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/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 if (msecs) {
126 blk_mq_requeue_request(rq, false);
127 blk_mq_delay_kick_requeue_list(rq->q, msecs);
128 } else
129 blk_mq_requeue_request(rq, true);
130}
131
132/**
133 * __scsi_queue_insert - private queue insertion
134 * @cmd: The SCSI command being requeued
135 * @reason: The reason for the requeue
136 * @unbusy: Whether the queue should be unbusied
137 *
138 * This is a private queue insertion. The public interface
139 * scsi_queue_insert() always assumes the queue should be unbusied
140 * because it's always called before the completion. This function is
141 * for a requeue after completion, which should only occur in this
142 * file.
143 */
144static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, bool unbusy)
145{
146 struct scsi_device *device = cmd->device;
147
148 SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
149 "Inserting command %p into mlqueue\n", cmd));
150
151 scsi_set_blocked(cmd, reason);
152
153 /*
154 * Decrement the counters, since these commands are no longer
155 * active on the host/device.
156 */
157 if (unbusy)
158 scsi_device_unbusy(device, cmd);
159
160 /*
161 * Requeue this command. It will go before all other commands
162 * that are already in the queue. Schedule requeue work under
163 * lock such that the kblockd_schedule_work() call happens
164 * before blk_mq_destroy_queue() finishes.
165 */
166 cmd->result = 0;
167
168 blk_mq_requeue_request(scsi_cmd_to_rq(cmd), true);
169}
170
171/**
172 * scsi_queue_insert - Reinsert a command in the queue.
173 * @cmd: command that we are adding to queue.
174 * @reason: why we are inserting command to queue.
175 *
176 * We do this for one of two cases. Either the host is busy and it cannot accept
177 * any more commands for the time being, or the device returned QUEUE_FULL and
178 * can accept no more commands.
179 *
180 * Context: This could be called either from an interrupt context or a normal
181 * process context.
182 */
183void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
184{
185 __scsi_queue_insert(cmd, reason, true);
186}
187
188
189/**
190 * __scsi_execute - insert request and wait for the result
191 * @sdev: scsi device
192 * @cmd: scsi command
193 * @data_direction: data direction
194 * @buffer: data buffer
195 * @bufflen: len of buffer
196 * @sense: optional sense buffer
197 * @sshdr: optional decoded sense header
198 * @timeout: request timeout in HZ
199 * @retries: number of times to retry request
200 * @flags: flags for ->cmd_flags
201 * @rq_flags: flags for ->rq_flags
202 * @resid: optional residual length
203 *
204 * Returns the scsi_cmnd result field if a command was executed, or a negative
205 * Linux error code if we didn't get that far.
206 */
207int __scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
208 int data_direction, void *buffer, unsigned bufflen,
209 unsigned char *sense, struct scsi_sense_hdr *sshdr,
210 int timeout, int retries, blk_opf_t flags,
211 req_flags_t rq_flags, int *resid)
212{
213 struct request *req;
214 struct scsi_cmnd *scmd;
215 int ret;
216
217 req = scsi_alloc_request(sdev->request_queue,
218 data_direction == DMA_TO_DEVICE ?
219 REQ_OP_DRV_OUT : REQ_OP_DRV_IN,
220 rq_flags & RQF_PM ? BLK_MQ_REQ_PM : 0);
221 if (IS_ERR(req))
222 return PTR_ERR(req);
223
224 if (bufflen) {
225 ret = blk_rq_map_kern(sdev->request_queue, req,
226 buffer, bufflen, GFP_NOIO);
227 if (ret)
228 goto out;
229 }
230 scmd = blk_mq_rq_to_pdu(req);
231 scmd->cmd_len = COMMAND_SIZE(cmd[0]);
232 memcpy(scmd->cmnd, cmd, scmd->cmd_len);
233 scmd->allowed = 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(req, true);
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(scmd->resid_len > 0 && scmd->resid_len <= bufflen))
250 memset(buffer + bufflen - scmd->resid_len, 0, scmd->resid_len);
251
252 if (resid)
253 *resid = scmd->resid_len;
254 if (sense && scmd->sense_len)
255 memcpy(sense, scmd->sense_buffer, SCSI_SENSE_BUFFERSIZE);
256 if (sshdr)
257 scsi_normalize_sense(scmd->sense_buffer, scmd->sense_len,
258 sshdr);
259 ret = scmd->result;
260 out:
261 blk_mq_free_request(req);
262
263 return ret;
264}
265EXPORT_SYMBOL(__scsi_execute);
266
267/*
268 * Wake up the error handler if necessary. Avoid as follows that the error
269 * handler is not woken up if host in-flight requests number ==
270 * shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
271 * with an RCU read lock in this function to ensure that this function in
272 * its entirety either finishes before scsi_eh_scmd_add() increases the
273 * host_failed counter or that it notices the shost state change made by
274 * scsi_eh_scmd_add().
275 */
276static void scsi_dec_host_busy(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
277{
278 unsigned long flags;
279
280 rcu_read_lock();
281 __clear_bit(SCMD_STATE_INFLIGHT, &cmd->state);
282 if (unlikely(scsi_host_in_recovery(shost))) {
283 spin_lock_irqsave(shost->host_lock, flags);
284 if (shost->host_failed || shost->host_eh_scheduled)
285 scsi_eh_wakeup(shost);
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
305static void scsi_kick_queue(struct request_queue *q)
306{
307 blk_mq_run_hw_queues(q, false);
308}
309
310/*
311 * Kick the queue of SCSI device @sdev if @sdev != current_sdev. Called with
312 * interrupts disabled.
313 */
314static void scsi_kick_sdev_queue(struct scsi_device *sdev, void *data)
315{
316 struct scsi_device *current_sdev = data;
317
318 if (sdev != current_sdev)
319 blk_mq_run_hw_queues(sdev->request_queue, true);
320}
321
322/*
323 * Called for single_lun devices on IO completion. Clear starget_sdev_user,
324 * and call blk_run_queue for all the scsi_devices on the target -
325 * including current_sdev first.
326 *
327 * Called with *no* scsi locks held.
328 */
329static void scsi_single_lun_run(struct scsi_device *current_sdev)
330{
331 struct Scsi_Host *shost = current_sdev->host;
332 struct scsi_target *starget = scsi_target(current_sdev);
333 unsigned long flags;
334
335 spin_lock_irqsave(shost->host_lock, flags);
336 starget->starget_sdev_user = NULL;
337 spin_unlock_irqrestore(shost->host_lock, flags);
338
339 /*
340 * Call blk_run_queue for all LUNs on the target, starting with
341 * current_sdev. We race with others (to set starget_sdev_user),
342 * but in most cases, we will be first. Ideally, each LU on the
343 * target would get some limited time or requests on the target.
344 */
345 scsi_kick_queue(current_sdev->request_queue);
346
347 spin_lock_irqsave(shost->host_lock, flags);
348 if (!starget->starget_sdev_user)
349 __starget_for_each_device(starget, current_sdev,
350 scsi_kick_sdev_queue);
351 spin_unlock_irqrestore(shost->host_lock, flags);
352}
353
354static inline bool scsi_device_is_busy(struct scsi_device *sdev)
355{
356 if (scsi_device_busy(sdev) >= sdev->queue_depth)
357 return true;
358 if (atomic_read(&sdev->device_blocked) > 0)
359 return true;
360 return false;
361}
362
363static inline bool scsi_target_is_busy(struct scsi_target *starget)
364{
365 if (starget->can_queue > 0) {
366 if (atomic_read(&starget->target_busy) >= starget->can_queue)
367 return true;
368 if (atomic_read(&starget->target_blocked) > 0)
369 return true;
370 }
371 return false;
372}
373
374static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
375{
376 if (atomic_read(&shost->host_blocked) > 0)
377 return true;
378 if (shost->host_self_blocked)
379 return true;
380 return false;
381}
382
383static void scsi_starved_list_run(struct Scsi_Host *shost)
384{
385 LIST_HEAD(starved_list);
386 struct scsi_device *sdev;
387 unsigned long flags;
388
389 spin_lock_irqsave(shost->host_lock, flags);
390 list_splice_init(&shost->starved_list, &starved_list);
391
392 while (!list_empty(&starved_list)) {
393 struct request_queue *slq;
394
395 /*
396 * As long as shost is accepting commands and we have
397 * starved queues, call blk_run_queue. scsi_request_fn
398 * drops the queue_lock and can add us back to the
399 * starved_list.
400 *
401 * host_lock protects the starved_list and starved_entry.
402 * scsi_request_fn must get the host_lock before checking
403 * or modifying starved_list or starved_entry.
404 */
405 if (scsi_host_is_busy(shost))
406 break;
407
408 sdev = list_entry(starved_list.next,
409 struct scsi_device, starved_entry);
410 list_del_init(&sdev->starved_entry);
411 if (scsi_target_is_busy(scsi_target(sdev))) {
412 list_move_tail(&sdev->starved_entry,
413 &shost->starved_list);
414 continue;
415 }
416
417 /*
418 * Once we drop the host lock, a racing scsi_remove_device()
419 * call may remove the sdev from the starved list and destroy
420 * it and the queue. Mitigate by taking a reference to the
421 * queue and never touching the sdev again after we drop the
422 * host lock. Note: if __scsi_remove_device() invokes
423 * blk_mq_destroy_queue() before the queue is run from this
424 * function then blk_run_queue() will return immediately since
425 * blk_mq_destroy_queue() marks the queue with QUEUE_FLAG_DYING.
426 */
427 slq = sdev->request_queue;
428 if (!blk_get_queue(slq))
429 continue;
430 spin_unlock_irqrestore(shost->host_lock, flags);
431
432 scsi_kick_queue(slq);
433 blk_put_queue(slq);
434
435 spin_lock_irqsave(shost->host_lock, flags);
436 }
437 /* put any unprocessed entries back */
438 list_splice(&starved_list, &shost->starved_list);
439 spin_unlock_irqrestore(shost->host_lock, flags);
440}
441
442/**
443 * scsi_run_queue - Select a proper request queue to serve next.
444 * @q: last request's queue
445 *
446 * The previous command was completely finished, start a new one if possible.
