsi_mode_sense);
2462
2463/**
2464 *	scsi_test_unit_ready - test if unit is ready
2465 *	@sdev:	scsi device to change the state of.
2466 *	@timeout: command timeout
2467 *	@retries: number of retries before failing
2468 *	@sshdr: outpout pointer for decoded sense information.
2469 *
2470 *	Returns zero if successful or an error if TUR failed.  For
2471 *	removable media, UNIT_ATTENTION sets ->changed flag.
2472 **/
2473int
2474scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2475		     struct scsi_sense_hdr *sshdr)
2476{
2477	char cmd[] = {
2478		TEST_UNIT_READY, 0, 0, 0, 0, 0,
2479	};
2480	const struct scsi_exec_args exec_args = {
2481		.sshdr = sshdr,
2482	};
2483	int result;
2484
2485	/* try to eat the UNIT_ATTENTION if there are enough retries */
2486	do {
2487		result = scsi_execute_cmd(sdev, cmd, REQ_OP_DRV_IN, NULL, 0,
2488					  timeout, 1, &exec_args);
2489		if (sdev->removable && result > 0 && scsi_sense_valid(sshdr) &&
2490		    sshdr->sense_key == UNIT_ATTENTION)
2491			sdev->changed = 1;
2492	} while (result > 0 && scsi_sense_valid(sshdr) &&
2493		 sshdr->sense_key == UNIT_ATTENTION && --retries);
2494
2495	return result;
2496}
2497EXPORT_SYMBOL(scsi_test_unit_ready);
2498
2499/**
2500 *	scsi_device_set_state - Take the given device through the device state model.
2501 *	@sdev:	scsi device to change the state of.
2502 *	@state:	state to change to.
2503 *
2504 *	Returns zero if successful or an error if the requested
2505 *	transition is illegal.
2506 */
2507int
2508scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2509{
2510	enum scsi_device_state oldstate = sdev->sdev_state;
2511
2512	if (state == oldstate)
2513		return 0;
2514
2515	switch (state) {
2516	case SDEV_CREATED:
2517		switch (oldstate) {
2518		case SDEV_CREATED_BLOCK:
2519			break;
2520		default:
2521			goto illegal;
2522		}
2523		break;
2524
2525	case SDEV_RUNNING:
2526		switch (oldstate) {
2527		case SDEV_CREATED:
2528		case SDEV_OFFLINE:
2529		case SDEV_TRANSPORT_OFFLINE:
2530		case SDEV_QUIESCE:
2531		case SDEV_BLOCK:
2532			break;
2533		default:
2534			goto illegal;
2535		}
2536		break;
2537
2538	case SDEV_QUIESCE:
2539		switch (oldstate) {
2540		case SDEV_RUNNING:
2541		case SDEV_OFFLINE:
2542		case SDEV_TRANSPORT_OFFLINE:
2543			break;
2544		default:
2545			goto illegal;
2546		}
2547		break;
2548
2549	case SDEV_OFFLINE:
2550	case SDEV_TRANSPORT_OFFLINE:
2551		switch (oldstate) {
2552		case SDEV_CREATED:
2553		case SDEV_RUNNING:
2554		case SDEV_QUIESCE:
2555		case SDEV_BLOCK:
2556			break;
2557		default:
2558			goto illegal;
2559		}
2560		break;
2561
2562	case SDEV_BLOCK:
2563		switch (oldstate) {
2564		case SDEV_RUNNING:
2565		case SDEV_CREATED_BLOCK:
2566		case SDEV_QUIESCE:
2567		case SDEV_OFFLINE:
2568			break;
2569		default:
2570			goto illegal;
2571		}
2572		break;
2573
2574	case SDEV_CREATED_BLOCK:
2575		switch (oldstate) {
2576		case SDEV_CREATED:
2577			break;
2578		default:
2579			goto illegal;
2580		}
2581		break;
2582
2583	case SDEV_CANCEL:
2584		switch (oldstate) {
2585		case SDEV_CREATED:
2586		case SDEV_RUNNING:
2587		case SDEV_QUIESCE:
2588		case SDEV_OFFLINE:
2589		case SDEV_TRANSPORT_OFFLINE:
2590			break;
2591		default:
2592			goto illegal;
2593		}
2594		break;
2595
2596	case SDEV_DEL:
2597		switch (oldstate) {
2598		case SDEV_CREATED:
2599		case SDEV_RUNNING:
2600		case SDEV_OFFLINE:
2601		case SDEV_TRANSPORT_OFFLINE:
2602		case SDEV_CANCEL:
2603		case SDEV_BLOCK:
2604		case SDEV_CREATED_BLOCK:
2605			break;
2606		default:
2607			goto illegal;
2608		}
2609		break;
2610
2611	}
2612	sdev->offline_already = false;
2613	sdev->sdev_state = state;
2614	return 0;
2615
2616 illegal:
2617	SCSI_LOG_ERROR_RECOVERY(1,
2618				sdev_printk(KERN_ERR, sdev,
2619					    "Illegal state transition %s->%s",
2620					    scsi_device_state_name(oldstate),
2621					    scsi_device_state_name(state))
2622				);
2623	return -EINVAL;
2624}
2625EXPORT_SYMBOL(scsi_device_set_state);
2626
2627/**
2628 *	scsi_evt_emit - emit a single SCSI device uevent
2629 *	@sdev: associated SCSI device
2630 *	@evt: event to emit
2631 *
2632 *	Send a single uevent (scsi_event) to the associated scsi_device.
