Loading...
1/*
2 * Copyright (C) 1999 Eric Youngdale
3 * Copyright (C) 2014 Christoph Hellwig
4 *
5 * SCSI queueing library.
6 * Initial versions: Eric Youngdale (eric@andante.org).
7 * Based upon conversations with large numbers
8 * of people at Linux Expo.
9 */
10
11#include <linux/bio.h>
12#include <linux/bitops.h>
13#include <linux/blkdev.h>
14#include <linux/completion.h>
15#include <linux/kernel.h>
16#include <linux/export.h>
17#include <linux/init.h>
18#include <linux/pci.h>
19#include <linux/delay.h>
20#include <linux/hardirq.h>
21#include <linux/scatterlist.h>
22#include <linux/blk-mq.h>
23#include <linux/ratelimit.h>
24#include <asm/unaligned.h>
25
26#include <scsi/scsi.h>
27#include <scsi/scsi_cmnd.h>
28#include <scsi/scsi_dbg.h>
29#include <scsi/scsi_device.h>
30#include <scsi/scsi_driver.h>
31#include <scsi/scsi_eh.h>
32#include <scsi/scsi_host.h>
33#include <scsi/scsi_transport.h> /* __scsi_init_queue() */
34#include <scsi/scsi_dh.h>
35
36#include <trace/events/scsi.h>
37
38#include "scsi_debugfs.h"
39#include "scsi_priv.h"
40#include "scsi_logging.h"
41
42static struct kmem_cache *scsi_sdb_cache;
43static struct kmem_cache *scsi_sense_cache;
44static struct kmem_cache *scsi_sense_isadma_cache;
45static DEFINE_MUTEX(scsi_sense_cache_mutex);
46
47static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd);
48
49static inline struct kmem_cache *
50scsi_select_sense_cache(bool unchecked_isa_dma)
51{
52 return unchecked_isa_dma ? scsi_sense_isadma_cache : scsi_sense_cache;
53}
54
55static void scsi_free_sense_buffer(bool unchecked_isa_dma,
56 unsigned char *sense_buffer)
57{
58 kmem_cache_free(scsi_select_sense_cache(unchecked_isa_dma),
59 sense_buffer);
60}
61
62static unsigned char *scsi_alloc_sense_buffer(bool unchecked_isa_dma,
63 gfp_t gfp_mask, int numa_node)
64{
65 return kmem_cache_alloc_node(scsi_select_sense_cache(unchecked_isa_dma),
66 gfp_mask, numa_node);
67}
68
69int scsi_init_sense_cache(struct Scsi_Host *shost)
70{
71 struct kmem_cache *cache;
72 int ret = 0;
73
74 cache = scsi_select_sense_cache(shost->unchecked_isa_dma);
75 if (cache)
76 return 0;
77
78 mutex_lock(&scsi_sense_cache_mutex);
79 if (shost->unchecked_isa_dma) {
80 scsi_sense_isadma_cache =
81 kmem_cache_create("scsi_sense_cache(DMA)",
82 SCSI_SENSE_BUFFERSIZE, 0,
83 SLAB_HWCACHE_ALIGN | SLAB_CACHE_DMA, NULL);
84 if (!scsi_sense_isadma_cache)
85 ret = -ENOMEM;
86 } else {
87 scsi_sense_cache =
88 kmem_cache_create_usercopy("scsi_sense_cache",
89 SCSI_SENSE_BUFFERSIZE, 0, SLAB_HWCACHE_ALIGN,
90 0, SCSI_SENSE_BUFFERSIZE, NULL);
91 if (!scsi_sense_cache)
92 ret = -ENOMEM;
93 }
94
95 mutex_unlock(&scsi_sense_cache_mutex);
96 return ret;
97}
98
99/*
100 * When to reinvoke queueing after a resource shortage. It's 3 msecs to
101 * not change behaviour from the previous unplug mechanism, experimentation
102 * may prove this needs changing.
103 */
104#define SCSI_QUEUE_DELAY 3
105
106static void
107scsi_set_blocked(struct scsi_cmnd *cmd, int reason)
108{
109 struct Scsi_Host *host = cmd->device->host;
110 struct scsi_device *device = cmd->device;
111 struct scsi_target *starget = scsi_target(device);
112
113 /*
114 * Set the appropriate busy bit for the device/host.
115 *
116 * If the host/device isn't busy, assume that something actually
117 * completed, and that we should be able to queue a command now.
118 *
119 * Note that the prior mid-layer assumption that any host could
120 * always queue at least one command is now broken. The mid-layer
121 * will implement a user specifiable stall (see
122 * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
123 * if a command is requeued with no other commands outstanding
124 * either for the device or for the host.
125 */
126 switch (reason) {
127 case SCSI_MLQUEUE_HOST_BUSY:
128 atomic_set(&host->host_blocked, host->max_host_blocked);
129 break;
130 case SCSI_MLQUEUE_DEVICE_BUSY:
131 case SCSI_MLQUEUE_EH_RETRY:
132 atomic_set(&device->device_blocked,
133 device->max_device_blocked);
134 break;
135 case SCSI_MLQUEUE_TARGET_BUSY:
136 atomic_set(&starget->target_blocked,
137 starget->max_target_blocked);
138 break;
139 }
140}
141
142static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd)
143{
144 struct scsi_device *sdev = cmd->device;
145
146 if (cmd->request->rq_flags & RQF_DONTPREP) {
147 cmd->request->rq_flags &= ~RQF_DONTPREP;
148 scsi_mq_uninit_cmd(cmd);
149 } else {
150 WARN_ON_ONCE(true);
151 }
152 blk_mq_requeue_request(cmd->request, true);
153 put_device(&sdev->sdev_gendev);
154}
155
156/**
157 * __scsi_queue_insert - private queue insertion
158 * @cmd: The SCSI command being requeued
159 * @reason: The reason for the requeue
160 * @unbusy: Whether the queue should be unbusied
161 *
162 * This is a private queue insertion. The public interface
163 * scsi_queue_insert() always assumes the queue should be unbusied
164 * because it's always called before the completion. This function is
165 * for a requeue after completion, which should only occur in this
166 * file.
167 */
168static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, bool unbusy)
169{
170 struct scsi_device *device = cmd->device;
171 struct request_queue *q = device->request_queue;
172 unsigned long flags;
173
174 SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
175 "Inserting command %p into mlqueue\n", cmd));
176
177 scsi_set_blocked(cmd, reason);
178
179 /*
180 * Decrement the counters, since these commands are no longer
181 * active on the host/device.
182 */
183 if (unbusy)
184 scsi_device_unbusy(device);
185
186 /*
187 * Requeue this command. It will go before all other commands
188 * that are already in the queue. Schedule requeue work under
189 * lock such that the kblockd_schedule_work() call happens
190 * before blk_cleanup_queue() finishes.