447 */
448static void scsi_run_queue(struct request_queue *q)
449{
450 struct scsi_device *sdev = q->queuedata;
451
452 if (scsi_target(sdev)->single_lun)
453 scsi_single_lun_run(sdev);
454 if (!list_empty(&sdev->host->starved_list))
455 scsi_starved_list_run(sdev->host);
456
457 blk_mq_run_hw_queues(q, false);
458}
459
460void scsi_requeue_run_queue(struct work_struct *work)
461{
462 struct scsi_device *sdev;
463 struct request_queue *q;
464
465 sdev = container_of(work, struct scsi_device, requeue_work);
466 q = sdev->request_queue;
467 scsi_run_queue(q);
468}
469
470void scsi_run_host_queues(struct Scsi_Host *shost)
471{
472 struct scsi_device *sdev;
473
474 shost_for_each_device(sdev, shost)
475 scsi_run_queue(sdev->request_queue);
476}
477
478static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
479{
480 if (!blk_rq_is_passthrough(scsi_cmd_to_rq(cmd))) {
481 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
482
483 if (drv->uninit_command)
484 drv->uninit_command(cmd);
485 }
486}
487
488void scsi_free_sgtables(struct scsi_cmnd *cmd)
489{
490 if (cmd->sdb.table.nents)
491 sg_free_table_chained(&cmd->sdb.table,
492 SCSI_INLINE_SG_CNT);
493 if (scsi_prot_sg_count(cmd))
494 sg_free_table_chained(&cmd->prot_sdb->table,
495 SCSI_INLINE_PROT_SG_CNT);
496}
497EXPORT_SYMBOL_GPL(scsi_free_sgtables);
498
499static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
500{
501 scsi_free_sgtables(cmd);
502 scsi_uninit_cmd(cmd);
503}
504
505static void scsi_run_queue_async(struct scsi_device *sdev)
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
582static inline u8 get_scsi_ml_byte(int result)
583{
584 return (result >> 8) & 0xff;
585}
586
587/**
588 * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
589 * @result: scsi error code
590 *
591 * Translate a SCSI result code into a blk_status_t value.
592 */
593static blk_status_t scsi_result_to_blk_status(int result)
594{
595 /*
596 * Check the scsi-ml byte first in case we converted a host or status
597 * byte.
598 */
599 switch (get_scsi_ml_byte(result)) {
600 case SCSIML_STAT_OK:
601 break;
602 case SCSIML_STAT_RESV_CONFLICT:
603 return BLK_STS_NEXUS;
604 case SCSIML_STAT_NOSPC:
605 return BLK_STS_NOSPC;
606 case SCSIML_STAT_MED_ERROR:
607 return BLK_STS_MEDIUM;
608 case SCSIML_STAT_TGT_FAILURE:
609 return BLK_STS_TARGET;
610 }
611
612 switch (host_byte(result)) {
613 case DID_OK:
614 if (scsi_status_is_good(result))
615 return BLK_STS_OK;
616 return BLK_STS_IOERR;
617 case DID_TRANSPORT_FAILFAST:
618 case DID_TRANSPORT_MARGINAL:
619 return BLK_STS_TRANSPORT;
620 default:
621 return BLK_STS_IOERR;
622 }
623}
624
625/**
626 * scsi_rq_err_bytes - determine number of bytes till the next failure boundary
627 * @rq: request to examine
628 *
629 * Description:
630 * A request could be merge of IOs which require different failure
631 * handling. This function determines the number of bytes which
632 * can be failed from the beginning of the request without
633 * crossing into area which need to be retried further.
634 *
635 * Return:
636 * The number of bytes to fail.
637 */
638static unsigned int scsi_rq_err_bytes(const struct request *rq)
639{
640 blk_opf_t ff = rq->cmd_flags & REQ_FAILFAST_MASK;
641 unsigned int bytes = 0;
642 struct bio *bio;
643
644 if (!(rq->rq_flags & RQF_MIXED_MERGE))
645 return blk_rq_bytes(rq);
646
647 /*
648 * Currently the only 'mixing' which can happen is between
649 * different fastfail types. We can safely fail portions
650 * which have all the failfast bits that the first one has -
651 * the ones which are at least as eager to fail as the first
652 * one.
653 */
654 for (bio = rq->bio; bio; bio = bio->bi_next) {
655 if ((bio->bi_opf & ff) != ff)
656 break;
657 bytes += bio->bi_iter.bi_size;
658 }
659
660 /* this could lead to infinite loop */
661 BUG_ON(blk_rq_bytes(rq) && !bytes);
662 return bytes;
663}
664
665static bool scsi_cmd_runtime_exceeced(struct scsi_cmnd *cmd)
666{
667 struct request *req = scsi_cmd_to_rq(cmd);
668 unsigned long wait_for;
669
670 if (cmd->allowed == SCSI_CMD_RETRIES_NO_LIMIT)
671 return false;
672
673 wait_for = (cmd->allowed + 1) * req->timeout;
674 if (time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
675 scmd_printk(KERN_ERR, cmd, "timing out command, waited %lus\n",
676 wait_for/HZ);
677 return true;
678 }
679 return false;
680}
681
682/*
683 * When ALUA transition state is returned, reprep the cmd to
684 * use the ALUA handler's transition timeout. Delay the reprep
685 * 1 sec to avoid aggressive retries of the target in that
686 * state.
687 */
688#define ALUA_TRANSITION_REPREP_DELAY 1000
689
690/* Helper for scsi_io_completion() when special action required. */
691static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
692{
693 struct request *req = scsi_cmd_to_rq(cmd);
694 int level = 0;
695 enum {ACTION_FAIL, ACTION_REPREP, ACTION_DELAYED_REPREP,
696 ACTION_RETRY, ACTION_DELAYED_RETRY} action;
697 struct scsi_sense_hdr sshdr;
698 bool sense_valid;
699 bool sense_current = true; /* false implies "deferred sense" */
700 blk_status_t blk_stat;
701
702 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
703 if (sense_valid)
704 sense_current = !scsi_sense_is_deferred(&sshdr);
705
706 blk_stat = scsi_result_to_blk_status(result);
707
708 if (host_byte(result) == DID_RESET) {
709 /* Third party bus reset or reset for error recovery
710 * reasons. Just retry the command and see what
711 * happens.
712 */
713 action = ACTION_RETRY;
714 } else if (sense_valid && sense_current) {
715 switch (sshdr.sense_key) {
716 case UNIT_ATTENTION:
717 if (cmd->device->removable) {
718 /* Detected disc change. Set a bit
719 * and quietly refuse further access.
720 */
721 cmd->device->changed = 1;
722 action = ACTION_FAIL;
723 } else {
724 /* Must have been a power glitch, or a
725 * bus reset. Could not have been a
726 * media change, so we just retry the
727 * command and see what happens.
728 */
729 action = ACTION_RETRY;
730 }
731 break;
732 case ILLEGAL_REQUEST:
733 /* If we had an ILLEGAL REQUEST returned, then
734 * we may have performed an unsupported
735 * command. The only thing this should be
736 * would be a ten byte read where only a six
737 * byte read was supported. Also, on a system
738 * where READ CAPACITY failed, we may have
739 * read past the end of the disk.
740 */
741 if ((cmd->device->use_10_for_rw &&
742 sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
743 (cmd->cmnd[0] == READ_10 ||
744 cmd->cmnd[0] == WRITE_10)) {
745 /* This will issue a new 6-byte command. */
746 cmd->device->use_10_for_rw = 0;
747 action = ACTION_REPREP;
748 } else if (sshdr.asc == 0x10) /* DIX */ {
749 action = ACTION_FAIL;
750 blk_stat = BLK_STS_PROTECTION;
751 /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
752 } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
753 action = ACTION_FAIL;
754 blk_stat = BLK_STS_TARGET;
755 } else
756 action = ACTION_FAIL;
757 break;
758 case ABORTED_COMMAND:
759 action = ACTION_FAIL;
760 if (sshdr.asc == 0x10) /* DIF */
761 blk_stat = BLK_STS_PROTECTION;
762 break;
763 case NOT_READY:
764 /* If the device is in the process of becoming
765 * ready, or has a temporary blockage, retry.
766 */
767 if (sshdr.asc == 0x04) {
768 switch (sshdr.ascq) {
769 case 0x01: /* becoming ready */
770 case 0x04: /* format in progress */
771 case 0x05: /* rebuild in progress */
772 case 0x06: /* recalculation in progress */
773 case 0x07: /* operation in progress */
774 case 0x08: /* Long write in progress */
775 case 0x09: /* self test in progress */
776 case 0x11: /* notify (enable spinup) required */
777 case 0x14: /* space allocation in progress */
778 case 0x1a: /* start stop unit in progress */
779 case 0x1b: /* sanitize in progress */
780 case 0x1d: /* configuration in progress */
781 case 0x24: /* depopulation in progress */
782 action = ACTION_DELAYED_RETRY;
783 break;
784 case 0x0a: /* ALUA state transition */
785 action = ACTION_DELAYED_REPREP;
786 break;
787 default:
788 action = ACTION_FAIL;
789 break;
790 }
791 } else
792 action = ACTION_FAIL;
793 break;
794 case VOLUME_OVERFLOW:
795 /* See SSC3rXX or current. */
796 action = ACTION_FAIL;
797 break;
798 case DATA_PROTECT:
799 action = ACTION_FAIL;
800 if ((sshdr.asc == 0x0C && sshdr.ascq == 0x12) ||
801 (sshdr.asc == 0x55 &&
802 (sshdr.ascq == 0x0E || sshdr.ascq == 0x0F))) {
803 /* Insufficient zone resources */
804 blk_stat = BLK_STS_ZONE_OPEN_RESOURCE;
805 }
806 break;
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 (atomic_read(&sdev->device_blocked)) {
1257 if (token < 0)
1258 goto out;
1259
1260 if (scsi_device_busy(sdev) > 1)
1261 goto out_dec;
1262
1263 /*
1264 * unblock after device_blocked iterates to zero
1265 */
1266 if (atomic_dec_return(&sdev->device_blocked) > 0)
1267 goto out_dec;
1268 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1269 "unblocking device at zero depth\n"));
1270 }
1271
1272 return token;
1273out_dec:
1274 if (token >= 0)
1275 sbitmap_put(&sdev->budget_map, token);
1276out:
1277 return -1;
1278}
1279
1280/*
1281 * scsi_target_queue_ready: checks if there we can send commands to target
1282 * @sdev: scsi device on starget to check.
1283 */
1284static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1285 struct scsi_device *sdev)
1286{
1287 struct scsi_target *starget = scsi_target(sdev);
1288 unsigned int busy;
1289
1290 if (starget->single_lun) {
1291 spin_lock_irq(shost->host_lock);
1292 if (starget->starget_sdev_user &&
1293 starget->starget_sdev_user != sdev) {
1294 spin_unlock_irq(shost->host_lock);
1295 return 0;
1296 }
1297 starget->starget_sdev_user = sdev;
1298 spin_unlock_irq(shost->host_lock);
1299 }
1300
1301 if (starget->can_queue <= 0)
1302 return 1;
1303
1304 busy = atomic_inc_return(&starget->target_busy) - 1;
1305 if (atomic_read(&starget->target_blocked) > 0) {
1306 if (busy)
1307 goto starved;
1308
1309 /*
1310 * unblock after target_blocked iterates to zero
1311 */
1312 if (atomic_dec_return(&starget->target_blocked) > 0)
1313 goto out_dec;
1314
1315 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1316 "unblocking target at zero depth\n"));
1317 }
1318
1319 if (busy >= starget->can_queue)
1320 goto starved;
1321
1322 return 1;
1323
1324starved:
1325 spin_lock_irq(shost->host_lock);
1326 list_move_tail(&sdev->starved_entry, &shost->starved_list);
1327 spin_unlock_irq(shost->host_lock);
1328out_dec:
1329 if (starget->can_queue > 0)
1330 atomic_dec(&starget->target_busy);
1331 return 0;
1332}
1333
1334/*
1335 * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1336 * return 0. We must end up running the queue again whenever 0 is
1337 * returned, else IO can hang.