2633 */
2634static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2635{
2636	int idx = 0;
2637	char *envp[3];
2638
2639	switch (evt->evt_type) {
2640	case SDEV_EVT_MEDIA_CHANGE:
2641		envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2642		break;
2643	case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2644		scsi_rescan_device(sdev);
2645		envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2646		break;
2647	case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2648		envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2649		break;
2650	case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2651	       envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2652		break;
2653	case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2654		envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2655		break;
2656	case SDEV_EVT_LUN_CHANGE_REPORTED:
2657		envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2658		break;
2659	case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2660		envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2661		break;
2662	case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2663		envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2664		break;
2665	default:
2666		/* do nothing */
2667		break;
2668	}
2669
2670	envp[idx++] = NULL;
2671
2672	kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2673}
2674
2675/**
2676 *	scsi_evt_thread - send a uevent for each scsi event
2677 *	@work: work struct for scsi_device
2678 *
2679 *	Dispatch queued events to their associated scsi_device kobjects
2680 *	as uevents.
2681 */
2682void scsi_evt_thread(struct work_struct *work)
2683{
2684	struct scsi_device *sdev;
2685	enum scsi_device_event evt_type;
2686	LIST_HEAD(event_list);
2687
2688	sdev = container_of(work, struct scsi_device, event_work);
2689
2690	for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2691		if (test_and_clear_bit(evt_type, sdev->pending_events))
2692			sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2693
2694	while (1) {
2695		struct scsi_event *evt;
2696		struct list_head *this, *tmp;
2697		unsigned long flags;
2698
2699		spin_lock_irqsave(&sdev->list_lock, flags);
2700		list_splice_init(&sdev->event_list, &event_list);
2701		spin_unlock_irqrestore(&sdev->list_lock, flags);
2702
2703		if (list_empty(&event_list))
2704			break;
2705
2706		list_for_each_safe(this, tmp, &event_list) {
2707			evt = list_entry(this, struct scsi_event, node);
2708			list_del(&evt->node);
2709			scsi_evt_emit(sdev, evt);
2710			kfree(evt);
2711		}
2712	}
2713}
2714
2715/**
2716 * 	sdev_evt_send - send asserted event to uevent thread
2717 *	@sdev: scsi_device event occurred on
2718 *	@evt: event to send
2719 *
2720 *	Assert scsi device event asynchronously.
2721 */
2722void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2723{
2724	unsigned long flags;
2725
2726#if 0
2727	/* FIXME: currently this check eliminates all media change events
2728	 * for polled devices.  Need to update to discriminate between AN
2729	 * and polled events */
2730	if (!test_bit(evt->evt_type, sdev->supported_events)) {
2731		kfree(evt);
2732		return;
2733	}
2734#endif
2735
2736	spin_lock_irqsave(&sdev->list_lock, flags);
2737	list_add_tail(&evt->node, &sdev->event_list);
2738	schedule_work(&sdev->event_work);
2739	spin_unlock_irqrestore(&sdev->list_lock, flags);
2740}
2741EXPORT_SYMBOL_GPL(sdev_evt_send);
2742
2743/**
2744 * 	sdev_evt_alloc - allocate a new scsi event
2745 *	@evt_type: type of event to allocate
2746 *	@gfpflags: GFP flags for allocation
2747 *
2748 *	Allocates and returns a new scsi_event.