191 */
192 cmd->result = 0;
193 if (q->mq_ops) {
194 /*
195 * Before a SCSI command is dispatched,
196 * get_device(&sdev->sdev_gendev) is called and the host,
197 * target and device busy counters are increased. Since
198 * requeuing a request causes these actions to be repeated and
199 * since scsi_device_unbusy() has already been called,
200 * put_device(&device->sdev_gendev) must still be called. Call
201 * put_device() after blk_mq_requeue_request() to avoid that
202 * removal of the SCSI device can start before requeueing has
203 * happened.
204 */
205 blk_mq_requeue_request(cmd->request, true);
206 put_device(&device->sdev_gendev);
207 return;
208 }
209 spin_lock_irqsave(q->queue_lock, flags);
210 blk_requeue_request(q, cmd->request);
211 kblockd_schedule_work(&device->requeue_work);
212 spin_unlock_irqrestore(q->queue_lock, flags);
213}
214
215/*
216 * Function: scsi_queue_insert()
217 *
218 * Purpose: Insert a command in the midlevel queue.
219 *
220 * Arguments: cmd - command that we are adding to queue.
221 * reason - why we are inserting command to queue.
222 *
223 * Lock status: Assumed that lock is not held upon entry.
224 *
225 * Returns: Nothing.
226 *
227 * Notes: We do this for one of two cases. Either the host is busy
228 * and it cannot accept any more commands for the time being,
229 * or the device returned QUEUE_FULL and can accept no more
230 * commands.
231 * Notes: This could be called either from an interrupt context or a
232 * normal process context.
233 */
234void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
235{
236 __scsi_queue_insert(cmd, reason, true);
237}
238
239
240/**
241 * scsi_execute - insert request and wait for the result
242 * @sdev: scsi device
243 * @cmd: scsi command
244 * @data_direction: data direction
245 * @buffer: data buffer
246 * @bufflen: len of buffer
247 * @sense: optional sense buffer
248 * @sshdr: optional decoded sense header
249 * @timeout: request timeout in seconds
250 * @retries: number of times to retry request
251 * @flags: flags for ->cmd_flags
252 * @rq_flags: flags for ->rq_flags
253 * @resid: optional residual length
254 *
255 * Returns the scsi_cmnd result field if a command was executed, or a negative
256 * Linux error code if we didn't get that far.
257 */
258int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
259 int data_direction, void *buffer, unsigned bufflen,
260 unsigned char *sense, struct scsi_sense_hdr *sshdr,
261 int timeout, int retries, u64 flags, req_flags_t rq_flags,
262 int *resid)
263{
264 struct request *req;
265 struct scsi_request *rq;
266 int ret = DRIVER_ERROR << 24;
267
268 req = blk_get_request_flags(sdev->request_queue,
269 data_direction == DMA_TO_DEVICE ?
270 REQ_OP_SCSI_OUT : REQ_OP_SCSI_IN, BLK_MQ_REQ_PREEMPT);
271 if (IS_ERR(req))
272 return ret;
273 rq = scsi_req(req);
274
275 if (bufflen && blk_rq_map_kern(sdev->request_queue, req,
276 buffer, bufflen, __GFP_RECLAIM))
277 goto out;
278
279 rq->cmd_len = COMMAND_SIZE(cmd[0]);
280 memcpy(rq->cmd, cmd, rq->cmd_len);
281 rq->retries = retries;
282 req->timeout = timeout;
283 req->cmd_flags |= flags;
284 req->rq_flags |= rq_flags | RQF_QUIET;
285
286 /*
287 * head injection *required* here otherwise quiesce won't work
288 */
289 blk_execute_rq(req->q, NULL, req, 1);
290
291 /*
292 * Some devices (USB mass-storage in particular) may transfer
293 * garbage data together with a residue indicating that the data
294 * is invalid. Prevent the garbage from being misinterpreted
295 * and prevent security leaks by zeroing out the excess data.
296 */
297 if (unlikely(rq->resid_len > 0 && rq->resid_len <= bufflen))
298 memset(buffer + (bufflen - rq->resid_len), 0, rq->resid_len);
299
300 if (resid)
301 *resid = rq->resid_len;
302 if (sense && rq->sense_len)
303 memcpy(sense, rq->sense, SCSI_SENSE_BUFFERSIZE);
304 if (sshdr)
305 scsi_normalize_sense(rq->sense, rq->sense_len, sshdr);
306 ret = rq->result;
307 out:
308 blk_put_request(req);
309
310 return ret;
311}
312EXPORT_SYMBOL(scsi_execute);
313
314/*
315 * Function: scsi_init_cmd_errh()
316 *
317 * Purpose: Initialize cmd fields related to error handling.
318 *
319 * Arguments: cmd - command that is ready to be queued.
320 *
321 * Notes: This function has the job of initializing a number of
322 * fields related to error handling. Typically this will
323 * be called once for each command, as required.
324 */
325static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
326{
327 cmd->serial_number = 0;
328 scsi_set_resid(cmd, 0);
329 memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
330 if (cmd->cmd_len == 0)
331 cmd->cmd_len = scsi_command_size(cmd->cmnd);
332}
333
334/*
335 * Decrement the host_busy counter and wake up the error handler if necessary.
336 * Avoid as follows that the error handler is not woken up if shost->host_busy
337 * == shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
338 * with an RCU read lock in this function to ensure that this function in its
339 * entirety either finishes before scsi_eh_scmd_add() increases the
340 * host_failed counter or that it notices the shost state change made by
341 * scsi_eh_scmd_add().
342 */
343static void scsi_dec_host_busy(struct Scsi_Host *shost)
344{
345 unsigned long flags;
346
347 rcu_read_lock();
348 atomic_dec(&shost->host_busy);
349 if (unlikely(scsi_host_in_recovery(shost))) {
350 spin_lock_irqsave(shost->host_lock, flags);
351 if (shost->host_failed || shost->host_eh_scheduled)
352 scsi_eh_wakeup(shost);
353 spin_unlock_irqrestore(shost->host_lock, flags);
354 }
355 rcu_read_unlock();
356}
357
358void scsi_device_unbusy(struct scsi_device *sdev)
359{
360 struct Scsi_Host *shost = sdev->host;
361 struct scsi_target *starget = scsi_target(sdev);
362
363 scsi_dec_host_busy(shost);
364
365 if (starget->can_queue > 0)
366 atomic_dec(&starget->target_busy);
367
368 atomic_dec(&sdev->device_busy);
369}
370
371static void scsi_kick_queue(struct request_queue *q)
372{
373 if (q->mq_ops)
374 blk_mq_start_hw_queues(q);
375 else
376 blk_run_queue(q);
377}
378
379/*
380 * Called for single_lun devices on IO completion. Clear starget_sdev_user,
381 * and call blk_run_queue for all the scsi_devices on the target -
382 * including current_sdev first.
383 *
384 * Called with *no* scsi locks held.
