Loading...
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 <asm/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_queue() */
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_execute_cmd - insert request and wait for the result
189 * @sdev: scsi_device
190 * @cmd: scsi command
191 * @opf: block layer request cmd_flags
192 * @buffer: data buffer
193 * @bufflen: len of buffer
194 * @timeout: request timeout in HZ
195 * @retries: number of times to retry request
196 * @args: Optional args. See struct definition for field descriptions
197 *
198 * Returns the scsi_cmnd result field if a command was executed, or a negative
199 * Linux error code if we didn't get that far.
200 */
201int scsi_execute_cmd(struct scsi_device *sdev, const unsigned char *cmd,
202 blk_opf_t opf, void *buffer, unsigned int bufflen,
203 int timeout, int retries,
204 const struct scsi_exec_args *args)
205{
206 static const struct scsi_exec_args default_args;
207 struct request *req;
208 struct scsi_cmnd *scmd;
209 int ret;
210
211 if (!args)
212 args = &default_args;
213 else if (WARN_ON_ONCE(args->sense &&
214 args->sense_len != SCSI_SENSE_BUFFERSIZE))
215 return -EINVAL;
216
217 req = scsi_alloc_request(sdev->request_queue, opf, args->req_flags);
218 if (IS_ERR(req))
219 return PTR_ERR(req);
220
221 if (bufflen) {
222 ret = blk_rq_map_kern(sdev->request_queue, req,
223 buffer, bufflen, GFP_NOIO);
224 if (ret)
225 goto out;
226 }
227 scmd = blk_mq_rq_to_pdu(req);
228 scmd->cmd_len = COMMAND_SIZE(cmd[0]);
229 memcpy(scmd->cmnd, cmd, scmd->cmd_len);
230 scmd->allowed = retries;
231 scmd->flags |= args->scmd_flags;
232 req->timeout = timeout;
233 req->rq_flags |= RQF_QUIET;
234
235 /*
236 * head injection *required* here otherwise quiesce won't work
237 */
238 blk_execute_rq(req, true);
239
240 /*
241 * Some devices (USB mass-storage in particular) may transfer
242 * garbage data together with a residue indicating that the data
243 * is invalid. Prevent the garbage from being misinterpreted
244 * and prevent security leaks by zeroing out the excess data.
245 */
246 if (unlikely(scmd->resid_len > 0 && scmd->resid_len <= bufflen))
247 memset(buffer + bufflen - scmd->resid_len, 0, scmd->resid_len);
248
249 if (args->resid)
250 *args->resid = scmd->resid_len;
251 if (args->sense)
252 memcpy(args->sense, scmd->sense_buffer, SCSI_SENSE_BUFFERSIZE);
253 if (args->sshdr)
254 scsi_normalize_sense(scmd->sense_buffer, scmd->sense_len,
255 args->sshdr);
256
257 ret = scmd->result;
258 out:
259 blk_mq_free_request(req);
260
261 return ret;
262}
263EXPORT_SYMBOL(scsi_execute_cmd);
264
265/*
266 * Wake up the error handler if necessary. Avoid as follows that the error
267 * handler is not woken up if host in-flight requests number ==
268 * shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
269 * with an RCU read lock in this function to ensure that this function in
270 * its entirety either finishes before scsi_eh_scmd_add() increases the
271 * host_failed counter or that it notices the shost state change made by
272 * scsi_eh_scmd_add().
273 */
274static void scsi_dec_host_busy(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
275{
276 unsigned long flags;
277
278 rcu_read_lock();
279 __clear_bit(SCMD_STATE_INFLIGHT, &cmd->state);
280 if (unlikely(scsi_host_in_recovery(shost))) {
281 unsigned int busy = scsi_host_busy(shost);
282
283 spin_lock_irqsave(shost->host_lock, flags);
284 if (shost->host_failed || shost->host_eh_scheduled)
285 scsi_eh_wakeup(shost, busy);
286 spin_unlock_irqrestore(shost->host_lock, flags);
287 }
288 rcu_read_unlock();
289}
290
291void scsi_device_unbusy(struct scsi_device *sdev, struct scsi_cmnd *cmd)
292{
293 struct Scsi_Host *shost = sdev->host;
294 struct scsi_target *starget = scsi_target(sdev);
295
296 scsi_dec_host_busy(shost, cmd);
297
298 if (starget->can_queue > 0)
299 atomic_dec(&starget->target_busy);
300
301 sbitmap_put(&sdev->budget_map, cmd->budget_token);
302 cmd->budget_token = -1;
303}
304
305/*
306 * Kick the queue of SCSI device @sdev if @sdev != current_sdev. Called with
307 * interrupts disabled.
308 */
309static void scsi_kick_sdev_queue(struct scsi_device *sdev, void *data)
310{
311 struct scsi_device *current_sdev = data;
312
313 if (sdev != current_sdev)
314 blk_mq_run_hw_queues(sdev->request_queue, true);
315}
316
317/*
318 * Called for single_lun devices on IO completion. Clear starget_sdev_user,
319 * and call blk_run_queue for all the scsi_devices on the target -
320 * including current_sdev first.
321 *
322 * Called with *no* scsi locks held.
323 */
324static void scsi_single_lun_run(struct scsi_device *current_sdev)
325{
326 struct Scsi_Host *shost = current_sdev->host;
327 struct scsi_target *starget = scsi_target(current_sdev);
328 unsigned long flags;
329
330 spin_lock_irqsave(shost->host_lock, flags);
331 starget->starget_sdev_user = NULL;
332 spin_unlock_irqrestore(shost->host_lock, flags);
333
334 /*
335 * Call blk_run_queue for all LUNs on the target, starting with
336 * current_sdev. We race with others (to set starget_sdev_user),
337 * but in most cases, we will be first. Ideally, each LU on the
338 * target would get some limited time or requests on the target.
339 */
340 blk_mq_run_hw_queues(current_sdev->request_queue,
341 shost->queuecommand_may_block);
342
343 spin_lock_irqsave(shost->host_lock, flags);
344 if (!starget->starget_sdev_user)
345 __starget_for_each_device(starget, current_sdev,
346 scsi_kick_sdev_queue);
347 spin_unlock_irqrestore(shost->host_lock, flags);
348}
349
350static inline bool scsi_device_is_busy(struct scsi_device *sdev)
351{
352 if (scsi_device_busy(sdev) >= sdev->queue_depth)
353 return true;
354 if (atomic_read(&sdev->device_blocked) > 0)
355 return true;
356 return false;
357}
358
359static inline bool scsi_target_is_busy(struct scsi_target *starget)
360{
361 if (starget->can_queue > 0) {
362 if (atomic_read(&starget->target_busy) >= starget->can_queue)
363 return true;
364 if (atomic_read(&starget->target_blocked) > 0)
365 return true;
366 }
367 return false;
368}
369
370static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
371{
372 if (atomic_read(&shost->host_blocked) > 0)
373 return true;
374 if (shost->host_self_blocked)
375 return true;
376 return false;
377}
378
379static void scsi_starved_list_run(struct Scsi_Host *shost)
380{
381 LIST_HEAD(starved_list);
382 struct scsi_device *sdev;
383 unsigned long flags;
384
385 spin_lock_irqsave(shost->host_lock, flags);
386 list_splice_init(&shost->starved_list, &starved_list);
387
388 while (!list_empty(&starved_list)) {
389 struct request_queue *slq;
390
391 /*
392 * As long as shost is accepting commands and we have
393 * starved queues, call blk_run_queue. scsi_request_fn
394 * drops the queue_lock and can add us back to the
395 * starved_list.
396 *
397 * host_lock protects the starved_list and starved_entry.
398 * scsi_request_fn must get the host_lock before checking
399 * or modifying starved_list or starved_entry.
400 */
401 if (scsi_host_is_busy(shost))
402 break;
403
404 sdev = list_entry(starved_list.next,
405 struct scsi_device, starved_entry);
406 list_del_init(&sdev->starved_entry);
407 if (scsi_target_is_busy(scsi_target(sdev))) {
408 list_move_tail(&sdev->starved_entry,
409 &shost->starved_list);
410 continue;
411 }
412
413 /*
414 * Once we drop the host lock, a racing scsi_remove_device()
415 * call may remove the sdev from the starved list and destroy
416 * it and the queue. Mitigate by taking a reference to the
417 * queue and never touching the sdev again after we drop the
418 * host lock. Note: if __scsi_remove_device() invokes
419 * blk_mq_destroy_queue() before the queue is run from this
420 * function then blk_run_queue() will return immediately since
421 * blk_mq_destroy_queue() marks the queue with QUEUE_FLAG_DYING.
422 */
423 slq = sdev->request_queue;
424 if (!blk_get_queue(slq))
425 continue;
426 spin_unlock_irqrestore(shost->host_lock, flags);
427
428 blk_mq_run_hw_queues(slq, false);
429 blk_put_queue(slq);
430
431 spin_lock_irqsave(shost->host_lock, flags);
432 }
433 /* put any unprocessed entries back */
434 list_splice(&starved_list, &shost->starved_list);
435 spin_unlock_irqrestore(shost->host_lock, flags);
436}
437
438/**
439 * scsi_run_queue - Select a proper request queue to serve next.
440 * @q: last request's queue
441 *
442 * The previous command was completely finished, start a new one if possible.
443 */
444static void scsi_run_queue(struct request_queue *q)
445{
446 struct scsi_device *sdev = q->queuedata;
447
448 if (scsi_target(sdev)->single_lun)
449 scsi_single_lun_run(sdev);
450 if (!list_empty(&sdev->host->starved_list))
451 scsi_starved_list_run(sdev->host);
452
453 /* Note: blk_mq_kick_requeue_list() runs the queue asynchronously. */
454 blk_mq_kick_requeue_list(q);
455}
456
457void scsi_requeue_run_queue(struct work_struct *work)
458{
459 struct scsi_device *sdev;
460 struct request_queue *q;
461
462 sdev = container_of(work, struct scsi_device, requeue_work);
463 q = sdev->request_queue;
464 scsi_run_queue(q);
465}
466
467void scsi_run_host_queues(struct Scsi_Host *shost)
468{
469 struct scsi_device *sdev;
470
471 shost_for_each_device(sdev, shost)
472 scsi_run_queue(sdev->request_queue);
473}
474
475static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
476{
477 if (!blk_rq_is_passthrough(scsi_cmd_to_rq(cmd))) {
478 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
479
480 if (drv->uninit_command)
481 drv->uninit_command(cmd);
482 }
483}
484
485void scsi_free_sgtables(struct scsi_cmnd *cmd)
486{
487 if (cmd->sdb.table.nents)
488 sg_free_table_chained(&cmd->sdb.table,
489 SCSI_INLINE_SG_CNT);
490 if (scsi_prot_sg_count(cmd))
491 sg_free_table_chained(&cmd->prot_sdb->table,
492 SCSI_INLINE_PROT_SG_CNT);
493}
494EXPORT_SYMBOL_GPL(scsi_free_sgtables);
495
496static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
497{
498 scsi_free_sgtables(cmd);
499 scsi_uninit_cmd(cmd);
500}
501
502static void scsi_run_queue_async(struct scsi_device *sdev)
503{
504 if (scsi_host_in_recovery(sdev->host))
505 return;
506
507 if (scsi_target(sdev)->single_lun ||
508 !list_empty(&sdev->host->starved_list)) {
509 kblockd_schedule_work(&sdev->requeue_work);
510 } else {
511 /*
512 * smp_mb() present in sbitmap_queue_clear() or implied in
513 * .end_io is for ordering writing .device_busy in
514 * scsi_device_unbusy() and reading sdev->restarts.
515 */
516 int old = atomic_read(&sdev->restarts);
517
518 /*
519 * ->restarts has to be kept as non-zero if new budget
520 * contention occurs.
521 *
522 * No need to run queue when either another re-run
523 * queue wins in updating ->restarts or a new budget
524 * contention occurs.
525 */
526 if (old && atomic_cmpxchg(&sdev->restarts, old, 0) == old)
527 blk_mq_run_hw_queues(sdev->request_queue, true);
528 }
529}
530
531/* Returns false when no more bytes to process, true if there are more */
532static bool scsi_end_request(struct request *req, blk_status_t error,
533 unsigned int bytes)
534{
535 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
536 struct scsi_device *sdev = cmd->device;
537 struct request_queue *q = sdev->request_queue;
538
539 if (blk_update_request(req, error, bytes))
540 return true;
541
542 // XXX:
543 if (blk_queue_add_random(q))
544 add_disk_randomness(req->q->disk);
545
546 if (!blk_rq_is_passthrough(req)) {
547 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
548 cmd->flags &= ~SCMD_INITIALIZED;
549 }
550
551 /*
552 * Calling rcu_barrier() is not necessary here because the
553 * SCSI error handler guarantees that the function called by
554 * call_rcu() has been called before scsi_end_request() is
555 * called.
556 */
557 destroy_rcu_head(&cmd->rcu);
558
559 /*
560 * In the MQ case the command gets freed by __blk_mq_end_request,
561 * so we have to do all cleanup that depends on it earlier.
562 *
563 * We also can't kick the queues from irq context, so we
564 * will have to defer it to a workqueue.
565 */
566 scsi_mq_uninit_cmd(cmd);
567
568 /*
569 * queue is still alive, so grab the ref for preventing it
570 * from being cleaned up during running queue.
571 */
572 percpu_ref_get(&q->q_usage_counter);
573
574 __blk_mq_end_request(req, error);
575
576 scsi_run_queue_async(sdev);
577
578 percpu_ref_put(&q->q_usage_counter);
579 return false;
580}
581
582/**
583 * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
584 * @result: scsi error code
585 *
586 * Translate a SCSI result code into a blk_status_t value.
587 */
588static blk_status_t scsi_result_to_blk_status(int result)
589{
590 /*
591 * Check the scsi-ml byte first in case we converted a host or status
592 * byte.
593 */
594 switch (scsi_ml_byte(result)) {
595 case SCSIML_STAT_OK:
596 break;
597 case SCSIML_STAT_RESV_CONFLICT:
598 return BLK_STS_RESV_CONFLICT;
599 case SCSIML_STAT_NOSPC:
600 return BLK_STS_NOSPC;
601 case SCSIML_STAT_MED_ERROR:
602 return BLK_STS_MEDIUM;
603 case SCSIML_STAT_TGT_FAILURE:
604 return BLK_STS_TARGET;
605 case SCSIML_STAT_DL_TIMEOUT:
606 return BLK_STS_DURATION_LIMIT;
607 }
608
609 switch (host_byte(result)) {
610 case DID_OK:
611 if (scsi_status_is_good(result))
612 return BLK_STS_OK;
613 return BLK_STS_IOERR;
614 case DID_TRANSPORT_FAILFAST:
615 case DID_TRANSPORT_MARGINAL:
616 return BLK_STS_TRANSPORT;
617 default:
618 return BLK_STS_IOERR;
619 }
620}
621
622/**
623 * scsi_rq_err_bytes - determine number of bytes till the next failure boundary
624 * @rq: request to examine
625 *
626 * Description:
627 * A request could be merge of IOs which require different failure
628 * handling. This function determines the number of bytes which
629 * can be failed from the beginning of the request without
630 * crossing into area which need to be retried further.
631 *
632 * Return:
633 * The number of bytes to fail.
634 */
635static unsigned int scsi_rq_err_bytes(const struct request *rq)
636{
637 blk_opf_t ff = rq->cmd_flags & REQ_FAILFAST_MASK;
638 unsigned int bytes = 0;
639 struct bio *bio;
640
641 if (!(rq->rq_flags & RQF_MIXED_MERGE))
642 return blk_rq_bytes(rq);
643
644 /*
645 * Currently the only 'mixing' which can happen is between
646 * different fastfail types. We can safely fail portions
647 * which have all the failfast bits that the first one has -
648 * the ones which are at least as eager to fail as the first
649 * one.
650 */
651 for (bio = rq->bio; bio; bio = bio->bi_next) {
652 if ((bio->bi_opf & ff) != ff)
653 break;
654 bytes += bio->bi_iter.bi_size;
655 }
656
657 /* this could lead to infinite loop */
658 BUG_ON(blk_rq_bytes(rq) && !bytes);
659 return bytes;
660}
661
662static bool scsi_cmd_runtime_exceeced(struct scsi_cmnd *cmd)
663{
664 struct request *req = scsi_cmd_to_rq(cmd);
665 unsigned long wait_for;
666
667 if (cmd->allowed == SCSI_CMD_RETRIES_NO_LIMIT)
668 return false;
669
670 wait_for = (cmd->allowed + 1) * req->timeout;
671 if (time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
672 scmd_printk(KERN_ERR, cmd, "timing out command, waited %lus\n",
673 wait_for/HZ);
674 return true;
675 }
676 return false;
677}
678
679/*
680 * When ALUA transition state is returned, reprep the cmd to
681 * use the ALUA handler's transition timeout. Delay the reprep
682 * 1 sec to avoid aggressive retries of the target in that
683 * state.
684 */
685#define ALUA_TRANSITION_REPREP_DELAY 1000
686
687/* Helper for scsi_io_completion() when special action required. */
688static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
689{
690 struct request *req = scsi_cmd_to_rq(cmd);
691 int level = 0;
692 enum {ACTION_FAIL, ACTION_REPREP, ACTION_DELAYED_REPREP,
693 ACTION_RETRY, ACTION_DELAYED_RETRY} action;
694 struct scsi_sense_hdr sshdr;
695 bool sense_valid;
696 bool sense_current = true; /* false implies "deferred sense" */
697 blk_status_t blk_stat;
698
699 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
700 if (sense_valid)
701 sense_current = !scsi_sense_is_deferred(&sshdr);
702
703 blk_stat = scsi_result_to_blk_status(result);
704
705 if (host_byte(result) == DID_RESET) {
706 /* Third party bus reset or reset for error recovery
707 * reasons. Just retry the command and see what
708 * happens.
709 */
710 action = ACTION_RETRY;
711 } else if (sense_valid && sense_current) {
712 switch (sshdr.sense_key) {
713 case UNIT_ATTENTION:
714 if (cmd->device->removable) {
715 /* Detected disc change. Set a bit
716 * and quietly refuse further access.
717 */
718 cmd->device->changed = 1;
719 action = ACTION_FAIL;
720 } else {
721 /* Must have been a power glitch, or a
722 * bus reset. Could not have been a
723 * media change, so we just retry the
724 * command and see what happens.
725 */
726 action = ACTION_RETRY;
727 }
728 break;
729 case ILLEGAL_REQUEST:
730 /* If we had an ILLEGAL REQUEST returned, then
731 * we may have performed an unsupported
732 * command. The only thing this should be
733 * would be a ten byte read where only a six
734 * byte read was supported. Also, on a system
735 * where READ CAPACITY failed, we may have
736 * read past the end of the disk.
737 */
738 if ((cmd->device->use_10_for_rw &&
739 sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
740 (cmd->cmnd[0] == READ_10 ||
741 cmd->cmnd[0] == WRITE_10)) {
742 /* This will issue a new 6-byte command. */
743 cmd->device->use_10_for_rw = 0;
744 action = ACTION_REPREP;
745 } else if (sshdr.asc == 0x10) /* DIX */ {
746 action = ACTION_FAIL;
747 blk_stat = BLK_STS_PROTECTION;
748 /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
749 } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
750 action = ACTION_FAIL;
751 blk_stat = BLK_STS_TARGET;
752 } else
753 action = ACTION_FAIL;
754 break;
755 case ABORTED_COMMAND:
756 action = ACTION_FAIL;
757 if (sshdr.asc == 0x10) /* DIF */
758 blk_stat = BLK_STS_PROTECTION;
759 break;
760 case NOT_READY:
761 /* If the device is in the process of becoming
762 * ready, or has a temporary blockage, retry.
763 */
764 if (sshdr.asc == 0x04) {
765 switch (sshdr.ascq) {
766 case 0x01: /* becoming ready */
767 case 0x04: /* format in progress */
768 case 0x05: /* rebuild in progress */
769 case 0x06: /* recalculation in progress */
770 case 0x07: /* operation in progress */
771 case 0x08: /* Long write in progress */
772 case 0x09: /* self test in progress */
773 case 0x11: /* notify (enable spinup) required */
774 case 0x14: /* space allocation in progress */
775 case 0x1a: /* start stop unit in progress */
776 case 0x1b: /* sanitize in progress */
777 case 0x1d: /* configuration in progress */
778 case 0x24: /* depopulation in progress */
779 case 0x25: /* depopulation restore in progress */
780 action = ACTION_DELAYED_RETRY;
781 break;
782 case 0x0a: /* ALUA state transition */
783 action = ACTION_DELAYED_REPREP;
784 break;
785 default:
786 action = ACTION_FAIL;
787 break;
788 }
789 } else
790 action = ACTION_FAIL;
791 break;
792 case VOLUME_OVERFLOW:
793 /* See SSC3rXX or current. */
794 action = ACTION_FAIL;
795 break;
796 case DATA_PROTECT:
797 action = ACTION_FAIL;
798 if ((sshdr.asc == 0x0C && sshdr.ascq == 0x12) ||
799 (sshdr.asc == 0x55 &&
800 (sshdr.ascq == 0x0E || sshdr.ascq == 0x0F))) {
801 /* Insufficient zone resources */
802 blk_stat = BLK_STS_ZONE_OPEN_RESOURCE;
803 }
804 break;
805 case COMPLETED:
806 fallthrough;
807 default:
808 action = ACTION_FAIL;
809 break;
810 }
811 } else
812 action = ACTION_FAIL;
813
814 if (action != ACTION_FAIL && scsi_cmd_runtime_exceeced(cmd))
815 action = ACTION_FAIL;
816
817 switch (action) {
818 case ACTION_FAIL:
819 /* Give up and fail the remainder of the request */
820 if (!(req->rq_flags & RQF_QUIET)) {
821 static DEFINE_RATELIMIT_STATE(_rs,
822 DEFAULT_RATELIMIT_INTERVAL,
823 DEFAULT_RATELIMIT_BURST);
824
825 if (unlikely(scsi_logging_level))
826 level =
827 SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
828 SCSI_LOG_MLCOMPLETE_BITS);
829
830 /*
831 * if logging is enabled the failure will be printed
832 * in scsi_log_completion(), so avoid duplicate messages
833 */
834 if (!level && __ratelimit(&_rs)) {
835 scsi_print_result(cmd, NULL, FAILED);
836 if (sense_valid)
837 scsi_print_sense(cmd);
838 scsi_print_command(cmd);
839 }
840 }
841 if (!scsi_end_request(req, blk_stat, scsi_rq_err_bytes(req)))
842 return;
843 fallthrough;
844 case ACTION_REPREP:
845 scsi_mq_requeue_cmd(cmd, 0);
846 break;
847 case ACTION_DELAYED_REPREP:
848 scsi_mq_requeue_cmd(cmd, ALUA_TRANSITION_REPREP_DELAY);
849 break;
850 case ACTION_RETRY:
851 /* Retry the same command immediately */
852 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
853 break;
854 case ACTION_DELAYED_RETRY:
855 /* Retry the same command after a delay */
856 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
857 break;
858 }
859}
860
861/*
862 * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
863 * new result that may suppress further error checking. Also modifies
864 * *blk_statp in some cases.
865 */
866static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
867 blk_status_t *blk_statp)
868{
869 bool sense_valid;
870 bool sense_current = true; /* false implies "deferred sense" */
871 struct request *req = scsi_cmd_to_rq(cmd);
872 struct scsi_sense_hdr sshdr;
873
874 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
875 if (sense_valid)
876 sense_current = !scsi_sense_is_deferred(&sshdr);
877
878 if (blk_rq_is_passthrough(req)) {
879 if (sense_valid) {
880 /*
881 * SG_IO wants current and deferred errors
882 */
883 cmd->sense_len = min(8 + cmd->sense_buffer[7],
884 SCSI_SENSE_BUFFERSIZE);
885 }
886 if (sense_current)
887 *blk_statp = scsi_result_to_blk_status(result);
888 } else if (blk_rq_bytes(req) == 0 && sense_current) {
889 /*
890 * Flush commands do not transfers any data, and thus cannot use
891 * good_bytes != blk_rq_bytes(req) as the signal for an error.
892 * This sets *blk_statp explicitly for the problem case.
893 */
894 *blk_statp = scsi_result_to_blk_status(result);
895 }
896 /*
897 * Recovered errors need reporting, but they're always treated as
898 * success, so fiddle the result code here. For passthrough requests
899 * we already took a copy of the original into sreq->result which
900 * is what gets returned to the user
901 */
902 if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
903 bool do_print = true;
904 /*
905 * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
906 * skip print since caller wants ATA registers. Only occurs
907 * on SCSI ATA PASS_THROUGH commands when CK_COND=1
908 */
909 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
910 do_print = false;
911 else if (req->rq_flags & RQF_QUIET)
912 do_print = false;
913 if (do_print)
914 scsi_print_sense(cmd);
915 result = 0;
916 /* for passthrough, *blk_statp may be set */
917 *blk_statp = BLK_STS_OK;
918 }
919 /*
920 * Another corner case: the SCSI status byte is non-zero but 'good'.
921 * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
922 * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
923 * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
924 * intermediate statuses (both obsolete in SAM-4) as good.
925 */
926 if ((result & 0xff) && scsi_status_is_good(result)) {
927 result = 0;
928 *blk_statp = BLK_STS_OK;
929 }
930 return result;
931}
932
933/**
934 * scsi_io_completion - Completion processing for SCSI commands.
935 * @cmd: command that is finished.
936 * @good_bytes: number of processed bytes.
937 *
938 * We will finish off the specified number of sectors. If we are done, the
939 * command block will be released and the queue function will be goosed. If we
940 * are not done then we have to figure out what to do next:
941 *
942 * a) We can call scsi_mq_requeue_cmd(). The request will be
943 * unprepared and put back on the queue. Then a new command will
944 * be created for it. This should be used if we made forward
945 * progress, or if we want to switch from READ(10) to READ(6) for
946 * example.
947 *
948 * b) We can call scsi_io_completion_action(). The request will be
949 * put back on the queue and retried using the same command as
950 * before, possibly after a delay.