1338 */
1339static inline int scsi_host_queue_ready(struct request_queue *q,
1340 struct Scsi_Host *shost,
1341 struct scsi_device *sdev,
1342 struct scsi_cmnd *cmd)
1343{
1344 if (atomic_read(&shost->host_blocked) > 0) {
1345 if (scsi_host_busy(shost) > 0)
1346 goto starved;
1347
1348 /*
1349 * unblock after host_blocked iterates to zero
1350 */
1351 if (atomic_dec_return(&shost->host_blocked) > 0)
1352 goto out_dec;
1353
1354 SCSI_LOG_MLQUEUE(3,
1355 shost_printk(KERN_INFO, shost,
1356 "unblocking host at zero depth\n"));
1357 }
1358
1359 if (shost->host_self_blocked)
1360 goto starved;
1361
1362 /* We're OK to process the command, so we can't be starved */
1363 if (!list_empty(&sdev->starved_entry)) {
1364 spin_lock_irq(shost->host_lock);
1365 if (!list_empty(&sdev->starved_entry))
1366 list_del_init(&sdev->starved_entry);
1367 spin_unlock_irq(shost->host_lock);
1368 }
1369
1370 __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1371
1372 return 1;
1373
1374starved:
1375 spin_lock_irq(shost->host_lock);
1376 if (list_empty(&sdev->starved_entry))
1377 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1378 spin_unlock_irq(shost->host_lock);
1379out_dec:
1380 scsi_dec_host_busy(shost, cmd);
1381 return 0;
1382}
1383
1384/*
1385 * Busy state exporting function for request stacking drivers.
1386 *
1387 * For efficiency, no lock is taken to check the busy state of
1388 * shost/starget/sdev, since the returned value is not guaranteed and
1389 * may be changed after request stacking drivers call the function,
1390 * regardless of taking lock or not.
1391 *
1392 * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1393 * needs to return 'not busy'. Otherwise, request stacking drivers
1394 * may hold requests forever.
1395 */
1396static bool scsi_mq_lld_busy(struct request_queue *q)
1397{
1398 struct scsi_device *sdev = q->queuedata;
1399 struct Scsi_Host *shost;
1400
1401 if (blk_queue_dying(q))
1402 return false;
1403
1404 shost = sdev->host;
1405
1406 /*
1407 * Ignore host/starget busy state.
1408 * Since block layer does not have a concept of fairness across
1409 * multiple queues, congestion of host/starget needs to be handled
1410 * in SCSI layer.
1411 */
1412 if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1413 return true;
1414
1415 return false;
1416}
1417
1418/*
1419 * Block layer request completion callback. May be called from interrupt
1420 * context.
1421 */
1422static void scsi_complete(struct request *rq)
1423{
1424 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1425 enum scsi_disposition disposition;
1426
1427 INIT_LIST_HEAD(&cmd->eh_entry);
1428
1429 atomic_inc(&cmd->device->iodone_cnt);
1430 if (cmd->result)
1431 atomic_inc(&cmd->device->ioerr_cnt);
1432
1433 disposition = scsi_decide_disposition(cmd);
1434 if (disposition != SUCCESS && scsi_cmd_runtime_exceeced(cmd))
1435 disposition = SUCCESS;
1436
1437 scsi_log_completion(cmd, disposition);
1438
1439 switch (disposition) {
1440 case SUCCESS:
1441 scsi_finish_command(cmd);
1442 break;
1443 case NEEDS_RETRY:
1444 scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1445 break;
1446 case ADD_TO_MLQUEUE:
1447 scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1448 break;
1449 default:
1450 scsi_eh_scmd_add(cmd);
1451 break;
1452 }
1453}
1454
1455/**
1456 * scsi_dispatch_cmd - Dispatch a command to the low-level driver.
1457 * @cmd: command block we are dispatching.
1458 *
1459 * Return: nonzero return request was rejected and device's queue needs to be
1460 * plugged.
1461 */
1462static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1463{
1464 struct Scsi_Host *host = cmd->device->host;
1465 int rtn = 0;
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 return SCSI_MLQUEUE_DEVICE_BUSY;
1488 }
1489
1490 /* Store the LUN value in cmnd, if needed. */
1491 if (cmd->device->lun_in_cdb)
1492 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1493 (cmd->device->lun << 5 & 0xe0);
1494
1495 scsi_log_send(cmd);
1496
1497 /*
1498 * Before we queue this command, check if the command
1499 * length exceeds what the host adapter can handle.
1500 */
1501 if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1502 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1503 "queuecommand : command too long. "
1504 "cdb_size=%d host->max_cmd_len=%d\n",
1505 cmd->cmd_len, cmd->device->host->max_cmd_len));
1506 cmd->result = (DID_ABORT << 16);
1507 goto done;
1508 }
1509
1510 if (unlikely(host->shost_state == SHOST_DEL)) {
1511 cmd->result = (DID_NO_CONNECT << 16);
1512 goto done;
1513
1514 }
1515
1516 trace_scsi_dispatch_cmd_start(cmd);
1517 rtn = host->hostt->queuecommand(host, cmd);
1518 if (rtn) {
1519 trace_scsi_dispatch_cmd_error(cmd, rtn);
1520 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1521 rtn != SCSI_MLQUEUE_TARGET_BUSY)
1522 rtn = SCSI_MLQUEUE_HOST_BUSY;
1523
1524 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1525 "queuecommand : request rejected\n"));
1526 }
1527
1528 return rtn;
1529 done:
1530 scsi_done(cmd);
1531 return 0;
1532}
1533
1534/* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1535static unsigned int scsi_mq_inline_sgl_size(struct Scsi_Host *shost)
1536{
1537 return min_t(unsigned int, shost->sg_tablesize, SCSI_INLINE_SG_CNT) *
1538 sizeof(struct scatterlist);
1539}
1540
1541static blk_status_t scsi_prepare_cmd(struct request *req)
1542{
1543 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1544 struct scsi_device *sdev = req->q->queuedata;
1545 struct Scsi_Host *shost = sdev->host;
1546 bool in_flight = test_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1547 struct scatterlist *sg;
1548
1549 scsi_init_command(sdev, cmd);
1550
1551 cmd->eh_eflags = 0;
1552 cmd->prot_type = 0;
1553 cmd->prot_flags = 0;
1554 cmd->submitter = 0;
1555 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1556 cmd->underflow = 0;
1557 cmd->transfersize = 0;
1558 cmd->host_scribble = NULL;
1559 cmd->result = 0;
1560 cmd->extra_len = 0;
1561 cmd->state = 0;
1562 if (in_flight)
1563 __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1564
1565 /*
1566 * Only clear the driver-private command data if the LLD does not supply
1567 * a function to initialize that data.
1568 */
1569 if (!shost->hostt->init_cmd_priv)
1570 memset(cmd + 1, 0, shost->hostt->cmd_size);
1571
1572 cmd->prot_op = SCSI_PROT_NORMAL;
1573 if (blk_rq_bytes(req))
1574 cmd->sc_data_direction = rq_dma_dir(req);
1575 else
1576 cmd->sc_data_direction = DMA_NONE;
1577
1578 sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1579 cmd->sdb.table.sgl = sg;
1580
1581 if (scsi_host_get_prot(shost)) {
1582 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1583
1584 cmd->prot_sdb->table.sgl =
1585 (struct scatterlist *)(cmd->prot_sdb + 1);
1586 }
1587
1588 /*
1589 * Special handling for passthrough commands, which don't go to the ULP
1590 * at all:
1591 */
1592 if (blk_rq_is_passthrough(req))
1593 return scsi_setup_scsi_cmnd(sdev, req);
1594
1595 if (sdev->handler && sdev->handler->prep_fn) {
1596 blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1597
1598 if (ret != BLK_STS_OK)
1599 return ret;
1600 }
1601
1602 /* Usually overridden by the ULP */
1603 cmd->allowed = 0;
1604 memset(cmd->cmnd, 0, sizeof(cmd->cmnd));
1605 return scsi_cmd_to_driver(cmd)->init_command(cmd);
1606}
1607
1608static void scsi_done_internal(struct scsi_cmnd *cmd, bool complete_directly)
1609{
1610 struct request *req = scsi_cmd_to_rq(cmd);
1611
1612 switch (cmd->submitter) {
1613 case SUBMITTED_BY_BLOCK_LAYER:
1614 break;
1615 case SUBMITTED_BY_SCSI_ERROR_HANDLER:
1616 return scsi_eh_done(cmd);
1617 case SUBMITTED_BY_SCSI_RESET_IOCTL:
1618 return;
1619 }
1620
1621 if (unlikely(blk_should_fake_timeout(scsi_cmd_to_rq(cmd)->q)))
1622 return;
1623 if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1624 return;
1625 trace_scsi_dispatch_cmd_done(cmd);
1626
1627 if (complete_directly)
1628 blk_mq_complete_request_direct(req, scsi_complete);
1629 else
1630 blk_mq_complete_request(req);
1631}
1632
1633void scsi_done(struct scsi_cmnd *cmd)
1634{
1635 scsi_done_internal(cmd, false);
1636}
1637EXPORT_SYMBOL(scsi_done);
1638
1639void scsi_done_direct(struct scsi_cmnd *cmd)
1640{
1641 scsi_done_internal(cmd, true);
1642}
1643EXPORT_SYMBOL(scsi_done_direct);
1644
1645static void scsi_mq_put_budget(struct request_queue *q, int budget_token)
1646{
1647 struct scsi_device *sdev = q->queuedata;
1648
1649 sbitmap_put(&sdev->budget_map, budget_token);
1650}
1651
1652/*
1653 * When to reinvoke queueing after a resource shortage. It's 3 msecs to
1654 * not change behaviour from the previous unplug mechanism, experimentation
1655 * may prove this needs changing.
1656 */
1657#define SCSI_QUEUE_DELAY 3
1658
1659static int scsi_mq_get_budget(struct request_queue *q)
1660{
1661 struct scsi_device *sdev = q->queuedata;
1662 int token = scsi_dev_queue_ready(q, sdev);
1663
1664 if (token >= 0)
1665 return token;
1666
1667 atomic_inc(&sdev->restarts);
1668
1669 /*
1670 * Orders atomic_inc(&sdev->restarts) and atomic_read(&sdev->device_busy).
1671 * .restarts must be incremented before .device_busy is read because the
1672 * code in scsi_run_queue_async() depends on the order of these operations.