385 */
386static void scsi_single_lun_run(struct scsi_device *current_sdev)
387{
388 struct Scsi_Host *shost = current_sdev->host;
389 struct scsi_device *sdev, *tmp;
390 struct scsi_target *starget = scsi_target(current_sdev);
391 unsigned long flags;
392
393 spin_lock_irqsave(shost->host_lock, flags);
394 starget->starget_sdev_user = NULL;
395 spin_unlock_irqrestore(shost->host_lock, flags);
396
397 /*
398 * Call blk_run_queue for all LUNs on the target, starting with
399 * current_sdev. We race with others (to set starget_sdev_user),
400 * but in most cases, we will be first. Ideally, each LU on the
401 * target would get some limited time or requests on the target.
402 */
403 scsi_kick_queue(current_sdev->request_queue);
404
405 spin_lock_irqsave(shost->host_lock, flags);
406 if (starget->starget_sdev_user)
407 goto out;
408 list_for_each_entry_safe(sdev, tmp, &starget->devices,
409 same_target_siblings) {
410 if (sdev == current_sdev)
411 continue;
412 if (scsi_device_get(sdev))
413 continue;
414
415 spin_unlock_irqrestore(shost->host_lock, flags);
416 scsi_kick_queue(sdev->request_queue);
417 spin_lock_irqsave(shost->host_lock, flags);
418
419 scsi_device_put(sdev);
420 }
421 out:
422 spin_unlock_irqrestore(shost->host_lock, flags);
423}
424
425static inline bool scsi_device_is_busy(struct scsi_device *sdev)
426{
427 if (atomic_read(&sdev->device_busy) >= sdev->queue_depth)
428 return true;
429 if (atomic_read(&sdev->device_blocked) > 0)
430 return true;
431 return false;
432}
433
434static inline bool scsi_target_is_busy(struct scsi_target *starget)
435{
436 if (starget->can_queue > 0) {
437 if (atomic_read(&starget->target_busy) >= starget->can_queue)
438 return true;
439 if (atomic_read(&starget->target_blocked) > 0)
440 return true;
441 }
442 return false;
443}
444
445static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
446{
447 if (shost->can_queue > 0 &&
448 atomic_read(&shost->host_busy) >= shost->can_queue)
449 return true;
450 if (atomic_read(&shost->host_blocked) > 0)
451 return true;
452 if (shost->host_self_blocked)
453 return true;
454 return false;
455}
456
457static void scsi_starved_list_run(struct Scsi_Host *shost)
458{
459 LIST_HEAD(starved_list);
460 struct scsi_device *sdev;
461 unsigned long flags;
462
463 spin_lock_irqsave(shost->host_lock, flags);
464 list_splice_init(&shost->starved_list, &starved_list);
465
466 while (!list_empty(&starved_list)) {
467 struct request_queue *slq;
468
469 /*
470 * As long as shost is accepting commands and we have
471 * starved queues, call blk_run_queue. scsi_request_fn
472 * drops the queue_lock and can add us back to the
473 * starved_list.
474 *
475 * host_lock protects the starved_list and starved_entry.
476 * scsi_request_fn must get the host_lock before checking
477 * or modifying starved_list or starved_entry.
478 */
479 if (scsi_host_is_busy(shost))
480 break;
481
482 sdev = list_entry(starved_list.next,
483 struct scsi_device, starved_entry);
484 list_del_init(&sdev->starved_entry);
485 if (scsi_target_is_busy(scsi_target(sdev))) {
486 list_move_tail(&sdev->starved_entry,
487 &shost->starved_list);
488 continue;
489 }
490
491 /*
492 * Once we drop the host lock, a racing scsi_remove_device()
493 * call may remove the sdev from the starved list and destroy
494 * it and the queue. Mitigate by taking a reference to the
495 * queue and never touching the sdev again after we drop the
496 * host lock. Note: if __scsi_remove_device() invokes
497 * blk_cleanup_queue() before the queue is run from this
498 * function then blk_run_queue() will return immediately since
499 * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
500 */
501 slq = sdev->request_queue;
502 if (!blk_get_queue(slq))
503 continue;
504 spin_unlock_irqrestore(shost->host_lock, flags);
505
506 scsi_kick_queue(slq);
507 blk_put_queue(slq);
508
509 spin_lock_irqsave(shost->host_lock, flags);
510 }
511 /* put any unprocessed entries back */
512 list_splice(&starved_list, &shost->starved_list);
513 spin_unlock_irqrestore(shost->host_lock, flags);
514}
515
516/*
517 * Function: scsi_run_queue()
518 *
519 * Purpose: Select a proper request queue to serve next
520 *
521 * Arguments: q - last request's queue
522 *
523 * Returns: Nothing
524 *
525 * Notes: The previous command was completely finished, start
526 * a new one if possible.
527 */
528static void scsi_run_queue(struct request_queue *q)
529{
530 struct scsi_device *sdev = q->queuedata;
531
532 if (scsi_target(sdev)->single_lun)
533 scsi_single_lun_run(sdev);
534 if (!list_empty(&sdev->host->starved_list))
535 scsi_starved_list_run(sdev->host);
536
537 if (q->mq_ops)
538 blk_mq_run_hw_queues(q, false);
539 else
540 blk_run_queue(q);
541}
542
543void scsi_requeue_run_queue(struct work_struct *work)
544{
545 struct scsi_device *sdev;
546 struct request_queue *q;
547
548 sdev = container_of(work, struct scsi_device, requeue_work);
549 q = sdev->request_queue;
550 scsi_run_queue(q);
551}
552
553/*
554 * Function: scsi_requeue_command()
555 *
556 * Purpose: Handle post-processing of completed commands.
557 *
558 * Arguments: q - queue to operate on
559 * cmd - command that may need to be requeued.
560 *
561 * Returns: Nothing
562 *
563 * Notes: After command completion, there may be blocks left
564 * over which weren't finished by the previous command
565 * this can be for a number of reasons - the main one is
566 * I/O errors in the middle of the request, in which case
567 * we need to request the blocks that come after the bad
568 * sector.
569 * Notes: Upon return, cmd is a stale pointer.