951 *
952 * c) We can call scsi_end_request() with blk_stat other than
953 * BLK_STS_OK, to fail the remainder of the request.
954 */
955void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
956{
957 int result = cmd->result;
958 struct request *req = scsi_cmd_to_rq(cmd);
959 blk_status_t blk_stat = BLK_STS_OK;
960
961 if (unlikely(result)) /* a nz result may or may not be an error */
962 result = scsi_io_completion_nz_result(cmd, result, &blk_stat);
963
964 /*
965 * Next deal with any sectors which we were able to correctly
966 * handle.
967 */
968 SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
969 "%u sectors total, %d bytes done.\n",
970 blk_rq_sectors(req), good_bytes));
971
972 /*
973 * Failed, zero length commands always need to drop down
974 * to retry code. Fast path should return in this block.
975 */
976 if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
977 if (likely(!scsi_end_request(req, blk_stat, good_bytes)))
978 return; /* no bytes remaining */
979 }
980
981 /* Kill remainder if no retries. */
982 if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
983 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req)))
984 WARN_ONCE(true,
985 "Bytes remaining after failed, no-retry command");
986 return;
987 }
988
989 /*
990 * If there had been no error, but we have leftover bytes in the
991 * request just queue the command up again.
992 */
993 if (likely(result == 0))
994 scsi_mq_requeue_cmd(cmd, 0);
995 else
996 scsi_io_completion_action(cmd, result);
997}
998
999static inline bool scsi_cmd_needs_dma_drain(struct scsi_device *sdev,
1000 struct request *rq)
1001{
1002 return sdev->dma_drain_len && blk_rq_is_passthrough(rq) &&
1003 !op_is_write(req_op(rq)) &&
1004 sdev->host->hostt->dma_need_drain(rq);
1005}
1006
1007/**
1008 * scsi_alloc_sgtables - Allocate and initialize data and integrity scatterlists
1009 * @cmd: SCSI command data structure to initialize.
1010 *
1011 * Initializes @cmd->sdb and also @cmd->prot_sdb if data integrity is enabled
1012 * for @cmd.
1013 *
1014 * Returns:
1015 * * BLK_STS_OK - on success
1016 * * BLK_STS_RESOURCE - if the failure is retryable
1017 * * BLK_STS_IOERR - if the failure is fatal
1018 */
1019blk_status_t scsi_alloc_sgtables(struct scsi_cmnd *cmd)
1020{
1021 struct scsi_device *sdev = cmd->device;
1022 struct request *rq = scsi_cmd_to_rq(cmd);
1023 unsigned short nr_segs = blk_rq_nr_phys_segments(rq);
1024 struct scatterlist *last_sg = NULL;
1025 blk_status_t ret;
1026 bool need_drain = scsi_cmd_needs_dma_drain(sdev, rq);
1027 int count;
1028
1029 if (WARN_ON_ONCE(!nr_segs))
1030 return BLK_STS_IOERR;
1031
1032 /*
1033 * Make sure there is space for the drain. The driver must adjust
1034 * max_hw_segments to be prepared for this.
1035 */
1036 if (need_drain)
1037 nr_segs++;
1038
1039 /*
1040 * If sg table allocation fails, requeue request later.
1041 */
1042 if (unlikely(sg_alloc_table_chained(&cmd->sdb.table, nr_segs,
1043 cmd->sdb.table.sgl, SCSI_INLINE_SG_CNT)))
1044 return BLK_STS_RESOURCE;
1045
1046 /*
1047 * Next, walk the list, and fill in the addresses and sizes of
1048 * each segment.
1049 */
1050 count = __blk_rq_map_sg(rq->q, rq, cmd->sdb.table.sgl, &last_sg);
1051
1052 if (blk_rq_bytes(rq) & rq->q->dma_pad_mask) {
1053 unsigned int pad_len =
1054 (rq->q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
1055
1056 last_sg->length += pad_len;
1057 cmd->extra_len += pad_len;
1058 }
1059
1060 if (need_drain) {
1061 sg_unmark_end(last_sg);
1062 last_sg = sg_next(last_sg);
1063 sg_set_buf(last_sg, sdev->dma_drain_buf, sdev->dma_drain_len);
1064 sg_mark_end(last_sg);
1065
1066 cmd->extra_len += sdev->dma_drain_len;
1067 count++;
1068 }
1069
1070 BUG_ON(count > cmd->sdb.table.nents);
1071 cmd->sdb.table.nents = count;
1072 cmd->sdb.length = blk_rq_payload_bytes(rq);
1073
1074 if (blk_integrity_rq(rq)) {
1075 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1076 int ivecs;
1077
1078 if (WARN_ON_ONCE(!prot_sdb)) {
1079 /*
1080 * This can happen if someone (e.g. multipath)
1081 * queues a command to a device on an adapter
1082 * that does not support DIX.
1083 */
1084 ret = BLK_STS_IOERR;
1085 goto out_free_sgtables;
1086 }
1087
1088 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1089
1090 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1091 prot_sdb->table.sgl,
1092 SCSI_INLINE_PROT_SG_CNT)) {
1093 ret = BLK_STS_RESOURCE;
1094 goto out_free_sgtables;
1095 }
1096
1097 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1098 prot_sdb->table.sgl);
1099 BUG_ON(count > ivecs);
1100 BUG_ON(count > queue_max_integrity_segments(rq->q));
1101
1102 cmd->prot_sdb = prot_sdb;
1103 cmd->prot_sdb->table.nents = count;
1104 }
1105
1106 return BLK_STS_OK;
1107out_free_sgtables:
1108 scsi_free_sgtables(cmd);
1109 return ret;
1110}
1111EXPORT_SYMBOL(scsi_alloc_sgtables);
1112
1113/**
1114 * scsi_initialize_rq - initialize struct scsi_cmnd partially
1115 * @rq: Request associated with the SCSI command to be initialized.
1116 *
1117 * This function initializes the members of struct scsi_cmnd that must be
1118 * initialized before request processing starts and that won't be
1119 * reinitialized if a SCSI command is requeued.
1120 */
1121static void scsi_initialize_rq(struct request *rq)
1122{
1123 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1124
1125 memset(cmd->cmnd, 0, sizeof(cmd->cmnd));
1126 cmd->cmd_len = MAX_COMMAND_SIZE;
1127 cmd->sense_len = 0;
1128 init_rcu_head(&cmd->rcu);
1129 cmd->jiffies_at_alloc = jiffies;
1130 cmd->retries = 0;
1131}
1132
1133struct request *scsi_alloc_request(struct request_queue *q, blk_opf_t opf,
1134 blk_mq_req_flags_t flags)
1135{
1136 struct request *rq;
1137
1138 rq = blk_mq_alloc_request(q, opf, flags);
1139 if (!IS_ERR(rq))
1140 scsi_initialize_rq(rq);
1141 return rq;
1142}
1143EXPORT_SYMBOL_GPL(scsi_alloc_request);
1144
1145/*
1146 * Only called when the request isn't completed by SCSI, and not freed by
1147 * SCSI
1148 */
1149static void scsi_cleanup_rq(struct request *rq)
1150{
1151 if (rq->rq_flags & RQF_DONTPREP) {
1152 scsi_mq_uninit_cmd(blk_mq_rq_to_pdu(rq));
1153 rq->rq_flags &= ~RQF_DONTPREP;
1154 }
1155}
1156
1157/* Called before a request is prepared. See also scsi_mq_prep_fn(). */
1158void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1159{
1160 struct request *rq = scsi_cmd_to_rq(cmd);
1161
1162 if (!blk_rq_is_passthrough(rq) && !(cmd->flags & SCMD_INITIALIZED)) {
1163 cmd->flags |= SCMD_INITIALIZED;
1164 scsi_initialize_rq(rq);
1165 }
1166
1167 cmd->device = dev;
1168 INIT_LIST_HEAD(&cmd->eh_entry);
1169 INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1170}
1171
1172static blk_status_t scsi_setup_scsi_cmnd(struct scsi_device *sdev,
1173 struct request *req)
1174{
1175 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1176
1177 /*
1178 * Passthrough requests may transfer data, in which case they must
1179 * a bio attached to them. Or they might contain a SCSI command
1180 * that does not transfer data, in which case they may optionally
1181 * submit a request without an attached bio.
1182 */
1183 if (req->bio) {
1184 blk_status_t ret = scsi_alloc_sgtables(cmd);
1185 if (unlikely(ret != BLK_STS_OK))
1186 return ret;
1187 } else {
1188 BUG_ON(blk_rq_bytes(req));
1189
1190 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1191 }
1192
1193 cmd->transfersize = blk_rq_bytes(req);
1194 return BLK_STS_OK;
1195}
1196
1197static blk_status_t
1198scsi_device_state_check(struct scsi_device *sdev, struct request *req)
1199{
1200 switch (sdev->sdev_state) {
1201 case SDEV_CREATED:
1202 return BLK_STS_OK;
1203 case SDEV_OFFLINE:
1204 case SDEV_TRANSPORT_OFFLINE:
1205 /*
1206 * If the device is offline we refuse to process any
1207 * commands. The device must be brought online
1208 * before trying any recovery commands.
1209 */
1210 if (!sdev->offline_already) {
1211 sdev->offline_already = true;
1212 sdev_printk(KERN_ERR, sdev,
1213 "rejecting I/O to offline device\n");
1214 }
1215 return BLK_STS_IOERR;
1216 case SDEV_DEL:
1217 /*
1218 * If the device is fully deleted, we refuse to
1219 * process any commands as well.
1220 */
1221 sdev_printk(KERN_ERR, sdev,
1222 "rejecting I/O to dead device\n");
1223 return BLK_STS_IOERR;
1224 case SDEV_BLOCK:
1225 case SDEV_CREATED_BLOCK:
1226 return BLK_STS_RESOURCE;
1227 case SDEV_QUIESCE:
1228 /*
1229 * If the device is blocked we only accept power management
1230 * commands.
1231 */
1232 if (req && WARN_ON_ONCE(!(req->rq_flags & RQF_PM)))
1233 return BLK_STS_RESOURCE;
1234 return BLK_STS_OK;
1235 default:
1236 /*
1237 * For any other not fully online state we only allow
1238 * power management commands.
1239 */
1240 if (req && !(req->rq_flags & RQF_PM))
1241 return BLK_STS_OFFLINE;
1242 return BLK_STS_OK;
1243 }
1244}
1245
1246/*
1247 * scsi_dev_queue_ready: if we can send requests to sdev, assign one token
1248 * and return the token else return -1.
1249 */
1250static inline int scsi_dev_queue_ready(struct request_queue *q,
1251 struct scsi_device *sdev)
1252{
1253 int token;
1254
1255 token = sbitmap_get(&sdev->budget_map);
1256 if (token < 0)
1257 return -1;
1258
1259 if (!atomic_read(&sdev->device_blocked))
1260 return token;
1261
1262 /*
1263 * Only unblock if no other commands are pending and
1264 * if device_blocked has decreased to zero
1265 */
1266 if (scsi_device_busy(sdev) > 1 ||
1267 atomic_dec_return(&sdev->device_blocked) > 0) {
1268 sbitmap_put(&sdev->budget_map, token);
1269 return -1;
1270 }
1271
1272 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1273 "unblocking device at zero depth\n"));
1274
1275 return token;
1276}
1277
1278/*
1279 * scsi_target_queue_ready: checks if there we can send commands to target
1280 * @sdev: scsi device on starget to check.
1281 */
1282static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1283 struct scsi_device *sdev)
1284{
1285 struct scsi_target *starget = scsi_target(sdev);
1286 unsigned int busy;
1287
1288 if (starget->single_lun) {
1289 spin_lock_irq(shost->host_lock);
1290 if (starget->starget_sdev_user &&
1291 starget->starget_sdev_user != sdev) {
1292 spin_unlock_irq(shost->host_lock);
1293 return 0;
1294 }
1295 starget->starget_sdev_user = sdev;
1296 spin_unlock_irq(shost->host_lock);
1297 }
1298
1299 if (starget->can_queue <= 0)
1300 return 1;
1301
1302 busy = atomic_inc_return(&starget->target_busy) - 1;
1303 if (atomic_read(&starget->target_blocked) > 0) {
1304 if (busy)
1305 goto starved;
1306
1307 /*
1308 * unblock after target_blocked iterates to zero
1309 */
1310 if (atomic_dec_return(&starget->target_blocked) > 0)
1311 goto out_dec;
1312
1313 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1314 "unblocking target at zero depth\n"));
1315 }
1316
1317 if (busy >= starget->can_queue)
1318 goto starved;
1319
1320 return 1;
1321
1322starved:
1323 spin_lock_irq(shost->host_lock);
1324 list_move_tail(&sdev->starved_entry, &shost->starved_list);
1325 spin_unlock_irq(shost->host_lock);
1326out_dec:
1327 if (starget->can_queue > 0)
1328 atomic_dec(&starget->target_busy);
1329 return 0;
1330}
1331
1332/*
1333 * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1334 * return 0. We must end up running the queue again whenever 0 is
1335 * returned, else IO can hang.
1336 */
1337static inline int scsi_host_queue_ready(struct request_queue *q,
1338 struct Scsi_Host *shost,
1339 struct scsi_device *sdev,
1340 struct scsi_cmnd *cmd)
1341{
1342 if (atomic_read(&shost->host_blocked) > 0) {
1343 if (scsi_host_busy(shost) > 0)
1344 goto starved;
1345
1346 /*
1347 * unblock after host_blocked iterates to zero
1348 */
1349 if (atomic_dec_return(&shost->host_blocked) > 0)
1350 goto out_dec;
1351
1352 SCSI_LOG_MLQUEUE(3,
1353 shost_printk(KERN_INFO, shost,
1354 "unblocking host at zero depth\n"));
1355 }
1356
1357 if (shost->host_self_blocked)
1358 goto starved;
1359
1360 /* We're OK to process the command, so we can't be starved */
1361 if (!list_empty(&sdev->starved_entry)) {
1362 spin_lock_irq(shost->host_lock);
1363 if (!list_empty(&sdev->starved_entry))
1364 list_del_init(&sdev->starved_entry);
1365 spin_unlock_irq(shost->host_lock);
1366 }
1367
1368 __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1369
1370 return 1;
1371
1372starved:
1373 spin_lock_irq(shost->host_lock);
1374 if (list_empty(&sdev->starved_entry))
1375 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1376 spin_unlock_irq(shost->host_lock);
1377out_dec:
1378 scsi_dec_host_busy(shost, cmd);
1379 return 0;
1380}
1381
1382/*
1383 * Busy state exporting function for request stacking drivers.
1384 *
1385 * For efficiency, no lock is taken to check the busy state of
1386 * shost/starget/sdev, since the returned value is not guaranteed and
1387 * may be changed after request stacking drivers call the function,
1388 * regardless of taking lock or not.
1389 *
1390 * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1391 * needs to return 'not busy'. Otherwise, request stacking drivers
1392 * may hold requests forever.
1393 */
1394static bool scsi_mq_lld_busy(struct request_queue *q)
1395{
1396 struct scsi_device *sdev = q->queuedata;
1397 struct Scsi_Host *shost;
1398
1399 if (blk_queue_dying(q))
1400 return false;
1401
1402 shost = sdev->host;
1403
1404 /*
1405 * Ignore host/starget busy state.
1406 * Since block layer does not have a concept of fairness across
1407 * multiple queues, congestion of host/starget needs to be handled
1408 * in SCSI layer.
1409 */
1410 if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1411 return true;
1412
1413 return false;
1414}
1415
1416/*
1417 * Block layer request completion callback. May be called from interrupt
1418 * context.
1419 */
1420static void scsi_complete(struct request *rq)
1421{
1422 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1423 enum scsi_disposition disposition;
1424
1425 INIT_LIST_HEAD(&cmd->eh_entry);
1426
1427 atomic_inc(&cmd->device->iodone_cnt);
1428 if (cmd->result)
1429 atomic_inc(&cmd->device->ioerr_cnt);
1430
1431 disposition = scsi_decide_disposition(cmd);
1432 if (disposition != SUCCESS && scsi_cmd_runtime_exceeced(cmd))
1433 disposition = SUCCESS;
1434
1435 scsi_log_completion(cmd, disposition);
1436
1437 switch (disposition) {
1438 case SUCCESS:
1439 scsi_finish_command(cmd);
1440 break;
1441 case NEEDS_RETRY:
1442 scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1443 break;
1444 case ADD_TO_MLQUEUE:
1445 scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1446 break;
1447 default:
1448 scsi_eh_scmd_add(cmd);
1449 break;
1450 }
1451}
1452
1453/**
1454 * scsi_dispatch_cmd - Dispatch a command to the low-level driver.
1455 * @cmd: command block we are dispatching.
1456 *
1457 * Return: nonzero return request was rejected and device's queue needs to be
1458 * plugged.
1459 */
1460static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1461{
1462 struct Scsi_Host *host = cmd->device->host;
1463 int rtn = 0;
1464
1465 atomic_inc(&cmd->device->iorequest_cnt);
1466
1467 /* check if the device is still usable */
1468 if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1469 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1470 * returns an immediate error upwards, and signals
1471 * that the device is no longer present */
1472 cmd->result = DID_NO_CONNECT << 16;
1473 goto done;
1474 }
1475
1476 /* Check to see if the scsi lld made this device blocked. */
1477 if (unlikely(scsi_device_blocked(cmd->device))) {
1478 /*
1479 * in blocked state, the command is just put back on
1480 * the device queue. The suspend state has already
1481 * blocked the queue so future requests should not
1482 * occur until the device transitions out of the
1483 * suspend state.
1484 */
1485 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1486 "queuecommand : device blocked\n"));
1487 atomic_dec(&cmd->device->iorequest_cnt);
1488 return SCSI_MLQUEUE_DEVICE_BUSY;
1489 }
1490
1491 /* Store the LUN value in cmnd, if needed. */
1492 if (cmd->device->lun_in_cdb)
1493 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1494 (cmd->device->lun << 5 & 0xe0);
1495
1496 scsi_log_send(cmd);
1497
1498 /*
1499 * Before we queue this command, check if the command
1500 * length exceeds what the host adapter can handle.
1501 */
1502 if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1503 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1504 "queuecommand : command too long. "
1505 "cdb_size=%d host->max_cmd_len=%d\n",
1506 cmd->cmd_len, cmd->device->host->max_cmd_len));
1507 cmd->result = (DID_ABORT << 16);
1508 goto done;
1509 }
1510
1511 if (unlikely(host->shost_state == SHOST_DEL)) {
1512 cmd->result = (DID_NO_CONNECT << 16);
1513 goto done;
1514
1515 }
1516
1517 trace_scsi_dispatch_cmd_start(cmd);
1518 rtn = host->hostt->queuecommand(host, cmd);
1519 if (rtn) {
1520 atomic_dec(&cmd->device->iorequest_cnt);
1521 trace_scsi_dispatch_cmd_error(cmd, rtn);
1522 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1523 rtn != SCSI_MLQUEUE_TARGET_BUSY)
1524 rtn = SCSI_MLQUEUE_HOST_BUSY;
1525
1526 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1527 "queuecommand : request rejected\n"));
1528 }
1529
1530 return rtn;
1531 done:
1532 scsi_done(cmd);
1533 return 0;
1534}
1535
1536/* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1537static unsigned int scsi_mq_inline_sgl_size(struct Scsi_Host *shost)
1538{
1539 return min_t(unsigned int, shost->sg_tablesize, SCSI_INLINE_SG_CNT) *
1540 sizeof(struct scatterlist);
1541}
1542
1543static blk_status_t scsi_prepare_cmd(struct request *req)
1544{
1545 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1546 struct scsi_device *sdev = req->q->queuedata;
1547 struct Scsi_Host *shost = sdev->host;
1548 bool in_flight = test_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1549 struct scatterlist *sg;
1550
1551 scsi_init_command(sdev, cmd);
1552
1553 cmd->eh_eflags = 0;
1554 cmd->prot_type = 0;
1555 cmd->prot_flags = 0;
1556 cmd->submitter = 0;
1557 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1558 cmd->underflow = 0;
1559 cmd->transfersize = 0;
1560 cmd->host_scribble = NULL;
1561 cmd->result = 0;
1562 cmd->extra_len = 0;
1563 cmd->state = 0;
1564 if (in_flight)
1565 __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1566
1567 /*
1568 * Only clear the driver-private command data if the LLD does not supply
1569 * a function to initialize that data.
1570 */
1571 if (!shost->hostt->init_cmd_priv)
1572 memset(cmd + 1, 0, shost->hostt->cmd_size);
1573
1574 cmd->prot_op = SCSI_PROT_NORMAL;
1575 if (blk_rq_bytes(req))
1576 cmd->sc_data_direction = rq_dma_dir(req);
1577 else
1578 cmd->sc_data_direction = DMA_NONE;
1579
1580 sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1581 cmd->sdb.table.sgl = sg;
1582
1583 if (scsi_host_get_prot(shost)) {
1584 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1585
1586 cmd->prot_sdb->table.sgl =
1587 (struct scatterlist *)(cmd->prot_sdb + 1);
1588 }
1589
1590 /*
1591 * Special handling for passthrough commands, which don't go to the ULP
1592 * at all:
1593 */
1594 if (blk_rq_is_passthrough(req))
1595 return scsi_setup_scsi_cmnd(sdev, req);
1596
1597 if (sdev->handler && sdev->handler->prep_fn) {
1598 blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1599
1600 if (ret != BLK_STS_OK)
1601 return ret;
1602 }
1603
1604 /* Usually overridden by the ULP */
1605 cmd->allowed = 0;
1606 memset(cmd->cmnd, 0, sizeof(cmd->cmnd));
1607 return scsi_cmd_to_driver(cmd)->init_command(cmd);
1608}
1609
1610static void scsi_done_internal(struct scsi_cmnd *cmd, bool complete_directly)
1611{
1612 struct request *req = scsi_cmd_to_rq(cmd);
1613
1614 switch (cmd->submitter) {
1615 case SUBMITTED_BY_BLOCK_LAYER:
1616 break;
1617 case SUBMITTED_BY_SCSI_ERROR_HANDLER:
1618 return scsi_eh_done(cmd);
1619 case SUBMITTED_BY_SCSI_RESET_IOCTL:
1620 return;
1621 }
1622
1623 if (unlikely(blk_should_fake_timeout(scsi_cmd_to_rq(cmd)->q)))
1624 return;
1625 if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1626 return;
1627 trace_scsi_dispatch_cmd_done(cmd);
1628
1629 if (complete_directly)
1630 blk_mq_complete_request_direct(req, scsi_complete);
1631 else
1632 blk_mq_complete_request(req);
1633}
1634
1635void scsi_done(struct scsi_cmnd *cmd)
1636{
1637 scsi_done_internal(cmd, false);
1638}
1639EXPORT_SYMBOL(scsi_done);
1640
1641void scsi_done_direct(struct scsi_cmnd *cmd)
1642{
1643 scsi_done_internal(cmd, true);
1644}
1645EXPORT_SYMBOL(scsi_done_direct);
1646
1647static void scsi_mq_put_budget(struct request_queue *q, int budget_token)
1648{
1649 struct scsi_device *sdev = q->queuedata;
1650
1651 sbitmap_put(&sdev->budget_map, budget_token);
1652}
1653
1654/*
1655 * When to reinvoke queueing after a resource shortage. It's 3 msecs to
1656 * not change behaviour from the previous unplug mechanism, experimentation
1657 * may prove this needs changing.
1658 */
1659#define SCSI_QUEUE_DELAY 3
1660
1661static int scsi_mq_get_budget(struct request_queue *q)
1662{
1663 struct scsi_device *sdev = q->queuedata;
1664 int token = scsi_dev_queue_ready(q, sdev);
1665
1666 if (token >= 0)
1667 return token;
1668
1669 atomic_inc(&sdev->restarts);
1670
1671 /*
1672 * Orders atomic_inc(&sdev->restarts) and atomic_read(&sdev->device_busy).
1673 * .restarts must be incremented before .device_busy is read because the
1674 * code in scsi_run_queue_async() depends on the order of these operations.
1675 */
1676 smp_mb__after_atomic();
1677
1678 /*
1679 * If all in-flight requests originated from this LUN are completed
1680 * before reading .device_busy, sdev->device_busy will be observed as
1681 * zero, then blk_mq_delay_run_hw_queues() will dispatch this request
1682 * soon. Otherwise, completion of one of these requests will observe
1683 * the .restarts flag, and the request queue will be run for handling
1684 * this request, see scsi_end_request().
1685 */
1686 if (unlikely(scsi_device_busy(sdev) == 0 &&
1687 !scsi_device_blocked(sdev)))
1688 blk_mq_delay_run_hw_queues(sdev->request_queue, SCSI_QUEUE_DELAY);
1689 return -1;
1690}
1691
1692static void scsi_mq_set_rq_budget_token(struct request *req, int token)
1693{
1694 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1695
1696 cmd->budget_token = token;
1697}
1698
1699static int scsi_mq_get_rq_budget_token(struct request *req)
1700{
1701 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1702
1703 return cmd->budget_token;
1704}
1705
1706static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1707 const struct blk_mq_queue_data *bd)
1708{
1709 struct request *req = bd->rq;
1710 struct request_queue *q = req->q;
1711 struct scsi_device *sdev = q->queuedata;
1712 struct Scsi_Host *shost = sdev->host;
1713 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1714 blk_status_t ret;
1715 int reason;
1716
1717 WARN_ON_ONCE(cmd->budget_token < 0);
1718
1719 /*
1720 * If the device is not in running state we will reject some or all
1721 * commands.