2749 */
2750struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2751				  gfp_t gfpflags)
2752{
2753	struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2754	if (!evt)
2755		return NULL;
2756
2757	evt->evt_type = evt_type;
2758	INIT_LIST_HEAD(&evt->node);
2759
2760	/* evt_type-specific initialization, if any */
2761	switch (evt_type) {
2762	case SDEV_EVT_MEDIA_CHANGE:
2763	case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2764	case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2765	case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2766	case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2767	case SDEV_EVT_LUN_CHANGE_REPORTED:
2768	case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2769	case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2770	default:
2771		/* do nothing */
2772		break;
2773	}
2774
2775	return evt;
2776}
2777EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2778
2779/**
2780 * 	sdev_evt_send_simple - send asserted event to uevent thread
2781 *	@sdev: scsi_device event occurred on
2782 *	@evt_type: type of event to send
2783 *	@gfpflags: GFP flags for allocation
2784 *
2785 *	Assert scsi device event asynchronously, given an event type.
2786 */
2787void sdev_evt_send_simple(struct scsi_device *sdev,
2788			  enum scsi_device_event evt_type, gfp_t gfpflags)
2789{
2790	struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2791	if (!evt) {
2792		sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2793			    evt_type);
2794		return;
2795	}
2796
2797	sdev_evt_send(sdev, evt);
2798}
2799EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2800
2801/**
2802 *	scsi_device_quiesce - Block all commands except power management.
2803 *	@sdev:	scsi device to quiesce.
2804 *
2805 *	This works by trying to transition to the SDEV_QUIESCE state
2806 *	(which must be a legal transition).  When the device is in this
2807 *	state, only power management requests will be accepted, all others will
2808 *	be deferred.
2809 *
2810 *	Must be called with user context, may sleep.
2811 *
2812 *	Returns zero if successful or an error if not.
2813 */
2814int
2815scsi_device_quiesce(struct scsi_device *sdev)
2816{
2817	struct request_queue *q = sdev->request_queue;
2818	unsigned int memflags;
2819	int err;
2820
2821	/*
2822	 * It is allowed to call scsi_device_quiesce() multiple times from
2823	 * the same context but concurrent scsi_device_quiesce() calls are
2824	 * not allowed.
2825	 */
2826	WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2827
2828	if (sdev->quiesced_by == current)
2829		return 0;
2830
2831	blk_set_pm_only(q);
2832
2833	memflags = blk_mq_freeze_queue(q);
2834	/*
2835	 * Ensure that the effect of blk_set_pm_only() will be visible
2836	 * for percpu_ref_tryget() callers that occur after the queue
2837	 * unfreeze even if the queue was already frozen before this function
2838	 * was called. See also https://lwn.net/Articles/573497/.
2839	 */
2840	synchronize_rcu();
2841	blk_mq_unfreeze_queue(q, memflags);
2842
2843	mutex_lock(&sdev->state_mutex);
2844	err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2845	if (err == 0)
2846		sdev->quiesced_by = current;
2847	else
2848		blk_clear_pm_only(q);
2849	mutex_unlock(&sdev->state_mutex);
2850
2851	return err;
2852}
2853EXPORT_SYMBOL(scsi_device_quiesce);
2854
2855/**
2856 *	scsi_device_resume - Restart user issued commands to a quiesced device.
2857 *	@sdev:	scsi device to resume.
2858 *
2859 *	Moves the device from quiesced back to running and restarts the
2860 *	queues.
2861 *
2862 *	Must be called with user context, may sleep.