1673 */
1674 smp_mb__after_atomic();
1675
1676 /*
1677 * If all in-flight requests originated from this LUN are completed
1678 * before reading .device_busy, sdev->device_busy will be observed as
1679 * zero, then blk_mq_delay_run_hw_queues() will dispatch this request
1680 * soon. Otherwise, completion of one of these requests will observe
1681 * the .restarts flag, and the request queue will be run for handling
1682 * this request, see scsi_end_request().
1683 */
1684 if (unlikely(scsi_device_busy(sdev) == 0 &&
1685 !scsi_device_blocked(sdev)))
1686 blk_mq_delay_run_hw_queues(sdev->request_queue, SCSI_QUEUE_DELAY);
1687 return -1;
1688}
1689
1690static void scsi_mq_set_rq_budget_token(struct request *req, int token)
1691{
1692 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1693
1694 cmd->budget_token = token;
1695}
1696
1697static int scsi_mq_get_rq_budget_token(struct request *req)
1698{
1699 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1700
1701 return cmd->budget_token;
1702}
1703
1704static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1705 const struct blk_mq_queue_data *bd)
1706{
1707 struct request *req = bd->rq;
1708 struct request_queue *q = req->q;
1709 struct scsi_device *sdev = q->queuedata;
1710 struct Scsi_Host *shost = sdev->host;
1711 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1712 blk_status_t ret;
1713 int reason;
1714
1715 WARN_ON_ONCE(cmd->budget_token < 0);
1716
1717 /*
1718 * If the device is not in running state we will reject some or all
1719 * commands.
1720 */
1721 if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1722 ret = scsi_device_state_check(sdev, req);
1723 if (ret != BLK_STS_OK)
1724 goto out_put_budget;
1725 }
1726
1727 ret = BLK_STS_RESOURCE;
1728 if (!scsi_target_queue_ready(shost, sdev))
1729 goto out_put_budget;
1730 if (unlikely(scsi_host_in_recovery(shost))) {
1731 if (cmd->flags & SCMD_FAIL_IF_RECOVERING)
1732 ret = BLK_STS_OFFLINE;
1733 goto out_dec_target_busy;
1734 }
1735 if (!scsi_host_queue_ready(q, shost, sdev, cmd))
1736 goto out_dec_target_busy;
1737
1738 if (!(req->rq_flags & RQF_DONTPREP)) {
1739 ret = scsi_prepare_cmd(req);
1740 if (ret != BLK_STS_OK)
1741 goto out_dec_host_busy;
1742 req->rq_flags |= RQF_DONTPREP;
1743 } else {
1744 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1745 }
1746
1747 cmd->flags &= SCMD_PRESERVED_FLAGS;
1748 if (sdev->simple_tags)
1749 cmd->flags |= SCMD_TAGGED;
1750 if (bd->last)
1751 cmd->flags |= SCMD_LAST;
1752
1753 scsi_set_resid(cmd, 0);
1754 memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
1755 cmd->submitter = SUBMITTED_BY_BLOCK_LAYER;
1756
1757 blk_mq_start_request(req);
1758 reason = scsi_dispatch_cmd(cmd);
1759 if (reason) {
1760 scsi_set_blocked(cmd, reason);
1761 ret = BLK_STS_RESOURCE;
1762 goto out_dec_host_busy;
1763 }
1764
1765 atomic_inc(&cmd->device->iorequest_cnt);
1766 return BLK_STS_OK;
1767
1768out_dec_host_busy:
1769 scsi_dec_host_busy(shost, cmd);
1770out_dec_target_busy:
1771 if (scsi_target(sdev)->can_queue > 0)
1772 atomic_dec(&scsi_target(sdev)->target_busy);
1773out_put_budget:
1774 scsi_mq_put_budget(q, cmd->budget_token);
1775 cmd->budget_token = -1;
1776 switch (ret) {
1777 case BLK_STS_OK:
1778 break;
1779 case BLK_STS_RESOURCE:
1780 case BLK_STS_ZONE_RESOURCE:
1781 if (scsi_device_blocked(sdev))
1782 ret = BLK_STS_DEV_RESOURCE;
1783 break;
1784 case BLK_STS_AGAIN:
1785 cmd->result = DID_BUS_BUSY << 16;
1786 if (req->rq_flags & RQF_DONTPREP)
1787 scsi_mq_uninit_cmd(cmd);
1788 break;
1789 default:
1790 if (unlikely(!scsi_device_online(sdev)))
1791 cmd->result = DID_NO_CONNECT << 16;
1792 else
1793 cmd->result = DID_ERROR << 16;
1794 /*
1795 * Make sure to release all allocated resources when
1796 * we hit an error, as we will never see this command
1797 * again.
1798 */
1799 if (req->rq_flags & RQF_DONTPREP)
1800 scsi_mq_uninit_cmd(cmd);
1801 scsi_run_queue_async(sdev);
1802 break;
1803 }
1804 return ret;
1805}
1806
1807static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1808 unsigned int hctx_idx, unsigned int numa_node)
1809{
1810 struct Scsi_Host *shost = set->driver_data;
1811 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1812 struct scatterlist *sg;
1813 int ret = 0;
1814
1815 cmd->sense_buffer =
1816 kmem_cache_alloc_node(scsi_sense_cache, GFP_KERNEL, numa_node);
1817 if (!cmd->sense_buffer)
1818 return -ENOMEM;
1819
1820 if (scsi_host_get_prot(shost)) {
1821 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1822 shost->hostt->cmd_size;
1823 cmd->prot_sdb = (void *)sg + scsi_mq_inline_sgl_size(shost);
1824 }
1825
1826 if (shost->hostt->init_cmd_priv) {
1827 ret = shost->hostt->init_cmd_priv(shost, cmd);
1828 if (ret < 0)
1829 kmem_cache_free(scsi_sense_cache, cmd->sense_buffer);
1830 }
1831
1832 return ret;
1833}
1834
1835static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1836 unsigned int hctx_idx)
1837{
1838 struct Scsi_Host *shost = set->driver_data;
1839 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1840
1841 if (shost->hostt->exit_cmd_priv)
1842 shost->hostt->exit_cmd_priv(shost, cmd);
1843 kmem_cache_free(scsi_sense_cache, cmd->sense_buffer);
1844}
1845
1846
1847static int scsi_mq_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
1848{
1849 struct Scsi_Host *shost = hctx->driver_data;
1850
1851 if (shost->hostt->mq_poll)
1852 return shost->hostt->mq_poll(shost, hctx->queue_num);
1853
1854 return 0;
1855}
1856
1857static int scsi_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
1858 unsigned int hctx_idx)
1859{
1860 struct Scsi_Host *shost = data;
1861
1862 hctx->driver_data = shost;
1863 return 0;
1864}
1865
1866static void scsi_map_queues(struct blk_mq_tag_set *set)
1867{
1868 struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
1869
1870 if (shost->hostt->map_queues)
1871 return shost->hostt->map_queues(shost);
1872 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
1873}
1874
1875void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
1876{
1877 struct device *dev = shost->dma_dev;
1878
1879 /*
1880 * this limit is imposed by hardware restrictions
1881 */
1882 blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1883 SG_MAX_SEGMENTS));
1884
1885 if (scsi_host_prot_dma(shost)) {
1886 shost->sg_prot_tablesize =
1887 min_not_zero(shost->sg_prot_tablesize,
1888 (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1889 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1890 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1891 }
1892
1893 blk_queue_max_hw_sectors(q, shost->max_sectors);
1894 blk_queue_segment_boundary(q, shost->dma_boundary);
1895 dma_set_seg_boundary(dev, shost->dma_boundary);
1896
1897 blk_queue_max_segment_size(q, shost->max_segment_size);
1898 blk_queue_virt_boundary(q, shost->virt_boundary_mask);
1899 dma_set_max_seg_size(dev, queue_max_segment_size(q));
1900
1901 /*
1902 * Set a reasonable default alignment: The larger of 32-byte (dword),
1903 * which is a common minimum for HBAs, and the minimum DMA alignment,
1904 * which is set by the platform.
1905 *
1906 * Devices that require a bigger alignment can increase it later.
1907 */
1908 blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
1909}
1910EXPORT_SYMBOL_GPL(__scsi_init_queue);
1911
1912static const struct blk_mq_ops scsi_mq_ops_no_commit = {
1913 .get_budget = scsi_mq_get_budget,
1914 .put_budget = scsi_mq_put_budget,
1915 .queue_rq = scsi_queue_rq,
1916 .complete = scsi_complete,
1917 .timeout = scsi_timeout,
1918#ifdef CONFIG_BLK_DEBUG_FS
1919 .show_rq = scsi_show_rq,
1920#endif
1921 .init_request = scsi_mq_init_request,
1922 .exit_request = scsi_mq_exit_request,
1923 .cleanup_rq = scsi_cleanup_rq,
1924 .busy = scsi_mq_lld_busy,
1925 .map_queues = scsi_map_queues,
1926 .init_hctx = scsi_init_hctx,
1927 .poll = scsi_mq_poll,
1928 .set_rq_budget_token = scsi_mq_set_rq_budget_token,
1929 .get_rq_budget_token = scsi_mq_get_rq_budget_token,
1930};
1931
1932
1933static void scsi_commit_rqs(struct blk_mq_hw_ctx *hctx)
1934{
1935 struct Scsi_Host *shost = hctx->driver_data;
1936
1937 shost->hostt->commit_rqs(shost, hctx->queue_num);
1938}
1939
1940static const struct blk_mq_ops scsi_mq_ops = {
1941 .get_budget = scsi_mq_get_budget,
1942 .put_budget = scsi_mq_put_budget,
1943 .queue_rq = scsi_queue_rq,
1944 .commit_rqs = scsi_commit_rqs,
1945 .complete = scsi_complete,
1946 .timeout = scsi_timeout,
1947#ifdef CONFIG_BLK_DEBUG_FS
1948 .show_rq = scsi_show_rq,
1949#endif
1950 .init_request = scsi_mq_init_request,
1951 .exit_request = scsi_mq_exit_request,
1952 .cleanup_rq = scsi_cleanup_rq,
1953 .busy = scsi_mq_lld_busy,
1954 .map_queues = scsi_map_queues,
1955 .init_hctx = scsi_init_hctx,
1956 .poll = scsi_mq_poll,
1957 .set_rq_budget_token = scsi_mq_set_rq_budget_token,
1958 .get_rq_budget_token = scsi_mq_get_rq_budget_token,
1959};
1960
1961int scsi_mq_setup_tags(struct Scsi_Host *shost)
1962{
1963 unsigned int cmd_size, sgl_size;
1964 struct blk_mq_tag_set *tag_set = &shost->tag_set;
1965
1966 sgl_size = max_t(unsigned int, sizeof(struct scatterlist),
1967 scsi_mq_inline_sgl_size(shost));
1968 cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
1969 if (scsi_host_get_prot(shost))
1970 cmd_size += sizeof(struct scsi_data_buffer) +
1971 sizeof(struct scatterlist) * SCSI_INLINE_PROT_SG_CNT;
1972
1973 memset(tag_set, 0, sizeof(*tag_set));
1974 if (shost->hostt->commit_rqs)
1975 tag_set->ops = &scsi_mq_ops;
1976 else
1977 tag_set->ops = &scsi_mq_ops_no_commit;
1978 tag_set->nr_hw_queues = shost->nr_hw_queues ? : 1;
1979 tag_set->nr_maps = shost->nr_maps ? : 1;
1980 tag_set->queue_depth = shost->can_queue;
1981 tag_set->cmd_size = cmd_size;
1982 tag_set->numa_node = dev_to_node(shost->dma_dev);
1983 tag_set->flags = BLK_MQ_F_SHOULD_MERGE;
1984 tag_set->flags |=
1985 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
1986 tag_set->driver_data = shost;
1987 if (shost->host_tagset)
1988 tag_set->flags |= BLK_MQ_F_TAG_HCTX_SHARED;
1989
1990 return blk_mq_alloc_tag_set(tag_set);
1991}
1992
1993void scsi_mq_free_tags(struct kref *kref)
1994{
1995 struct Scsi_Host *shost = container_of(kref, typeof(*shost),
1996 tagset_refcnt);
1997
1998 blk_mq_free_tag_set(&shost->tag_set);
1999 complete(&shost->tagset_freed);
2000}
2001
2002/**
2003 * scsi_device_from_queue - return sdev associated with a request_queue
2004 * @q: The request queue to return the sdev from
2005 *
2006 * Return the sdev associated with a request queue or NULL if the
2007 * request_queue does not reference a SCSI device.