1722 */
1723 if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1724 ret = scsi_device_state_check(sdev, req);
1725 if (ret != BLK_STS_OK)
1726 goto out_put_budget;
1727 }
1728
1729 ret = BLK_STS_RESOURCE;
1730 if (!scsi_target_queue_ready(shost, sdev))
1731 goto out_put_budget;
1732 if (unlikely(scsi_host_in_recovery(shost))) {
1733 if (cmd->flags & SCMD_FAIL_IF_RECOVERING)
1734 ret = BLK_STS_OFFLINE;
1735 goto out_dec_target_busy;
1736 }
1737 if (!scsi_host_queue_ready(q, shost, sdev, cmd))
1738 goto out_dec_target_busy;
1739
1740 if (!(req->rq_flags & RQF_DONTPREP)) {
1741 ret = scsi_prepare_cmd(req);
1742 if (ret != BLK_STS_OK)
1743 goto out_dec_host_busy;
1744 req->rq_flags |= RQF_DONTPREP;
1745 } else {
1746 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1747 }
1748
1749 cmd->flags &= SCMD_PRESERVED_FLAGS;
1750 if (sdev->simple_tags)
1751 cmd->flags |= SCMD_TAGGED;
1752 if (bd->last)
1753 cmd->flags |= SCMD_LAST;
1754
1755 scsi_set_resid(cmd, 0);
1756 memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
1757 cmd->submitter = SUBMITTED_BY_BLOCK_LAYER;
1758
1759 blk_mq_start_request(req);
1760 reason = scsi_dispatch_cmd(cmd);
1761 if (reason) {
1762 scsi_set_blocked(cmd, reason);
1763 ret = BLK_STS_RESOURCE;
1764 goto out_dec_host_busy;
1765 }
1766
1767 return BLK_STS_OK;
1768
1769out_dec_host_busy:
1770 scsi_dec_host_busy(shost, cmd);
1771out_dec_target_busy:
1772 if (scsi_target(sdev)->can_queue > 0)
1773 atomic_dec(&scsi_target(sdev)->target_busy);
1774out_put_budget:
1775 scsi_mq_put_budget(q, cmd->budget_token);
1776 cmd->budget_token = -1;
1777 switch (ret) {
1778 case BLK_STS_OK:
1779 break;
1780 case BLK_STS_RESOURCE:
1781 case BLK_STS_ZONE_RESOURCE:
1782 if (scsi_device_blocked(sdev))
1783 ret = BLK_STS_DEV_RESOURCE;
1784 break;
1785 case BLK_STS_AGAIN:
1786 cmd->result = DID_BUS_BUSY << 16;
1787 if (req->rq_flags & RQF_DONTPREP)
1788 scsi_mq_uninit_cmd(cmd);
1789 break;
1790 default:
1791 if (unlikely(!scsi_device_online(sdev)))
1792 cmd->result = DID_NO_CONNECT << 16;
1793 else
1794 cmd->result = DID_ERROR << 16;
1795 /*
1796 * Make sure to release all allocated resources when
1797 * we hit an error, as we will never see this command
1798 * again.
1799 */
1800 if (req->rq_flags & RQF_DONTPREP)
1801 scsi_mq_uninit_cmd(cmd);
1802 scsi_run_queue_async(sdev);
1803 break;
1804 }
1805 return ret;
1806}
1807
1808static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1809 unsigned int hctx_idx, unsigned int numa_node)
1810{
1811 struct Scsi_Host *shost = set->driver_data;
1812 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1813 struct scatterlist *sg;
1814 int ret = 0;
1815
1816 cmd->sense_buffer =
1817 kmem_cache_alloc_node(scsi_sense_cache, GFP_KERNEL, numa_node);
1818 if (!cmd->sense_buffer)
1819 return -ENOMEM;
1820
1821 if (scsi_host_get_prot(shost)) {
1822 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1823 shost->hostt->cmd_size;
1824 cmd->prot_sdb = (void *)sg + scsi_mq_inline_sgl_size(shost);
1825 }
1826
1827 if (shost->hostt->init_cmd_priv) {
1828 ret = shost->hostt->init_cmd_priv(shost, cmd);
1829 if (ret < 0)
1830 kmem_cache_free(scsi_sense_cache, cmd->sense_buffer);
1831 }
1832
1833 return ret;
1834}
1835
1836static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1837 unsigned int hctx_idx)
1838{
1839 struct Scsi_Host *shost = set->driver_data;
1840 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1841
1842 if (shost->hostt->exit_cmd_priv)
1843 shost->hostt->exit_cmd_priv(shost, cmd);
1844 kmem_cache_free(scsi_sense_cache, cmd->sense_buffer);
1845}
1846
1847
1848static int scsi_mq_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
1849{
1850 struct Scsi_Host *shost = hctx->driver_data;
1851
1852 if (shost->hostt->mq_poll)
1853 return shost->hostt->mq_poll(shost, hctx->queue_num);
1854
1855 return 0;
1856}
1857
1858static int scsi_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
1859 unsigned int hctx_idx)
1860{
1861 struct Scsi_Host *shost = data;
1862
1863 hctx->driver_data = shost;
1864 return 0;
1865}
1866
1867static void scsi_map_queues(struct blk_mq_tag_set *set)
1868{
1869 struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
1870
1871 if (shost->hostt->map_queues)
1872 return shost->hostt->map_queues(shost);
1873 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
1874}
1875
1876void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
1877{
1878 struct device *dev = shost->dma_dev;
1879
1880 /*
1881 * this limit is imposed by hardware restrictions
1882 */
1883 blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1884 SG_MAX_SEGMENTS));
1885
1886 if (scsi_host_prot_dma(shost)) {
1887 shost->sg_prot_tablesize =
1888 min_not_zero(shost->sg_prot_tablesize,
1889 (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1890 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1891 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1892 }
1893
1894 blk_queue_max_hw_sectors(q, shost->max_sectors);
1895 blk_queue_segment_boundary(q, shost->dma_boundary);
1896 dma_set_seg_boundary(dev, shost->dma_boundary);
1897
1898 blk_queue_max_segment_size(q, shost->max_segment_size);
1899 blk_queue_virt_boundary(q, shost->virt_boundary_mask);
1900 dma_set_max_seg_size(dev, queue_max_segment_size(q));
1901
1902 /*
1903 * Set a reasonable default alignment: The larger of 32-byte (dword),
1904 * which is a common minimum for HBAs, and the minimum DMA alignment,
1905 * which is set by the platform.
1906 *
1907 * Devices that require a bigger alignment can increase it later.
1908 */
1909 blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
1910}
1911EXPORT_SYMBOL_GPL(__scsi_init_queue);
1912
1913static const struct blk_mq_ops scsi_mq_ops_no_commit = {
1914 .get_budget = scsi_mq_get_budget,
1915 .put_budget = scsi_mq_put_budget,
1916 .queue_rq = scsi_queue_rq,
1917 .complete = scsi_complete,
1918 .timeout = scsi_timeout,
1919#ifdef CONFIG_BLK_DEBUG_FS
1920 .show_rq = scsi_show_rq,
1921#endif
1922 .init_request = scsi_mq_init_request,
1923 .exit_request = scsi_mq_exit_request,
1924 .cleanup_rq = scsi_cleanup_rq,
1925 .busy = scsi_mq_lld_busy,
1926 .map_queues = scsi_map_queues,
1927 .init_hctx = scsi_init_hctx,
1928 .poll = scsi_mq_poll,
1929 .set_rq_budget_token = scsi_mq_set_rq_budget_token,
1930 .get_rq_budget_token = scsi_mq_get_rq_budget_token,
1931};
1932
1933
1934static void scsi_commit_rqs(struct blk_mq_hw_ctx *hctx)
1935{
1936 struct Scsi_Host *shost = hctx->driver_data;
1937
1938 shost->hostt->commit_rqs(shost, hctx->queue_num);
1939}
1940
1941static const struct blk_mq_ops scsi_mq_ops = {
1942 .get_budget = scsi_mq_get_budget,
1943 .put_budget = scsi_mq_put_budget,
1944 .queue_rq = scsi_queue_rq,
1945 .commit_rqs = scsi_commit_rqs,
1946 .complete = scsi_complete,
1947 .timeout = scsi_timeout,
1948#ifdef CONFIG_BLK_DEBUG_FS
1949 .show_rq = scsi_show_rq,
1950#endif
1951 .init_request = scsi_mq_init_request,
1952 .exit_request = scsi_mq_exit_request,
1953 .cleanup_rq = scsi_cleanup_rq,
1954 .busy = scsi_mq_lld_busy,
1955 .map_queues = scsi_map_queues,
1956 .init_hctx = scsi_init_hctx,
1957 .poll = scsi_mq_poll,
1958 .set_rq_budget_token = scsi_mq_set_rq_budget_token,
1959 .get_rq_budget_token = scsi_mq_get_rq_budget_token,
1960};
1961
1962int scsi_mq_setup_tags(struct Scsi_Host *shost)
1963{
1964 unsigned int cmd_size, sgl_size;
1965 struct blk_mq_tag_set *tag_set = &shost->tag_set;
1966
1967 sgl_size = max_t(unsigned int, sizeof(struct scatterlist),
1968 scsi_mq_inline_sgl_size(shost));
1969 cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
1970 if (scsi_host_get_prot(shost))
1971 cmd_size += sizeof(struct scsi_data_buffer) +
1972 sizeof(struct scatterlist) * SCSI_INLINE_PROT_SG_CNT;
1973
1974 memset(tag_set, 0, sizeof(*tag_set));
1975 if (shost->hostt->commit_rqs)
1976 tag_set->ops = &scsi_mq_ops;
1977 else
1978 tag_set->ops = &scsi_mq_ops_no_commit;
1979 tag_set->nr_hw_queues = shost->nr_hw_queues ? : 1;
1980 tag_set->nr_maps = shost->nr_maps ? : 1;
1981 tag_set->queue_depth = shost->can_queue;
1982 tag_set->cmd_size = cmd_size;
1983 tag_set->numa_node = dev_to_node(shost->dma_dev);
1984 tag_set->flags = BLK_MQ_F_SHOULD_MERGE;
1985 tag_set->flags |=
1986 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
1987 if (shost->queuecommand_may_block)
1988 tag_set->flags |= BLK_MQ_F_BLOCKING;
1989 tag_set->driver_data = shost;
1990 if (shost->host_tagset)
1991 tag_set->flags |= BLK_MQ_F_TAG_HCTX_SHARED;
1992
1993 return blk_mq_alloc_tag_set(tag_set);
1994}
1995
1996void scsi_mq_free_tags(struct kref *kref)
1997{
1998 struct Scsi_Host *shost = container_of(kref, typeof(*shost),
1999 tagset_refcnt);
2000
2001 blk_mq_free_tag_set(&shost->tag_set);
2002 complete(&shost->tagset_freed);
2003}
2004
2005/**
2006 * scsi_device_from_queue - return sdev associated with a request_queue
2007 * @q: The request queue to return the sdev from
2008 *
2009 * Return the sdev associated with a request queue or NULL if the
2010 * request_queue does not reference a SCSI device.
2011 */
2012struct scsi_device *scsi_device_from_queue(struct request_queue *q)
2013{
2014 struct scsi_device *sdev = NULL;
2015
2016 if (q->mq_ops == &scsi_mq_ops_no_commit ||
2017 q->mq_ops == &scsi_mq_ops)
2018 sdev = q->queuedata;
2019 if (!sdev || !get_device(&sdev->sdev_gendev))
2020 sdev = NULL;
2021
2022 return sdev;
2023}
2024/*
2025 * pktcdvd should have been integrated into the SCSI layers, but for historical
2026 * reasons like the old IDE driver it isn't. This export allows it to safely
2027 * probe if a given device is a SCSI one and only attach to that.
2028 */
2029#ifdef CONFIG_CDROM_PKTCDVD_MODULE
2030EXPORT_SYMBOL_GPL(scsi_device_from_queue);
2031#endif
2032
2033/**
2034 * scsi_block_requests - Utility function used by low-level drivers to prevent
2035 * further commands from being queued to the device.
2036 * @shost: host in question
2037 *
2038 * There is no timer nor any other means by which the requests get unblocked
2039 * other than the low-level driver calling scsi_unblock_requests().
2040 */
2041void scsi_block_requests(struct Scsi_Host *shost)
2042{
2043 shost->host_self_blocked = 1;
2044}
2045EXPORT_SYMBOL(scsi_block_requests);
2046
2047/**
2048 * scsi_unblock_requests - Utility function used by low-level drivers to allow
2049 * further commands to be queued to the device.
2050 * @shost: host in question
2051 *
2052 * There is no timer nor any other means by which the requests get unblocked
2053 * other than the low-level driver calling scsi_unblock_requests(). This is done
2054 * as an API function so that changes to the internals of the scsi mid-layer
2055 * won't require wholesale changes to drivers that use this feature.
2056 */
2057void scsi_unblock_requests(struct Scsi_Host *shost)
2058{
2059 shost->host_self_blocked = 0;
2060 scsi_run_host_queues(shost);
2061}
2062EXPORT_SYMBOL(scsi_unblock_requests);
2063
2064void scsi_exit_queue(void)
2065{
2066 kmem_cache_destroy(scsi_sense_cache);
2067}
2068
2069/**
2070 * scsi_mode_select - issue a mode select
2071 * @sdev: SCSI device to be queried
2072 * @pf: Page format bit (1 == standard, 0 == vendor specific)
2073 * @sp: Save page bit (0 == don't save, 1 == save)
2074 * @buffer: request buffer (may not be smaller than eight bytes)
2075 * @len: length of request buffer.
2076 * @timeout: command timeout
2077 * @retries: number of retries before failing
2078 * @data: returns a structure abstracting the mode header data
2079 * @sshdr: place to put sense data (or NULL if no sense to be collected).
2080 * must be SCSI_SENSE_BUFFERSIZE big.
2081 *
2082 * Returns zero if successful; negative error number or scsi
2083 * status on error
2084 *
2085 */
2086int scsi_mode_select(struct scsi_device *sdev, int pf, int sp,
2087 unsigned char *buffer, int len, int timeout, int retries,
2088 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2089{
2090 unsigned char cmd[10];
2091 unsigned char *real_buffer;
2092 const struct scsi_exec_args exec_args = {
2093 .sshdr = sshdr,
2094 };
2095 int ret;
2096
2097 memset(cmd, 0, sizeof(cmd));
2098 cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2099
2100 /*
2101 * Use MODE SELECT(10) if the device asked for it or if the mode page
2102 * and the mode select header cannot fit within the maximumm 255 bytes
2103 * of the MODE SELECT(6) command.
2104 */
2105 if (sdev->use_10_for_ms ||
2106 len + 4 > 255 ||
2107 data->block_descriptor_length > 255) {
2108 if (len > 65535 - 8)
2109 return -EINVAL;
2110 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2111 if (!real_buffer)
2112 return -ENOMEM;
2113 memcpy(real_buffer + 8, buffer, len);
2114 len += 8;
2115 real_buffer[0] = 0;
2116 real_buffer[1] = 0;
2117 real_buffer[2] = data->medium_type;
2118 real_buffer[3] = data->device_specific;
2119 real_buffer[4] = data->longlba ? 0x01 : 0;
2120 real_buffer[5] = 0;
2121 put_unaligned_be16(data->block_descriptor_length,
2122 &real_buffer[6]);
2123
2124 cmd[0] = MODE_SELECT_10;
2125 put_unaligned_be16(len, &cmd[7]);
2126 } else {
2127 if (data->longlba)
2128 return -EINVAL;
2129
2130 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2131 if (!real_buffer)
2132 return -ENOMEM;
2133 memcpy(real_buffer + 4, buffer, len);
2134 len += 4;
2135 real_buffer[0] = 0;
2136 real_buffer[1] = data->medium_type;
2137 real_buffer[2] = data->device_specific;
2138 real_buffer[3] = data->block_descriptor_length;
2139
2140 cmd[0] = MODE_SELECT;
2141 cmd[4] = len;
2142 }
2143
2144 ret = scsi_execute_cmd(sdev, cmd, REQ_OP_DRV_OUT, real_buffer, len,
2145 timeout, retries, &exec_args);
2146 kfree(real_buffer);
2147 return ret;
2148}
2149EXPORT_SYMBOL_GPL(scsi_mode_select);
2150
2151/**
2152 * scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2153 * @sdev: SCSI device to be queried
2154 * @dbd: set to prevent mode sense from returning block descriptors
2155 * @modepage: mode page being requested
2156 * @subpage: sub-page of the mode page being requested
2157 * @buffer: request buffer (may not be smaller than eight bytes)
2158 * @len: length of request buffer.
2159 * @timeout: command timeout
2160 * @retries: number of retries before failing
2161 * @data: returns a structure abstracting the mode header data
2162 * @sshdr: place to put sense data (or NULL if no sense to be collected).
2163 * must be SCSI_SENSE_BUFFERSIZE big.
2164 *
2165 * Returns zero if successful, or a negative error number on failure
2166 */
2167int
2168scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage, int subpage,
2169 unsigned char *buffer, int len, int timeout, int retries,
2170 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2171{
2172 unsigned char cmd[12];
2173 int use_10_for_ms;
2174 int header_length;
2175 int result, retry_count = retries;
2176 struct scsi_sense_hdr my_sshdr;
2177 const struct scsi_exec_args exec_args = {
2178 /* caller might not be interested in sense, but we need it */
2179 .sshdr = sshdr ? : &my_sshdr,
2180 };
2181
2182 memset(data, 0, sizeof(*data));
2183 memset(&cmd[0], 0, 12);
2184
2185 dbd = sdev->set_dbd_for_ms ? 8 : dbd;
2186 cmd[1] = dbd & 0x18; /* allows DBD and LLBA bits */
2187 cmd[2] = modepage;
2188 cmd[3] = subpage;
2189
2190 sshdr = exec_args.sshdr;
2191
2192 retry:
2193 use_10_for_ms = sdev->use_10_for_ms || len > 255;
2194
2195 if (use_10_for_ms) {
2196 if (len < 8 || len > 65535)
2197 return -EINVAL;
2198
2199 cmd[0] = MODE_SENSE_10;
2200 put_unaligned_be16(len, &cmd[7]);
2201 header_length = 8;
2202 } else {
2203 if (len < 4)
2204 return -EINVAL;
2205
2206 cmd[0] = MODE_SENSE;
2207 cmd[4] = len;
2208 header_length = 4;
2209 }
2210
2211 memset(buffer, 0, len);
2212
2213 result = scsi_execute_cmd(sdev, cmd, REQ_OP_DRV_IN, buffer, len,
2214 timeout, retries, &exec_args);
2215 if (result < 0)
2216 return result;
2217
2218 /* This code looks awful: what it's doing is making sure an
2219 * ILLEGAL REQUEST sense return identifies the actual command
2220 * byte as the problem. MODE_SENSE commands can return
2221 * ILLEGAL REQUEST if the code page isn't supported */
2222
2223 if (!scsi_status_is_good(result)) {
2224 if (scsi_sense_valid(sshdr)) {
2225 if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2226 (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2227 /*
2228 * Invalid command operation code: retry using
2229 * MODE SENSE(6) if this was a MODE SENSE(10)
2230 * request, except if the request mode page is
2231 * too large for MODE SENSE single byte
2232 * allocation length field.
2233 */
2234 if (use_10_for_ms) {
2235 if (len > 255)
2236 return -EIO;
2237 sdev->use_10_for_ms = 0;
2238 goto retry;
2239 }
2240 }
2241 if (scsi_status_is_check_condition(result) &&
2242 sshdr->sense_key == UNIT_ATTENTION &&
2243 retry_count) {
2244 retry_count--;
2245 goto retry;
2246 }
2247 }
2248 return -EIO;
2249 }
2250 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2251 (modepage == 6 || modepage == 8))) {
2252 /* Initio breakage? */
2253 header_length = 0;
2254 data->length = 13;
2255 data->medium_type = 0;
2256 data->device_specific = 0;
2257 data->longlba = 0;
2258 data->block_descriptor_length = 0;
2259 } else if (use_10_for_ms) {
2260 data->length = get_unaligned_be16(&buffer[0]) + 2;
2261 data->medium_type = buffer[2];
2262 data->device_specific = buffer[3];
2263 data->longlba = buffer[4] & 0x01;
2264 data->block_descriptor_length = get_unaligned_be16(&buffer[6]);
2265 } else {
2266 data->length = buffer[0] + 1;
2267 data->medium_type = buffer[1];
2268 data->device_specific = buffer[2];
2269 data->block_descriptor_length = buffer[3];
2270 }
2271 data->header_length = header_length;
2272
2273 return 0;
2274}
2275EXPORT_SYMBOL(scsi_mode_sense);
2276
2277/**
2278 * scsi_test_unit_ready - test if unit is ready
2279 * @sdev: scsi device to change the state of.
2280 * @timeout: command timeout
2281 * @retries: number of retries before failing
2282 * @sshdr: outpout pointer for decoded sense information.
2283 *
2284 * Returns zero if unsuccessful or an error if TUR failed. For
2285 * removable media, UNIT_ATTENTION sets ->changed flag.
2286 **/
2287int
2288scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2289 struct scsi_sense_hdr *sshdr)
2290{
2291 char cmd[] = {
2292 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2293 };
2294 const struct scsi_exec_args exec_args = {
2295 .sshdr = sshdr,
2296 };
2297 int result;
2298
2299 /* try to eat the UNIT_ATTENTION if there are enough retries */
2300 do {
2301 result = scsi_execute_cmd(sdev, cmd, REQ_OP_DRV_IN, NULL, 0,
2302 timeout, 1, &exec_args);
2303 if (sdev->removable && result > 0 && scsi_sense_valid(sshdr) &&
2304 sshdr->sense_key == UNIT_ATTENTION)
2305 sdev->changed = 1;
2306 } while (result > 0 && scsi_sense_valid(sshdr) &&
2307 sshdr->sense_key == UNIT_ATTENTION && --retries);
2308
2309 return result;
2310}
2311EXPORT_SYMBOL(scsi_test_unit_ready);
2312
2313/**
2314 * scsi_device_set_state - Take the given device through the device state model.
2315 * @sdev: scsi device to change the state of.
2316 * @state: state to change to.
2317 *
2318 * Returns zero if successful or an error if the requested
2319 * transition is illegal.
2320 */
2321int
2322scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2323{
2324 enum scsi_device_state oldstate = sdev->sdev_state;
2325
2326 if (state == oldstate)
2327 return 0;
2328
2329 switch (state) {
2330 case SDEV_CREATED:
2331 switch (oldstate) {
2332 case SDEV_CREATED_BLOCK:
2333 break;
2334 default:
2335 goto illegal;
2336 }
2337 break;
2338
2339 case SDEV_RUNNING:
2340 switch (oldstate) {
2341 case SDEV_CREATED:
2342 case SDEV_OFFLINE:
2343 case SDEV_TRANSPORT_OFFLINE:
2344 case SDEV_QUIESCE:
2345 case SDEV_BLOCK:
2346 break;
2347 default:
2348 goto illegal;
2349 }
2350 break;
2351
2352 case SDEV_QUIESCE:
2353 switch (oldstate) {
2354 case SDEV_RUNNING:
2355 case SDEV_OFFLINE:
2356 case SDEV_TRANSPORT_OFFLINE:
2357 break;
2358 default:
2359 goto illegal;
2360 }
2361 break;
2362
2363 case SDEV_OFFLINE:
2364 case SDEV_TRANSPORT_OFFLINE:
2365 switch (oldstate) {
2366 case SDEV_CREATED:
2367 case SDEV_RUNNING:
2368 case SDEV_QUIESCE:
2369 case SDEV_BLOCK:
2370 break;
2371 default:
2372 goto illegal;
2373 }
2374 break;
2375
2376 case SDEV_BLOCK:
2377 switch (oldstate) {
2378 case SDEV_RUNNING:
2379 case SDEV_CREATED_BLOCK:
2380 case SDEV_QUIESCE:
2381 case SDEV_OFFLINE:
2382 break;
2383 default:
2384 goto illegal;
2385 }
2386 break;
2387
2388 case SDEV_CREATED_BLOCK:
2389 switch (oldstate) {
2390 case SDEV_CREATED:
2391 break;
2392 default:
2393 goto illegal;
2394 }
2395 break;
2396
2397 case SDEV_CANCEL:
2398 switch (oldstate) {
2399 case SDEV_CREATED:
2400 case SDEV_RUNNING:
2401 case SDEV_QUIESCE:
2402 case SDEV_OFFLINE:
2403 case SDEV_TRANSPORT_OFFLINE:
2404 break;
2405 default:
2406 goto illegal;
2407 }
2408 break;
2409
2410 case SDEV_DEL:
2411 switch (oldstate) {
2412 case SDEV_CREATED:
2413 case SDEV_RUNNING:
2414 case SDEV_OFFLINE:
2415 case SDEV_TRANSPORT_OFFLINE:
2416 case SDEV_CANCEL:
2417 case SDEV_BLOCK:
2418 case SDEV_CREATED_BLOCK:
2419 break;
2420 default:
2421 goto illegal;
2422 }
2423 break;
2424
2425 }
2426 sdev->offline_already = false;
2427 sdev->sdev_state = state;
2428 return 0;
2429
2430 illegal:
2431 SCSI_LOG_ERROR_RECOVERY(1,
2432 sdev_printk(KERN_ERR, sdev,
2433 "Illegal state transition %s->%s",
2434 scsi_device_state_name(oldstate),
2435 scsi_device_state_name(state))
2436 );
2437 return -EINVAL;
2438}
2439EXPORT_SYMBOL(scsi_device_set_state);
2440
2441/**
2442 * scsi_evt_emit - emit a single SCSI device uevent
2443 * @sdev: associated SCSI device
2444 * @evt: event to emit
2445 *
2446 * Send a single uevent (scsi_event) to the associated scsi_device.
2447 */
2448static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2449{
2450 int idx = 0;
2451 char *envp[3];
2452
2453 switch (evt->evt_type) {
2454 case SDEV_EVT_MEDIA_CHANGE:
2455 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2456 break;
2457 case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2458 scsi_rescan_device(sdev);
2459 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2460 break;
2461 case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2462 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2463 break;
2464 case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2465 envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2466 break;
2467 case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2468 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2469 break;
2470 case SDEV_EVT_LUN_CHANGE_REPORTED:
2471 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2472 break;
2473 case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2474 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2475 break;
2476 case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2477 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2478 break;
2479 default:
2480 /* do nothing */
2481 break;
2482 }
2483
2484 envp[idx++] = NULL;
2485
2486 kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2487}
2488
2489/**
2490 * scsi_evt_thread - send a uevent for each scsi event
2491 * @work: work struct for scsi_device
2492 *
2493 * Dispatch queued events to their associated scsi_device kobjects
2494 * as uevents.