2863 */
2864void scsi_device_resume(struct scsi_device *sdev)
2865{
2866	/* check if the device state was mutated prior to resume, and if
2867	 * so assume the state is being managed elsewhere (for example
2868	 * device deleted during suspend)
2869	 */
2870	mutex_lock(&sdev->state_mutex);
2871	if (sdev->sdev_state == SDEV_QUIESCE)
2872		scsi_device_set_state(sdev, SDEV_RUNNING);
2873	if (sdev->quiesced_by) {
2874		sdev->quiesced_by = NULL;
2875		blk_clear_pm_only(sdev->request_queue);
2876	}
2877	mutex_unlock(&sdev->state_mutex);
2878}
2879EXPORT_SYMBOL(scsi_device_resume);
2880
2881static void
2882device_quiesce_fn(struct scsi_device *sdev, void *data)
2883{
2884	scsi_device_quiesce(sdev);
2885}
2886
2887void
2888scsi_target_quiesce(struct scsi_target *starget)
2889{
2890	starget_for_each_device(starget, NULL, device_quiesce_fn);
2891}
2892EXPORT_SYMBOL(scsi_target_quiesce);
2893
2894static void
2895device_resume_fn(struct scsi_device *sdev, void *data)
2896{
2897	scsi_device_resume(sdev);
2898}
2899
2900void
2901scsi_target_resume(struct scsi_target *starget)
2902{
2903	starget_for_each_device(starget, NULL, device_resume_fn);
2904}
2905EXPORT_SYMBOL(scsi_target_resume);
2906
2907static int __scsi_internal_device_block_nowait(struct scsi_device *sdev)
2908{
2909	if (scsi_device_set_state(sdev, SDEV_BLOCK))
2910		return scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2911
2912	return 0;
2913}
2914
2915void scsi_start_queue(struct scsi_device *sdev)
2916{
2917	if (cmpxchg(&sdev->queue_stopped, 1, 0))
2918		blk_mq_unquiesce_queue(sdev->request_queue);
2919}
2920
2921static void scsi_stop_queue(struct scsi_device *sdev)
2922{
2923	/*
2924	 * The atomic variable of ->queue_stopped covers that
2925	 * blk_mq_quiesce_queue* is balanced with blk_mq_unquiesce_queue.
2926	 *
2927	 * The caller needs to wait until quiesce is done.
2928	 */
2929	if (!cmpxchg(&sdev->queue_stopped, 0, 1))
2930		blk_mq_quiesce_queue_nowait(sdev->request_queue);
2931}
2932
2933/**
2934 * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2935 * @sdev: device to block
2936 *
2937 * Pause SCSI command processing on the specified device. Does not sleep.
2938 *
2939 * Returns zero if successful or a negative error code upon failure.
2940 *
2941 * Notes:
2942 * This routine transitions the device to the SDEV_BLOCK state (which must be
2943 * a legal transition). When the device is in this state, command processing
2944 * is paused until the device leaves the SDEV_BLOCK state. See also
2945 * scsi_internal_device_unblock_nowait().
2946 */
2947int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2948{
2949	int ret = __scsi_internal_device_block_nowait(sdev);
2950
2951	/*
2952	 * The device has transitioned to SDEV_BLOCK.  Stop the
2953	 * block layer from calling the midlayer with this device's
2954	 * request queue.
2955	 */
2956	if (!ret)
2957		scsi_stop_queue(sdev);
2958	return ret;
2959}
2960EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2961
2962/**
2963 * scsi_device_block - try to transition to the SDEV_BLOCK state
2964 * @sdev: device to block
2965 * @data: dummy argument, ignored
2966 *
2967 * Pause SCSI command processing on the specified device. Callers must wait
2968 * until all ongoing scsi_queue_rq() calls have finished after this function
2969 * returns.
2970 *
2971 * Note:
2972 * This routine transitions the device to the SDEV_BLOCK state (which must be
2973 * a legal transition). When the device is in this state, command processing
2974 * is paused until the device leaves the SDEV_BLOCK state. See also
2975 * scsi_internal_device_unblock().
2976 */
2977static void scsi_device_block(struct scsi_device *sdev, void *data)
2978{
2979	int err;
2980	enum scsi_device_state state;
2981
2982	mutex_lock(&sdev->state_mutex);
2983	err = __scsi_internal_device_block_nowait(sdev);
2984	state = sdev->sdev_state;
2985	if (err == 0)
2986		/*
2987		 * scsi_stop_queue() must be called with the state_mutex
2988		 * held. Otherwise a simultaneous scsi_start_queue() call
2989		 * might unquiesce the queue before we quiesce it.
2990		 */
2991		scsi_stop_queue(sdev);
2992
2993	mutex_unlock(&sdev->state_mutex);
2994
2995	WARN_ONCE(err, "%s: failed to block %s in state %d\n",
2996		  __func__, dev_name(&sdev->sdev_gendev), state);
2997}
2998
2999/**
3000 * scsi_internal_device_unblock_nowait - resume a device after a block request
3001 * @sdev:	device to resume
3002 * @new_state:	state to set the device to after unblocking
3003 *
3004 * Restart the device queue for a previously suspended SCSI device. Does not
3005 * sleep.