570 */
571static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
572{
573 struct scsi_device *sdev = cmd->device;
574 struct request *req = cmd->request;
575 unsigned long flags;
576
577 spin_lock_irqsave(q->queue_lock, flags);
578 blk_unprep_request(req);
579 req->special = NULL;
580 scsi_put_command(cmd);
581 blk_requeue_request(q, req);
582 spin_unlock_irqrestore(q->queue_lock, flags);
583
584 scsi_run_queue(q);
585
586 put_device(&sdev->sdev_gendev);
587}
588
589void scsi_run_host_queues(struct Scsi_Host *shost)
590{
591 struct scsi_device *sdev;
592
593 shost_for_each_device(sdev, shost)
594 scsi_run_queue(sdev->request_queue);
595}
596
597static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
598{
599 if (!blk_rq_is_passthrough(cmd->request)) {
600 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
601
602 if (drv->uninit_command)
603 drv->uninit_command(cmd);
604 }
605}
606
607static void scsi_mq_free_sgtables(struct scsi_cmnd *cmd)
608{
609 struct scsi_data_buffer *sdb;
610
611 if (cmd->sdb.table.nents)
612 sg_free_table_chained(&cmd->sdb.table, true);
613 if (cmd->request->next_rq) {
614 sdb = cmd->request->next_rq->special;
615 if (sdb)
616 sg_free_table_chained(&sdb->table, true);
617 }
618 if (scsi_prot_sg_count(cmd))
619 sg_free_table_chained(&cmd->prot_sdb->table, true);
620}
621
622static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
623{
624 scsi_mq_free_sgtables(cmd);
625 scsi_uninit_cmd(cmd);
626 scsi_del_cmd_from_list(cmd);
627}
628
629/*
630 * Function: scsi_release_buffers()
631 *
632 * Purpose: Free resources allocate for a scsi_command.
633 *
634 * Arguments: cmd - command that we are bailing.
635 *
636 * Lock status: Assumed that no lock is held upon entry.
637 *
638 * Returns: Nothing
639 *
640 * Notes: In the event that an upper level driver rejects a
641 * command, we must release resources allocated during
642 * the __init_io() function. Primarily this would involve
643 * the scatter-gather table.
644 */
645static void scsi_release_buffers(struct scsi_cmnd *cmd)
646{
647 if (cmd->sdb.table.nents)
648 sg_free_table_chained(&cmd->sdb.table, false);
649
650 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
651
652 if (scsi_prot_sg_count(cmd))
653 sg_free_table_chained(&cmd->prot_sdb->table, false);
654}
655
656static void scsi_release_bidi_buffers(struct scsi_cmnd *cmd)
657{
658 struct scsi_data_buffer *bidi_sdb = cmd->request->next_rq->special;
659
660 sg_free_table_chained(&bidi_sdb->table, false);
661 kmem_cache_free(scsi_sdb_cache, bidi_sdb);
662 cmd->request->next_rq->special = NULL;
663}
664
665static bool scsi_end_request(struct request *req, blk_status_t error,
666 unsigned int bytes, unsigned int bidi_bytes)
667{
668 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
669 struct scsi_device *sdev = cmd->device;
670 struct request_queue *q = sdev->request_queue;
671
672 if (blk_update_request(req, error, bytes))
673 return true;
674
675 /* Bidi request must be completed as a whole */
676 if (unlikely(bidi_bytes) &&
677 blk_update_request(req->next_rq, error, bidi_bytes))
678 return true;
679
680 if (blk_queue_add_random(q))
681 add_disk_randomness(req->rq_disk);
682
683 if (!blk_rq_is_scsi(req)) {
684 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
685 cmd->flags &= ~SCMD_INITIALIZED;
686 destroy_rcu_head(&cmd->rcu);
687 }
688
689 if (req->mq_ctx) {
690 /*
691 * In the MQ case the command gets freed by __blk_mq_end_request,
692 * so we have to do all cleanup that depends on it earlier.
693 *
694 * We also can't kick the queues from irq context, so we
695 * will have to defer it to a workqueue.
696 */
697 scsi_mq_uninit_cmd(cmd);
698
699 __blk_mq_end_request(req, error);
700
701 if (scsi_target(sdev)->single_lun ||
702 !list_empty(&sdev->host->starved_list))
703 kblockd_schedule_work(&sdev->requeue_work);
704 else
705 blk_mq_run_hw_queues(q, true);
706 } else {
707 unsigned long flags;
708
709 if (bidi_bytes)
710 scsi_release_bidi_buffers(cmd);
711 scsi_release_buffers(cmd);
712 scsi_put_command(cmd);
713
714 spin_lock_irqsave(q->queue_lock, flags);
715 blk_finish_request(req, error);
716 spin_unlock_irqrestore(q->queue_lock, flags);
717
718 scsi_run_queue(q);
719 }
720
721 put_device(&sdev->sdev_gendev);
722 return false;
723}
724
725/**
726 * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
727 * @cmd: SCSI command
728 * @result: scsi error code
729 *
730 * Translate a SCSI result code into a blk_status_t value. May reset the host
731 * byte of @cmd->result.
732 */
733static blk_status_t scsi_result_to_blk_status(struct scsi_cmnd *cmd, int result)
734{
735 switch (host_byte(result)) {
736 case DID_OK:
737 /*
738 * Also check the other bytes than the status byte in result
739 * to handle the case when a SCSI LLD sets result to
740 * DRIVER_SENSE << 24 without setting SAM_STAT_CHECK_CONDITION.
741 */
742 if (scsi_status_is_good(result) && (result & ~0xff) == 0)
743 return BLK_STS_OK;
744 return BLK_STS_IOERR;
745 case DID_TRANSPORT_FAILFAST:
746 return BLK_STS_TRANSPORT;
747 case DID_TARGET_FAILURE:
748 set_host_byte(cmd, DID_OK);
749 return BLK_STS_TARGET;
750 case DID_NEXUS_FAILURE:
751 return BLK_STS_NEXUS;
752 case DID_ALLOC_FAILURE:
753 set_host_byte(cmd, DID_OK);
754 return BLK_STS_NOSPC;
755 case DID_MEDIUM_ERROR:
756 set_host_byte(cmd, DID_OK);
757 return BLK_STS_MEDIUM;
758 default:
759 return BLK_STS_IOERR;
760 }
761}
762
763/*
764 * Function: scsi_io_completion()
765 *
766 * Purpose: Completion processing for block device I/O requests.
767 *
768 * Arguments: cmd - command that is finished.
769 *
770 * Lock status: Assumed that no lock is held upon entry.
771 *
772 * Returns: Nothing
773 *
774 * Notes: We will finish off the specified number of sectors. If we
775 * are done, the command block will be released and the queue
776 * function will be goosed. If we are not done then we have to
777 * figure out what to do next:
778 *
779 * a) We can call scsi_requeue_command(). The request
780 * will be unprepared and put back on the queue. Then
781 * a new command will be created for it. This should
782 * be used if we made forward progress, or if we want
783 * to switch from READ(10) to READ(6) for example.
784 *
785 * b) We can call __scsi_queue_insert(). The request will
786 * be put back on the queue and retried using the same
787 * command as before, possibly after a delay.
788 *
789 * c) We can call scsi_end_request() with -EIO to fail
790 * the remainder of the request.