2495 */
2496void scsi_evt_thread(struct work_struct *work)
2497{
2498 struct scsi_device *sdev;
2499 enum scsi_device_event evt_type;
2500 LIST_HEAD(event_list);
2501
2502 sdev = container_of(work, struct scsi_device, event_work);
2503
2504 for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2505 if (test_and_clear_bit(evt_type, sdev->pending_events))
2506 sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2507
2508 while (1) {
2509 struct scsi_event *evt;
2510 struct list_head *this, *tmp;
2511 unsigned long flags;
2512
2513 spin_lock_irqsave(&sdev->list_lock, flags);
2514 list_splice_init(&sdev->event_list, &event_list);
2515 spin_unlock_irqrestore(&sdev->list_lock, flags);
2516
2517 if (list_empty(&event_list))
2518 break;
2519
2520 list_for_each_safe(this, tmp, &event_list) {
2521 evt = list_entry(this, struct scsi_event, node);
2522 list_del(&evt->node);
2523 scsi_evt_emit(sdev, evt);
2524 kfree(evt);
2525 }
2526 }
2527}
2528
2529/**
2530 * sdev_evt_send - send asserted event to uevent thread
2531 * @sdev: scsi_device event occurred on
2532 * @evt: event to send
2533 *
2534 * Assert scsi device event asynchronously.
2535 */
2536void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2537{
2538 unsigned long flags;
2539
2540#if 0
2541 /* FIXME: currently this check eliminates all media change events
2542 * for polled devices. Need to update to discriminate between AN
2543 * and polled events */
2544 if (!test_bit(evt->evt_type, sdev->supported_events)) {
2545 kfree(evt);
2546 return;
2547 }
2548#endif
2549
2550 spin_lock_irqsave(&sdev->list_lock, flags);
2551 list_add_tail(&evt->node, &sdev->event_list);
2552 schedule_work(&sdev->event_work);
2553 spin_unlock_irqrestore(&sdev->list_lock, flags);
2554}
2555EXPORT_SYMBOL_GPL(sdev_evt_send);
2556
2557/**
2558 * sdev_evt_alloc - allocate a new scsi event
2559 * @evt_type: type of event to allocate
2560 * @gfpflags: GFP flags for allocation
2561 *
2562 * Allocates and returns a new scsi_event.
2563 */
2564struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2565 gfp_t gfpflags)
2566{
2567 struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2568 if (!evt)
2569 return NULL;
2570
2571 evt->evt_type = evt_type;
2572 INIT_LIST_HEAD(&evt->node);
2573
2574 /* evt_type-specific initialization, if any */
2575 switch (evt_type) {
2576 case SDEV_EVT_MEDIA_CHANGE:
2577 case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2578 case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2579 case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2580 case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2581 case SDEV_EVT_LUN_CHANGE_REPORTED:
2582 case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2583 case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2584 default:
2585 /* do nothing */
2586 break;
2587 }
2588
2589 return evt;
2590}
2591EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2592
2593/**
2594 * sdev_evt_send_simple - send asserted event to uevent thread
2595 * @sdev: scsi_device event occurred on
2596 * @evt_type: type of event to send
2597 * @gfpflags: GFP flags for allocation
2598 *
2599 * Assert scsi device event asynchronously, given an event type.
2600 */
2601void sdev_evt_send_simple(struct scsi_device *sdev,
2602 enum scsi_device_event evt_type, gfp_t gfpflags)
2603{
2604 struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2605 if (!evt) {
2606 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2607 evt_type);
2608 return;
2609 }
2610
2611 sdev_evt_send(sdev, evt);
2612}
2613EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2614
2615/**
2616 * scsi_device_quiesce - Block all commands except power management.
2617 * @sdev: scsi device to quiesce.
2618 *
2619 * This works by trying to transition to the SDEV_QUIESCE state
2620 * (which must be a legal transition). When the device is in this
2621 * state, only power management requests will be accepted, all others will
2622 * be deferred.
2623 *
2624 * Must be called with user context, may sleep.
2625 *
2626 * Returns zero if unsuccessful or an error if not.
2627 */
2628int
2629scsi_device_quiesce(struct scsi_device *sdev)
2630{
2631 struct request_queue *q = sdev->request_queue;
2632 int err;
2633
2634 /*
2635 * It is allowed to call scsi_device_quiesce() multiple times from
2636 * the same context but concurrent scsi_device_quiesce() calls are
2637 * not allowed.
2638 */
2639 WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2640
2641 if (sdev->quiesced_by == current)
2642 return 0;
2643
2644 blk_set_pm_only(q);
2645
2646 blk_mq_freeze_queue(q);
2647 /*
2648 * Ensure that the effect of blk_set_pm_only() will be visible
2649 * for percpu_ref_tryget() callers that occur after the queue
2650 * unfreeze even if the queue was already frozen before this function
2651 * was called. See also https://lwn.net/Articles/573497/.
2652 */
2653 synchronize_rcu();
2654 blk_mq_unfreeze_queue(q);
2655
2656 mutex_lock(&sdev->state_mutex);
2657 err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2658 if (err == 0)
2659 sdev->quiesced_by = current;
2660 else
2661 blk_clear_pm_only(q);
2662 mutex_unlock(&sdev->state_mutex);
2663
2664 return err;
2665}
2666EXPORT_SYMBOL(scsi_device_quiesce);
2667
2668/**
2669 * scsi_device_resume - Restart user issued commands to a quiesced device.
2670 * @sdev: scsi device to resume.
2671 *
2672 * Moves the device from quiesced back to running and restarts the
2673 * queues.
2674 *
2675 * Must be called with user context, may sleep.
2676 */
2677void scsi_device_resume(struct scsi_device *sdev)
2678{
2679 /* check if the device state was mutated prior to resume, and if
2680 * so assume the state is being managed elsewhere (for example
2681 * device deleted during suspend)
2682 */
2683 mutex_lock(&sdev->state_mutex);
2684 if (sdev->sdev_state == SDEV_QUIESCE)
2685 scsi_device_set_state(sdev, SDEV_RUNNING);
2686 if (sdev->quiesced_by) {
2687 sdev->quiesced_by = NULL;
2688 blk_clear_pm_only(sdev->request_queue);
2689 }
2690 mutex_unlock(&sdev->state_mutex);
2691}
2692EXPORT_SYMBOL(scsi_device_resume);
2693
2694static void
2695device_quiesce_fn(struct scsi_device *sdev, void *data)
2696{
2697 scsi_device_quiesce(sdev);
2698}
2699
2700void
2701scsi_target_quiesce(struct scsi_target *starget)
2702{
2703 starget_for_each_device(starget, NULL, device_quiesce_fn);
2704}
2705EXPORT_SYMBOL(scsi_target_quiesce);
2706
2707static void
2708device_resume_fn(struct scsi_device *sdev, void *data)
2709{
2710 scsi_device_resume(sdev);
2711}
2712
2713void
2714scsi_target_resume(struct scsi_target *starget)
2715{
2716 starget_for_each_device(starget, NULL, device_resume_fn);
2717}
2718EXPORT_SYMBOL(scsi_target_resume);
2719
2720static int __scsi_internal_device_block_nowait(struct scsi_device *sdev)
2721{
2722 if (scsi_device_set_state(sdev, SDEV_BLOCK))
2723 return scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2724
2725 return 0;
2726}
2727
2728void scsi_start_queue(struct scsi_device *sdev)
2729{
2730 if (cmpxchg(&sdev->queue_stopped, 1, 0))
2731 blk_mq_unquiesce_queue(sdev->request_queue);
2732}
2733
2734static void scsi_stop_queue(struct scsi_device *sdev)
2735{
2736 /*
2737 * The atomic variable of ->queue_stopped covers that
2738 * blk_mq_quiesce_queue* is balanced with blk_mq_unquiesce_queue.
2739 *
2740 * The caller needs to wait until quiesce is done.
2741 */
2742 if (!cmpxchg(&sdev->queue_stopped, 0, 1))
2743 blk_mq_quiesce_queue_nowait(sdev->request_queue);
2744}
2745
2746/**
2747 * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2748 * @sdev: device to block
2749 *
2750 * Pause SCSI command processing on the specified device. Does not sleep.
2751 *
2752 * Returns zero if successful or a negative error code upon failure.
2753 *
2754 * Notes:
2755 * This routine transitions the device to the SDEV_BLOCK state (which must be
2756 * a legal transition). When the device is in this state, command processing
2757 * is paused until the device leaves the SDEV_BLOCK state. See also
2758 * scsi_internal_device_unblock_nowait().
2759 */
2760int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2761{
2762 int ret = __scsi_internal_device_block_nowait(sdev);
2763
2764 /*
2765 * The device has transitioned to SDEV_BLOCK. Stop the
2766 * block layer from calling the midlayer with this device's
2767 * request queue.
2768 */
2769 if (!ret)
2770 scsi_stop_queue(sdev);
2771 return ret;
2772}
2773EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2774
2775/**
2776 * scsi_device_block - try to transition to the SDEV_BLOCK state
2777 * @sdev: device to block
2778 * @data: dummy argument, ignored
2779 *
2780 * Pause SCSI command processing on the specified device. Callers must wait
2781 * until all ongoing scsi_queue_rq() calls have finished after this function
2782 * returns.
2783 *
2784 * Note:
2785 * This routine transitions the device to the SDEV_BLOCK state (which must be
2786 * a legal transition). When the device is in this state, command processing
2787 * is paused until the device leaves the SDEV_BLOCK state. See also
2788 * scsi_internal_device_unblock().
2789 */
2790static void scsi_device_block(struct scsi_device *sdev, void *data)
2791{
2792 int err;
2793 enum scsi_device_state state;
2794
2795 mutex_lock(&sdev->state_mutex);
2796 err = __scsi_internal_device_block_nowait(sdev);
2797 state = sdev->sdev_state;
2798 if (err == 0)
2799 /*
2800 * scsi_stop_queue() must be called with the state_mutex
2801 * held. Otherwise a simultaneous scsi_start_queue() call
2802 * might unquiesce the queue before we quiesce it.
2803 */
2804 scsi_stop_queue(sdev);
2805
2806 mutex_unlock(&sdev->state_mutex);
2807
2808 WARN_ONCE(err, "%s: failed to block %s in state %d\n",
2809 __func__, dev_name(&sdev->sdev_gendev), state);
2810}
2811
2812/**
2813 * scsi_internal_device_unblock_nowait - resume a device after a block request
2814 * @sdev: device to resume
2815 * @new_state: state to set the device to after unblocking
2816 *
2817 * Restart the device queue for a previously suspended SCSI device. Does not
2818 * sleep.
2819 *
2820 * Returns zero if successful or a negative error code upon failure.
2821 *
2822 * Notes:
2823 * This routine transitions the device to the SDEV_RUNNING state or to one of
2824 * the offline states (which must be a legal transition) allowing the midlayer
2825 * to goose the queue for this device.
2826 */
2827int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
2828 enum scsi_device_state new_state)
2829{
2830 switch (new_state) {
2831 case SDEV_RUNNING:
2832 case SDEV_TRANSPORT_OFFLINE:
2833 break;
2834 default:
2835 return -EINVAL;
2836 }
2837
2838 /*
2839 * Try to transition the scsi device to SDEV_RUNNING or one of the
2840 * offlined states and goose the device queue if successful.
2841 */
2842 switch (sdev->sdev_state) {
2843 case SDEV_BLOCK:
2844 case SDEV_TRANSPORT_OFFLINE:
2845 sdev->sdev_state = new_state;
2846 break;
2847 case SDEV_CREATED_BLOCK:
2848 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2849 new_state == SDEV_OFFLINE)
2850 sdev->sdev_state = new_state;
2851 else
2852 sdev->sdev_state = SDEV_CREATED;
2853 break;
2854 case SDEV_CANCEL:
2855 case SDEV_OFFLINE:
2856 break;
2857 default:
2858 return -EINVAL;
2859 }
2860 scsi_start_queue(sdev);
2861
2862 return 0;
2863}
2864EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
2865
2866/**
2867 * scsi_internal_device_unblock - resume a device after a block request
2868 * @sdev: device to resume
2869 * @new_state: state to set the device to after unblocking
2870 *
2871 * Restart the device queue for a previously suspended SCSI device. May sleep.
2872 *
2873 * Returns zero if successful or a negative error code upon failure.
2874 *
2875 * Notes:
2876 * This routine transitions the device to the SDEV_RUNNING state or to one of
2877 * the offline states (which must be a legal transition) allowing the midlayer
2878 * to goose the queue for this device.
2879 */
2880static int scsi_internal_device_unblock(struct scsi_device *sdev,
2881 enum scsi_device_state new_state)
2882{
2883 int ret;
2884
2885 mutex_lock(&sdev->state_mutex);
2886 ret = scsi_internal_device_unblock_nowait(sdev, new_state);
2887 mutex_unlock(&sdev->state_mutex);
2888
2889 return ret;
2890}
2891
2892static int
2893target_block(struct device *dev, void *data)
2894{
2895 if (scsi_is_target_device(dev))
2896 starget_for_each_device(to_scsi_target(dev), NULL,
2897 scsi_device_block);
2898 return 0;
2899}
2900
2901/**
2902 * scsi_block_targets - transition all SCSI child devices to SDEV_BLOCK state
2903 * @dev: a parent device of one or more scsi_target devices
2904 * @shost: the Scsi_Host to which this device belongs
2905 *
2906 * Iterate over all children of @dev, which should be scsi_target devices,
2907 * and switch all subordinate scsi devices to SDEV_BLOCK state. Wait for
2908 * ongoing scsi_queue_rq() calls to finish. May sleep.
2909 *
2910 * Note:
2911 * @dev must not itself be a scsi_target device.
2912 */
2913void
2914scsi_block_targets(struct Scsi_Host *shost, struct device *dev)
2915{
2916 WARN_ON_ONCE(scsi_is_target_device(dev));
2917 device_for_each_child(dev, NULL, target_block);
2918 blk_mq_wait_quiesce_done(&shost->tag_set);
2919}
2920EXPORT_SYMBOL_GPL(scsi_block_targets);
2921
2922static void
2923device_unblock(struct scsi_device *sdev, void *data)
2924{
2925 scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2926}
2927
2928static int
2929target_unblock(struct device *dev, void *data)
2930{
2931 if (scsi_is_target_device(dev))
2932 starget_for_each_device(to_scsi_target(dev), data,
2933 device_unblock);
2934 return 0;
2935}
2936
2937void
2938scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2939{
2940 if (scsi_is_target_device(dev))
2941 starget_for_each_device(to_scsi_target(dev), &new_state,
2942 device_unblock);
2943 else
2944 device_for_each_child(dev, &new_state, target_unblock);
2945}
2946EXPORT_SYMBOL_GPL(scsi_target_unblock);
2947
2948/**
2949 * scsi_host_block - Try to transition all logical units to the SDEV_BLOCK state
2950 * @shost: device to block
2951 *
2952 * Pause SCSI command processing for all logical units associated with the SCSI
2953 * host and wait until pending scsi_queue_rq() calls have finished.
2954 *
2955 * Returns zero if successful or a negative error code upon failure.
2956 */
2957int
2958scsi_host_block(struct Scsi_Host *shost)
2959{
2960 struct scsi_device *sdev;
2961 int ret;
2962
2963 /*
2964 * Call scsi_internal_device_block_nowait so we can avoid
2965 * calling synchronize_rcu() for each LUN.
2966 */
2967 shost_for_each_device(sdev, shost) {
2968 mutex_lock(&sdev->state_mutex);
2969 ret = scsi_internal_device_block_nowait(sdev);
2970 mutex_unlock(&sdev->state_mutex);
2971 if (ret) {
2972 scsi_device_put(sdev);
2973 return ret;
2974 }
2975 }
2976
2977 /* Wait for ongoing scsi_queue_rq() calls to finish. */
2978 blk_mq_wait_quiesce_done(&shost->tag_set);
2979
2980 return 0;
2981}
2982EXPORT_SYMBOL_GPL(scsi_host_block);
2983
2984int
2985scsi_host_unblock(struct Scsi_Host *shost, int new_state)
2986{
2987 struct scsi_device *sdev;
2988 int ret = 0;
2989
2990 shost_for_each_device(sdev, shost) {
2991 ret = scsi_internal_device_unblock(sdev, new_state);
2992 if (ret) {
2993 scsi_device_put(sdev);
2994 break;
2995 }
2996 }
2997 return ret;
2998}
2999EXPORT_SYMBOL_GPL(scsi_host_unblock);
3000
3001/**
3002 * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
3003 * @sgl: scatter-gather list
3004 * @sg_count: number of segments in sg
3005 * @offset: offset in bytes into sg, on return offset into the mapped area
3006 * @len: bytes to map, on return number of bytes mapped
3007 *
3008 * Returns virtual address of the start of the mapped page
3009 */
3010void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
3011 size_t *offset, size_t *len)
3012{
3013 int i;
3014 size_t sg_len = 0, len_complete = 0;
3015 struct scatterlist *sg;
3016 struct page *page;
3017
3018 WARN_ON(!irqs_disabled());
3019
3020 for_each_sg(sgl, sg, sg_count, i) {
3021 len_complete = sg_len; /* Complete sg-entries */
3022 sg_len += sg->length;
3023 if (sg_len > *offset)
3024 break;
3025 }
3026
3027 if (unlikely(i == sg_count)) {
3028 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
3029 "elements %d\n",
3030 __func__, sg_len, *offset, sg_count);
3031 WARN_ON(1);
3032 return NULL;
3033 }
3034
3035 /* Offset starting from the beginning of first page in this sg-entry */
3036 *offset = *offset - len_complete + sg->offset;
3037
3038 /* Assumption: contiguous pages can be accessed as "page + i" */
3039 page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
3040 *offset &= ~PAGE_MASK;
3041
3042 /* Bytes in this sg-entry from *offset to the end of the page */
3043 sg_len = PAGE_SIZE - *offset;
3044 if (*len > sg_len)
3045 *len = sg_len;
3046
3047 return kmap_atomic(page);
3048}
3049EXPORT_SYMBOL(scsi_kmap_atomic_sg);
3050
3051/**
3052 * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
3053 * @virt: virtual address to be unmapped
3054 */
3055void scsi_kunmap_atomic_sg(void *virt)
3056{
3057 kunmap_atomic(virt);
3058}
3059EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
3060
3061void sdev_disable_disk_events(struct scsi_device *sdev)
3062{
3063 atomic_inc(&sdev->disk_events_disable_depth);
3064}
3065EXPORT_SYMBOL(sdev_disable_disk_events);
3066
3067void sdev_enable_disk_events(struct scsi_device *sdev)
3068{
3069 if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
3070 return;
3071 atomic_dec(&sdev->disk_events_disable_depth);
3072}
3073EXPORT_SYMBOL(sdev_enable_disk_events);
3074
3075static unsigned char designator_prio(const unsigned char *d)
3076{
3077 if (d[1] & 0x30)
3078 /* not associated with LUN */
3079 return 0;
3080
3081 if (d[3] == 0)
3082 /* invalid length */
3083 return 0;
3084
3085 /*
3086 * Order of preference for lun descriptor:
3087 * - SCSI name string
3088 * - NAA IEEE Registered Extended
3089 * - EUI-64 based 16-byte
3090 * - EUI-64 based 12-byte
3091 * - NAA IEEE Registered
3092 * - NAA IEEE Extended
3093 * - EUI-64 based 8-byte
3094 * - SCSI name string (truncated)
3095 * - T10 Vendor ID
3096 * as longer descriptors reduce the likelyhood
3097 * of identification clashes.
3098 */
3099
3100 switch (d[1] & 0xf) {
3101 case 8:
3102 /* SCSI name string, variable-length UTF-8 */
3103 return 9;
3104 case 3:
3105 switch (d[4] >> 4) {
3106 case 6:
3107 /* NAA registered extended */
3108 return 8;
3109 case 5:
3110 /* NAA registered */
3111 return 5;
3112 case 4:
3113 /* NAA extended */
3114 return 4;
3115 case 3:
3116 /* NAA locally assigned */
3117 return 1;
3118 default:
3119 break;
3120 }
3121 break;
3122 case 2:
3123 switch (d[3]) {
3124 case 16:
3125 /* EUI64-based, 16 byte */
3126 return 7;
3127 case 12:
3128 /* EUI64-based, 12 byte */
3129 return 6;
3130 case 8:
3131 /* EUI64-based, 8 byte */
3132 return 3;
3133 default:
3134 break;
3135 }
3136 break;
3137 case 1:
3138 /* T10 vendor ID */
3139 return 1;
3140 default:
3141 break;
3142 }
3143
3144 return 0;
3145}
3146
3147/**
3148 * scsi_vpd_lun_id - return a unique device identification
3149 * @sdev: SCSI device
3150 * @id: buffer for the identification
3151 * @id_len: length of the buffer
3152 *
3153 * Copies a unique device identification into @id based
3154 * on the information in the VPD page 0x83 of the device.
3155 * The string will be formatted as a SCSI name string.
3156 *
3157 * Returns the length of the identification or error on failure.
3158 * If the identifier is longer than the supplied buffer the actual
3159 * identifier length is returned and the buffer is not zero-padded.
3160 */
3161int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
3162{
3163 u8 cur_id_prio = 0;
3164 u8 cur_id_size = 0;
3165 const unsigned char *d, *cur_id_str;
3166 const struct scsi_vpd *vpd_pg83;
3167 int id_size = -EINVAL;
3168
3169 rcu_read_lock();
3170 vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3171 if (!vpd_pg83) {
3172 rcu_read_unlock();
3173 return -ENXIO;
3174 }
3175
3176 /* The id string must be at least 20 bytes + terminating NULL byte */
3177 if (id_len < 21) {
3178 rcu_read_unlock();
3179 return -EINVAL;
3180 }
3181
3182 memset(id, 0, id_len);
3183 for (d = vpd_pg83->data + 4;
3184 d < vpd_pg83->data + vpd_pg83->len;
3185 d += d[3] + 4) {
3186 u8 prio = designator_prio(d);
3187
3188 if (prio == 0 || cur_id_prio > prio)
3189 continue;
3190
3191 switch (d[1] & 0xf) {
3192 case 0x1:
3193 /* T10 Vendor ID */
3194 if (cur_id_size > d[3])
3195 break;
3196 cur_id_prio = prio;
3197 cur_id_size = d[3];
3198 if (cur_id_size + 4 > id_len)
3199 cur_id_size = id_len - 4;
3200 cur_id_str = d + 4;
3201 id_size = snprintf(id, id_len, "t10.%*pE",
3202 cur_id_size, cur_id_str);
3203 break;
3204 case 0x2:
3205 /* EUI-64 */
3206 cur_id_prio = prio;
3207 cur_id_size = d[3];
3208 cur_id_str = d + 4;
3209 switch (cur_id_size) {
3210 case 8:
3211 id_size = snprintf(id, id_len,
3212 "eui.%8phN",
3213 cur_id_str);
3214 break;
3215 case 12:
3216 id_size = snprintf(id, id_len,
3217 "eui.%12phN",
3218 cur_id_str);
3219 break;
3220 case 16:
3221 id_size = snprintf(id, id_len,
3222 "eui.%16phN",
3223 cur_id_str);
3224 break;
3225 default:
3226 break;
3227 }
3228 break;
3229 case 0x3:
3230 /* NAA */
3231 cur_id_prio = prio;
3232 cur_id_size = d[3];
3233 cur_id_str = d + 4;
3234 switch (cur_id_size) {
3235 case 8:
3236 id_size = snprintf(id, id_len,
3237 "naa.%8phN",
3238 cur_id_str);
3239 break;
3240 case 16:
3241 id_size = snprintf(id, id_len,
3242 "naa.%16phN",
3243 cur_id_str);
3244 break;
3245 default:
3246 break;
3247 }
3248 break;
3249 case 0x8:
3250 /* SCSI name string */
3251 if (cur_id_size > d[3])
3252 break;
3253 /* Prefer others for truncated descriptor */
3254 if (d[3] > id_len) {
3255 prio = 2;
3256 if (cur_id_prio > prio)
3257 break;
3258 }
3259 cur_id_prio = prio;
3260 cur_id_size = id_size = d[3];
3261 cur_id_str = d + 4;
3262 if (cur_id_size >= id_len)
3263 cur_id_size = id_len - 1;
3264 memcpy(id, cur_id_str, cur_id_size);
3265 break;
3266 default:
3267 break;
3268 }
3269 }
3270 rcu_read_unlock();
3271
3272 return id_size;
3273}
3274EXPORT_SYMBOL(scsi_vpd_lun_id);
3275
3276/*
3277 * scsi_vpd_tpg_id - return a target port group identifier
3278 * @sdev: SCSI device
3279 *
3280 * Returns the Target Port Group identifier from the information
3281 * froom VPD page 0x83 of the device.
3282 *
3283 * Returns the identifier or error on failure.
3284 */
3285int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3286{
3287 const unsigned char *d;
3288 const struct scsi_vpd *vpd_pg83;
3289 int group_id = -EAGAIN, rel_port = -1;
3290
3291 rcu_read_lock();
3292 vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3293 if (!vpd_pg83) {
3294 rcu_read_unlock();
3295 return -ENXIO;
3296 }
3297
3298 d = vpd_pg83->data + 4;
3299 while (d < vpd_pg83->data + vpd_pg83->len) {
3300 switch (d[1] & 0xf) {
3301 case 0x4:
3302 /* Relative target port */
3303 rel_port = get_unaligned_be16(&d[6]);
3304 break;
3305 case 0x5:
3306 /* Target port group */
3307 group_id = get_unaligned_be16(&d[6]);
3308 break;
3309 default:
3310 break;
3311 }
3312 d += d[3] + 4;
3313 }
3314 rcu_read_unlock();
3315
3316 if (group_id >= 0 && rel_id && rel_port != -1)
3317 *rel_id = rel_port;
3318
3319 return group_id;
3320}
3321EXPORT_SYMBOL(scsi_vpd_tpg_id);
3322
3323/**
3324 * scsi_build_sense - build sense data for a command
3325 * @scmd: scsi command for which the sense should be formatted
3326 * @desc: Sense format (non-zero == descriptor format,
3327 * 0 == fixed format)
3328 * @key: Sense key
3329 * @asc: Additional sense code
3330 * @ascq: Additional sense code qualifier
3331 *
3332 **/
3333void scsi_build_sense(struct scsi_cmnd *scmd, int desc, u8 key, u8 asc, u8 ascq)
3334{
3335 scsi_build_sense_buffer(desc, scmd->sense_buffer, key, asc, ascq);
3336 scmd->result = SAM_STAT_CHECK_CONDITION;
3337}
3338EXPORT_SYMBOL_GPL(scsi_build_sense);
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/ratelimit.h>
25#include <asm/unaligned.h>
26
27#include <scsi/scsi.h>
28#include <scsi/scsi_cmnd.h>
29#include <scsi/scsi_dbg.h>
30#include <scsi/scsi_device.h>
31#include <scsi/scsi_driver.h>
32#include <scsi/scsi_eh.h>
33#include <scsi/scsi_host.h>
34#include <scsi/scsi_transport.h> /* __scsi_init_queue() */
35#include <scsi/scsi_dh.h>
36
37#include <trace/events/scsi.h>
38
39#include "scsi_debugfs.h"
40#include "scsi_priv.h"
41#include "scsi_logging.h"
42
43/*
44 * Size of integrity metadata is usually small, 1 inline sg should
45 * cover normal cases.