3006 *
3007 * Returns zero if successful or a negative error code upon failure.
3008 *
3009 * Notes:
3010 * This routine transitions the device to the SDEV_RUNNING state or to one of
3011 * the offline states (which must be a legal transition) allowing the midlayer
3012 * to goose the queue for this device.
3013 */
3014int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
3015					enum scsi_device_state new_state)
3016{
3017	switch (new_state) {
3018	case SDEV_RUNNING:
3019	case SDEV_TRANSPORT_OFFLINE:
3020		break;
3021	default:
3022		return -EINVAL;
3023	}
3024
3025	/*
3026	 * Try to transition the scsi device to SDEV_RUNNING or one of the
3027	 * offlined states and goose the device queue if successful.
3028	 */
3029	switch (sdev->sdev_state) {
3030	case SDEV_BLOCK:
3031	case SDEV_TRANSPORT_OFFLINE:
3032		sdev->sdev_state = new_state;
3033		break;
3034	case SDEV_CREATED_BLOCK:
3035		if (new_state == SDEV_TRANSPORT_OFFLINE ||
3036		    new_state == SDEV_OFFLINE)
3037			sdev->sdev_state = new_state;
3038		else
3039			sdev->sdev_state = SDEV_CREATED;
3040		break;
3041	case SDEV_CANCEL:
3042	case SDEV_OFFLINE:
3043		break;
3044	default:
3045		return -EINVAL;
3046	}
3047	scsi_start_queue(sdev);
3048
3049	return 0;
3050}
3051EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
3052
3053/**
3054 * scsi_internal_device_unblock - resume a device after a block request
3055 * @sdev:	device to resume
3056 * @new_state:	state to set the device to after unblocking
3057 *
3058 * Restart the device queue for a previously suspended SCSI device. May sleep.
3059 *
3060 * Returns zero if successful or a negative error code upon failure.
3061 *
3062 * Notes:
3063 * This routine transitions the device to the SDEV_RUNNING state or to one of
3064 * the offline states (which must be a legal transition) allowing the midlayer
3065 * to goose the queue for this device.
3066 */
3067static int scsi_internal_device_unblock(struct scsi_device *sdev,
3068					enum scsi_device_state new_state)
3069{
3070	int ret;
3071
3072	mutex_lock(&sdev->state_mutex);
3073	ret = scsi_internal_device_unblock_nowait(sdev, new_state);
3074	mutex_unlock(&sdev->state_mutex);
3075
3076	return ret;
3077}
3078
3079static int
3080target_block(struct device *dev, void *data)
3081{
3082	if (scsi_is_target_device(dev))
3083		starget_for_each_device(to_scsi_target(dev), NULL,
3084					scsi_device_block);
3085	return 0;
3086}
3087
3088/**
3089 * scsi_block_targets - transition all SCSI child devices to SDEV_BLOCK state
3090 * @dev: a parent device of one or more scsi_target devices
3091 * @shost: the Scsi_Host to which this device belongs
3092 *
3093 * Iterate over all children of @dev, which should be scsi_target devices,
3094 * and switch all subordinate scsi devices to SDEV_BLOCK state. Wait for
3095 * ongoing scsi_queue_rq() calls to finish. May sleep.
3096 *
3097 * Note:
3098 * @dev must not itself be a scsi_target device.
3099 */
3100void
3101scsi_block_targets(struct Scsi_Host *shost, struct device *dev)
3102{
3103	WARN_ON_ONCE(scsi_is_target_device(dev));
3104	device_for_each_child(dev, NULL, target_block);
3105	blk_mq_wait_quiesce_done(&shost->tag_set);
3106}
3107EXPORT_SYMBOL_GPL(scsi_block_targets);
3108
3109static void
3110device_unblock(struct scsi_device *sdev, void *data)
3111{
3112	scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
3113}
3114
3115static int
3116target_unblock(struct device *dev, void *data)
3117{
3118	if (scsi_is_target_device(dev))
3119		starget_for_each_device(to_scsi_target(dev), data,
3120					device_unblock);
3121	return 0;
3122}
3123
3124void
3125scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
3126{
3127	if (scsi_is_target_device(dev))
3128		starget_for_each_device(to_scsi_target(dev), &new_state,
3129					device_unblock);
3130	else
3131		device_for_each_child(dev, &new_state, target_unblock);
3132}
3133EXPORT_SYMBOL_GPL(scsi_target_unblock);
3134
3135/**
3136 * scsi_host_block - Try to transition all logical units to the SDEV_BLOCK state
3137 * @shost: device to block
3138 *
3139 * Pause SCSI command processing for all logical units associated with the SCSI
3140 * host and wait until pending scsi_queue_rq() calls have finished.