791 */
792void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
793{
794 int result = cmd->result;
795 struct request_queue *q = cmd->device->request_queue;
796 struct request *req = cmd->request;
797 blk_status_t error = BLK_STS_OK;
798 struct scsi_sense_hdr sshdr;
799 bool sense_valid = false;
800 int sense_deferred = 0, level = 0;
801 enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
802 ACTION_DELAYED_RETRY} action;
803 unsigned long wait_for = (cmd->allowed + 1) * req->timeout;
804
805 if (result) {
806 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
807 if (sense_valid)
808 sense_deferred = scsi_sense_is_deferred(&sshdr);
809 }
810
811 if (blk_rq_is_passthrough(req)) {
812 if (result) {
813 if (sense_valid) {
814 /*
815 * SG_IO wants current and deferred errors
816 */
817 scsi_req(req)->sense_len =
818 min(8 + cmd->sense_buffer[7],
819 SCSI_SENSE_BUFFERSIZE);
820 }
821 if (!sense_deferred)
822 error = scsi_result_to_blk_status(cmd, result);
823 }
824 /*
825 * scsi_result_to_blk_status may have reset the host_byte
826 */
827 scsi_req(req)->result = cmd->result;
828 scsi_req(req)->resid_len = scsi_get_resid(cmd);
829
830 if (scsi_bidi_cmnd(cmd)) {
831 /*
832 * Bidi commands Must be complete as a whole,
833 * both sides at once.
834 */
835 scsi_req(req->next_rq)->resid_len = scsi_in(cmd)->resid;
836 if (scsi_end_request(req, BLK_STS_OK, blk_rq_bytes(req),
837 blk_rq_bytes(req->next_rq)))
838 BUG();
839 return;
840 }
841 } else if (blk_rq_bytes(req) == 0 && result && !sense_deferred) {
842 /*
843 * Flush commands do not transfers any data, and thus cannot use
844 * good_bytes != blk_rq_bytes(req) as the signal for an error.
845 * This sets the error explicitly for the problem case.
846 */
847 error = scsi_result_to_blk_status(cmd, result);
848 }
849
850 /* no bidi support for !blk_rq_is_passthrough yet */
851 BUG_ON(blk_bidi_rq(req));
852
853 /*
854 * Next deal with any sectors which we were able to correctly
855 * handle.
856 */
857 SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
858 "%u sectors total, %d bytes done.\n",
859 blk_rq_sectors(req), good_bytes));
860
861 /*
862 * Recovered errors need reporting, but they're always treated as
863 * success, so fiddle the result code here. For passthrough requests
864 * we already took a copy of the original into sreq->result which
865 * is what gets returned to the user
866 */
867 if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
868 /* if ATA PASS-THROUGH INFORMATION AVAILABLE skip
869 * print since caller wants ATA registers. Only occurs on
870 * SCSI ATA PASS_THROUGH commands when CK_COND=1
871 */
872 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
873 ;
874 else if (!(req->rq_flags & RQF_QUIET))
875 scsi_print_sense(cmd);
876 result = 0;
877 /* for passthrough error may be set */
878 error = BLK_STS_OK;
879 }
880 /*
881 * Another corner case: the SCSI status byte is non-zero but 'good'.
882 * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
883 * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
884 * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
885 * intermediate statuses (both obsolete in SAM-4) as good.
886 */
887 if (status_byte(result) && scsi_status_is_good(result)) {
888 result = 0;
889 error = BLK_STS_OK;
890 }
891
892 /*
893 * special case: failed zero length commands always need to
894 * drop down into the retry code. Otherwise, if we finished
895 * all bytes in the request we are done now.
896 */
897 if (!(blk_rq_bytes(req) == 0 && error) &&
898 !scsi_end_request(req, error, good_bytes, 0))
899 return;
900
901 /*
902 * Kill remainder if no retrys.
903 */
904 if (error && scsi_noretry_cmd(cmd)) {
905 if (scsi_end_request(req, error, blk_rq_bytes(req), 0))
906 BUG();
907 return;
908 }
909
910 /*
911 * If there had been no error, but we have leftover bytes in the
912 * requeues just queue the command up again.
913 */
914 if (result == 0)
915 goto requeue;
916
917 error = scsi_result_to_blk_status(cmd, result);
918
919 if (host_byte(result) == DID_RESET) {
920 /* Third party bus reset or reset for error recovery
921 * reasons. Just retry the command and see what
922 * happens.
923 */
924 action = ACTION_RETRY;
925 } else if (sense_valid && !sense_deferred) {
926 switch (sshdr.sense_key) {
927 case UNIT_ATTENTION:
928 if (cmd->device->removable) {
929 /* Detected disc change. Set a bit
930 * and quietly refuse further access.
931 */
932 cmd->device->changed = 1;
933 action = ACTION_FAIL;
934 } else {
935 /* Must have been a power glitch, or a
936 * bus reset. Could not have been a
937 * media change, so we just retry the
938 * command and see what happens.
939 */
940 action = ACTION_RETRY;
941 }
942 break;
943 case ILLEGAL_REQUEST:
944 /* If we had an ILLEGAL REQUEST returned, then
945 * we may have performed an unsupported
946 * command. The only thing this should be
947 * would be a ten byte read where only a six
948 * byte read was supported. Also, on a system
949 * where READ CAPACITY failed, we may have
950 * read past the end of the disk.
951 */
952 if ((cmd->device->use_10_for_rw &&
953 sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
954 (cmd->cmnd[0] == READ_10 ||
955 cmd->cmnd[0] == WRITE_10)) {
956 /* This will issue a new 6-byte command. */
957 cmd->device->use_10_for_rw = 0;
958 action = ACTION_REPREP;
959 } else if (sshdr.asc == 0x10) /* DIX */ {
960 action = ACTION_FAIL;
961 error = BLK_STS_PROTECTION;
962 /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
963 } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
964 action = ACTION_FAIL;
965 error = BLK_STS_TARGET;
966 } else
967 action = ACTION_FAIL;
968 break;
969 case ABORTED_COMMAND:
970 action = ACTION_FAIL;
971 if (sshdr.asc == 0x10) /* DIF */
972 error = BLK_STS_PROTECTION;
973 break;
974 case NOT_READY:
975 /* If the device is in the process of becoming
976 * ready, or has a temporary blockage, retry.