46 */
47#ifdef CONFIG_ARCH_NO_SG_CHAIN
48#define SCSI_INLINE_PROT_SG_CNT 0
49#define SCSI_INLINE_SG_CNT 0
50#else
51#define SCSI_INLINE_PROT_SG_CNT 1
52#define SCSI_INLINE_SG_CNT 2
53#endif
54
55static struct kmem_cache *scsi_sense_cache;
56static DEFINE_MUTEX(scsi_sense_cache_mutex);
57
58static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd);
59
60int scsi_init_sense_cache(struct Scsi_Host *shost)
61{
62 int ret = 0;
63
64 mutex_lock(&scsi_sense_cache_mutex);
65 if (!scsi_sense_cache) {
66 scsi_sense_cache =
67 kmem_cache_create_usercopy("scsi_sense_cache",
68 SCSI_SENSE_BUFFERSIZE, 0, SLAB_HWCACHE_ALIGN,
69 0, SCSI_SENSE_BUFFERSIZE, NULL);
70 if (!scsi_sense_cache)
71 ret = -ENOMEM;
72 }
73 mutex_unlock(&scsi_sense_cache_mutex);
74 return ret;
75}
76
77/*
78 * When to reinvoke queueing after a resource shortage. It's 3 msecs to
79 * not change behaviour from the previous unplug mechanism, experimentation
80 * may prove this needs changing.
81 */
82#define SCSI_QUEUE_DELAY 3
83
84static void
85scsi_set_blocked(struct scsi_cmnd *cmd, int reason)
86{
87 struct Scsi_Host *host = cmd->device->host;
88 struct scsi_device *device = cmd->device;
89 struct scsi_target *starget = scsi_target(device);
90
91 /*
92 * Set the appropriate busy bit for the device/host.
93 *
94 * If the host/device isn't busy, assume that something actually
95 * completed, and that we should be able to queue a command now.
96 *
97 * Note that the prior mid-layer assumption that any host could
98 * always queue at least one command is now broken. The mid-layer
99 * will implement a user specifiable stall (see
100 * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
101 * if a command is requeued with no other commands outstanding
102 * either for the device or for the host.
103 */
104 switch (reason) {
105 case SCSI_MLQUEUE_HOST_BUSY:
106 atomic_set(&host->host_blocked, host->max_host_blocked);
107 break;
108 case SCSI_MLQUEUE_DEVICE_BUSY:
109 case SCSI_MLQUEUE_EH_RETRY:
110 atomic_set(&device->device_blocked,
111 device->max_device_blocked);
112 break;
113 case SCSI_MLQUEUE_TARGET_BUSY:
114 atomic_set(&starget->target_blocked,
115 starget->max_target_blocked);
116 break;
117 }
118}
119
120static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd)
121{
122 if (cmd->request->rq_flags & RQF_DONTPREP) {
123 cmd->request->rq_flags &= ~RQF_DONTPREP;
124 scsi_mq_uninit_cmd(cmd);
125 } else {
126 WARN_ON_ONCE(true);
127 }
128 blk_mq_requeue_request(cmd->request, true);
129}
130
131/**
132 * __scsi_queue_insert - private queue insertion
133 * @cmd: The SCSI command being requeued
134 * @reason: The reason for the requeue
135 * @unbusy: Whether the queue should be unbusied
136 *
137 * This is a private queue insertion. The public interface
138 * scsi_queue_insert() always assumes the queue should be unbusied
139 * because it's always called before the completion. This function is
140 * for a requeue after completion, which should only occur in this
141 * file.
142 */
143static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, bool unbusy)
144{
145 struct scsi_device *device = cmd->device;
146
147 SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
148 "Inserting command %p into mlqueue\n", cmd));
149
150 scsi_set_blocked(cmd, reason);
151
152 /*
153 * Decrement the counters, since these commands are no longer
154 * active on the host/device.
155 */
156 if (unbusy)
157 scsi_device_unbusy(device, cmd);
158
159 /*
160 * Requeue this command. It will go before all other commands
161 * that are already in the queue. Schedule requeue work under
162 * lock such that the kblockd_schedule_work() call happens
163 * before blk_cleanup_queue() finishes.
164 */
165 cmd->result = 0;
166
167 blk_mq_requeue_request(cmd->request, true);
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/**
189 * __scsi_execute - insert request and wait for the result
190 * @sdev: scsi device
191 * @cmd: scsi command
192 * @data_direction: data direction
193 * @buffer: data buffer
194 * @bufflen: len of buffer
195 * @sense: optional sense buffer
196 * @sshdr: optional decoded sense header
197 * @timeout: request timeout in HZ
198 * @retries: number of times to retry request
199 * @flags: flags for ->cmd_flags
200 * @rq_flags: flags for ->rq_flags
201 * @resid: optional residual length
202 *
203 * Returns the scsi_cmnd result field if a command was executed, or a negative
204 * Linux error code if we didn't get that far.
205 */
206int __scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
207 int data_direction, void *buffer, unsigned bufflen,
208 unsigned char *sense, struct scsi_sense_hdr *sshdr,
209 int timeout, int retries, u64 flags, req_flags_t rq_flags,
210 int *resid)
211{
212 struct request *req;
213 struct scsi_request *rq;
214 int ret;
215
216 req = blk_get_request(sdev->request_queue,
217 data_direction == DMA_TO_DEVICE ?
218 REQ_OP_DRV_OUT : REQ_OP_DRV_IN,
219 rq_flags & RQF_PM ? BLK_MQ_REQ_PM : 0);
220 if (IS_ERR(req))
221 return PTR_ERR(req);
222
223 rq = scsi_req(req);
224
225 if (bufflen) {
226 ret = blk_rq_map_kern(sdev->request_queue, req,
227 buffer, bufflen, GFP_NOIO);
228 if (ret)
229 goto out;
230 }
231 rq->cmd_len = COMMAND_SIZE(cmd[0]);
232 memcpy(rq->cmd, cmd, rq->cmd_len);
233 rq->retries = retries;
234 req->timeout = timeout;
235 req->cmd_flags |= flags;
236 req->rq_flags |= rq_flags | RQF_QUIET;
237
238 /*
239 * head injection *required* here otherwise quiesce won't work
240 */
241 blk_execute_rq(NULL, req, 1);
242
243 /*
244 * Some devices (USB mass-storage in particular) may transfer
245 * garbage data together with a residue indicating that the data
246 * is invalid. Prevent the garbage from being misinterpreted
247 * and prevent security leaks by zeroing out the excess data.
248 */
249 if (unlikely(rq->resid_len > 0 && rq->resid_len <= bufflen))
250 memset(buffer + (bufflen - rq->resid_len), 0, rq->resid_len);
251
252 if (resid)
253 *resid = rq->resid_len;
254 if (sense && rq->sense_len)
255 memcpy(sense, rq->sense, SCSI_SENSE_BUFFERSIZE);
256 if (sshdr)
257 scsi_normalize_sense(rq->sense, rq->sense_len, sshdr);
258 ret = rq->result;
259 out:
260 blk_put_request(req);
261
262 return ret;
263}
264EXPORT_SYMBOL(__scsi_execute);
265
266/*
267 * Wake up the error handler if necessary. Avoid as follows that the error
268 * handler is not woken up if host in-flight requests number ==
269 * shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
270 * with an RCU read lock in this function to ensure that this function in
271 * its entirety either finishes before scsi_eh_scmd_add() increases the
272 * host_failed counter or that it notices the shost state change made by
273 * scsi_eh_scmd_add().
274 */
275static void scsi_dec_host_busy(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
276{
277 unsigned long flags;
278
279 rcu_read_lock();
280 __clear_bit(SCMD_STATE_INFLIGHT, &cmd->state);
281 if (unlikely(scsi_host_in_recovery(shost))) {
282 spin_lock_irqsave(shost->host_lock, flags);
283 if (shost->host_failed || shost->host_eh_scheduled)
284 scsi_eh_wakeup(shost);
285 spin_unlock_irqrestore(shost->host_lock, flags);
286 }
287 rcu_read_unlock();
288}
289
290void scsi_device_unbusy(struct scsi_device *sdev, struct scsi_cmnd *cmd)
291{
292 struct Scsi_Host *shost = sdev->host;
293 struct scsi_target *starget = scsi_target(sdev);
294
295 scsi_dec_host_busy(shost, cmd);
296
297 if (starget->can_queue > 0)
298 atomic_dec(&starget->target_busy);
299
300 sbitmap_put(&sdev->budget_map, cmd->budget_token);
301 cmd->budget_token = -1;
302}
303
304static void scsi_kick_queue(struct request_queue *q)
305{
306 blk_mq_run_hw_queues(q, false);
307}
308
309/*
310 * Called for single_lun devices on IO completion. Clear starget_sdev_user,
311 * and call blk_run_queue for all the scsi_devices on the target -
312 * including current_sdev first.
313 *
314 * Called with *no* scsi locks held.
315 */
316static void scsi_single_lun_run(struct scsi_device *current_sdev)
317{
318 struct Scsi_Host *shost = current_sdev->host;
319 struct scsi_device *sdev, *tmp;
320 struct scsi_target *starget = scsi_target(current_sdev);
321 unsigned long flags;
322
323 spin_lock_irqsave(shost->host_lock, flags);
324 starget->starget_sdev_user = NULL;
325 spin_unlock_irqrestore(shost->host_lock, flags);
326
327 /*
328 * Call blk_run_queue for all LUNs on the target, starting with
329 * current_sdev. We race with others (to set starget_sdev_user),
330 * but in most cases, we will be first. Ideally, each LU on the
331 * target would get some limited time or requests on the target.
332 */
333 scsi_kick_queue(current_sdev->request_queue);
334
335 spin_lock_irqsave(shost->host_lock, flags);
336 if (starget->starget_sdev_user)
337 goto out;
338 list_for_each_entry_safe(sdev, tmp, &starget->devices,
339 same_target_siblings) {
340 if (sdev == current_sdev)
341 continue;
342 if (scsi_device_get(sdev))
343 continue;
344
345 spin_unlock_irqrestore(shost->host_lock, flags);
346 scsi_kick_queue(sdev->request_queue);
347 spin_lock_irqsave(shost->host_lock, flags);
348
349 scsi_device_put(sdev);
350 }
351 out:
352 spin_unlock_irqrestore(shost->host_lock, flags);
353}
354
355static inline bool scsi_device_is_busy(struct scsi_device *sdev)
356{
357 if (scsi_device_busy(sdev) >= sdev->queue_depth)
358 return true;
359 if (atomic_read(&sdev->device_blocked) > 0)
360 return true;
361 return false;
362}
363
364static inline bool scsi_target_is_busy(struct scsi_target *starget)
365{
366 if (starget->can_queue > 0) {
367 if (atomic_read(&starget->target_busy) >= starget->can_queue)
368 return true;
369 if (atomic_read(&starget->target_blocked) > 0)
370 return true;
371 }
372 return false;
373}
374
375static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
376{
377 if (atomic_read(&shost->host_blocked) > 0)
378 return true;
379 if (shost->host_self_blocked)
380 return true;
381 return false;
382}
383
384static void scsi_starved_list_run(struct Scsi_Host *shost)
385{
386 LIST_HEAD(starved_list);
387 struct scsi_device *sdev;
388 unsigned long flags;
389
390 spin_lock_irqsave(shost->host_lock, flags);
391 list_splice_init(&shost->starved_list, &starved_list);
392
393 while (!list_empty(&starved_list)) {
394 struct request_queue *slq;
395
396 /*
397 * As long as shost is accepting commands and we have
398 * starved queues, call blk_run_queue. scsi_request_fn
399 * drops the queue_lock and can add us back to the
400 * starved_list.
401 *
402 * host_lock protects the starved_list and starved_entry.
403 * scsi_request_fn must get the host_lock before checking
404 * or modifying starved_list or starved_entry.
405 */
406 if (scsi_host_is_busy(shost))
407 break;
408
409 sdev = list_entry(starved_list.next,
410 struct scsi_device, starved_entry);
411 list_del_init(&sdev->starved_entry);
412 if (scsi_target_is_busy(scsi_target(sdev))) {
413 list_move_tail(&sdev->starved_entry,
414 &shost->starved_list);
415 continue;
416 }
417
418 /*
419 * Once we drop the host lock, a racing scsi_remove_device()
420 * call may remove the sdev from the starved list and destroy
421 * it and the queue. Mitigate by taking a reference to the
422 * queue and never touching the sdev again after we drop the
423 * host lock. Note: if __scsi_remove_device() invokes
424 * blk_cleanup_queue() before the queue is run from this
425 * function then blk_run_queue() will return immediately since
426 * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
427 */
428 slq = sdev->request_queue;
429 if (!blk_get_queue(slq))
430 continue;
431 spin_unlock_irqrestore(shost->host_lock, flags);
432
433 scsi_kick_queue(slq);
434 blk_put_queue(slq);
435
436 spin_lock_irqsave(shost->host_lock, flags);
437 }
438 /* put any unprocessed entries back */
439 list_splice(&starved_list, &shost->starved_list);
440 spin_unlock_irqrestore(shost->host_lock, flags);
441}
442
443/**
444 * scsi_run_queue - Select a proper request queue to serve next.
445 * @q: last request's queue
446 *
447 * The previous command was completely finished, start a new one if possible.
448 */
449static void scsi_run_queue(struct request_queue *q)
450{
451 struct scsi_device *sdev = q->queuedata;
452
453 if (scsi_target(sdev)->single_lun)
454 scsi_single_lun_run(sdev);
455 if (!list_empty(&sdev->host->starved_list))
456 scsi_starved_list_run(sdev->host);
457
458 blk_mq_run_hw_queues(q, false);
459}
460
461void scsi_requeue_run_queue(struct work_struct *work)
462{
463 struct scsi_device *sdev;
464 struct request_queue *q;
465
466 sdev = container_of(work, struct scsi_device, requeue_work);
467 q = sdev->request_queue;
468 scsi_run_queue(q);
469}
470
471void scsi_run_host_queues(struct Scsi_Host *shost)
472{
473 struct scsi_device *sdev;
474
475 shost_for_each_device(sdev, shost)
476 scsi_run_queue(sdev->request_queue);
477}
478
479static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
480{
481 if (!blk_rq_is_passthrough(cmd->request)) {
482 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
483
484 if (drv->uninit_command)
485 drv->uninit_command(cmd);
486 }
487}
488
489void scsi_free_sgtables(struct scsi_cmnd *cmd)
490{
491 if (cmd->sdb.table.nents)
492 sg_free_table_chained(&cmd->sdb.table,
493 SCSI_INLINE_SG_CNT);
494 if (scsi_prot_sg_count(cmd))
495 sg_free_table_chained(&cmd->prot_sdb->table,
496 SCSI_INLINE_PROT_SG_CNT);
497}
498EXPORT_SYMBOL_GPL(scsi_free_sgtables);
499
500static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
501{
502 scsi_free_sgtables(cmd);
503 scsi_uninit_cmd(cmd);
504}
505
506static void scsi_run_queue_async(struct scsi_device *sdev)
507{
508 if (scsi_target(sdev)->single_lun ||
509 !list_empty(&sdev->host->starved_list)) {
510 kblockd_schedule_work(&sdev->requeue_work);
511 } else {
512 /*
513 * smp_mb() present in sbitmap_queue_clear() or implied in
514 * .end_io is for ordering writing .device_busy in
515 * scsi_device_unbusy() and reading sdev->restarts.
516 */
517 int old = atomic_read(&sdev->restarts);
518
519 /*
520 * ->restarts has to be kept as non-zero if new budget
521 * contention occurs.
522 *
523 * No need to run queue when either another re-run
524 * queue wins in updating ->restarts or a new budget
525 * contention occurs.
526 */
527 if (old && atomic_cmpxchg(&sdev->restarts, old, 0) == old)
528 blk_mq_run_hw_queues(sdev->request_queue, true);
529 }
530}
531
532/* Returns false when no more bytes to process, true if there are more */
533static bool scsi_end_request(struct request *req, blk_status_t error,
534 unsigned int bytes)
535{
536 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
537 struct scsi_device *sdev = cmd->device;
538 struct request_queue *q = sdev->request_queue;
539
540 if (blk_update_request(req, error, bytes))
541 return true;
542
543 if (blk_queue_add_random(q))
544 add_disk_randomness(req->rq_disk);
545
546 if (!blk_rq_is_passthrough(req)) {
547 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
548 cmd->flags &= ~SCMD_INITIALIZED;
549 }
550
551 /*
552 * Calling rcu_barrier() is not necessary here because the
553 * SCSI error handler guarantees that the function called by
554 * call_rcu() has been called before scsi_end_request() is
555 * called.
556 */
557 destroy_rcu_head(&cmd->rcu);
558
559 /*
560 * In the MQ case the command gets freed by __blk_mq_end_request,
561 * so we have to do all cleanup that depends on it earlier.
562 *
563 * We also can't kick the queues from irq context, so we
564 * will have to defer it to a workqueue.
565 */
566 scsi_mq_uninit_cmd(cmd);
567
568 /*
569 * queue is still alive, so grab the ref for preventing it
570 * from being cleaned up during running queue.
571 */
572 percpu_ref_get(&q->q_usage_counter);
573
574 __blk_mq_end_request(req, error);
575
576 scsi_run_queue_async(sdev);
577
578 percpu_ref_put(&q->q_usage_counter);
579 return false;
580}
581
582/**
583 * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
584 * @cmd: SCSI command
585 * @result: scsi error code
586 *
587 * Translate a SCSI result code into a blk_status_t value. May reset the host
588 * byte of @cmd->result.
589 */
590static blk_status_t scsi_result_to_blk_status(struct scsi_cmnd *cmd, int result)
591{
592 switch (host_byte(result)) {
593 case DID_OK:
594 if (scsi_status_is_good(result))
595 return BLK_STS_OK;
596 return BLK_STS_IOERR;
597 case DID_TRANSPORT_FAILFAST:
598 case DID_TRANSPORT_MARGINAL:
599 return BLK_STS_TRANSPORT;
600 case DID_TARGET_FAILURE:
601 set_host_byte(cmd, DID_OK);
602 return BLK_STS_TARGET;
603 case DID_NEXUS_FAILURE:
604 set_host_byte(cmd, DID_OK);
605 return BLK_STS_NEXUS;
606 case DID_ALLOC_FAILURE:
607 set_host_byte(cmd, DID_OK);
608 return BLK_STS_NOSPC;
609 case DID_MEDIUM_ERROR:
610 set_host_byte(cmd, DID_OK);
611 return BLK_STS_MEDIUM;
612 default:
613 return BLK_STS_IOERR;
614 }
615}
616
617/* Helper for scsi_io_completion() when "reprep" action required. */
618static void scsi_io_completion_reprep(struct scsi_cmnd *cmd,
619 struct request_queue *q)
620{
621 /* A new command will be prepared and issued. */
622 scsi_mq_requeue_cmd(cmd);
623}
624
625static bool scsi_cmd_runtime_exceeced(struct scsi_cmnd *cmd)
626{
627 struct request *req = cmd->request;
628 unsigned long wait_for;
629
630 if (cmd->allowed == SCSI_CMD_RETRIES_NO_LIMIT)
631 return false;
632
633 wait_for = (cmd->allowed + 1) * req->timeout;
634 if (time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
635 scmd_printk(KERN_ERR, cmd, "timing out command, waited %lus\n",
636 wait_for/HZ);
637 return true;
638 }
639 return false;
640}
641
642/* Helper for scsi_io_completion() when special action required. */
643static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
644{
645 struct request_queue *q = cmd->device->request_queue;
646 struct request *req = cmd->request;
647 int level = 0;
648 enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
649 ACTION_DELAYED_RETRY} action;
650 struct scsi_sense_hdr sshdr;
651 bool sense_valid;
652 bool sense_current = true; /* false implies "deferred sense" */
653 blk_status_t blk_stat;
654
655 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
656 if (sense_valid)
657 sense_current = !scsi_sense_is_deferred(&sshdr);
658
659 blk_stat = scsi_result_to_blk_status(cmd, result);
660
661 if (host_byte(result) == DID_RESET) {
662 /* Third party bus reset or reset for error recovery
663 * reasons. Just retry the command and see what
664 * happens.
665 */
666 action = ACTION_RETRY;
667 } else if (sense_valid && sense_current) {
668 switch (sshdr.sense_key) {
669 case UNIT_ATTENTION:
670 if (cmd->device->removable) {
671 /* Detected disc change. Set a bit
672 * and quietly refuse further access.
673 */
674 cmd->device->changed = 1;
675 action = ACTION_FAIL;
676 } else {
677 /* Must have been a power glitch, or a
678 * bus reset. Could not have been a
679 * media change, so we just retry the
680 * command and see what happens.
681 */
682 action = ACTION_RETRY;
683 }
684 break;
685 case ILLEGAL_REQUEST:
686 /* If we had an ILLEGAL REQUEST returned, then
687 * we may have performed an unsupported
688 * command. The only thing this should be
689 * would be a ten byte read where only a six
690 * byte read was supported. Also, on a system
691 * where READ CAPACITY failed, we may have
692 * read past the end of the disk.
693 */
694 if ((cmd->device->use_10_for_rw &&
695 sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
696 (cmd->cmnd[0] == READ_10 ||
697 cmd->cmnd[0] == WRITE_10)) {
698 /* This will issue a new 6-byte command. */
699 cmd->device->use_10_for_rw = 0;
700 action = ACTION_REPREP;
701 } else if (sshdr.asc == 0x10) /* DIX */ {
702 action = ACTION_FAIL;
703 blk_stat = BLK_STS_PROTECTION;
704 /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
705 } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
706 action = ACTION_FAIL;
707 blk_stat = BLK_STS_TARGET;
708 } else
709 action = ACTION_FAIL;
710 break;
711 case ABORTED_COMMAND:
712 action = ACTION_FAIL;
713 if (sshdr.asc == 0x10) /* DIF */
714 blk_stat = BLK_STS_PROTECTION;
715 break;
716 case NOT_READY:
717 /* If the device is in the process of becoming
718 * ready, or has a temporary blockage, retry.
719 */
720 if (sshdr.asc == 0x04) {
721 switch (sshdr.ascq) {
722 case 0x01: /* becoming ready */
723 case 0x04: /* format in progress */
724 case 0x05: /* rebuild in progress */
725 case 0x06: /* recalculation in progress */
726 case 0x07: /* operation in progress */
727 case 0x08: /* Long write in progress */
728 case 0x09: /* self test in progress */
729 case 0x11: /* notify (enable spinup) required */
730 case 0x14: /* space allocation in progress */
731 case 0x1a: /* start stop unit in progress */
732 case 0x1b: /* sanitize in progress */
733 case 0x1d: /* configuration in progress */
734 case 0x24: /* depopulation in progress */
735 action = ACTION_DELAYED_RETRY;
736 break;
737 case 0x0a: /* ALUA state transition */
738 blk_stat = BLK_STS_AGAIN;
739 fallthrough;
740 default:
741 action = ACTION_FAIL;
742 break;
743 }
744 } else
745 action = ACTION_FAIL;
746 break;
747 case VOLUME_OVERFLOW:
748 /* See SSC3rXX or current. */
749 action = ACTION_FAIL;
750 break;
751 case DATA_PROTECT:
752 action = ACTION_FAIL;
753 if ((sshdr.asc == 0x0C && sshdr.ascq == 0x12) ||
754 (sshdr.asc == 0x55 &&
755 (sshdr.ascq == 0x0E || sshdr.ascq == 0x0F))) {
756 /* Insufficient zone resources */
757 blk_stat = BLK_STS_ZONE_OPEN_RESOURCE;
758 }
759 break;
760 default:
761 action = ACTION_FAIL;
762 break;
763 }
764 } else
765 action = ACTION_FAIL;
766
767 if (action != ACTION_FAIL && scsi_cmd_runtime_exceeced(cmd))
768 action = ACTION_FAIL;
769
770 switch (action) {
771 case ACTION_FAIL:
772 /* Give up and fail the remainder of the request */
773 if (!(req->rq_flags & RQF_QUIET)) {
774 static DEFINE_RATELIMIT_STATE(_rs,
775 DEFAULT_RATELIMIT_INTERVAL,
776 DEFAULT_RATELIMIT_BURST);
777
778 if (unlikely(scsi_logging_level))
779 level =
780 SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
781 SCSI_LOG_MLCOMPLETE_BITS);
782
783 /*
784 * if logging is enabled the failure will be printed
785 * in scsi_log_completion(), so avoid duplicate messages
786 */
787 if (!level && __ratelimit(&_rs)) {
788 scsi_print_result(cmd, NULL, FAILED);
789 if (sense_valid)
790 scsi_print_sense(cmd);
791 scsi_print_command(cmd);
792 }
793 }
794 if (!scsi_end_request(req, blk_stat, blk_rq_err_bytes(req)))
795 return;
796 fallthrough;
797 case ACTION_REPREP:
798 scsi_io_completion_reprep(cmd, q);
799 break;
800 case ACTION_RETRY:
801 /* Retry the same command immediately */
802 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
803 break;
804 case ACTION_DELAYED_RETRY:
805 /* Retry the same command after a delay */
806 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
807 break;
808 }
809}
810
811/*
812 * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
813 * new result that may suppress further error checking. Also modifies
814 * *blk_statp in some cases.