3141 *
3142 * Returns zero if successful or a negative error code upon failure.
3143 */
3144int
3145scsi_host_block(struct Scsi_Host *shost)
3146{
3147	struct scsi_device *sdev;
3148	int ret;
3149
3150	/*
3151	 * Call scsi_internal_device_block_nowait so we can avoid
3152	 * calling synchronize_rcu() for each LUN.
3153	 */
3154	shost_for_each_device(sdev, shost) {
3155		mutex_lock(&sdev->state_mutex);
3156		ret = scsi_internal_device_block_nowait(sdev);
3157		mutex_unlock(&sdev->state_mutex);
3158		if (ret) {
3159			scsi_device_put(sdev);
3160			return ret;
3161		}
3162	}
3163
3164	/* Wait for ongoing scsi_queue_rq() calls to finish. */
3165	blk_mq_wait_quiesce_done(&shost->tag_set);
3166
3167	return 0;
3168}
3169EXPORT_SYMBOL_GPL(scsi_host_block);
3170
3171int
3172scsi_host_unblock(struct Scsi_Host *shost, int new_state)
3173{
3174	struct scsi_device *sdev;
3175	int ret = 0;
3176
3177	shost_for_each_device(sdev, shost) {
3178		ret = scsi_internal_device_unblock(sdev, new_state);
3179		if (ret) {
3180			scsi_device_put(sdev);
3181			break;
3182		}
3183	}
3184	return ret;
3185}
3186EXPORT_SYMBOL_GPL(scsi_host_unblock);
3187
3188/**
3189 * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
3190 * @sgl:	scatter-gather list
3191 * @sg_count:	number of segments in sg
3192 * @offset:	offset in bytes into sg, on return offset into the mapped area
3193 * @len:	bytes to map, on return number of bytes mapped
3194 *
3195 * Returns virtual address of the start of the mapped page
3196 */
3197void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
3198			  size_t *offset, size_t *len)
3199{
3200	int i;
3201	size_t sg_len = 0, len_complete = 0;
3202	struct scatterlist *sg;
3203	struct page *page;
3204
3205	WARN_ON(!irqs_disabled());
3206
3207	for_each_sg(sgl, sg, sg_count, i) {
3208		len_complete = sg_len; /* Complete sg-entries */
3209		sg_len += sg->length;
3210		if (sg_len > *offset)
3211			break;
3212	}
3213
3214	if (unlikely(i == sg_count)) {
3215		printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
3216			"elements %d\n",
3217		       __func__, sg_len, *offset, sg_count);
3218		WARN_ON(1);
3219		return NULL;
3220	}
3221
3222	/* Offset starting from the beginning of first page in this sg-entry */
3223	*offset = *offset - len_complete + sg->offset;
3224
3225	page = sg_page(sg) + (*offset >> PAGE_SHIFT);
3226	*offset &= ~PAGE_MASK;
3227
3228	/* Bytes in this sg-entry from *offset to the end of the page */
3229	sg_len = PAGE_SIZE - *offset;
3230	if (*len > sg_len)
3231		*len = sg_len;
3232
3233	return kmap_atomic(page);
3234}
3235EXPORT_SYMBOL(scsi_kmap_atomic_sg);
3236
3237/**
3238 * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
3239 * @virt:	virtual address to be unmapped
3240 */
3241void scsi_kunmap_atomic_sg(void *virt)
3242{
3243	kunmap_atomic(virt);
3244}
3245EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
3246
3247void sdev_disable_disk_events(struct scsi_device *sdev)
3248{
3249	atomic_inc(&sdev->disk_events_disable_depth);
3250}
3251EXPORT_SYMBOL(sdev_disable_disk_events);
3252
3253void sdev_enable_disk_events(struct scsi_device *sdev)
3254{
3255	if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
3256		return;
3257	atomic_dec(&sdev->disk_events_disable_depth);
3258}
3259EXPORT_SYMBOL(sdev_enable_disk_events);
3260
3261static unsigned char designator_prio(const unsigned char *d)
3262{
3263	if (d[1] & 0x30)
3264		/* not associated with LUN */
3265		return 0;
3266
3267	if (d[3] == 0)
3268		/* invalid length */
3269		return 0;
3270
3271	/*
3272	 * Order of preference for lun descriptor:
3273	 * - SCSI name string
3274	 * - NAA IEEE Registered Extended
3275	 * - EUI-64 based 16-byte
3276	 * - EUI-64 based 12-byte
3277	 * - NAA IEEE Registered
3278	 * - NAA IEEE Extended
3279	 * - EUI-64 based 8-byte
3280	 * - SCSI name string (truncated)
3281	 * - T10 Vendor ID
3282	 * as longer descriptors reduce the likelyhood
3283	 * of identification clashes.