2008 */
2009struct scsi_device *scsi_device_from_queue(struct request_queue *q)
2010{
2011 struct scsi_device *sdev = NULL;
2012
2013 if (q->mq_ops == &scsi_mq_ops_no_commit ||
2014 q->mq_ops == &scsi_mq_ops)
2015 sdev = q->queuedata;
2016 if (!sdev || !get_device(&sdev->sdev_gendev))
2017 sdev = NULL;
2018
2019 return sdev;
2020}
2021/*
2022 * pktcdvd should have been integrated into the SCSI layers, but for historical
2023 * reasons like the old IDE driver it isn't. This export allows it to safely
2024 * probe if a given device is a SCSI one and only attach to that.
2025 */
2026#ifdef CONFIG_CDROM_PKTCDVD_MODULE
2027EXPORT_SYMBOL_GPL(scsi_device_from_queue);
2028#endif
2029
2030/**
2031 * scsi_block_requests - Utility function used by low-level drivers to prevent
2032 * further commands from being queued to the device.
2033 * @shost: host in question
2034 *
2035 * There is no timer nor any other means by which the requests get unblocked
2036 * other than the low-level driver calling scsi_unblock_requests().
2037 */
2038void scsi_block_requests(struct Scsi_Host *shost)
2039{
2040 shost->host_self_blocked = 1;
2041}
2042EXPORT_SYMBOL(scsi_block_requests);
2043
2044/**
2045 * scsi_unblock_requests - Utility function used by low-level drivers to allow
2046 * further commands to be queued to the device.
2047 * @shost: host in question
2048 *
2049 * There is no timer nor any other means by which the requests get unblocked
2050 * other than the low-level driver calling scsi_unblock_requests(). This is done
2051 * as an API function so that changes to the internals of the scsi mid-layer
2052 * won't require wholesale changes to drivers that use this feature.
2053 */
2054void scsi_unblock_requests(struct Scsi_Host *shost)
2055{
2056 shost->host_self_blocked = 0;
2057 scsi_run_host_queues(shost);
2058}
2059EXPORT_SYMBOL(scsi_unblock_requests);
2060
2061void scsi_exit_queue(void)
2062{
2063 kmem_cache_destroy(scsi_sense_cache);
2064}
2065
2066/**
2067 * scsi_mode_select - issue a mode select
2068 * @sdev: SCSI device to be queried
2069 * @pf: Page format bit (1 == standard, 0 == vendor specific)
2070 * @sp: Save page bit (0 == don't save, 1 == save)
2071 * @buffer: request buffer (may not be smaller than eight bytes)
2072 * @len: length of request buffer.
2073 * @timeout: command timeout
2074 * @retries: number of retries before failing
2075 * @data: returns a structure abstracting the mode header data
2076 * @sshdr: place to put sense data (or NULL if no sense to be collected).
2077 * must be SCSI_SENSE_BUFFERSIZE big.
2078 *
2079 * Returns zero if successful; negative error number or scsi
2080 * status on error
2081 *
2082 */
2083int scsi_mode_select(struct scsi_device *sdev, int pf, int sp,
2084 unsigned char *buffer, int len, int timeout, int retries,
2085 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2086{
2087 unsigned char cmd[10];
2088 unsigned char *real_buffer;
2089 int ret;
2090
2091 memset(cmd, 0, sizeof(cmd));
2092 cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2093
2094 /*
2095 * Use MODE SELECT(10) if the device asked for it or if the mode page
2096 * and the mode select header cannot fit within the maximumm 255 bytes
2097 * of the MODE SELECT(6) command.
2098 */
2099 if (sdev->use_10_for_ms ||
2100 len + 4 > 255 ||
2101 data->block_descriptor_length > 255) {
2102 if (len > 65535 - 8)
2103 return -EINVAL;
2104 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2105 if (!real_buffer)
2106 return -ENOMEM;
2107 memcpy(real_buffer + 8, buffer, len);
2108 len += 8;
2109 real_buffer[0] = 0;
2110 real_buffer[1] = 0;
2111 real_buffer[2] = data->medium_type;
2112 real_buffer[3] = data->device_specific;
2113 real_buffer[4] = data->longlba ? 0x01 : 0;
2114 real_buffer[5] = 0;
2115 put_unaligned_be16(data->block_descriptor_length,
2116 &real_buffer[6]);
2117
2118 cmd[0] = MODE_SELECT_10;
2119 put_unaligned_be16(len, &cmd[7]);
2120 } else {
2121 if (data->longlba)
2122 return -EINVAL;
2123
2124 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2125 if (!real_buffer)
2126 return -ENOMEM;
2127 memcpy(real_buffer + 4, buffer, len);
2128 len += 4;
2129 real_buffer[0] = 0;
2130 real_buffer[1] = data->medium_type;
2131 real_buffer[2] = data->device_specific;
2132 real_buffer[3] = data->block_descriptor_length;
2133
2134 cmd[0] = MODE_SELECT;
2135 cmd[4] = len;
2136 }
2137
2138 ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2139 sshdr, timeout, retries, NULL);
2140 kfree(real_buffer);
2141 return ret;
2142}
2143EXPORT_SYMBOL_GPL(scsi_mode_select);
2144
2145/**
2146 * scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2147 * @sdev: SCSI device to be queried
2148 * @dbd: set to prevent mode sense from returning block descriptors
2149 * @modepage: mode page being requested
2150 * @buffer: request buffer (may not be smaller than eight bytes)
2151 * @len: length of request buffer.
2152 * @timeout: command timeout
2153 * @retries: number of retries before failing
2154 * @data: returns a structure abstracting the mode header data
2155 * @sshdr: place to put sense data (or NULL if no sense to be collected).
2156 * must be SCSI_SENSE_BUFFERSIZE big.
2157 *
2158 * Returns zero if successful, or a negative error number on failure
2159 */
2160int
2161scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2162 unsigned char *buffer, int len, int timeout, int retries,
2163 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2164{
2165 unsigned char cmd[12];
2166 int use_10_for_ms;
2167 int header_length;
2168 int result, retry_count = retries;
2169 struct scsi_sense_hdr my_sshdr;
2170
2171 memset(data, 0, sizeof(*data));
2172 memset(&cmd[0], 0, 12);
2173
2174 dbd = sdev->set_dbd_for_ms ? 8 : dbd;
2175 cmd[1] = dbd & 0x18; /* allows DBD and LLBA bits */
2176 cmd[2] = modepage;
2177
2178 /* caller might not be interested in sense, but we need it */
2179 if (!sshdr)
2180 sshdr = &my_sshdr;
2181
2182 retry:
2183 use_10_for_ms = sdev->use_10_for_ms || len > 255;
2184
2185 if (use_10_for_ms) {
2186 if (len < 8 || len > 65535)
2187 return -EINVAL;
2188
2189 cmd[0] = MODE_SENSE_10;
2190 put_unaligned_be16(len, &cmd[7]);
2191 header_length = 8;
2192 } else {
2193 if (len < 4)
2194 return -EINVAL;
2195
2196 cmd[0] = MODE_SENSE;
2197 cmd[4] = len;
2198 header_length = 4;
2199 }
2200
2201 memset(buffer, 0, len);
2202
2203 result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2204 sshdr, timeout, retries, NULL);
2205 if (result < 0)
2206 return result;
2207
2208 /* This code looks awful: what it's doing is making sure an
2209 * ILLEGAL REQUEST sense return identifies the actual command
2210 * byte as the problem. MODE_SENSE commands can return
2211 * ILLEGAL REQUEST if the code page isn't supported */
2212
2213 if (!scsi_status_is_good(result)) {
2214 if (scsi_sense_valid(sshdr)) {
2215 if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2216 (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2217 /*
2218 * Invalid command operation code: retry using
2219 * MODE SENSE(6) if this was a MODE SENSE(10)
2220 * request, except if the request mode page is
2221 * too large for MODE SENSE single byte
2222 * allocation length field.
2223 */
2224 if (use_10_for_ms) {
2225 if (len > 255)
2226 return -EIO;
2227 sdev->use_10_for_ms = 0;
2228 goto retry;
2229 }
2230 }
2231 if (scsi_status_is_check_condition(result) &&
2232 sshdr->sense_key == UNIT_ATTENTION &&
2233 retry_count) {
2234 retry_count--;
2235 goto retry;
2236 }
2237 }
2238 return -EIO;
2239 }
2240 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2241 (modepage == 6 || modepage == 8))) {
2242 /* Initio breakage? */
2243 header_length = 0;
2244 data->length = 13;
2245 data->medium_type = 0;
2246 data->device_specific = 0;
2247 data->longlba = 0;
2248 data->block_descriptor_length = 0;
2249 } else if (use_10_for_ms) {
2250 data->length = get_unaligned_be16(&buffer[0]) + 2;
2251 data->medium_type = buffer[2];
2252 data->device_specific = buffer[3];
2253 data->longlba = buffer[4] & 0x01;
2254 data->block_descriptor_length = get_unaligned_be16(&buffer[6]);
2255 } else {
2256 data->length = buffer[0] + 1;
2257 data->medium_type = buffer[1];
2258 data->device_specific = buffer[2];
2259 data->block_descriptor_length = buffer[3];
2260 }
2261 data->header_length = header_length;
2262
2263 return 0;
2264}
2265EXPORT_SYMBOL(scsi_mode_sense);
2266
2267/**
2268 * scsi_test_unit_ready - test if unit is ready
2269 * @sdev: scsi device to change the state of.