815 */
816static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
817 blk_status_t *blk_statp)
818{
819 bool sense_valid;
820 bool sense_current = true; /* false implies "deferred sense" */
821 struct request *req = cmd->request;
822 struct scsi_sense_hdr sshdr;
823
824 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
825 if (sense_valid)
826 sense_current = !scsi_sense_is_deferred(&sshdr);
827
828 if (blk_rq_is_passthrough(req)) {
829 if (sense_valid) {
830 /*
831 * SG_IO wants current and deferred errors
832 */
833 scsi_req(req)->sense_len =
834 min(8 + cmd->sense_buffer[7],
835 SCSI_SENSE_BUFFERSIZE);
836 }
837 if (sense_current)
838 *blk_statp = scsi_result_to_blk_status(cmd, result);
839 } else if (blk_rq_bytes(req) == 0 && sense_current) {
840 /*
841 * Flush commands do not transfers any data, and thus cannot use
842 * good_bytes != blk_rq_bytes(req) as the signal for an error.
843 * This sets *blk_statp explicitly for the problem case.
844 */
845 *blk_statp = scsi_result_to_blk_status(cmd, result);
846 }
847 /*
848 * Recovered errors need reporting, but they're always treated as
849 * success, so fiddle the result code here. For passthrough requests
850 * we already took a copy of the original into sreq->result which
851 * is what gets returned to the user
852 */
853 if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
854 bool do_print = true;
855 /*
856 * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
857 * skip print since caller wants ATA registers. Only occurs
858 * on SCSI ATA PASS_THROUGH commands when CK_COND=1
859 */
860 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
861 do_print = false;
862 else if (req->rq_flags & RQF_QUIET)
863 do_print = false;
864 if (do_print)
865 scsi_print_sense(cmd);
866 result = 0;
867 /* for passthrough, *blk_statp may be set */
868 *blk_statp = BLK_STS_OK;
869 }
870 /*
871 * Another corner case: the SCSI status byte is non-zero but 'good'.
872 * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
873 * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
874 * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
875 * intermediate statuses (both obsolete in SAM-4) as good.
876 */
877 if ((result & 0xff) && scsi_status_is_good(result)) {
878 result = 0;
879 *blk_statp = BLK_STS_OK;
880 }
881 return result;
882}
883
884/**
885 * scsi_io_completion - Completion processing for SCSI commands.
886 * @cmd: command that is finished.
887 * @good_bytes: number of processed bytes.
888 *
889 * We will finish off the specified number of sectors. If we are done, the
890 * command block will be released and the queue function will be goosed. If we
891 * are not done then we have to figure out what to do next:
892 *
893 * a) We can call scsi_io_completion_reprep(). The request will be
894 * unprepared and put back on the queue. Then a new command will
895 * be created for it. This should be used if we made forward
896 * progress, or if we want to switch from READ(10) to READ(6) for
897 * example.
898 *
899 * b) We can call scsi_io_completion_action(). The request will be
900 * put back on the queue and retried using the same command as
901 * before, possibly after a delay.
902 *
903 * c) We can call scsi_end_request() with blk_stat other than
904 * BLK_STS_OK, to fail the remainder of the request.
905 */
906void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
907{
908 int result = cmd->result;
909 struct request_queue *q = cmd->device->request_queue;
910 struct request *req = cmd->request;
911 blk_status_t blk_stat = BLK_STS_OK;
912
913 if (unlikely(result)) /* a nz result may or may not be an error */
914 result = scsi_io_completion_nz_result(cmd, result, &blk_stat);
915
916 if (unlikely(blk_rq_is_passthrough(req))) {
917 /*
918 * scsi_result_to_blk_status may have reset the host_byte
919 */
920 scsi_req(req)->result = cmd->result;
921 }
922
923 /*
924 * Next deal with any sectors which we were able to correctly
925 * handle.
926 */
927 SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
928 "%u sectors total, %d bytes done.\n",
929 blk_rq_sectors(req), good_bytes));
930
931 /*
932 * Failed, zero length commands always need to drop down
933 * to retry code. Fast path should return in this block.
934 */
935 if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
936 if (likely(!scsi_end_request(req, blk_stat, good_bytes)))
937 return; /* no bytes remaining */
938 }
939
940 /* Kill remainder if no retries. */
941 if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
942 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req)))
943 WARN_ONCE(true,
944 "Bytes remaining after failed, no-retry command");
945 return;
946 }
947
948 /*
949 * If there had been no error, but we have leftover bytes in the
950 * requeues just queue the command up again.
951 */
952 if (likely(result == 0))
953 scsi_io_completion_reprep(cmd, q);
954 else
955 scsi_io_completion_action(cmd, result);
956}
957
958static inline bool scsi_cmd_needs_dma_drain(struct scsi_device *sdev,
959 struct request *rq)
960{
961 return sdev->dma_drain_len && blk_rq_is_passthrough(rq) &&
962 !op_is_write(req_op(rq)) &&
963 sdev->host->hostt->dma_need_drain(rq);
964}
965
966/**
967 * scsi_alloc_sgtables - Allocate and initialize data and integrity scatterlists
968 * @cmd: SCSI command data structure to initialize.
969 *
970 * Initializes @cmd->sdb and also @cmd->prot_sdb if data integrity is enabled
971 * for @cmd.
972 *
973 * Returns:
974 * * BLK_STS_OK - on success
975 * * BLK_STS_RESOURCE - if the failure is retryable
976 * * BLK_STS_IOERR - if the failure is fatal
977 */
978blk_status_t scsi_alloc_sgtables(struct scsi_cmnd *cmd)
979{
980 struct scsi_device *sdev = cmd->device;
981 struct request *rq = cmd->request;
982 unsigned short nr_segs = blk_rq_nr_phys_segments(rq);
983 struct scatterlist *last_sg = NULL;
984 blk_status_t ret;
985 bool need_drain = scsi_cmd_needs_dma_drain(sdev, rq);
986 int count;
987
988 if (WARN_ON_ONCE(!nr_segs))
989 return BLK_STS_IOERR;
990
991 /*
992 * Make sure there is space for the drain. The driver must adjust
993 * max_hw_segments to be prepared for this.
994 */
995 if (need_drain)
996 nr_segs++;
997
998 /*
999 * If sg table allocation fails, requeue request later.
1000 */
1001 if (unlikely(sg_alloc_table_chained(&cmd->sdb.table, nr_segs,
1002 cmd->sdb.table.sgl, SCSI_INLINE_SG_CNT)))
1003 return BLK_STS_RESOURCE;
1004
1005 /*
1006 * Next, walk the list, and fill in the addresses and sizes of
1007 * each segment.
1008 */
1009 count = __blk_rq_map_sg(rq->q, rq, cmd->sdb.table.sgl, &last_sg);
1010
1011 if (blk_rq_bytes(rq) & rq->q->dma_pad_mask) {
1012 unsigned int pad_len =
1013 (rq->q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
1014
1015 last_sg->length += pad_len;
1016 cmd->extra_len += pad_len;
1017 }
1018
1019 if (need_drain) {
1020 sg_unmark_end(last_sg);
1021 last_sg = sg_next(last_sg);
1022 sg_set_buf(last_sg, sdev->dma_drain_buf, sdev->dma_drain_len);
1023 sg_mark_end(last_sg);
1024
1025 cmd->extra_len += sdev->dma_drain_len;
1026 count++;
1027 }
1028
1029 BUG_ON(count > cmd->sdb.table.nents);
1030 cmd->sdb.table.nents = count;
1031 cmd->sdb.length = blk_rq_payload_bytes(rq);
1032
1033 if (blk_integrity_rq(rq)) {
1034 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1035 int ivecs;
1036
1037 if (WARN_ON_ONCE(!prot_sdb)) {
1038 /*
1039 * This can happen if someone (e.g. multipath)
1040 * queues a command to a device on an adapter
1041 * that does not support DIX.
1042 */
1043 ret = BLK_STS_IOERR;
1044 goto out_free_sgtables;
1045 }
1046
1047 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1048
1049 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1050 prot_sdb->table.sgl,
1051 SCSI_INLINE_PROT_SG_CNT)) {
1052 ret = BLK_STS_RESOURCE;
1053 goto out_free_sgtables;
1054 }
1055
1056 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1057 prot_sdb->table.sgl);
1058 BUG_ON(count > ivecs);
1059 BUG_ON(count > queue_max_integrity_segments(rq->q));
1060
1061 cmd->prot_sdb = prot_sdb;
1062 cmd->prot_sdb->table.nents = count;
1063 }
1064
1065 return BLK_STS_OK;
1066out_free_sgtables:
1067 scsi_free_sgtables(cmd);
1068 return ret;
1069}
1070EXPORT_SYMBOL(scsi_alloc_sgtables);
1071
1072/**
1073 * scsi_initialize_rq - initialize struct scsi_cmnd partially
1074 * @rq: Request associated with the SCSI command to be initialized.
1075 *
1076 * This function initializes the members of struct scsi_cmnd that must be
1077 * initialized before request processing starts and that won't be
1078 * reinitialized if a SCSI command is requeued.
1079 *
1080 * Called from inside blk_get_request() for pass-through requests and from
1081 * inside scsi_init_command() for filesystem requests.
1082 */
1083static void scsi_initialize_rq(struct request *rq)
1084{
1085 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1086
1087 scsi_req_init(&cmd->req);
1088 init_rcu_head(&cmd->rcu);
1089 cmd->jiffies_at_alloc = jiffies;
1090 cmd->retries = 0;
1091}
1092
1093/*
1094 * Only called when the request isn't completed by SCSI, and not freed by
1095 * SCSI
1096 */
1097static void scsi_cleanup_rq(struct request *rq)
1098{
1099 if (rq->rq_flags & RQF_DONTPREP) {
1100 scsi_mq_uninit_cmd(blk_mq_rq_to_pdu(rq));
1101 rq->rq_flags &= ~RQF_DONTPREP;
1102 }
1103}
1104
1105/* Called before a request is prepared. See also scsi_mq_prep_fn(). */
1106void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1107{
1108 void *buf = cmd->sense_buffer;
1109 void *prot = cmd->prot_sdb;
1110 struct request *rq = blk_mq_rq_from_pdu(cmd);
1111 unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1112 unsigned long jiffies_at_alloc;
1113 int retries, to_clear;
1114 bool in_flight;
1115 int budget_token = cmd->budget_token;
1116
1117 if (!blk_rq_is_passthrough(rq) && !(flags & SCMD_INITIALIZED)) {
1118 flags |= SCMD_INITIALIZED;
1119 scsi_initialize_rq(rq);
1120 }
1121
1122 jiffies_at_alloc = cmd->jiffies_at_alloc;
1123 retries = cmd->retries;
1124 in_flight = test_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1125 /*
1126 * Zero out the cmd, except for the embedded scsi_request. Only clear
1127 * the driver-private command data if the LLD does not supply a
1128 * function to initialize that data.
1129 */
1130 to_clear = sizeof(*cmd) - sizeof(cmd->req);
1131 if (!dev->host->hostt->init_cmd_priv)
1132 to_clear += dev->host->hostt->cmd_size;
1133 memset((char *)cmd + sizeof(cmd->req), 0, to_clear);
1134
1135 cmd->device = dev;
1136 cmd->sense_buffer = buf;
1137 cmd->prot_sdb = prot;
1138 cmd->flags = flags;
1139 INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1140 cmd->jiffies_at_alloc = jiffies_at_alloc;
1141 cmd->retries = retries;
1142 if (in_flight)
1143 __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1144 cmd->budget_token = budget_token;
1145
1146}
1147
1148static blk_status_t scsi_setup_scsi_cmnd(struct scsi_device *sdev,
1149 struct request *req)
1150{
1151 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1152
1153 /*
1154 * Passthrough requests may transfer data, in which case they must
1155 * a bio attached to them. Or they might contain a SCSI command
1156 * that does not transfer data, in which case they may optionally
1157 * submit a request without an attached bio.
1158 */
1159 if (req->bio) {
1160 blk_status_t ret = scsi_alloc_sgtables(cmd);
1161 if (unlikely(ret != BLK_STS_OK))
1162 return ret;
1163 } else {
1164 BUG_ON(blk_rq_bytes(req));
1165
1166 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1167 }
1168
1169 cmd->cmd_len = scsi_req(req)->cmd_len;
1170 if (cmd->cmd_len == 0)
1171 cmd->cmd_len = scsi_command_size(cmd->cmnd);
1172 cmd->cmnd = scsi_req(req)->cmd;
1173 cmd->transfersize = blk_rq_bytes(req);
1174 cmd->allowed = scsi_req(req)->retries;
1175 return BLK_STS_OK;
1176}
1177
1178static blk_status_t
1179scsi_device_state_check(struct scsi_device *sdev, struct request *req)
1180{
1181 switch (sdev->sdev_state) {
1182 case SDEV_CREATED:
1183 return BLK_STS_OK;
1184 case SDEV_OFFLINE:
1185 case SDEV_TRANSPORT_OFFLINE:
1186 /*
1187 * If the device is offline we refuse to process any
1188 * commands. The device must be brought online
1189 * before trying any recovery commands.
1190 */
1191 if (!sdev->offline_already) {
1192 sdev->offline_already = true;
1193 sdev_printk(KERN_ERR, sdev,
1194 "rejecting I/O to offline device\n");
1195 }
1196 return BLK_STS_IOERR;
1197 case SDEV_DEL:
1198 /*
1199 * If the device is fully deleted, we refuse to
1200 * process any commands as well.
1201 */
1202 sdev_printk(KERN_ERR, sdev,
1203 "rejecting I/O to dead device\n");
1204 return BLK_STS_IOERR;
1205 case SDEV_BLOCK:
1206 case SDEV_CREATED_BLOCK:
1207 return BLK_STS_RESOURCE;
1208 case SDEV_QUIESCE:
1209 /*
1210 * If the device is blocked we only accept power management
1211 * commands.
1212 */
1213 if (req && WARN_ON_ONCE(!(req->rq_flags & RQF_PM)))
1214 return BLK_STS_RESOURCE;
1215 return BLK_STS_OK;
1216 default:
1217 /*
1218 * For any other not fully online state we only allow
1219 * power management commands.
1220 */
1221 if (req && !(req->rq_flags & RQF_PM))
1222 return BLK_STS_IOERR;
1223 return BLK_STS_OK;
1224 }
1225}
1226
1227/*
1228 * scsi_dev_queue_ready: if we can send requests to sdev, assign one token
1229 * and return the token else return -1.
1230 */
1231static inline int scsi_dev_queue_ready(struct request_queue *q,
1232 struct scsi_device *sdev)
1233{
1234 int token;
1235
1236 token = sbitmap_get(&sdev->budget_map);
1237 if (atomic_read(&sdev->device_blocked)) {
1238 if (token < 0)
1239 goto out;
1240
1241 if (scsi_device_busy(sdev) > 1)
1242 goto out_dec;
1243
1244 /*
1245 * unblock after device_blocked iterates to zero
1246 */
1247 if (atomic_dec_return(&sdev->device_blocked) > 0)
1248 goto out_dec;
1249 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1250 "unblocking device at zero depth\n"));
1251 }
1252
1253 return token;
1254out_dec:
1255 if (token >= 0)
1256 sbitmap_put(&sdev->budget_map, token);
1257out:
1258 return -1;
1259}
1260
1261/*
1262 * scsi_target_queue_ready: checks if there we can send commands to target
1263 * @sdev: scsi device on starget to check.
1264 */
1265static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1266 struct scsi_device *sdev)
1267{
1268 struct scsi_target *starget = scsi_target(sdev);
1269 unsigned int busy;
1270
1271 if (starget->single_lun) {
1272 spin_lock_irq(shost->host_lock);
1273 if (starget->starget_sdev_user &&
1274 starget->starget_sdev_user != sdev) {
1275 spin_unlock_irq(shost->host_lock);
1276 return 0;
1277 }
1278 starget->starget_sdev_user = sdev;
1279 spin_unlock_irq(shost->host_lock);
1280 }
1281
1282 if (starget->can_queue <= 0)
1283 return 1;
1284
1285 busy = atomic_inc_return(&starget->target_busy) - 1;
1286 if (atomic_read(&starget->target_blocked) > 0) {
1287 if (busy)
1288 goto starved;
1289
1290 /*
1291 * unblock after target_blocked iterates to zero
1292 */
1293 if (atomic_dec_return(&starget->target_blocked) > 0)
1294 goto out_dec;
1295
1296 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1297 "unblocking target at zero depth\n"));
1298 }
1299
1300 if (busy >= starget->can_queue)
1301 goto starved;
1302
1303 return 1;
1304
1305starved:
1306 spin_lock_irq(shost->host_lock);
1307 list_move_tail(&sdev->starved_entry, &shost->starved_list);
1308 spin_unlock_irq(shost->host_lock);
1309out_dec:
1310 if (starget->can_queue > 0)
1311 atomic_dec(&starget->target_busy);
1312 return 0;
1313}
1314
1315/*
1316 * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1317 * return 0. We must end up running the queue again whenever 0 is
1318 * returned, else IO can hang.
1319 */
1320static inline int scsi_host_queue_ready(struct request_queue *q,
1321 struct Scsi_Host *shost,
1322 struct scsi_device *sdev,
1323 struct scsi_cmnd *cmd)
1324{
1325 if (scsi_host_in_recovery(shost))
1326 return 0;
1327
1328 if (atomic_read(&shost->host_blocked) > 0) {
1329 if (scsi_host_busy(shost) > 0)
1330 goto starved;
1331
1332 /*
1333 * unblock after host_blocked iterates to zero
1334 */
1335 if (atomic_dec_return(&shost->host_blocked) > 0)
1336 goto out_dec;
1337
1338 SCSI_LOG_MLQUEUE(3,
1339 shost_printk(KERN_INFO, shost,
1340 "unblocking host at zero depth\n"));
1341 }
1342
1343 if (shost->host_self_blocked)
1344 goto starved;
1345
1346 /* We're OK to process the command, so we can't be starved */
1347 if (!list_empty(&sdev->starved_entry)) {
1348 spin_lock_irq(shost->host_lock);
1349 if (!list_empty(&sdev->starved_entry))
1350 list_del_init(&sdev->starved_entry);
1351 spin_unlock_irq(shost->host_lock);
1352 }
1353
1354 __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1355
1356 return 1;
1357
1358starved:
1359 spin_lock_irq(shost->host_lock);
1360 if (list_empty(&sdev->starved_entry))
1361 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1362 spin_unlock_irq(shost->host_lock);
1363out_dec:
1364 scsi_dec_host_busy(shost, cmd);
1365 return 0;
1366}
1367
1368/*
1369 * Busy state exporting function for request stacking drivers.
1370 *
1371 * For efficiency, no lock is taken to check the busy state of
1372 * shost/starget/sdev, since the returned value is not guaranteed and
1373 * may be changed after request stacking drivers call the function,
1374 * regardless of taking lock or not.
1375 *
1376 * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1377 * needs to return 'not busy'. Otherwise, request stacking drivers
1378 * may hold requests forever.
1379 */
1380static bool scsi_mq_lld_busy(struct request_queue *q)
1381{
1382 struct scsi_device *sdev = q->queuedata;
1383 struct Scsi_Host *shost;
1384
1385 if (blk_queue_dying(q))
1386 return false;
1387
1388 shost = sdev->host;
1389
1390 /*
1391 * Ignore host/starget busy state.
1392 * Since block layer does not have a concept of fairness across
1393 * multiple queues, congestion of host/starget needs to be handled
1394 * in SCSI layer.
1395 */
1396 if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1397 return true;
1398
1399 return false;
1400}
1401
1402/*
1403 * Block layer request completion callback. May be called from interrupt
1404 * context.
1405 */
1406static void scsi_complete(struct request *rq)
1407{
1408 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1409 enum scsi_disposition disposition;
1410
1411 INIT_LIST_HEAD(&cmd->eh_entry);
1412
1413 atomic_inc(&cmd->device->iodone_cnt);
1414 if (cmd->result)
1415 atomic_inc(&cmd->device->ioerr_cnt);
1416
1417 disposition = scsi_decide_disposition(cmd);
1418 if (disposition != SUCCESS && scsi_cmd_runtime_exceeced(cmd))
1419 disposition = SUCCESS;
1420
1421 scsi_log_completion(cmd, disposition);
1422
1423 switch (disposition) {
1424 case SUCCESS:
1425 scsi_finish_command(cmd);
1426 break;
1427 case NEEDS_RETRY:
1428 scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1429 break;
1430 case ADD_TO_MLQUEUE:
1431 scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1432 break;
1433 default:
1434 scsi_eh_scmd_add(cmd);
1435 break;
1436 }
1437}
1438
1439/**
1440 * scsi_dispatch_cmd - Dispatch a command to the low-level driver.
1441 * @cmd: command block we are dispatching.
1442 *
1443 * Return: nonzero return request was rejected and device's queue needs to be
1444 * plugged.
1445 */
1446static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1447{
1448 struct Scsi_Host *host = cmd->device->host;
1449 int rtn = 0;
1450
1451 atomic_inc(&cmd->device->iorequest_cnt);
1452
1453 /* check if the device is still usable */
1454 if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1455 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1456 * returns an immediate error upwards, and signals
1457 * that the device is no longer present */
1458 cmd->result = DID_NO_CONNECT << 16;
1459 goto done;
1460 }
1461
1462 /* Check to see if the scsi lld made this device blocked. */
1463 if (unlikely(scsi_device_blocked(cmd->device))) {
1464 /*
1465 * in blocked state, the command is just put back on
1466 * the device queue. The suspend state has already
1467 * blocked the queue so future requests should not
1468 * occur until the device transitions out of the
1469 * suspend state.
1470 */
1471 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1472 "queuecommand : device blocked\n"));
1473 return SCSI_MLQUEUE_DEVICE_BUSY;
1474 }
1475
1476 /* Store the LUN value in cmnd, if needed. */
1477 if (cmd->device->lun_in_cdb)
1478 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1479 (cmd->device->lun << 5 & 0xe0);
1480
1481 scsi_log_send(cmd);
1482
1483 /*
1484 * Before we queue this command, check if the command
1485 * length exceeds what the host adapter can handle.
1486 */
1487 if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1488 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1489 "queuecommand : command too long. "
1490 "cdb_size=%d host->max_cmd_len=%d\n",
1491 cmd->cmd_len, cmd->device->host->max_cmd_len));
1492 cmd->result = (DID_ABORT << 16);
1493 goto done;
1494 }
1495
1496 if (unlikely(host->shost_state == SHOST_DEL)) {
1497 cmd->result = (DID_NO_CONNECT << 16);
1498 goto done;
1499
1500 }
1501
1502 trace_scsi_dispatch_cmd_start(cmd);
1503 rtn = host->hostt->queuecommand(host, cmd);
1504 if (rtn) {
1505 trace_scsi_dispatch_cmd_error(cmd, rtn);
1506 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1507 rtn != SCSI_MLQUEUE_TARGET_BUSY)
1508 rtn = SCSI_MLQUEUE_HOST_BUSY;
1509
1510 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1511 "queuecommand : request rejected\n"));
1512 }
1513
1514 return rtn;
1515 done:
1516 cmd->scsi_done(cmd);
1517 return 0;
1518}
1519
1520/* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1521static unsigned int scsi_mq_inline_sgl_size(struct Scsi_Host *shost)
1522{
1523 return min_t(unsigned int, shost->sg_tablesize, SCSI_INLINE_SG_CNT) *
1524 sizeof(struct scatterlist);
1525}
1526
1527static blk_status_t scsi_prepare_cmd(struct request *req)
1528{
1529 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1530 struct scsi_device *sdev = req->q->queuedata;
1531 struct Scsi_Host *shost = sdev->host;
1532 struct scatterlist *sg;
1533
1534 scsi_init_command(sdev, cmd);
1535
1536 cmd->request = req;
1537 cmd->tag = req->tag;
1538 cmd->prot_op = SCSI_PROT_NORMAL;
1539 if (blk_rq_bytes(req))
1540 cmd->sc_data_direction = rq_dma_dir(req);
1541 else
1542 cmd->sc_data_direction = DMA_NONE;
1543
1544 sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1545 cmd->sdb.table.sgl = sg;
1546
1547 if (scsi_host_get_prot(shost)) {
1548 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1549
1550 cmd->prot_sdb->table.sgl =
1551 (struct scatterlist *)(cmd->prot_sdb + 1);
1552 }
1553
1554 /*
1555 * Special handling for passthrough commands, which don't go to the ULP
1556 * at all:
1557 */
1558 if (blk_rq_is_passthrough(req))
1559 return scsi_setup_scsi_cmnd(sdev, req);
1560
1561 if (sdev->handler && sdev->handler->prep_fn) {
1562 blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1563
1564 if (ret != BLK_STS_OK)
1565 return ret;
1566 }
1567
1568 cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1569 memset(cmd->cmnd, 0, BLK_MAX_CDB);
1570 return scsi_cmd_to_driver(cmd)->init_command(cmd);
1571}
1572
1573static void scsi_mq_done(struct scsi_cmnd *cmd)
1574{
1575 if (unlikely(blk_should_fake_timeout(cmd->request->q)))
1576 return;
1577 if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1578 return;
1579 trace_scsi_dispatch_cmd_done(cmd);
1580 blk_mq_complete_request(cmd->request);
1581}
1582
1583static void scsi_mq_put_budget(struct request_queue *q, int budget_token)
1584{
1585 struct scsi_device *sdev = q->queuedata;
1586
1587 sbitmap_put(&sdev->budget_map, budget_token);
1588}
1589
1590static int scsi_mq_get_budget(struct request_queue *q)
1591{
1592 struct scsi_device *sdev = q->queuedata;
1593 int token = scsi_dev_queue_ready(q, sdev);
1594
1595 if (token >= 0)
1596 return token;
1597
1598 atomic_inc(&sdev->restarts);
1599
1600 /*
1601 * Orders atomic_inc(&sdev->restarts) and atomic_read(&sdev->device_busy).