977 */
978 if (sshdr.asc == 0x04) {
979 switch (sshdr.ascq) {
980 case 0x01: /* becoming ready */
981 case 0x04: /* format in progress */
982 case 0x05: /* rebuild in progress */
983 case 0x06: /* recalculation in progress */
984 case 0x07: /* operation in progress */
985 case 0x08: /* Long write in progress */
986 case 0x09: /* self test in progress */
987 case 0x14: /* space allocation in progress */
988 action = ACTION_DELAYED_RETRY;
989 break;
990 default:
991 action = ACTION_FAIL;
992 break;
993 }
994 } else
995 action = ACTION_FAIL;
996 break;
997 case VOLUME_OVERFLOW:
998 /* See SSC3rXX or current. */
999 action = ACTION_FAIL;
1000 break;
1001 default:
1002 action = ACTION_FAIL;
1003 break;
1004 }
1005 } else
1006 action = ACTION_FAIL;
1007
1008 if (action != ACTION_FAIL &&
1009 time_before(cmd->jiffies_at_alloc + wait_for, jiffies))
1010 action = ACTION_FAIL;
1011
1012 switch (action) {
1013 case ACTION_FAIL:
1014 /* Give up and fail the remainder of the request */
1015 if (!(req->rq_flags & RQF_QUIET)) {
1016 static DEFINE_RATELIMIT_STATE(_rs,
1017 DEFAULT_RATELIMIT_INTERVAL,
1018 DEFAULT_RATELIMIT_BURST);
1019
1020 if (unlikely(scsi_logging_level))
1021 level = SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
1022 SCSI_LOG_MLCOMPLETE_BITS);
1023
1024 /*
1025 * if logging is enabled the failure will be printed
1026 * in scsi_log_completion(), so avoid duplicate messages
1027 */
1028 if (!level && __ratelimit(&_rs)) {
1029 scsi_print_result(cmd, NULL, FAILED);
1030 if (driver_byte(result) & DRIVER_SENSE)
1031 scsi_print_sense(cmd);
1032 scsi_print_command(cmd);
1033 }
1034 }
1035 if (!scsi_end_request(req, error, blk_rq_err_bytes(req), 0))
1036 return;
1037 /*FALLTHRU*/
1038 case ACTION_REPREP:
1039 requeue:
1040 /* Unprep the request and put it back at the head of the queue.
1041 * A new command will be prepared and issued.
1042 */
1043 if (q->mq_ops) {
1044 scsi_mq_requeue_cmd(cmd);
1045 } else {
1046 scsi_release_buffers(cmd);
1047 scsi_requeue_command(q, cmd);
1048 }
1049 break;
1050 case ACTION_RETRY:
1051 /* Retry the same command immediately */
1052 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
1053 break;
1054 case ACTION_DELAYED_RETRY:
1055 /* Retry the same command after a delay */
1056 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
1057 break;
1058 }
1059}
1060
1061static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb)
1062{
1063 int count;
1064
1065 /*
1066 * If sg table allocation fails, requeue request later.
1067 */
1068 if (unlikely(sg_alloc_table_chained(&sdb->table,
1069 blk_rq_nr_phys_segments(req), sdb->table.sgl)))
1070 return BLKPREP_DEFER;
1071
1072 /*
1073 * Next, walk the list, and fill in the addresses and sizes of
1074 * each segment.
1075 */
1076 count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
1077 BUG_ON(count > sdb->table.nents);
1078 sdb->table.nents = count;
1079 sdb->length = blk_rq_payload_bytes(req);
1080 return BLKPREP_OK;
1081}
1082
1083/*
1084 * Function: scsi_init_io()
1085 *
1086 * Purpose: SCSI I/O initialize function.
1087 *
1088 * Arguments: cmd - Command descriptor we wish to initialize
1089 *
1090 * Returns: 0 on success
1091 * BLKPREP_DEFER if the failure is retryable
1092 * BLKPREP_KILL if the failure is fatal
1093 */
1094int scsi_init_io(struct scsi_cmnd *cmd)
1095{
1096 struct scsi_device *sdev = cmd->device;
1097 struct request *rq = cmd->request;
1098 bool is_mq = (rq->mq_ctx != NULL);
1099 int error = BLKPREP_KILL;
1100
1101 if (WARN_ON_ONCE(!blk_rq_nr_phys_segments(rq)))
1102 goto err_exit;
1103
1104 error = scsi_init_sgtable(rq, &cmd->sdb);
1105 if (error)
1106 goto err_exit;
1107
1108 if (blk_bidi_rq(rq)) {
1109 if (!rq->q->mq_ops) {
1110 struct scsi_data_buffer *bidi_sdb =
1111 kmem_cache_zalloc(scsi_sdb_cache, GFP_ATOMIC);
1112 if (!bidi_sdb) {
1113 error = BLKPREP_DEFER;
1114 goto err_exit;
1115 }
1116
1117 rq->next_rq->special = bidi_sdb;
1118 }
1119
1120 error = scsi_init_sgtable(rq->next_rq, rq->next_rq->special);
1121 if (error)
1122 goto err_exit;
1123 }
1124
1125 if (blk_integrity_rq(rq)) {
1126 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1127 int ivecs, count;
1128
1129 if (prot_sdb == NULL) {
1130 /*
1131 * This can happen if someone (e.g. multipath)
1132 * queues a command to a device on an adapter
1133 * that does not support DIX.
1134 */
1135 WARN_ON_ONCE(1);
1136 error = BLKPREP_KILL;
1137 goto err_exit;
1138 }
1139
1140 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1141
1142 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1143 prot_sdb->table.sgl)) {
1144 error = BLKPREP_DEFER;
1145 goto err_exit;
1146 }
1147
1148 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1149 prot_sdb->table.sgl);
1150 BUG_ON(unlikely(count > ivecs));
1151 BUG_ON(unlikely(count > queue_max_integrity_segments(rq->q)));
1152
1153 cmd->prot_sdb = prot_sdb;
1154 cmd->prot_sdb->table.nents = count;
1155 }
1156
1157 return BLKPREP_OK;
1158err_exit:
1159 if (is_mq) {
1160 scsi_mq_free_sgtables(cmd);
1161 } else {
1162 scsi_release_buffers(cmd);
1163 cmd->request->special = NULL;
1164 scsi_put_command(cmd);
1165 put_device(&sdev->sdev_gendev);
1166 }
1167 return error;
1168}
1169EXPORT_SYMBOL(scsi_init_io);
1170
1171/**
1172 * scsi_initialize_rq - initialize struct scsi_cmnd partially
1173 * @rq: Request associated with the SCSI command to be initialized.
1174 *
1175 * This function initializes the members of struct scsi_cmnd that must be
1176 * initialized before request processing starts and that won't be
1177 * reinitialized if a SCSI command is requeued.
1178 *
1179 * Called from inside blk_get_request() for pass-through requests and from
1180 * inside scsi_init_command() for filesystem requests.