2270 * @timeout: command timeout
2271 * @retries: number of retries before failing
2272 * @sshdr: outpout pointer for decoded sense information.
2273 *
2274 * Returns zero if unsuccessful or an error if TUR failed. For
2275 * removable media, UNIT_ATTENTION sets ->changed flag.
2276 **/
2277int
2278scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2279 struct scsi_sense_hdr *sshdr)
2280{
2281 char cmd[] = {
2282 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2283 };
2284 int result;
2285
2286 /* try to eat the UNIT_ATTENTION if there are enough retries */
2287 do {
2288 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2289 timeout, 1, NULL);
2290 if (sdev->removable && scsi_sense_valid(sshdr) &&
2291 sshdr->sense_key == UNIT_ATTENTION)
2292 sdev->changed = 1;
2293 } while (scsi_sense_valid(sshdr) &&
2294 sshdr->sense_key == UNIT_ATTENTION && --retries);
2295
2296 return result;
2297}
2298EXPORT_SYMBOL(scsi_test_unit_ready);
2299
2300/**
2301 * scsi_device_set_state - Take the given device through the device state model.
2302 * @sdev: scsi device to change the state of.
2303 * @state: state to change to.
2304 *
2305 * Returns zero if successful or an error if the requested
2306 * transition is illegal.
2307 */
2308int
2309scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2310{
2311 enum scsi_device_state oldstate = sdev->sdev_state;
2312
2313 if (state == oldstate)
2314 return 0;
2315
2316 switch (state) {
2317 case SDEV_CREATED:
2318 switch (oldstate) {
2319 case SDEV_CREATED_BLOCK:
2320 break;
2321 default:
2322 goto illegal;
2323 }
2324 break;
2325
2326 case SDEV_RUNNING:
2327 switch (oldstate) {
2328 case SDEV_CREATED:
2329 case SDEV_OFFLINE:
2330 case SDEV_TRANSPORT_OFFLINE:
2331 case SDEV_QUIESCE:
2332 case SDEV_BLOCK:
2333 break;
2334 default:
2335 goto illegal;
2336 }
2337 break;
2338
2339 case SDEV_QUIESCE:
2340 switch (oldstate) {
2341 case SDEV_RUNNING:
2342 case SDEV_OFFLINE:
2343 case SDEV_TRANSPORT_OFFLINE:
2344 break;
2345 default:
2346 goto illegal;
2347 }
2348 break;
2349
2350 case SDEV_OFFLINE:
2351 case SDEV_TRANSPORT_OFFLINE:
2352 switch (oldstate) {
2353 case SDEV_CREATED:
2354 case SDEV_RUNNING:
2355 case SDEV_QUIESCE:
2356 case SDEV_BLOCK:
2357 break;
2358 default:
2359 goto illegal;
2360 }
2361 break;
2362
2363 case SDEV_BLOCK:
2364 switch (oldstate) {
2365 case SDEV_RUNNING:
2366 case SDEV_CREATED_BLOCK:
2367 case SDEV_QUIESCE:
2368 case SDEV_OFFLINE:
2369 break;
2370 default:
2371 goto illegal;
2372 }
2373 break;
2374
2375 case SDEV_CREATED_BLOCK:
2376 switch (oldstate) {
2377 case SDEV_CREATED:
2378 break;
2379 default:
2380 goto illegal;
2381 }
2382 break;
2383
2384 case SDEV_CANCEL:
2385 switch (oldstate) {
2386 case SDEV_CREATED:
2387 case SDEV_RUNNING:
2388 case SDEV_QUIESCE:
2389 case SDEV_OFFLINE:
2390 case SDEV_TRANSPORT_OFFLINE:
2391 break;
2392 default:
2393 goto illegal;
2394 }
2395 break;
2396
2397 case SDEV_DEL:
2398 switch (oldstate) {
2399 case SDEV_CREATED:
2400 case SDEV_RUNNING:
2401 case SDEV_OFFLINE:
2402 case SDEV_TRANSPORT_OFFLINE:
2403 case SDEV_CANCEL:
2404 case SDEV_BLOCK:
2405 case SDEV_CREATED_BLOCK:
2406 break;
2407 default:
2408 goto illegal;
2409 }
2410 break;
2411
2412 }
2413 sdev->offline_already = false;
2414 sdev->sdev_state = state;
2415 return 0;
2416
2417 illegal:
2418 SCSI_LOG_ERROR_RECOVERY(1,
2419 sdev_printk(KERN_ERR, sdev,
2420 "Illegal state transition %s->%s",
2421 scsi_device_state_name(oldstate),
2422 scsi_device_state_name(state))
2423 );
2424 return -EINVAL;
2425}
2426EXPORT_SYMBOL(scsi_device_set_state);
2427
2428/**
2429 * scsi_evt_emit - emit a single SCSI device uevent
2430 * @sdev: associated SCSI device
2431 * @evt: event to emit
2432 *
2433 * Send a single uevent (scsi_event) to the associated scsi_device.
2434 */
2435static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2436{
2437 int idx = 0;
2438 char *envp[3];
2439
2440 switch (evt->evt_type) {
2441 case SDEV_EVT_MEDIA_CHANGE:
2442 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2443 break;
2444 case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2445 scsi_rescan_device(&sdev->sdev_gendev);
2446 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2447 break;
2448 case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2449 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2450 break;
2451 case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2452 envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2453 break;
2454 case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2455 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2456 break;
2457 case SDEV_EVT_LUN_CHANGE_REPORTED:
2458 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2459 break;
2460 case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2461 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2462 break;
2463 case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2464 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2465 break;
2466 default:
2467 /* do nothing */
2468 break;
2469 }
2470
2471 envp[idx++] = NULL;
2472
2473 kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2474}
2475
2476/**
2477 * scsi_evt_thread - send a uevent for each scsi event
2478 * @work: work struct for scsi_device
2479 *
2480 * Dispatch queued events to their associated scsi_device kobjects
2481 * as uevents.
2482 */
2483void scsi_evt_thread(struct work_struct *work)
2484{
2485 struct scsi_device *sdev;
2486 enum scsi_device_event evt_type;
2487 LIST_HEAD(event_list);
2488
2489 sdev = container_of(work, struct scsi_device, event_work);
2490
2491 for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2492 if (test_and_clear_bit(evt_type, sdev->pending_events))
2493 sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2494
2495 while (1) {
2496 struct scsi_event *evt;
2497 struct list_head *this, *tmp;
2498 unsigned long flags;
2499
2500 spin_lock_irqsave(&sdev->list_lock, flags);
2501 list_splice_init(&sdev->event_list, &event_list);
2502 spin_unlock_irqrestore(&sdev->list_lock, flags);
2503
2504 if (list_empty(&event_list))
2505 break;
2506
2507 list_for_each_safe(this, tmp, &event_list) {
2508 evt = list_entry(this, struct scsi_event, node);
2509 list_del(&evt->node);
2510 scsi_evt_emit(sdev, evt);
2511 kfree(evt);
2512 }
2513 }
2514}
2515
2516/**
2517 * sdev_evt_send - send asserted event to uevent thread
2518 * @sdev: scsi_device event occurred on
2519 * @evt: event to send
2520 *
2521 * Assert scsi device event asynchronously.
2522 */
2523void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2524{
2525 unsigned long flags;
2526
2527#if 0
2528 /* FIXME: currently this check eliminates all media change events
2529 * for polled devices. Need to update to discriminate between AN
2530 * and polled events */
2531 if (!test_bit(evt->evt_type, sdev->supported_events)) {
2532 kfree(evt);
2533 return;
2534 }
2535#endif
2536
2537 spin_lock_irqsave(&sdev->list_lock, flags);
2538 list_add_tail(&evt->node, &sdev->event_list);
2539 schedule_work(&sdev->event_work);
2540 spin_unlock_irqrestore(&sdev->list_lock, flags);
2541}
2542EXPORT_SYMBOL_GPL(sdev_evt_send);
2543
2544/**
2545 * sdev_evt_alloc - allocate a new scsi event
2546 * @evt_type: type of event to allocate
2547 * @gfpflags: GFP flags for allocation
2548 *
2549 * Allocates and returns a new scsi_event.
2550 */
2551struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2552 gfp_t gfpflags)
2553{
2554 struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2555 if (!evt)
2556 return NULL;
2557
2558 evt->evt_type = evt_type;
2559 INIT_LIST_HEAD(&evt->node);
2560
2561 /* evt_type-specific initialization, if any */
2562 switch (evt_type) {
2563 case SDEV_EVT_MEDIA_CHANGE:
2564 case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2565 case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2566 case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2567 case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2568 case SDEV_EVT_LUN_CHANGE_REPORTED:
2569 case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2570 case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2571 default:
2572 /* do nothing */
2573 break;
2574 }
2575
2576 return evt;
2577}
2578EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2579
2580/**
2581 * sdev_evt_send_simple - send asserted event to uevent thread
2582 * @sdev: scsi_device event occurred on
2583 * @evt_type: type of event to send
2584 * @gfpflags: GFP flags for allocation
2585 *
2586 * Assert scsi device event asynchronously, given an event type.
2587 */
2588void sdev_evt_send_simple(struct scsi_device *sdev,
2589 enum scsi_device_event evt_type, gfp_t gfpflags)
2590{
2591 struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2592 if (!evt) {
2593 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2594 evt_type);
2595 return;
2596 }
2597
2598 sdev_evt_send(sdev, evt);
2599}
2600EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2601
2602/**
2603 * scsi_device_quiesce - Block all commands except power management.
2604 * @sdev: scsi device to quiesce.
2605 *
2606 * This works by trying to transition to the SDEV_QUIESCE state
2607 * (which must be a legal transition). When the device is in this
2608 * state, only power management requests will be accepted, all others will
2609 * be deferred.
2610 *
2611 * Must be called with user context, may sleep.
2612 *
2613 * Returns zero if unsuccessful or an error if not.
2614 */
2615int
2616scsi_device_quiesce(struct scsi_device *sdev)
2617{
2618 struct request_queue *q = sdev->request_queue;
2619 int err;
2620
2621 /*
2622 * It is allowed to call scsi_device_quiesce() multiple times from
2623 * the same context but concurrent scsi_device_quiesce() calls are
2624 * not allowed.
2625 */
2626 WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2627
2628 if (sdev->quiesced_by == current)
2629 return 0;
2630
2631 blk_set_pm_only(q);
2632
2633 blk_mq_freeze_queue(q);
2634 /*
2635 * Ensure that the effect of blk_set_pm_only() will be visible
2636 * for percpu_ref_tryget() callers that occur after the queue
2637 * unfreeze even if the queue was already frozen before this function
2638 * was called. See also https://lwn.net/Articles/573497/.