1602 * .restarts must be incremented before .device_busy is read because the
1603 * code in scsi_run_queue_async() depends on the order of these operations.
1604 */
1605 smp_mb__after_atomic();
1606
1607 /*
1608 * If all in-flight requests originated from this LUN are completed
1609 * before reading .device_busy, sdev->device_busy will be observed as
1610 * zero, then blk_mq_delay_run_hw_queues() will dispatch this request
1611 * soon. Otherwise, completion of one of these requests will observe
1612 * the .restarts flag, and the request queue will be run for handling
1613 * this request, see scsi_end_request().
1614 */
1615 if (unlikely(scsi_device_busy(sdev) == 0 &&
1616 !scsi_device_blocked(sdev)))
1617 blk_mq_delay_run_hw_queues(sdev->request_queue, SCSI_QUEUE_DELAY);
1618 return -1;
1619}
1620
1621static void scsi_mq_set_rq_budget_token(struct request *req, int token)
1622{
1623 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1624
1625 cmd->budget_token = token;
1626}
1627
1628static int scsi_mq_get_rq_budget_token(struct request *req)
1629{
1630 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1631
1632 return cmd->budget_token;
1633}
1634
1635static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1636 const struct blk_mq_queue_data *bd)
1637{
1638 struct request *req = bd->rq;
1639 struct request_queue *q = req->q;
1640 struct scsi_device *sdev = q->queuedata;
1641 struct Scsi_Host *shost = sdev->host;
1642 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1643 blk_status_t ret;
1644 int reason;
1645
1646 WARN_ON_ONCE(cmd->budget_token < 0);
1647
1648 /*
1649 * If the device is not in running state we will reject some or all
1650 * commands.
1651 */
1652 if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1653 ret = scsi_device_state_check(sdev, req);
1654 if (ret != BLK_STS_OK)
1655 goto out_put_budget;
1656 }
1657
1658 ret = BLK_STS_RESOURCE;
1659 if (!scsi_target_queue_ready(shost, sdev))
1660 goto out_put_budget;
1661 if (!scsi_host_queue_ready(q, shost, sdev, cmd))
1662 goto out_dec_target_busy;
1663
1664 if (!(req->rq_flags & RQF_DONTPREP)) {
1665 ret = scsi_prepare_cmd(req);
1666 if (ret != BLK_STS_OK)
1667 goto out_dec_host_busy;
1668 req->rq_flags |= RQF_DONTPREP;
1669 } else {
1670 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1671 }
1672
1673 cmd->flags &= SCMD_PRESERVED_FLAGS;
1674 if (sdev->simple_tags)
1675 cmd->flags |= SCMD_TAGGED;
1676 if (bd->last)
1677 cmd->flags |= SCMD_LAST;
1678
1679 scsi_set_resid(cmd, 0);
1680 memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
1681 cmd->scsi_done = scsi_mq_done;
1682
1683 blk_mq_start_request(req);
1684 reason = scsi_dispatch_cmd(cmd);
1685 if (reason) {
1686 scsi_set_blocked(cmd, reason);
1687 ret = BLK_STS_RESOURCE;
1688 goto out_dec_host_busy;
1689 }
1690
1691 return BLK_STS_OK;
1692
1693out_dec_host_busy:
1694 scsi_dec_host_busy(shost, cmd);
1695out_dec_target_busy:
1696 if (scsi_target(sdev)->can_queue > 0)
1697 atomic_dec(&scsi_target(sdev)->target_busy);
1698out_put_budget:
1699 scsi_mq_put_budget(q, cmd->budget_token);
1700 cmd->budget_token = -1;
1701 switch (ret) {
1702 case BLK_STS_OK:
1703 break;
1704 case BLK_STS_RESOURCE:
1705 case BLK_STS_ZONE_RESOURCE:
1706 if (scsi_device_blocked(sdev))
1707 ret = BLK_STS_DEV_RESOURCE;
1708 break;
1709 case BLK_STS_AGAIN:
1710 scsi_req(req)->result = DID_BUS_BUSY << 16;
1711 if (req->rq_flags & RQF_DONTPREP)
1712 scsi_mq_uninit_cmd(cmd);
1713 break;
1714 default:
1715 if (unlikely(!scsi_device_online(sdev)))
1716 scsi_req(req)->result = DID_NO_CONNECT << 16;
1717 else
1718 scsi_req(req)->result = DID_ERROR << 16;
1719 /*
1720 * Make sure to release all allocated resources when
1721 * we hit an error, as we will never see this command
1722 * again.
1723 */
1724 if (req->rq_flags & RQF_DONTPREP)
1725 scsi_mq_uninit_cmd(cmd);
1726 scsi_run_queue_async(sdev);
1727 break;
1728 }
1729 return ret;
1730}
1731
1732static enum blk_eh_timer_return scsi_timeout(struct request *req,
1733 bool reserved)
1734{
1735 if (reserved)
1736 return BLK_EH_RESET_TIMER;
1737 return scsi_times_out(req);
1738}
1739
1740static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1741 unsigned int hctx_idx, unsigned int numa_node)
1742{
1743 struct Scsi_Host *shost = set->driver_data;
1744 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1745 struct scatterlist *sg;
1746 int ret = 0;
1747
1748 cmd->sense_buffer =
1749 kmem_cache_alloc_node(scsi_sense_cache, GFP_KERNEL, numa_node);
1750 if (!cmd->sense_buffer)
1751 return -ENOMEM;
1752 cmd->req.sense = cmd->sense_buffer;
1753
1754 if (scsi_host_get_prot(shost)) {
1755 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1756 shost->hostt->cmd_size;
1757 cmd->prot_sdb = (void *)sg + scsi_mq_inline_sgl_size(shost);
1758 }
1759
1760 if (shost->hostt->init_cmd_priv) {
1761 ret = shost->hostt->init_cmd_priv(shost, cmd);
1762 if (ret < 0)
1763 kmem_cache_free(scsi_sense_cache, cmd->sense_buffer);
1764 }
1765
1766 return ret;
1767}
1768
1769static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1770 unsigned int hctx_idx)
1771{
1772 struct Scsi_Host *shost = set->driver_data;
1773 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1774
1775 if (shost->hostt->exit_cmd_priv)
1776 shost->hostt->exit_cmd_priv(shost, cmd);
1777 kmem_cache_free(scsi_sense_cache, cmd->sense_buffer);
1778}
1779
1780
1781static int scsi_mq_poll(struct blk_mq_hw_ctx *hctx)
1782{
1783 struct Scsi_Host *shost = hctx->driver_data;
1784
1785 if (shost->hostt->mq_poll)
1786 return shost->hostt->mq_poll(shost, hctx->queue_num);
1787
1788 return 0;
1789}
1790
1791static int scsi_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
1792 unsigned int hctx_idx)
1793{
1794 struct Scsi_Host *shost = data;
1795
1796 hctx->driver_data = shost;
1797 return 0;
1798}
1799
1800static int scsi_map_queues(struct blk_mq_tag_set *set)
1801{
1802 struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
1803
1804 if (shost->hostt->map_queues)
1805 return shost->hostt->map_queues(shost);
1806 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
1807}
1808
1809void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
1810{
1811 struct device *dev = shost->dma_dev;
1812
1813 /*
1814 * this limit is imposed by hardware restrictions
1815 */
1816 blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1817 SG_MAX_SEGMENTS));
1818
1819 if (scsi_host_prot_dma(shost)) {
1820 shost->sg_prot_tablesize =
1821 min_not_zero(shost->sg_prot_tablesize,
1822 (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1823 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1824 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1825 }
1826
1827 if (dev->dma_mask) {
1828 shost->max_sectors = min_t(unsigned int, shost->max_sectors,
1829 dma_max_mapping_size(dev) >> SECTOR_SHIFT);
1830 }
1831 blk_queue_max_hw_sectors(q, shost->max_sectors);
1832 blk_queue_segment_boundary(q, shost->dma_boundary);
1833 dma_set_seg_boundary(dev, shost->dma_boundary);
1834
1835 blk_queue_max_segment_size(q, shost->max_segment_size);
1836 blk_queue_virt_boundary(q, shost->virt_boundary_mask);
1837 dma_set_max_seg_size(dev, queue_max_segment_size(q));
1838
1839 /*
1840 * Set a reasonable default alignment: The larger of 32-byte (dword),
1841 * which is a common minimum for HBAs, and the minimum DMA alignment,
1842 * which is set by the platform.
1843 *
1844 * Devices that require a bigger alignment can increase it later.
1845 */
1846 blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
1847}
1848EXPORT_SYMBOL_GPL(__scsi_init_queue);
1849
1850static const struct blk_mq_ops scsi_mq_ops_no_commit = {
1851 .get_budget = scsi_mq_get_budget,
1852 .put_budget = scsi_mq_put_budget,
1853 .queue_rq = scsi_queue_rq,
1854 .complete = scsi_complete,
1855 .timeout = scsi_timeout,
1856#ifdef CONFIG_BLK_DEBUG_FS
1857 .show_rq = scsi_show_rq,
1858#endif
1859 .init_request = scsi_mq_init_request,
1860 .exit_request = scsi_mq_exit_request,
1861 .initialize_rq_fn = scsi_initialize_rq,
1862 .cleanup_rq = scsi_cleanup_rq,
1863 .busy = scsi_mq_lld_busy,
1864 .map_queues = scsi_map_queues,
1865 .init_hctx = scsi_init_hctx,
1866 .poll = scsi_mq_poll,
1867 .set_rq_budget_token = scsi_mq_set_rq_budget_token,
1868 .get_rq_budget_token = scsi_mq_get_rq_budget_token,
1869};
1870
1871
1872static void scsi_commit_rqs(struct blk_mq_hw_ctx *hctx)
1873{
1874 struct Scsi_Host *shost = hctx->driver_data;
1875
1876 shost->hostt->commit_rqs(shost, hctx->queue_num);
1877}
1878
1879static const struct blk_mq_ops scsi_mq_ops = {
1880 .get_budget = scsi_mq_get_budget,
1881 .put_budget = scsi_mq_put_budget,
1882 .queue_rq = scsi_queue_rq,
1883 .commit_rqs = scsi_commit_rqs,
1884 .complete = scsi_complete,
1885 .timeout = scsi_timeout,
1886#ifdef CONFIG_BLK_DEBUG_FS
1887 .show_rq = scsi_show_rq,
1888#endif
1889 .init_request = scsi_mq_init_request,
1890 .exit_request = scsi_mq_exit_request,
1891 .initialize_rq_fn = scsi_initialize_rq,
1892 .cleanup_rq = scsi_cleanup_rq,
1893 .busy = scsi_mq_lld_busy,
1894 .map_queues = scsi_map_queues,
1895 .init_hctx = scsi_init_hctx,
1896 .poll = scsi_mq_poll,
1897 .set_rq_budget_token = scsi_mq_set_rq_budget_token,
1898 .get_rq_budget_token = scsi_mq_get_rq_budget_token,
1899};
1900
1901int scsi_mq_setup_tags(struct Scsi_Host *shost)
1902{
1903 unsigned int cmd_size, sgl_size;
1904 struct blk_mq_tag_set *tag_set = &shost->tag_set;
1905
1906 sgl_size = max_t(unsigned int, sizeof(struct scatterlist),
1907 scsi_mq_inline_sgl_size(shost));
1908 cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
1909 if (scsi_host_get_prot(shost))
1910 cmd_size += sizeof(struct scsi_data_buffer) +
1911 sizeof(struct scatterlist) * SCSI_INLINE_PROT_SG_CNT;
1912
1913 memset(tag_set, 0, sizeof(*tag_set));
1914 if (shost->hostt->commit_rqs)
1915 tag_set->ops = &scsi_mq_ops;
1916 else
1917 tag_set->ops = &scsi_mq_ops_no_commit;
1918 tag_set->nr_hw_queues = shost->nr_hw_queues ? : 1;
1919 tag_set->nr_maps = shost->nr_maps ? : 1;
1920 tag_set->queue_depth = shost->can_queue;
1921 tag_set->cmd_size = cmd_size;
1922 tag_set->numa_node = NUMA_NO_NODE;
1923 tag_set->flags = BLK_MQ_F_SHOULD_MERGE;
1924 tag_set->flags |=
1925 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
1926 tag_set->driver_data = shost;
1927 if (shost->host_tagset)
1928 tag_set->flags |= BLK_MQ_F_TAG_HCTX_SHARED;
1929
1930 return blk_mq_alloc_tag_set(tag_set);
1931}
1932
1933void scsi_mq_destroy_tags(struct Scsi_Host *shost)
1934{
1935 blk_mq_free_tag_set(&shost->tag_set);
1936}
1937
1938/**
1939 * scsi_device_from_queue - return sdev associated with a request_queue
1940 * @q: The request queue to return the sdev from
1941 *
1942 * Return the sdev associated with a request queue or NULL if the
1943 * request_queue does not reference a SCSI device.
1944 */
1945struct scsi_device *scsi_device_from_queue(struct request_queue *q)
1946{
1947 struct scsi_device *sdev = NULL;
1948
1949 if (q->mq_ops == &scsi_mq_ops_no_commit ||
1950 q->mq_ops == &scsi_mq_ops)
1951 sdev = q->queuedata;
1952 if (!sdev || !get_device(&sdev->sdev_gendev))
1953 sdev = NULL;
1954
1955 return sdev;
1956}
1957
1958/**
1959 * scsi_block_requests - Utility function used by low-level drivers to prevent
1960 * further commands from being queued to the device.
1961 * @shost: host in question
1962 *
1963 * There is no timer nor any other means by which the requests get unblocked
1964 * other than the low-level driver calling scsi_unblock_requests().
1965 */
1966void scsi_block_requests(struct Scsi_Host *shost)
1967{
1968 shost->host_self_blocked = 1;
1969}
1970EXPORT_SYMBOL(scsi_block_requests);
1971
1972/**
1973 * scsi_unblock_requests - Utility function used by low-level drivers to allow
1974 * further commands to be queued to the device.
1975 * @shost: host in question
1976 *
1977 * There is no timer nor any other means by which the requests get unblocked
1978 * other than the low-level driver calling scsi_unblock_requests(). This is done
1979 * as an API function so that changes to the internals of the scsi mid-layer
1980 * won't require wholesale changes to drivers that use this feature.
1981 */
1982void scsi_unblock_requests(struct Scsi_Host *shost)
1983{
1984 shost->host_self_blocked = 0;
1985 scsi_run_host_queues(shost);
1986}
1987EXPORT_SYMBOL(scsi_unblock_requests);
1988
1989void scsi_exit_queue(void)
1990{
1991 kmem_cache_destroy(scsi_sense_cache);
1992}
1993
1994/**
1995 * scsi_mode_select - issue a mode select
1996 * @sdev: SCSI device to be queried
1997 * @pf: Page format bit (1 == standard, 0 == vendor specific)
1998 * @sp: Save page bit (0 == don't save, 1 == save)
1999 * @modepage: mode page being requested
2000 * @buffer: request buffer (may not be smaller than eight bytes)
2001 * @len: length of request buffer.
2002 * @timeout: command timeout
2003 * @retries: number of retries before failing
2004 * @data: returns a structure abstracting the mode header data
2005 * @sshdr: place to put sense data (or NULL if no sense to be collected).
2006 * must be SCSI_SENSE_BUFFERSIZE big.
2007 *
2008 * Returns zero if successful; negative error number or scsi
2009 * status on error
2010 *
2011 */
2012int
2013scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
2014 unsigned char *buffer, int len, int timeout, int retries,
2015 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2016{
2017 unsigned char cmd[10];
2018 unsigned char *real_buffer;
2019 int ret;
2020
2021 memset(cmd, 0, sizeof(cmd));
2022 cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2023
2024 if (sdev->use_10_for_ms) {
2025 if (len > 65535)
2026 return -EINVAL;
2027 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2028 if (!real_buffer)
2029 return -ENOMEM;
2030 memcpy(real_buffer + 8, buffer, len);
2031 len += 8;
2032 real_buffer[0] = 0;
2033 real_buffer[1] = 0;
2034 real_buffer[2] = data->medium_type;
2035 real_buffer[3] = data->device_specific;
2036 real_buffer[4] = data->longlba ? 0x01 : 0;
2037 real_buffer[5] = 0;
2038 real_buffer[6] = data->block_descriptor_length >> 8;
2039 real_buffer[7] = data->block_descriptor_length;
2040
2041 cmd[0] = MODE_SELECT_10;
2042 cmd[7] = len >> 8;
2043 cmd[8] = len;
2044 } else {
2045 if (len > 255 || data->block_descriptor_length > 255 ||
2046 data->longlba)
2047 return -EINVAL;
2048
2049 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2050 if (!real_buffer)
2051 return -ENOMEM;
2052 memcpy(real_buffer + 4, buffer, len);
2053 len += 4;
2054 real_buffer[0] = 0;
2055 real_buffer[1] = data->medium_type;
2056 real_buffer[2] = data->device_specific;
2057 real_buffer[3] = data->block_descriptor_length;
2058
2059 cmd[0] = MODE_SELECT;
2060 cmd[4] = len;
2061 }
2062
2063 ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2064 sshdr, timeout, retries, NULL);
2065 kfree(real_buffer);
2066 return ret;
2067}
2068EXPORT_SYMBOL_GPL(scsi_mode_select);
2069
2070/**
2071 * scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2072 * @sdev: SCSI device to be queried
2073 * @dbd: set if mode sense will allow block descriptors to be returned
2074 * @modepage: mode page being requested
2075 * @buffer: request buffer (may not be smaller than eight bytes)
2076 * @len: length of request buffer.
2077 * @timeout: command timeout
2078 * @retries: number of retries before failing
2079 * @data: returns a structure abstracting the mode header data
2080 * @sshdr: place to put sense data (or NULL if no sense to be collected).
2081 * must be SCSI_SENSE_BUFFERSIZE big.
2082 *
2083 * Returns zero if successful, or a negative error number on failure
2084 */
2085int
2086scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2087 unsigned char *buffer, int len, int timeout, int retries,
2088 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2089{
2090 unsigned char cmd[12];
2091 int use_10_for_ms;
2092 int header_length;
2093 int result, retry_count = retries;
2094 struct scsi_sense_hdr my_sshdr;
2095
2096 memset(data, 0, sizeof(*data));
2097 memset(&cmd[0], 0, 12);
2098
2099 dbd = sdev->set_dbd_for_ms ? 8 : dbd;
2100 cmd[1] = dbd & 0x18; /* allows DBD and LLBA bits */
2101 cmd[2] = modepage;
2102
2103 /* caller might not be interested in sense, but we need it */
2104 if (!sshdr)
2105 sshdr = &my_sshdr;
2106
2107 retry:
2108 use_10_for_ms = sdev->use_10_for_ms;
2109
2110 if (use_10_for_ms) {
2111 if (len < 8)
2112 len = 8;
2113
2114 cmd[0] = MODE_SENSE_10;
2115 cmd[8] = len;
2116 header_length = 8;
2117 } else {
2118 if (len < 4)
2119 len = 4;
2120
2121 cmd[0] = MODE_SENSE;
2122 cmd[4] = len;
2123 header_length = 4;
2124 }
2125
2126 memset(buffer, 0, len);
2127
2128 result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2129 sshdr, timeout, retries, NULL);
2130 if (result < 0)
2131 return result;
2132
2133 /* This code looks awful: what it's doing is making sure an
2134 * ILLEGAL REQUEST sense return identifies the actual command
2135 * byte as the problem. MODE_SENSE commands can return
2136 * ILLEGAL REQUEST if the code page isn't supported */
2137
2138 if (!scsi_status_is_good(result)) {
2139 if (scsi_sense_valid(sshdr)) {
2140 if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2141 (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2142 /*
2143 * Invalid command operation code
2144 */
2145 if (use_10_for_ms) {
2146 sdev->use_10_for_ms = 0;
2147 goto retry;
2148 }
2149 }
2150 if (scsi_status_is_check_condition(result) &&
2151 sshdr->sense_key == UNIT_ATTENTION &&
2152 retry_count) {
2153 retry_count--;
2154 goto retry;
2155 }
2156 }
2157 return -EIO;
2158 }
2159 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2160 (modepage == 6 || modepage == 8))) {
2161 /* Initio breakage? */
2162 header_length = 0;
2163 data->length = 13;
2164 data->medium_type = 0;
2165 data->device_specific = 0;
2166 data->longlba = 0;
2167 data->block_descriptor_length = 0;
2168 } else if (use_10_for_ms) {
2169 data->length = buffer[0]*256 + buffer[1] + 2;
2170 data->medium_type = buffer[2];
2171 data->device_specific = buffer[3];
2172 data->longlba = buffer[4] & 0x01;
2173 data->block_descriptor_length = buffer[6]*256
2174 + buffer[7];
2175 } else {
2176 data->length = buffer[0] + 1;
2177 data->medium_type = buffer[1];
2178 data->device_specific = buffer[2];
2179 data->block_descriptor_length = buffer[3];
2180 }
2181 data->header_length = header_length;
2182
2183 return 0;
2184}
2185EXPORT_SYMBOL(scsi_mode_sense);
2186
2187/**
2188 * scsi_test_unit_ready - test if unit is ready
2189 * @sdev: scsi device to change the state of.
2190 * @timeout: command timeout
2191 * @retries: number of retries before failing
2192 * @sshdr: outpout pointer for decoded sense information.
2193 *
2194 * Returns zero if unsuccessful or an error if TUR failed. For
2195 * removable media, UNIT_ATTENTION sets ->changed flag.
2196 **/
2197int
2198scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2199 struct scsi_sense_hdr *sshdr)
2200{
2201 char cmd[] = {
2202 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2203 };
2204 int result;
2205
2206 /* try to eat the UNIT_ATTENTION if there are enough retries */
2207 do {
2208 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2209 timeout, 1, NULL);
2210 if (sdev->removable && scsi_sense_valid(sshdr) &&
2211 sshdr->sense_key == UNIT_ATTENTION)
2212 sdev->changed = 1;
2213 } while (scsi_sense_valid(sshdr) &&
2214 sshdr->sense_key == UNIT_ATTENTION && --retries);
2215
2216 return result;
2217}
2218EXPORT_SYMBOL(scsi_test_unit_ready);
2219
2220/**
2221 * scsi_device_set_state - Take the given device through the device state model.
2222 * @sdev: scsi device to change the state of.
2223 * @state: state to change to.
2224 *
2225 * Returns zero if successful or an error if the requested
2226 * transition is illegal.
2227 */
2228int
2229scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2230{
2231 enum scsi_device_state oldstate = sdev->sdev_state;
2232
2233 if (state == oldstate)
2234 return 0;
2235
2236 switch (state) {
2237 case SDEV_CREATED:
2238 switch (oldstate) {
2239 case SDEV_CREATED_BLOCK:
2240 break;
2241 default:
2242 goto illegal;
2243 }
2244 break;
2245
2246 case SDEV_RUNNING:
2247 switch (oldstate) {
2248 case SDEV_CREATED:
2249 case SDEV_OFFLINE:
2250 case SDEV_TRANSPORT_OFFLINE:
2251 case SDEV_QUIESCE:
2252 case SDEV_BLOCK:
2253 break;
2254 default:
2255 goto illegal;
2256 }
2257 break;
2258
2259 case SDEV_QUIESCE:
2260 switch (oldstate) {
2261 case SDEV_RUNNING:
2262 case SDEV_OFFLINE:
2263 case SDEV_TRANSPORT_OFFLINE:
2264 break;
2265 default:
2266 goto illegal;
2267 }
2268 break;
2269
2270 case SDEV_OFFLINE:
2271 case SDEV_TRANSPORT_OFFLINE:
2272 switch (oldstate) {
2273 case SDEV_CREATED:
2274 case SDEV_RUNNING:
2275 case SDEV_QUIESCE:
2276 case SDEV_BLOCK:
2277 break;
2278 default:
2279 goto illegal;
2280 }
2281 break;
2282
2283 case SDEV_BLOCK:
2284 switch (oldstate) {
2285 case SDEV_RUNNING:
2286 case SDEV_CREATED_BLOCK:
2287 case SDEV_QUIESCE:
2288 case SDEV_OFFLINE:
2289 break;
2290 default:
2291 goto illegal;
2292 }
2293 break;
2294
2295 case SDEV_CREATED_BLOCK:
2296 switch (oldstate) {
2297 case SDEV_CREATED:
2298 break;
2299 default:
2300 goto illegal;
2301 }
2302 break;
2303
2304 case SDEV_CANCEL:
2305 switch (oldstate) {
2306 case SDEV_CREATED:
2307 case SDEV_RUNNING:
2308 case SDEV_QUIESCE:
2309 case SDEV_OFFLINE:
2310 case SDEV_TRANSPORT_OFFLINE:
2311 break;
2312 default:
2313 goto illegal;
2314 }
2315 break;
2316
2317 case SDEV_DEL:
2318 switch (oldstate) {
2319 case SDEV_CREATED:
2320 case SDEV_RUNNING:
2321 case SDEV_OFFLINE:
2322 case SDEV_TRANSPORT_OFFLINE:
2323 case SDEV_CANCEL:
2324 case SDEV_BLOCK:
2325 case SDEV_CREATED_BLOCK:
2326 break;
2327 default:
2328 goto illegal;
2329 }
2330 break;
2331
2332 }
2333 sdev->offline_already = false;
2334 sdev->sdev_state = state;
2335 return 0;
2336
2337 illegal:
2338 SCSI_LOG_ERROR_RECOVERY(1,
2339 sdev_printk(KERN_ERR, sdev,
2340 "Illegal state transition %s->%s",
2341 scsi_device_state_name(oldstate),
2342 scsi_device_state_name(state))
2343 );
2344 return -EINVAL;
2345}
2346EXPORT_SYMBOL(scsi_device_set_state);
2347
2348/**
2349 * scsi_evt_emit - emit a single SCSI device uevent
2350 * @sdev: associated SCSI device
2351 * @evt: event to emit
2352 *
2353 * Send a single uevent (scsi_event) to the associated scsi_device.