1181 */
1182static void scsi_initialize_rq(struct request *rq)
1183{
1184 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1185
1186 scsi_req_init(&cmd->req);
1187 init_rcu_head(&cmd->rcu);
1188 cmd->jiffies_at_alloc = jiffies;
1189 cmd->retries = 0;
1190}
1191
1192/* Add a command to the list used by the aacraid and dpt_i2o drivers */
1193void scsi_add_cmd_to_list(struct scsi_cmnd *cmd)
1194{
1195 struct scsi_device *sdev = cmd->device;
1196 struct Scsi_Host *shost = sdev->host;
1197 unsigned long flags;
1198
1199 if (shost->use_cmd_list) {
1200 spin_lock_irqsave(&sdev->list_lock, flags);
1201 list_add_tail(&cmd->list, &sdev->cmd_list);
1202 spin_unlock_irqrestore(&sdev->list_lock, flags);
1203 }
1204}
1205
1206/* Remove a command from the list used by the aacraid and dpt_i2o drivers */
1207void scsi_del_cmd_from_list(struct scsi_cmnd *cmd)
1208{
1209 struct scsi_device *sdev = cmd->device;
1210 struct Scsi_Host *shost = sdev->host;
1211 unsigned long flags;
1212
1213 if (shost->use_cmd_list) {
1214 spin_lock_irqsave(&sdev->list_lock, flags);
1215 BUG_ON(list_empty(&cmd->list));
1216 list_del_init(&cmd->list);
1217 spin_unlock_irqrestore(&sdev->list_lock, flags);
1218 }
1219}
1220
1221/* Called after a request has been started. */
1222void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1223{
1224 void *buf = cmd->sense_buffer;
1225 void *prot = cmd->prot_sdb;
1226 struct request *rq = blk_mq_rq_from_pdu(cmd);
1227 unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1228 unsigned long jiffies_at_alloc;
1229 int retries;
1230
1231 if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1232 flags |= SCMD_INITIALIZED;
1233 scsi_initialize_rq(rq);
1234 }
1235
1236 jiffies_at_alloc = cmd->jiffies_at_alloc;
1237 retries = cmd->retries;
1238 /* zero out the cmd, except for the embedded scsi_request */
1239 memset((char *)cmd + sizeof(cmd->req), 0,
1240 sizeof(*cmd) - sizeof(cmd->req) + dev->host->hostt->cmd_size);
1241
1242 cmd->device = dev;
1243 cmd->sense_buffer = buf;
1244 cmd->prot_sdb = prot;
1245 cmd->flags = flags;
1246 INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1247 cmd->jiffies_at_alloc = jiffies_at_alloc;
1248 cmd->retries = retries;
1249
1250 scsi_add_cmd_to_list(cmd);
1251}
1252
1253static int scsi_setup_scsi_cmnd(struct scsi_device *sdev, struct request *req)
1254{
1255 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1256
1257 /*
1258 * Passthrough requests may transfer data, in which case they must
1259 * a bio attached to them. Or they might contain a SCSI command
1260 * that does not transfer data, in which case they may optionally
1261 * submit a request without an attached bio.
1262 */
1263 if (req->bio) {
1264 int ret = scsi_init_io(cmd);
1265 if (unlikely(ret))
1266 return ret;
1267 } else {
1268 BUG_ON(blk_rq_bytes(req));
1269
1270 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1271 }
1272
1273 cmd->cmd_len = scsi_req(req)->cmd_len;
1274 cmd->cmnd = scsi_req(req)->cmd;
1275 cmd->transfersize = blk_rq_bytes(req);
1276 cmd->allowed = scsi_req(req)->retries;
1277 return BLKPREP_OK;
1278}
1279
1280/*
1281 * Setup a normal block command. These are simple request from filesystems
1282 * that still need to be translated to SCSI CDBs from the ULD.
1283 */
1284static int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1285{
1286 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1287
1288 if (unlikely(sdev->handler && sdev->handler->prep_fn)) {
1289 int ret = sdev->handler->prep_fn(sdev, req);
1290 if (ret != BLKPREP_OK)
1291 return ret;
1292 }
1293
1294 cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1295 memset(cmd->cmnd, 0, BLK_MAX_CDB);
1296 return scsi_cmd_to_driver(cmd)->init_command(cmd);
1297}
1298
1299static int scsi_setup_cmnd(struct scsi_device *sdev, struct request *req)
1300{
1301 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1302
1303 if (!blk_rq_bytes(req))
1304 cmd->sc_data_direction = DMA_NONE;
1305 else if (rq_data_dir(req) == WRITE)
1306 cmd->sc_data_direction = DMA_TO_DEVICE;
1307 else
1308 cmd->sc_data_direction = DMA_FROM_DEVICE;
1309
1310 if (blk_rq_is_scsi(req))
1311 return scsi_setup_scsi_cmnd(sdev, req);
1312 else
1313 return scsi_setup_fs_cmnd(sdev, req);
1314}
1315
1316static int
1317scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1318{
1319 int ret = BLKPREP_OK;
1320
1321 /*
1322 * If the device is not in running state we will reject some
1323 * or all commands.
1324 */
1325 if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1326 switch (sdev->sdev_state) {
1327 case SDEV_OFFLINE:
1328 case SDEV_TRANSPORT_OFFLINE:
1329 /*
1330 * If the device is offline we refuse to process any
1331 * commands. The device must be brought online
1332 * before trying any recovery commands.
1333 */
1334 sdev_printk(KERN_ERR, sdev,
1335 "rejecting I/O to offline device\n");
1336 ret = BLKPREP_KILL;
1337 break;
1338 case SDEV_DEL:
1339 /*
1340 * If the device is fully deleted, we refuse to
1341 * process any commands as well.
1342 */
1343 sdev_printk(KERN_ERR, sdev,
1344 "rejecting I/O to dead device\n");
1345 ret = BLKPREP_KILL;
1346 break;
1347 case SDEV_BLOCK:
1348 case SDEV_CREATED_BLOCK:
1349 ret = BLKPREP_DEFER;
1350 break;
1351 case SDEV_QUIESCE:
1352 /*
1353 * If the devices is blocked we defer normal commands.
1354 */
1355 if (req && !(req->rq_flags & RQF_PREEMPT))
1356 ret = BLKPREP_DEFER;
1357 break;
1358 default:
1359 /*
1360 * For any other not fully online state we only allow
1361 * special commands. In particular any user initiated
1362 * command is not allowed.
1363 */
1364 if (req && !(req->rq_flags & RQF_PREEMPT))
1365 ret = BLKPREP_KILL;
1366 break;
1367 }
1368 }
1369 return ret;
1370}
1371
1372static int
1373scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1374{
1375 struct scsi_device *sdev = q->queuedata;
1376
1377 switch (ret) {
1378 case BLKPREP_KILL:
1379 case BLKPREP_INVALID:
1380 scsi_req(req)->result = DID_NO_CONNECT << 16;
1381 /* release the command and kill it */
1382 if (req->special) {
1383 struct scsi_cmnd *cmd = req->special;
1384 scsi_release_buffers(cmd);
1385 scsi_put_command(cmd);
1386 put_device(&sdev->sdev_gendev);
1387 req->special = NULL;
1388 }
1389 break;
1390 case BLKPREP_DEFER:
1391 /*
1392 * If we defer, the blk_peek_request() returns NULL, but the
1393 * queue must be restarted, so we schedule a callback to happen
1394 * shortly.