2639 */
2640 synchronize_rcu();
2641 blk_mq_unfreeze_queue(q);
2642
2643 mutex_lock(&sdev->state_mutex);
2644 err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2645 if (err == 0)
2646 sdev->quiesced_by = current;
2647 else
2648 blk_clear_pm_only(q);
2649 mutex_unlock(&sdev->state_mutex);
2650
2651 return err;
2652}
2653EXPORT_SYMBOL(scsi_device_quiesce);
2654
2655/**
2656 * scsi_device_resume - Restart user issued commands to a quiesced device.
2657 * @sdev: scsi device to resume.
2658 *
2659 * Moves the device from quiesced back to running and restarts the
2660 * queues.
2661 *
2662 * Must be called with user context, may sleep.
2663 */
2664void scsi_device_resume(struct scsi_device *sdev)
2665{
2666 /* check if the device state was mutated prior to resume, and if
2667 * so assume the state is being managed elsewhere (for example
2668 * device deleted during suspend)
2669 */
2670 mutex_lock(&sdev->state_mutex);
2671 if (sdev->sdev_state == SDEV_QUIESCE)
2672 scsi_device_set_state(sdev, SDEV_RUNNING);
2673 if (sdev->quiesced_by) {
2674 sdev->quiesced_by = NULL;
2675 blk_clear_pm_only(sdev->request_queue);
2676 }
2677 mutex_unlock(&sdev->state_mutex);
2678}
2679EXPORT_SYMBOL(scsi_device_resume);
2680
2681static void
2682device_quiesce_fn(struct scsi_device *sdev, void *data)
2683{
2684 scsi_device_quiesce(sdev);
2685}
2686
2687void
2688scsi_target_quiesce(struct scsi_target *starget)
2689{
2690 starget_for_each_device(starget, NULL, device_quiesce_fn);
2691}
2692EXPORT_SYMBOL(scsi_target_quiesce);
2693
2694static void
2695device_resume_fn(struct scsi_device *sdev, void *data)
2696{
2697 scsi_device_resume(sdev);
2698}
2699
2700void
2701scsi_target_resume(struct scsi_target *starget)
2702{
2703 starget_for_each_device(starget, NULL, device_resume_fn);
2704}
2705EXPORT_SYMBOL(scsi_target_resume);
2706
2707static int __scsi_internal_device_block_nowait(struct scsi_device *sdev)
2708{
2709 if (scsi_device_set_state(sdev, SDEV_BLOCK))
2710 return scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2711
2712 return 0;
2713}
2714
2715void scsi_start_queue(struct scsi_device *sdev)
2716{
2717 if (cmpxchg(&sdev->queue_stopped, 1, 0))
2718 blk_mq_unquiesce_queue(sdev->request_queue);
2719}
2720
2721static void scsi_stop_queue(struct scsi_device *sdev, bool nowait)
2722{
2723 /*
2724 * The atomic variable of ->queue_stopped covers that
2725 * blk_mq_quiesce_queue* is balanced with blk_mq_unquiesce_queue.
2726 *
2727 * However, we still need to wait until quiesce is done
2728 * in case that queue has been stopped.
2729 */
2730 if (!cmpxchg(&sdev->queue_stopped, 0, 1)) {
2731 if (nowait)
2732 blk_mq_quiesce_queue_nowait(sdev->request_queue);
2733 else
2734 blk_mq_quiesce_queue(sdev->request_queue);
2735 } else {
2736 if (!nowait)
2737 blk_mq_wait_quiesce_done(sdev->request_queue->tag_set);
2738 }
2739}
2740
2741/**
2742 * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2743 * @sdev: device to block
2744 *
2745 * Pause SCSI command processing on the specified device. Does not sleep.
2746 *
2747 * Returns zero if successful or a negative error code upon failure.
2748 *
2749 * Notes:
2750 * This routine transitions the device to the SDEV_BLOCK state (which must be
2751 * a legal transition). When the device is in this state, command processing
2752 * is paused until the device leaves the SDEV_BLOCK state. See also
2753 * scsi_internal_device_unblock_nowait().
2754 */
2755int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2756{
2757 int ret = __scsi_internal_device_block_nowait(sdev);
2758
2759 /*
2760 * The device has transitioned to SDEV_BLOCK. Stop the
2761 * block layer from calling the midlayer with this device's
2762 * request queue.
2763 */
2764 if (!ret)
2765 scsi_stop_queue(sdev, true);
2766 return ret;
2767}
2768EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2769
2770/**
2771 * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
2772 * @sdev: device to block
2773 *
2774 * Pause SCSI command processing on the specified device and wait until all
2775 * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
2776 *
2777 * Returns zero if successful or a negative error code upon failure.
2778 *
2779 * Note:
2780 * This routine transitions the device to the SDEV_BLOCK state (which must be
2781 * a legal transition). When the device is in this state, command processing
2782 * is paused until the device leaves the SDEV_BLOCK state. See also
2783 * scsi_internal_device_unblock().
2784 */
2785static int scsi_internal_device_block(struct scsi_device *sdev)
2786{
2787 int err;
2788
2789 mutex_lock(&sdev->state_mutex);
2790 err = __scsi_internal_device_block_nowait(sdev);
2791 if (err == 0)
2792 scsi_stop_queue(sdev, false);
2793 mutex_unlock(&sdev->state_mutex);
2794
2795 return err;
2796}
2797
2798/**
2799 * scsi_internal_device_unblock_nowait - resume a device after a block request
2800 * @sdev: device to resume
2801 * @new_state: state to set the device to after unblocking
2802 *
2803 * Restart the device queue for a previously suspended SCSI device. Does not
2804 * sleep.
2805 *
2806 * Returns zero if successful or a negative error code upon failure.
2807 *
2808 * Notes:
2809 * This routine transitions the device to the SDEV_RUNNING state or to one of
2810 * the offline states (which must be a legal transition) allowing the midlayer
2811 * to goose the queue for this device.
2812 */
2813int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
2814 enum scsi_device_state new_state)
2815{
2816 switch (new_state) {
2817 case SDEV_RUNNING:
2818 case SDEV_TRANSPORT_OFFLINE:
2819 break;
2820 default:
2821 return -EINVAL;
2822 }
2823
2824 /*
2825 * Try to transition the scsi device to SDEV_RUNNING or one of the
2826 * offlined states and goose the device queue if successful.
2827 */
2828 switch (sdev->sdev_state) {
2829 case SDEV_BLOCK:
2830 case SDEV_TRANSPORT_OFFLINE:
2831 sdev->sdev_state = new_state;
2832 break;
2833 case SDEV_CREATED_BLOCK:
2834 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2835 new_state == SDEV_OFFLINE)
2836 sdev->sdev_state = new_state;
2837 else
2838 sdev->sdev_state = SDEV_CREATED;
2839 break;
2840 case SDEV_CANCEL:
2841 case SDEV_OFFLINE:
2842 break;
2843 default:
2844 return -EINVAL;
2845 }
2846 scsi_start_queue(sdev);
2847
2848 return 0;
2849}
2850EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
2851
2852/**
2853 * scsi_internal_device_unblock - resume a device after a block request
2854 * @sdev: device to resume
2855 * @new_state: state to set the device to after unblocking
2856 *
2857 * Restart the device queue for a previously suspended SCSI device. May sleep.
2858 *
2859 * Returns zero if successful or a negative error code upon failure.
2860 *
2861 * Notes:
2862 * This routine transitions the device to the SDEV_RUNNING state or to one of
2863 * the offline states (which must be a legal transition) allowing the midlayer
2864 * to goose the queue for this device.
2865 */
2866static int scsi_internal_device_unblock(struct scsi_device *sdev,
2867 enum scsi_device_state new_state)
2868{
2869 int ret;
2870
2871 mutex_lock(&sdev->state_mutex);
2872 ret = scsi_internal_device_unblock_nowait(sdev, new_state);
2873 mutex_unlock(&sdev->state_mutex);
2874
2875 return ret;
2876}
2877
2878static void
2879device_block(struct scsi_device *sdev, void *data)
2880{
2881 int ret;
2882
2883 ret = scsi_internal_device_block(sdev);
2884
2885 WARN_ONCE(ret, "scsi_internal_device_block(%s) failed: ret = %d\n",
2886 dev_name(&sdev->sdev_gendev), ret);
2887}
2888
2889static int
2890target_block(struct device *dev, void *data)
2891{
2892 if (scsi_is_target_device(dev))
2893 starget_for_each_device(to_scsi_target(dev), NULL,
2894 device_block);
2895 return 0;
2896}
2897
2898void
2899scsi_target_block(struct device *dev)
2900{
2901 if (scsi_is_target_device(dev))
2902 starget_for_each_device(to_scsi_target(dev), NULL,
2903 device_block);
2904 else
2905 device_for_each_child(dev, NULL, target_block);
2906}
2907EXPORT_SYMBOL_GPL(scsi_target_block);
2908
2909static void
2910device_unblock(struct scsi_device *sdev, void *data)
2911{
2912 scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2913}
2914
2915static int
2916target_unblock(struct device *dev, void *data)
2917{
2918 if (scsi_is_target_device(dev))
2919 starget_for_each_device(to_scsi_target(dev), data,
2920 device_unblock);
2921 return 0;
2922}
2923
2924void
2925scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2926{
2927 if (scsi_is_target_device(dev))
2928 starget_for_each_device(to_scsi_target(dev), &new_state,
2929 device_unblock);
2930 else
2931 device_for_each_child(dev, &new_state, target_unblock);
2932}
2933EXPORT_SYMBOL_GPL(scsi_target_unblock);
2934
2935int
2936scsi_host_block(struct Scsi_Host *shost)
2937{
2938 struct scsi_device *sdev;
2939 int ret = 0;
2940
2941 /*
2942 * Call scsi_internal_device_block_nowait so we can avoid
2943 * calling synchronize_rcu() for each LUN.
2944 */
2945 shost_for_each_device(sdev, shost) {
2946 mutex_lock(&sdev->state_mutex);
2947 ret = scsi_internal_device_block_nowait(sdev);
2948 mutex_unlock(&sdev->state_mutex);
2949 if (ret) {
2950 scsi_device_put(sdev);
2951 break;
2952 }
2953 }
2954
2955 /*
2956 * SCSI never enables blk-mq's BLK_MQ_F_BLOCKING flag so
2957 * calling synchronize_rcu() once is enough.