2354 */
2355static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2356{
2357 int idx = 0;
2358 char *envp[3];
2359
2360 switch (evt->evt_type) {
2361 case SDEV_EVT_MEDIA_CHANGE:
2362 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2363 break;
2364 case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2365 scsi_rescan_device(&sdev->sdev_gendev);
2366 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2367 break;
2368 case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2369 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2370 break;
2371 case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2372 envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2373 break;
2374 case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2375 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2376 break;
2377 case SDEV_EVT_LUN_CHANGE_REPORTED:
2378 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2379 break;
2380 case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2381 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2382 break;
2383 case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2384 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2385 break;
2386 default:
2387 /* do nothing */
2388 break;
2389 }
2390
2391 envp[idx++] = NULL;
2392
2393 kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2394}
2395
2396/**
2397 * scsi_evt_thread - send a uevent for each scsi event
2398 * @work: work struct for scsi_device
2399 *
2400 * Dispatch queued events to their associated scsi_device kobjects
2401 * as uevents.
2402 */
2403void scsi_evt_thread(struct work_struct *work)
2404{
2405 struct scsi_device *sdev;
2406 enum scsi_device_event evt_type;
2407 LIST_HEAD(event_list);
2408
2409 sdev = container_of(work, struct scsi_device, event_work);
2410
2411 for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2412 if (test_and_clear_bit(evt_type, sdev->pending_events))
2413 sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2414
2415 while (1) {
2416 struct scsi_event *evt;
2417 struct list_head *this, *tmp;
2418 unsigned long flags;
2419
2420 spin_lock_irqsave(&sdev->list_lock, flags);
2421 list_splice_init(&sdev->event_list, &event_list);
2422 spin_unlock_irqrestore(&sdev->list_lock, flags);
2423
2424 if (list_empty(&event_list))
2425 break;
2426
2427 list_for_each_safe(this, tmp, &event_list) {
2428 evt = list_entry(this, struct scsi_event, node);
2429 list_del(&evt->node);
2430 scsi_evt_emit(sdev, evt);
2431 kfree(evt);
2432 }
2433 }
2434}
2435
2436/**
2437 * sdev_evt_send - send asserted event to uevent thread
2438 * @sdev: scsi_device event occurred on
2439 * @evt: event to send
2440 *
2441 * Assert scsi device event asynchronously.
2442 */
2443void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2444{
2445 unsigned long flags;
2446
2447#if 0
2448 /* FIXME: currently this check eliminates all media change events
2449 * for polled devices. Need to update to discriminate between AN
2450 * and polled events */
2451 if (!test_bit(evt->evt_type, sdev->supported_events)) {
2452 kfree(evt);
2453 return;
2454 }
2455#endif
2456
2457 spin_lock_irqsave(&sdev->list_lock, flags);
2458 list_add_tail(&evt->node, &sdev->event_list);
2459 schedule_work(&sdev->event_work);
2460 spin_unlock_irqrestore(&sdev->list_lock, flags);
2461}
2462EXPORT_SYMBOL_GPL(sdev_evt_send);
2463
2464/**
2465 * sdev_evt_alloc - allocate a new scsi event
2466 * @evt_type: type of event to allocate
2467 * @gfpflags: GFP flags for allocation
2468 *
2469 * Allocates and returns a new scsi_event.
2470 */
2471struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2472 gfp_t gfpflags)
2473{
2474 struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2475 if (!evt)
2476 return NULL;
2477
2478 evt->evt_type = evt_type;
2479 INIT_LIST_HEAD(&evt->node);
2480
2481 /* evt_type-specific initialization, if any */
2482 switch (evt_type) {
2483 case SDEV_EVT_MEDIA_CHANGE:
2484 case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2485 case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2486 case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2487 case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2488 case SDEV_EVT_LUN_CHANGE_REPORTED:
2489 case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2490 case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2491 default:
2492 /* do nothing */
2493 break;
2494 }
2495
2496 return evt;
2497}
2498EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2499
2500/**
2501 * sdev_evt_send_simple - send asserted event to uevent thread
2502 * @sdev: scsi_device event occurred on
2503 * @evt_type: type of event to send
2504 * @gfpflags: GFP flags for allocation
2505 *
2506 * Assert scsi device event asynchronously, given an event type.
2507 */
2508void sdev_evt_send_simple(struct scsi_device *sdev,
2509 enum scsi_device_event evt_type, gfp_t gfpflags)
2510{
2511 struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2512 if (!evt) {
2513 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2514 evt_type);
2515 return;
2516 }
2517
2518 sdev_evt_send(sdev, evt);
2519}
2520EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2521
2522/**
2523 * scsi_device_quiesce - Block all commands except power management.
2524 * @sdev: scsi device to quiesce.
2525 *
2526 * This works by trying to transition to the SDEV_QUIESCE state
2527 * (which must be a legal transition). When the device is in this
2528 * state, only power management requests will be accepted, all others will
2529 * be deferred.
2530 *
2531 * Must be called with user context, may sleep.
2532 *
2533 * Returns zero if unsuccessful or an error if not.
2534 */
2535int
2536scsi_device_quiesce(struct scsi_device *sdev)
2537{
2538 struct request_queue *q = sdev->request_queue;
2539 int err;
2540
2541 /*
2542 * It is allowed to call scsi_device_quiesce() multiple times from
2543 * the same context but concurrent scsi_device_quiesce() calls are
2544 * not allowed.
2545 */
2546 WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2547
2548 if (sdev->quiesced_by == current)
2549 return 0;
2550
2551 blk_set_pm_only(q);
2552
2553 blk_mq_freeze_queue(q);
2554 /*
2555 * Ensure that the effect of blk_set_pm_only() will be visible
2556 * for percpu_ref_tryget() callers that occur after the queue
2557 * unfreeze even if the queue was already frozen before this function
2558 * was called. See also https://lwn.net/Articles/573497/.
2559 */
2560 synchronize_rcu();
2561 blk_mq_unfreeze_queue(q);
2562
2563 mutex_lock(&sdev->state_mutex);
2564 err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2565 if (err == 0)
2566 sdev->quiesced_by = current;
2567 else
2568 blk_clear_pm_only(q);
2569 mutex_unlock(&sdev->state_mutex);
2570
2571 return err;
2572}
2573EXPORT_SYMBOL(scsi_device_quiesce);
2574
2575/**
2576 * scsi_device_resume - Restart user issued commands to a quiesced device.
2577 * @sdev: scsi device to resume.
2578 *
2579 * Moves the device from quiesced back to running and restarts the
2580 * queues.
2581 *
2582 * Must be called with user context, may sleep.
2583 */
2584void scsi_device_resume(struct scsi_device *sdev)
2585{
2586 /* check if the device state was mutated prior to resume, and if
2587 * so assume the state is being managed elsewhere (for example
2588 * device deleted during suspend)
2589 */
2590 mutex_lock(&sdev->state_mutex);
2591 if (sdev->sdev_state == SDEV_QUIESCE)
2592 scsi_device_set_state(sdev, SDEV_RUNNING);
2593 if (sdev->quiesced_by) {
2594 sdev->quiesced_by = NULL;
2595 blk_clear_pm_only(sdev->request_queue);
2596 }
2597 mutex_unlock(&sdev->state_mutex);
2598}
2599EXPORT_SYMBOL(scsi_device_resume);
2600
2601static void
2602device_quiesce_fn(struct scsi_device *sdev, void *data)
2603{
2604 scsi_device_quiesce(sdev);
2605}
2606
2607void
2608scsi_target_quiesce(struct scsi_target *starget)
2609{
2610 starget_for_each_device(starget, NULL, device_quiesce_fn);
2611}
2612EXPORT_SYMBOL(scsi_target_quiesce);
2613
2614static void
2615device_resume_fn(struct scsi_device *sdev, void *data)
2616{
2617 scsi_device_resume(sdev);
2618}
2619
2620void
2621scsi_target_resume(struct scsi_target *starget)
2622{
2623 starget_for_each_device(starget, NULL, device_resume_fn);
2624}
2625EXPORT_SYMBOL(scsi_target_resume);
2626
2627/**
2628 * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2629 * @sdev: device to block
2630 *
2631 * Pause SCSI command processing on the specified device. Does not sleep.
2632 *
2633 * Returns zero if successful or a negative error code upon failure.
2634 *
2635 * Notes:
2636 * This routine transitions the device to the SDEV_BLOCK state (which must be
2637 * a legal transition). When the device is in this state, command processing
2638 * is paused until the device leaves the SDEV_BLOCK state. See also
2639 * scsi_internal_device_unblock_nowait().
2640 */
2641int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2642{
2643 struct request_queue *q = sdev->request_queue;
2644 int err = 0;
2645
2646 err = scsi_device_set_state(sdev, SDEV_BLOCK);
2647 if (err) {
2648 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2649
2650 if (err)
2651 return err;
2652 }
2653
2654 /*
2655 * The device has transitioned to SDEV_BLOCK. Stop the
2656 * block layer from calling the midlayer with this device's
2657 * request queue.
2658 */
2659 blk_mq_quiesce_queue_nowait(q);
2660 return 0;
2661}
2662EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2663
2664/**
2665 * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
2666 * @sdev: device to block
2667 *
2668 * Pause SCSI command processing on the specified device and wait until all
2669 * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
2670 *
2671 * Returns zero if successful or a negative error code upon failure.
2672 *
2673 * Note:
2674 * This routine transitions the device to the SDEV_BLOCK state (which must be
2675 * a legal transition). When the device is in this state, command processing
2676 * is paused until the device leaves the SDEV_BLOCK state. See also
2677 * scsi_internal_device_unblock().
2678 */
2679static int scsi_internal_device_block(struct scsi_device *sdev)
2680{
2681 struct request_queue *q = sdev->request_queue;
2682 int err;
2683
2684 mutex_lock(&sdev->state_mutex);
2685 err = scsi_internal_device_block_nowait(sdev);
2686 if (err == 0)
2687 blk_mq_quiesce_queue(q);
2688 mutex_unlock(&sdev->state_mutex);
2689
2690 return err;
2691}
2692
2693void scsi_start_queue(struct scsi_device *sdev)
2694{
2695 struct request_queue *q = sdev->request_queue;
2696
2697 blk_mq_unquiesce_queue(q);
2698}
2699
2700/**
2701 * scsi_internal_device_unblock_nowait - resume a device after a block request
2702 * @sdev: device to resume
2703 * @new_state: state to set the device to after unblocking
2704 *
2705 * Restart the device queue for a previously suspended SCSI device. Does not
2706 * sleep.
2707 *
2708 * Returns zero if successful or a negative error code upon failure.
2709 *
2710 * Notes:
2711 * This routine transitions the device to the SDEV_RUNNING state or to one of
2712 * the offline states (which must be a legal transition) allowing the midlayer
2713 * to goose the queue for this device.
2714 */
2715int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
2716 enum scsi_device_state new_state)
2717{
2718 switch (new_state) {
2719 case SDEV_RUNNING:
2720 case SDEV_TRANSPORT_OFFLINE:
2721 break;
2722 default:
2723 return -EINVAL;
2724 }
2725
2726 /*
2727 * Try to transition the scsi device to SDEV_RUNNING or one of the
2728 * offlined states and goose the device queue if successful.
2729 */
2730 switch (sdev->sdev_state) {
2731 case SDEV_BLOCK:
2732 case SDEV_TRANSPORT_OFFLINE:
2733 sdev->sdev_state = new_state;
2734 break;
2735 case SDEV_CREATED_BLOCK:
2736 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2737 new_state == SDEV_OFFLINE)
2738 sdev->sdev_state = new_state;
2739 else
2740 sdev->sdev_state = SDEV_CREATED;
2741 break;
2742 case SDEV_CANCEL:
2743 case SDEV_OFFLINE:
2744 break;
2745 default:
2746 return -EINVAL;
2747 }
2748 scsi_start_queue(sdev);
2749
2750 return 0;
2751}
2752EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
2753
2754/**
2755 * scsi_internal_device_unblock - resume a device after a block request
2756 * @sdev: device to resume
2757 * @new_state: state to set the device to after unblocking
2758 *
2759 * Restart the device queue for a previously suspended SCSI device. May sleep.
2760 *
2761 * Returns zero if successful or a negative error code upon failure.
2762 *
2763 * Notes:
2764 * This routine transitions the device to the SDEV_RUNNING state or to one of
2765 * the offline states (which must be a legal transition) allowing the midlayer
2766 * to goose the queue for this device.
2767 */
2768static int scsi_internal_device_unblock(struct scsi_device *sdev,
2769 enum scsi_device_state new_state)
2770{
2771 int ret;
2772
2773 mutex_lock(&sdev->state_mutex);
2774 ret = scsi_internal_device_unblock_nowait(sdev, new_state);
2775 mutex_unlock(&sdev->state_mutex);
2776
2777 return ret;
2778}
2779
2780static void
2781device_block(struct scsi_device *sdev, void *data)
2782{
2783 int ret;
2784
2785 ret = scsi_internal_device_block(sdev);
2786
2787 WARN_ONCE(ret, "scsi_internal_device_block(%s) failed: ret = %d\n",
2788 dev_name(&sdev->sdev_gendev), ret);
2789}
2790
2791static int
2792target_block(struct device *dev, void *data)
2793{
2794 if (scsi_is_target_device(dev))
2795 starget_for_each_device(to_scsi_target(dev), NULL,
2796 device_block);
2797 return 0;
2798}
2799
2800void
2801scsi_target_block(struct device *dev)
2802{
2803 if (scsi_is_target_device(dev))
2804 starget_for_each_device(to_scsi_target(dev), NULL,
2805 device_block);
2806 else
2807 device_for_each_child(dev, NULL, target_block);
2808}
2809EXPORT_SYMBOL_GPL(scsi_target_block);
2810
2811static void
2812device_unblock(struct scsi_device *sdev, void *data)
2813{
2814 scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2815}
2816
2817static int
2818target_unblock(struct device *dev, void *data)
2819{
2820 if (scsi_is_target_device(dev))
2821 starget_for_each_device(to_scsi_target(dev), data,
2822 device_unblock);
2823 return 0;
2824}
2825
2826void
2827scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2828{
2829 if (scsi_is_target_device(dev))
2830 starget_for_each_device(to_scsi_target(dev), &new_state,
2831 device_unblock);
2832 else
2833 device_for_each_child(dev, &new_state, target_unblock);
2834}
2835EXPORT_SYMBOL_GPL(scsi_target_unblock);
2836
2837int
2838scsi_host_block(struct Scsi_Host *shost)
2839{
2840 struct scsi_device *sdev;
2841 int ret = 0;
2842
2843 /*
2844 * Call scsi_internal_device_block_nowait so we can avoid
2845 * calling synchronize_rcu() for each LUN.
2846 */
2847 shost_for_each_device(sdev, shost) {
2848 mutex_lock(&sdev->state_mutex);
2849 ret = scsi_internal_device_block_nowait(sdev);
2850 mutex_unlock(&sdev->state_mutex);
2851 if (ret) {
2852 scsi_device_put(sdev);
2853 break;
2854 }
2855 }
2856
2857 /*
2858 * SCSI never enables blk-mq's BLK_MQ_F_BLOCKING flag so
2859 * calling synchronize_rcu() once is enough.
2860 */
2861 WARN_ON_ONCE(shost->tag_set.flags & BLK_MQ_F_BLOCKING);
2862
2863 if (!ret)
2864 synchronize_rcu();
2865
2866 return ret;
2867}
2868EXPORT_SYMBOL_GPL(scsi_host_block);
2869
2870int
2871scsi_host_unblock(struct Scsi_Host *shost, int new_state)
2872{
2873 struct scsi_device *sdev;
2874 int ret = 0;
2875
2876 shost_for_each_device(sdev, shost) {
2877 ret = scsi_internal_device_unblock(sdev, new_state);
2878 if (ret) {
2879 scsi_device_put(sdev);
2880 break;
2881 }
2882 }
2883 return ret;
2884}
2885EXPORT_SYMBOL_GPL(scsi_host_unblock);
2886
2887/**
2888 * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2889 * @sgl: scatter-gather list
2890 * @sg_count: number of segments in sg
2891 * @offset: offset in bytes into sg, on return offset into the mapped area
2892 * @len: bytes to map, on return number of bytes mapped
2893 *
2894 * Returns virtual address of the start of the mapped page
2895 */
2896void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2897 size_t *offset, size_t *len)
2898{
2899 int i;
2900 size_t sg_len = 0, len_complete = 0;
2901 struct scatterlist *sg;
2902 struct page *page;
2903
2904 WARN_ON(!irqs_disabled());
2905
2906 for_each_sg(sgl, sg, sg_count, i) {
2907 len_complete = sg_len; /* Complete sg-entries */
2908 sg_len += sg->length;
2909 if (sg_len > *offset)
2910 break;
2911 }
2912
2913 if (unlikely(i == sg_count)) {
2914 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2915 "elements %d\n",
2916 __func__, sg_len, *offset, sg_count);
2917 WARN_ON(1);
2918 return NULL;
2919 }
2920
2921 /* Offset starting from the beginning of first page in this sg-entry */
2922 *offset = *offset - len_complete + sg->offset;
2923
2924 /* Assumption: contiguous pages can be accessed as "page + i" */
2925 page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2926 *offset &= ~PAGE_MASK;
2927
2928 /* Bytes in this sg-entry from *offset to the end of the page */
2929 sg_len = PAGE_SIZE - *offset;
2930 if (*len > sg_len)
2931 *len = sg_len;
2932
2933 return kmap_atomic(page);
2934}
2935EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2936
2937/**
2938 * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2939 * @virt: virtual address to be unmapped
2940 */
2941void scsi_kunmap_atomic_sg(void *virt)
2942{
2943 kunmap_atomic(virt);
2944}
2945EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
2946
2947void sdev_disable_disk_events(struct scsi_device *sdev)
2948{
2949 atomic_inc(&sdev->disk_events_disable_depth);
2950}
2951EXPORT_SYMBOL(sdev_disable_disk_events);
2952
2953void sdev_enable_disk_events(struct scsi_device *sdev)
2954{
2955 if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
2956 return;
2957 atomic_dec(&sdev->disk_events_disable_depth);
2958}
2959EXPORT_SYMBOL(sdev_enable_disk_events);
2960
2961static unsigned char designator_prio(const unsigned char *d)
2962{
2963 if (d[1] & 0x30)
2964 /* not associated with LUN */
2965 return 0;
2966
2967 if (d[3] == 0)
2968 /* invalid length */
2969 return 0;
2970
2971 /*
2972 * Order of preference for lun descriptor:
2973 * - SCSI name string
2974 * - NAA IEEE Registered Extended
2975 * - EUI-64 based 16-byte
2976 * - EUI-64 based 12-byte
2977 * - NAA IEEE Registered
2978 * - NAA IEEE Extended
2979 * - EUI-64 based 8-byte
2980 * - SCSI name string (truncated)
2981 * - T10 Vendor ID
2982 * as longer descriptors reduce the likelyhood
2983 * of identification clashes.
2984 */
2985
2986 switch (d[1] & 0xf) {
2987 case 8:
2988 /* SCSI name string, variable-length UTF-8 */
2989 return 9;
2990 case 3:
2991 switch (d[4] >> 4) {
2992 case 6:
2993 /* NAA registered extended */
2994 return 8;
2995 case 5:
2996 /* NAA registered */
2997 return 5;
2998 case 4:
2999 /* NAA extended */
3000 return 4;
3001 case 3:
3002 /* NAA locally assigned */
3003 return 1;
3004 default:
3005 break;
3006 }
3007 break;
3008 case 2:
3009 switch (d[3]) {
3010 case 16:
3011 /* EUI64-based, 16 byte */
3012 return 7;
3013 case 12:
3014 /* EUI64-based, 12 byte */
3015 return 6;
3016 case 8:
3017 /* EUI64-based, 8 byte */
3018 return 3;
3019 default:
3020 break;
3021 }
3022 break;
3023 case 1:
3024 /* T10 vendor ID */
3025 return 1;
3026 default:
3027 break;
3028 }
3029
3030 return 0;
3031}
3032
3033/**
3034 * scsi_vpd_lun_id - return a unique device identification
3035 * @sdev: SCSI device
3036 * @id: buffer for the identification
3037 * @id_len: length of the buffer
3038 *
3039 * Copies a unique device identification into @id based
3040 * on the information in the VPD page 0x83 of the device.
3041 * The string will be formatted as a SCSI name string.
3042 *
3043 * Returns the length of the identification or error on failure.
3044 * If the identifier is longer than the supplied buffer the actual
3045 * identifier length is returned and the buffer is not zero-padded.
3046 */
3047int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
3048{
3049 u8 cur_id_prio = 0;
3050 u8 cur_id_size = 0;
3051 const unsigned char *d, *cur_id_str;
3052 const struct scsi_vpd *vpd_pg83;
3053 int id_size = -EINVAL;
3054
3055 rcu_read_lock();
3056 vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3057 if (!vpd_pg83) {
3058 rcu_read_unlock();
3059 return -ENXIO;
3060 }
3061
3062 /* The id string must be at least 20 bytes + terminating NULL byte */
3063 if (id_len < 21) {
3064 rcu_read_unlock();
3065 return -EINVAL;
3066 }
3067
3068 memset(id, 0, id_len);
3069 for (d = vpd_pg83->data + 4;
3070 d < vpd_pg83->data + vpd_pg83->len;
3071 d += d[3] + 4) {
3072 u8 prio = designator_prio(d);
3073
3074 if (prio == 0 || cur_id_prio > prio)
3075 continue;
3076
3077 switch (d[1] & 0xf) {
3078 case 0x1:
3079 /* T10 Vendor ID */
3080 if (cur_id_size > d[3])
3081 break;
3082 cur_id_prio = prio;
3083 cur_id_size = d[3];
3084 if (cur_id_size + 4 > id_len)
3085 cur_id_size = id_len - 4;
3086 cur_id_str = d + 4;
3087 id_size = snprintf(id, id_len, "t10.%*pE",
3088 cur_id_size, cur_id_str);
3089 break;
3090 case 0x2:
3091 /* EUI-64 */
3092 cur_id_prio = prio;
3093 cur_id_size = d[3];
3094 cur_id_str = d + 4;
3095 switch (cur_id_size) {
3096 case 8:
3097 id_size = snprintf(id, id_len,
3098 "eui.%8phN",
3099 cur_id_str);
3100 break;
3101 case 12:
3102 id_size = snprintf(id, id_len,
3103 "eui.%12phN",
3104 cur_id_str);
3105 break;
3106 case 16:
3107 id_size = snprintf(id, id_len,
3108 "eui.%16phN",
3109 cur_id_str);
3110 break;
3111 default:
3112 break;
3113 }
3114 break;
3115 case 0x3:
3116 /* NAA */
3117 cur_id_prio = prio;
3118 cur_id_size = d[3];
3119 cur_id_str = d + 4;
3120 switch (cur_id_size) {
3121 case 8:
3122 id_size = snprintf(id, id_len,
3123 "naa.%8phN",
3124 cur_id_str);
3125 break;
3126 case 16:
3127 id_size = snprintf(id, id_len,
3128 "naa.%16phN",
3129 cur_id_str);
3130 break;
3131 default:
3132 break;
3133 }
3134 break;
3135 case 0x8:
3136 /* SCSI name string */
3137 if (cur_id_size > d[3])
3138 break;
3139 /* Prefer others for truncated descriptor */
3140 if (d[3] > id_len) {
3141 prio = 2;
3142 if (cur_id_prio > prio)
3143 break;
3144 }
3145 cur_id_prio = prio;
3146 cur_id_size = id_size = d[3];
3147 cur_id_str = d + 4;
3148 if (cur_id_size >= id_len)
3149 cur_id_size = id_len - 1;
3150 memcpy(id, cur_id_str, cur_id_size);
3151 break;
3152 default:
3153 break;
3154 }
3155 }
3156 rcu_read_unlock();
3157
3158 return id_size;
3159}
3160EXPORT_SYMBOL(scsi_vpd_lun_id);
3161
3162/*
3163 * scsi_vpd_tpg_id - return a target port group identifier
3164 * @sdev: SCSI device
3165 *
3166 * Returns the Target Port Group identifier from the information
3167 * froom VPD page 0x83 of the device.
3168 *
3169 * Returns the identifier or error on failure.
3170 */
3171int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3172{
3173 const unsigned char *d;
3174 const struct scsi_vpd *vpd_pg83;
3175 int group_id = -EAGAIN, rel_port = -1;
3176
3177 rcu_read_lock();
3178 vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3179 if (!vpd_pg83) {
3180 rcu_read_unlock();
3181 return -ENXIO;
3182 }
3183
3184 d = vpd_pg83->data + 4;
3185 while (d < vpd_pg83->data + vpd_pg83->len) {
3186 switch (d[1] & 0xf) {
3187 case 0x4:
3188 /* Relative target port */
3189 rel_port = get_unaligned_be16(&d[6]);
3190 break;
3191 case 0x5:
3192 /* Target port group */
3193 group_id = get_unaligned_be16(&d[6]);
3194 break;
3195 default:
3196 break;
3197 }
3198 d += d[3] + 4;
3199 }
3200 rcu_read_unlock();
3201
3202 if (group_id >= 0 && rel_id && rel_port != -1)
3203 *rel_id = rel_port;
3204
3205 return group_id;
3206}
3207EXPORT_SYMBOL(scsi_vpd_tpg_id);
3208
3209/**
3210 * scsi_build_sense - build sense data for a command
3211 * @scmd: scsi command for which the sense should be formatted
3212 * @desc: Sense format (non-zero == descriptor format,
3213 * 0 == fixed format)
3214 * @key: Sense key
3215 * @asc: Additional sense code
3216 * @ascq: Additional sense code qualifier
3217 *
3218 **/
3219void scsi_build_sense(struct scsi_cmnd *scmd, int desc, u8 key, u8 asc, u8 ascq)
3220{
3221 scsi_build_sense_buffer(desc, scmd->sense_buffer, key, asc, ascq);
3222 scmd->result = SAM_STAT_CHECK_CONDITION;
3223}
3224EXPORT_SYMBOL_GPL(scsi_build_sense);