3284	 */
3285
3286	switch (d[1] & 0xf) {
3287	case 8:
3288		/* SCSI name string, variable-length UTF-8 */
3289		return 9;
3290	case 3:
3291		switch (d[4] >> 4) {
3292		case 6:
3293			/* NAA registered extended */
3294			return 8;
3295		case 5:
3296			/* NAA registered */
3297			return 5;
3298		case 4:
3299			/* NAA extended */
3300			return 4;
3301		case 3:
3302			/* NAA locally assigned */
3303			return 1;
3304		default:
3305			break;
3306		}
3307		break;
3308	case 2:
3309		switch (d[3]) {
3310		case 16:
3311			/* EUI64-based, 16 byte */
3312			return 7;
3313		case 12:
3314			/* EUI64-based, 12 byte */
3315			return 6;
3316		case 8:
3317			/* EUI64-based, 8 byte */
3318			return 3;
3319		default:
3320			break;
3321		}
3322		break;
3323	case 1:
3324		/* T10 vendor ID */
3325		return 1;
3326	default:
3327		break;
3328	}
3329
3330	return 0;
3331}
3332
3333/**
3334 * scsi_vpd_lun_id - return a unique device identification
3335 * @sdev: SCSI device
3336 * @id:   buffer for the identification
3337 * @id_len:  length of the buffer
3338 *
3339 * Copies a unique device identification into @id based
3340 * on the information in the VPD page 0x83 of the device.
3341 * The string will be formatted as a SCSI name string.
3342 *
3343 * Returns the length of the identification or error on failure.
3344 * If the identifier is longer than the supplied buffer the actual
3345 * identifier length is returned and the buffer is not zero-padded.
3346 */
3347int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
3348{
3349	u8 cur_id_prio = 0;
3350	u8 cur_id_size = 0;
3351	const unsigned char *d, *cur_id_str;
3352	const struct scsi_vpd *vpd_pg83;
3353	int id_size = -EINVAL;
3354
3355	rcu_read_lock();
3356	vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3357	if (!vpd_pg83) {
3358		rcu_read_unlock();
3359		return -ENXIO;
3360	}
3361
3362	/* The id string must be at least 20 bytes + terminating NULL byte */
3363	if (id_len < 21) {
3364		rcu_read_unlock();
3365		return -EINVAL;
3366	}
3367
3368	memset(id, 0, id_len);
3369	for (d = vpd_pg83->data + 4;
3370	     d < vpd_pg83->data + vpd_pg83->len;
3371	     d += d[3] + 4) {
3372		u8 prio = designator_prio(d);
3373
3374		if (prio == 0 || cur_id_prio > prio)
3375			continue;
3376
3377		switch (d[1] & 0xf) {
3378		case 0x1:
3379			/* T10 Vendor ID */
3380			if (cur_id_size > d[3])
3381				break;
3382			cur_id_prio = prio;
3383			cur_id_size = d[3];
3384			if (cur_id_size + 4 > id_len)
3385				cur_id_size = id_len - 4;
3386			cur_id_str = d + 4;
3387			id_size = snprintf(id, id_len, "t10.%*pE",
3388					   cur_id_size, cur_id_str);
3389			break;
3390		case 0x2:
3391			/* EUI-64 */
3392			cur_id_prio = prio;
3393			cur_id_size = d[3];
3394			cur_id_str = d + 4;
3395			switch (cur_id_size) {
3396			case 8:
3397				id_size = snprintf(id, id_len,
3398						   "eui.%8phN",
3399						   cur_id_str);
3400				break;
3401			case 12:
3402				id_size = snprintf(id, id_len,
3403						   "eui.%12phN",
3404						   cur_id_str);
3405				break;
3406			case 16:
3407				id_size = snprintf(id, id_len,
3408						   "eui.%16phN",
3409						   cur_id_str);
3410				break;
3411			default:
3412				break;
3413			}
3414			break;
3415		case 0x3:
3416			/* NAA */
3417			cur_id_prio = prio;
3418			cur_id_size = d[3];
3419			cur_id_str = d + 4;
3420			switch (cur_id_size) {
3421			case 8:
3422				id_size = snprintf(id, id_len,
3423						   "naa.%8phN",
3424						   cur_id_str);
3425				break;
3426			case 16:
3427				id_size = snprintf(id, id_len,
3428						   "naa.%16phN",
3429						   cur_id_str);
3430				break;
3431			default:
3432				break;
3433			}
3434			break;
3435		case 0x8:
3436			/* SCSI name string */
3437			if (cur_id_size > d[3])
3438				break;
3439			/* Prefer others for truncated descriptor */
3440			if (d[3] > id_len) {
3441				prio = 2;
3442				if (cur_id_prio > prio)
3443					break;
3444			}
3445			cur_id_prio = prio;
3446			cur_id_size = id_size = d[3];
3447			cur_id_str = d + 4;
3448			if (cur_id_size >= id_len)
3449				cur_id_size = id_len - 1;
3450			memcpy(id, cur_id_str, cur_id_size);
3451			break;
3452		default:
3453			break;
3454		}
3455	}
3456	rcu_read_unlock();
3457
3458	return id_size;
3459}
3460EXPORT_SYMBOL(scsi_vpd_lun_id);
3461
3462/**
3463 * scsi_vpd_tpg_id - return a target port group identifier
3464 * @sdev: SCSI device
3465 * @rel_id: pointer to return relative target port in if not %NULL
3466 *
3467 * Returns the Target Port Group identifier from the information
3468 * from VPD page 0x83 of the device.
3469 * Optionally sets @rel_id to the relative target port on success.
3470 *
3471 * Return: the identifier or error on failure.
3472 */
3473int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3474{
3475	const unsigned char *d;
3476	const struct scsi_vpd *vpd_pg83;
3477	int group_id = -EAGAIN, rel_port = -1;
3478
3479	rcu_read_lock();
3480	vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3481	if (!vpd_pg83) {
3482		rcu_read_unlock();
3483		return -ENXIO;
3484	}
3485
3486	d = vpd_pg83->data + 4;
3487	while (d < vpd_pg83->data + vpd_pg83->len) {
3488		switch (d[1] & 0xf) {
3489		case 0x4:
3490			/* Relative target port */
3491			rel_port = get_unaligned_be16(&d[6]);
3492			break;
3493		case 0x5:
3494			/* Target port group */
3495			group_id = get_unaligned_be16(&d[6]);
3496			break;
3497		default:
3498			break;
3499		}
3500		d += d[3] + 4;
3501	}
3502	rcu_read_unlock();
3503
3504	if (group_id >= 0 && rel_id && rel_port != -1)
3505		*rel_id = rel_port;
3506
3507	return group_id;
3508}
3509EXPORT_SYMBOL(scsi_vpd_tpg_id);
3510
3511/**
3512 * scsi_build_sense - build sense data for a command
3513 * @scmd:	scsi command for which the sense should be formatted
3514 * @desc:	Sense format (non-zero == descriptor format,
3515 *              0 == fixed format)
3516 * @key:	Sense key
3517 * @asc:	Additional sense code
3518 * @ascq:	Additional sense code qualifier
3519 *
3520 **/
3521void scsi_build_sense(struct scsi_cmnd *scmd, int desc, u8 key, u8 asc, u8 ascq)
3522{
3523	scsi_build_sense_buffer(desc, scmd->sense_buffer, key, asc, ascq);
3524	scmd->result = SAM_STAT_CHECK_CONDITION;
3525}
3526EXPORT_SYMBOL_GPL(scsi_build_sense);
3527
3528#ifdef CONFIG_SCSI_LIB_KUNIT_TEST
3529#include "scsi_lib_test.c"
3530#endif
~/ Read-only file preview.