2958 */
2959 WARN_ON_ONCE(shost->tag_set.flags & BLK_MQ_F_BLOCKING);
2960
2961 if (!ret)
2962 synchronize_rcu();
2963
2964 return ret;
2965}
2966EXPORT_SYMBOL_GPL(scsi_host_block);
2967
2968int
2969scsi_host_unblock(struct Scsi_Host *shost, int new_state)
2970{
2971 struct scsi_device *sdev;
2972 int ret = 0;
2973
2974 shost_for_each_device(sdev, shost) {
2975 ret = scsi_internal_device_unblock(sdev, new_state);
2976 if (ret) {
2977 scsi_device_put(sdev);
2978 break;
2979 }
2980 }
2981 return ret;
2982}
2983EXPORT_SYMBOL_GPL(scsi_host_unblock);
2984
2985/**
2986 * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2987 * @sgl: scatter-gather list
2988 * @sg_count: number of segments in sg
2989 * @offset: offset in bytes into sg, on return offset into the mapped area
2990 * @len: bytes to map, on return number of bytes mapped
2991 *
2992 * Returns virtual address of the start of the mapped page
2993 */
2994void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2995 size_t *offset, size_t *len)
2996{
2997 int i;
2998 size_t sg_len = 0, len_complete = 0;
2999 struct scatterlist *sg;
3000 struct page *page;
3001
3002 WARN_ON(!irqs_disabled());
3003
3004 for_each_sg(sgl, sg, sg_count, i) {
3005 len_complete = sg_len; /* Complete sg-entries */
3006 sg_len += sg->length;
3007 if (sg_len > *offset)
3008 break;
3009 }
3010
3011 if (unlikely(i == sg_count)) {
3012 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
3013 "elements %d\n",
3014 __func__, sg_len, *offset, sg_count);
3015 WARN_ON(1);
3016 return NULL;
3017 }
3018
3019 /* Offset starting from the beginning of first page in this sg-entry */
3020 *offset = *offset - len_complete + sg->offset;
3021
3022 /* Assumption: contiguous pages can be accessed as "page + i" */
3023 page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
3024 *offset &= ~PAGE_MASK;
3025
3026 /* Bytes in this sg-entry from *offset to the end of the page */
3027 sg_len = PAGE_SIZE - *offset;
3028 if (*len > sg_len)
3029 *len = sg_len;
3030
3031 return kmap_atomic(page);
3032}
3033EXPORT_SYMBOL(scsi_kmap_atomic_sg);
3034
3035/**
3036 * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
3037 * @virt: virtual address to be unmapped
3038 */
3039void scsi_kunmap_atomic_sg(void *virt)
3040{
3041 kunmap_atomic(virt);
3042}
3043EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
3044
3045void sdev_disable_disk_events(struct scsi_device *sdev)
3046{
3047 atomic_inc(&sdev->disk_events_disable_depth);
3048}
3049EXPORT_SYMBOL(sdev_disable_disk_events);
3050
3051void sdev_enable_disk_events(struct scsi_device *sdev)
3052{
3053 if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
3054 return;
3055 atomic_dec(&sdev->disk_events_disable_depth);
3056}
3057EXPORT_SYMBOL(sdev_enable_disk_events);
3058
3059static unsigned char designator_prio(const unsigned char *d)
3060{
3061 if (d[1] & 0x30)
3062 /* not associated with LUN */
3063 return 0;
3064
3065 if (d[3] == 0)
3066 /* invalid length */
3067 return 0;
3068
3069 /*
3070 * Order of preference for lun descriptor:
3071 * - SCSI name string
3072 * - NAA IEEE Registered Extended
3073 * - EUI-64 based 16-byte
3074 * - EUI-64 based 12-byte
3075 * - NAA IEEE Registered
3076 * - NAA IEEE Extended
3077 * - EUI-64 based 8-byte
3078 * - SCSI name string (truncated)
3079 * - T10 Vendor ID
3080 * as longer descriptors reduce the likelyhood
3081 * of identification clashes.
3082 */
3083
3084 switch (d[1] & 0xf) {
3085 case 8:
3086 /* SCSI name string, variable-length UTF-8 */
3087 return 9;
3088 case 3:
3089 switch (d[4] >> 4) {
3090 case 6:
3091 /* NAA registered extended */
3092 return 8;
3093 case 5:
3094 /* NAA registered */
3095 return 5;
3096 case 4:
3097 /* NAA extended */
3098 return 4;
3099 case 3:
3100 /* NAA locally assigned */
3101 return 1;
3102 default:
3103 break;
3104 }
3105 break;
3106 case 2:
3107 switch (d[3]) {
3108 case 16:
3109 /* EUI64-based, 16 byte */
3110 return 7;
3111 case 12:
3112 /* EUI64-based, 12 byte */
3113 return 6;
3114 case 8:
3115 /* EUI64-based, 8 byte */
3116 return 3;
3117 default:
3118 break;
3119 }
3120 break;
3121 case 1:
3122 /* T10 vendor ID */
3123 return 1;
3124 default:
3125 break;
3126 }
3127
3128 return 0;
3129}
3130
3131/**
3132 * scsi_vpd_lun_id - return a unique device identification
3133 * @sdev: SCSI device
3134 * @id: buffer for the identification
3135 * @id_len: length of the buffer
3136 *
3137 * Copies a unique device identification into @id based
3138 * on the information in the VPD page 0x83 of the device.
3139 * The string will be formatted as a SCSI name string.
3140 *
3141 * Returns the length of the identification or error on failure.
3142 * If the identifier is longer than the supplied buffer the actual
3143 * identifier length is returned and the buffer is not zero-padded.
3144 */
3145int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
3146{
3147 u8 cur_id_prio = 0;
3148 u8 cur_id_size = 0;
3149 const unsigned char *d, *cur_id_str;
3150 const struct scsi_vpd *vpd_pg83;
3151 int id_size = -EINVAL;
3152
3153 rcu_read_lock();
3154 vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3155 if (!vpd_pg83) {
3156 rcu_read_unlock();
3157 return -ENXIO;
3158 }
3159
3160 /* The id string must be at least 20 bytes + terminating NULL byte */
3161 if (id_len < 21) {
3162 rcu_read_unlock();
3163 return -EINVAL;
3164 }
3165
3166 memset(id, 0, id_len);
3167 for (d = vpd_pg83->data + 4;
3168 d < vpd_pg83->data + vpd_pg83->len;
3169 d += d[3] + 4) {
3170 u8 prio = designator_prio(d);
3171
3172 if (prio == 0 || cur_id_prio > prio)
3173 continue;
3174
3175 switch (d[1] & 0xf) {
3176 case 0x1:
3177 /* T10 Vendor ID */
3178 if (cur_id_size > d[3])
3179 break;
3180 cur_id_prio = prio;
3181 cur_id_size = d[3];
3182 if (cur_id_size + 4 > id_len)
3183 cur_id_size = id_len - 4;
3184 cur_id_str = d + 4;
3185 id_size = snprintf(id, id_len, "t10.%*pE",
3186 cur_id_size, cur_id_str);
3187 break;
3188 case 0x2:
3189 /* EUI-64 */
3190 cur_id_prio = prio;
3191 cur_id_size = d[3];
3192 cur_id_str = d + 4;
3193 switch (cur_id_size) {
3194 case 8:
3195 id_size = snprintf(id, id_len,
3196 "eui.%8phN",
3197 cur_id_str);
3198 break;
3199 case 12:
3200 id_size = snprintf(id, id_len,
3201 "eui.%12phN",
3202 cur_id_str);
3203 break;
3204 case 16:
3205 id_size = snprintf(id, id_len,
3206 "eui.%16phN",
3207 cur_id_str);
3208 break;
3209 default:
3210 break;
3211 }
3212 break;
3213 case 0x3:
3214 /* NAA */
3215 cur_id_prio = prio;
3216 cur_id_size = d[3];
3217 cur_id_str = d + 4;
3218 switch (cur_id_size) {
3219 case 8:
3220 id_size = snprintf(id, id_len,
3221 "naa.%8phN",
3222 cur_id_str);
3223 break;
3224 case 16:
3225 id_size = snprintf(id, id_len,
3226 "naa.%16phN",
3227 cur_id_str);
3228 break;
3229 default:
3230 break;
3231 }
3232 break;
3233 case 0x8:
3234 /* SCSI name string */
3235 if (cur_id_size > d[3])
3236 break;
3237 /* Prefer others for truncated descriptor */
3238 if (d[3] > id_len) {
3239 prio = 2;
3240 if (cur_id_prio > prio)
3241 break;
3242 }
3243 cur_id_prio = prio;
3244 cur_id_size = id_size = d[3];
3245 cur_id_str = d + 4;
3246 if (cur_id_size >= id_len)
3247 cur_id_size = id_len - 1;
3248 memcpy(id, cur_id_str, cur_id_size);
3249 break;
3250 default:
3251 break;
3252 }
3253 }
3254 rcu_read_unlock();
3255
3256 return id_size;
3257}
3258EXPORT_SYMBOL(scsi_vpd_lun_id);
3259
3260/*
3261 * scsi_vpd_tpg_id - return a target port group identifier
3262 * @sdev: SCSI device
3263 *
3264 * Returns the Target Port Group identifier from the information
3265 * froom VPD page 0x83 of the device.
3266 *
3267 * Returns the identifier or error on failure.
3268 */
3269int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3270{
3271 const unsigned char *d;
3272 const struct scsi_vpd *vpd_pg83;
3273 int group_id = -EAGAIN, rel_port = -1;
3274
3275 rcu_read_lock();
3276 vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3277 if (!vpd_pg83) {
3278 rcu_read_unlock();
3279 return -ENXIO;
3280 }
3281
3282 d = vpd_pg83->data + 4;
3283 while (d < vpd_pg83->data + vpd_pg83->len) {
3284 switch (d[1] & 0xf) {
3285 case 0x4:
3286 /* Relative target port */
3287 rel_port = get_unaligned_be16(&d[6]);
3288 break;
3289 case 0x5:
3290 /* Target port group */
3291 group_id = get_unaligned_be16(&d[6]);
3292 break;
3293 default:
3294 break;
3295 }
3296 d += d[3] + 4;
3297 }
3298 rcu_read_unlock();
3299
3300 if (group_id >= 0 && rel_id && rel_port != -1)
3301 *rel_id = rel_port;
3302
3303 return group_id;
3304}
3305EXPORT_SYMBOL(scsi_vpd_tpg_id);
3306
3307/**
3308 * scsi_build_sense - build sense data for a command
3309 * @scmd: scsi command for which the sense should be formatted
3310 * @desc: Sense format (non-zero == descriptor format,
3311 * 0 == fixed format)
3312 * @key: Sense key
3313 * @asc: Additional sense code
3314 * @ascq: Additional sense code qualifier
3315 *
3316 **/
3317void scsi_build_sense(struct scsi_cmnd *scmd, int desc, u8 key, u8 asc, u8 ascq)
3318{
3319 scsi_build_sense_buffer(desc, scmd->sense_buffer, key, asc, ascq);
3320 scmd->result = SAM_STAT_CHECK_CONDITION;
3321}
3322EXPORT_SYMBOL_GPL(scsi_build_sense);