1395 */
1396 if (atomic_read(&sdev->device_busy) == 0)
1397 blk_delay_queue(q, SCSI_QUEUE_DELAY);
1398 break;
1399 default:
1400 req->rq_flags |= RQF_DONTPREP;
1401 }
1402
1403 return ret;
1404}
1405
1406static int scsi_prep_fn(struct request_queue *q, struct request *req)
1407{
1408 struct scsi_device *sdev = q->queuedata;
1409 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1410 int ret;
1411
1412 ret = scsi_prep_state_check(sdev, req);
1413 if (ret != BLKPREP_OK)
1414 goto out;
1415
1416 if (!req->special) {
1417 /* Bail if we can't get a reference to the device */
1418 if (unlikely(!get_device(&sdev->sdev_gendev))) {
1419 ret = BLKPREP_DEFER;
1420 goto out;
1421 }
1422
1423 scsi_init_command(sdev, cmd);
1424 req->special = cmd;
1425 }
1426
1427 cmd->tag = req->tag;
1428 cmd->request = req;
1429 cmd->prot_op = SCSI_PROT_NORMAL;
1430
1431 ret = scsi_setup_cmnd(sdev, req);
1432out:
1433 return scsi_prep_return(q, req, ret);
1434}
1435
1436static void scsi_unprep_fn(struct request_queue *q, struct request *req)
1437{
1438 scsi_uninit_cmd(blk_mq_rq_to_pdu(req));
1439}
1440
1441/*
1442 * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1443 * return 0.
1444 *
1445 * Called with the queue_lock held.
1446 */
1447static inline int scsi_dev_queue_ready(struct request_queue *q,
1448 struct scsi_device *sdev)
1449{
1450 unsigned int busy;
1451
1452 busy = atomic_inc_return(&sdev->device_busy) - 1;
1453 if (atomic_read(&sdev->device_blocked)) {
1454 if (busy)
1455 goto out_dec;
1456
1457 /*
1458 * unblock after device_blocked iterates to zero
1459 */
1460 if (atomic_dec_return(&sdev->device_blocked) > 0) {
1461 /*
1462 * For the MQ case we take care of this in the caller.
1463 */
1464 if (!q->mq_ops)
1465 blk_delay_queue(q, SCSI_QUEUE_DELAY);
1466 goto out_dec;
1467 }
1468 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1469 "unblocking device at zero depth\n"));
1470 }
1471
1472 if (busy >= sdev->queue_depth)
1473 goto out_dec;
1474
1475 return 1;
1476out_dec:
1477 atomic_dec(&sdev->device_busy);
1478 return 0;
1479}
1480
1481/*
1482 * scsi_target_queue_ready: checks if there we can send commands to target
1483 * @sdev: scsi device on starget to check.
1484 */
1485static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1486 struct scsi_device *sdev)
1487{
1488 struct scsi_target *starget = scsi_target(sdev);
1489 unsigned int busy;
1490
1491 if (starget->single_lun) {
1492 spin_lock_irq(shost->host_lock);
1493 if (starget->starget_sdev_user &&
1494 starget->starget_sdev_user != sdev) {
1495 spin_unlock_irq(shost->host_lock);
1496 return 0;
1497 }
1498 starget->starget_sdev_user = sdev;
1499 spin_unlock_irq(shost->host_lock);
1500 }
1501
1502 if (starget->can_queue <= 0)
1503 return 1;
1504
1505 busy = atomic_inc_return(&starget->target_busy) - 1;
1506 if (atomic_read(&starget->target_blocked) > 0) {
1507 if (busy)
1508 goto starved;
1509
1510 /*
1511 * unblock after target_blocked iterates to zero
1512 */
1513 if (atomic_dec_return(&starget->target_blocked) > 0)
1514 goto out_dec;
1515
1516 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1517 "unblocking target at zero depth\n"));
1518 }
1519
1520 if (busy >= starget->can_queue)
1521 goto starved;
1522
1523 return 1;
1524
1525starved:
1526 spin_lock_irq(shost->host_lock);
1527 list_move_tail(&sdev->starved_entry, &shost->starved_list);
1528 spin_unlock_irq(shost->host_lock);
1529out_dec:
1530 if (starget->can_queue > 0)
1531 atomic_dec(&starget->target_busy);
1532 return 0;
1533}
1534
1535/*
1536 * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1537 * return 0. We must end up running the queue again whenever 0 is
1538 * returned, else IO can hang.
1539 */
1540static inline int scsi_host_queue_ready(struct request_queue *q,
1541 struct Scsi_Host *shost,
1542 struct scsi_device *sdev)
1543{
1544 unsigned int busy;
1545
1546 if (scsi_host_in_recovery(shost))
1547 return 0;
1548
1549 busy = atomic_inc_return(&shost->host_busy) - 1;
1550 if (atomic_read(&shost->host_blocked) > 0) {
1551 if (busy)
1552 goto starved;
1553
1554 /*
1555 * unblock after host_blocked iterates to zero
1556 */
1557 if (atomic_dec_return(&shost->host_blocked) > 0)
1558 goto out_dec;
1559
1560 SCSI_LOG_MLQUEUE(3,
1561 shost_printk(KERN_INFO, shost,
1562 "unblocking host at zero depth\n"));
1563 }
1564
1565 if (shost->can_queue > 0 && busy >= shost->can_queue)
1566 goto starved;
1567 if (shost->host_self_blocked)
1568 goto starved;
1569
1570 /* We're OK to process the command, so we can't be starved */
1571 if (!list_empty(&sdev->starved_entry)) {
1572 spin_lock_irq(shost->host_lock);
1573 if (!list_empty(&sdev->starved_entry))
1574 list_del_init(&sdev->starved_entry);
1575 spin_unlock_irq(shost->host_lock);
1576 }
1577
1578 return 1;
1579
1580starved:
1581 spin_lock_irq(shost->host_lock);
1582 if (list_empty(&sdev->starved_entry))
1583 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1584 spin_unlock_irq(shost->host_lock);
1585out_dec:
1586 scsi_dec_host_busy(shost);
1587 return 0;
1588}
1589
1590/*
1591 * Busy state exporting function for request stacking drivers.
1592 *
1593 * For efficiency, no lock is taken to check the busy state of
1594 * shost/starget/sdev, since the returned value is not guaranteed and
1595 * may be changed after request stacking drivers call the function,
1596 * regardless of taking lock or not.
1597 *
1598 * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1599 * needs to return 'not busy'. Otherwise, request stacking drivers
1600 * may hold requests forever.
1601 */
1602static int scsi_lld_busy(struct request_queue *q)
1603{
1604 struct scsi_device *sdev = q->queuedata;
1605 struct Scsi_Host *shost;
1606
1607 if (blk_queue_dying(q))
1608 return 0;
1609
1610 shost = sdev->host;
1611
1612 /*
1613 * Ignore host/starget busy state.
1614 * Since block layer does not have a concept of fairness across
1615 * multiple queues, congestion of host/starget needs to be handled
1616 * in SCSI layer.
1617 */
1618 if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1619 return 1;
1620
1621 return 0;
1622}
1623
1624/*
1625 * Kill a request for a dead device
1626 */
1627static void scsi_kill_request(struct request *req, struct request_queue *q)
1628{
1629 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1630 struct scsi_device *sdev;
1631 struct scsi_target *starget;
1632 struct Scsi_Host *shost;
1633
1634 blk_start_request(req);
1635
1636 scmd_printk(KERN_INFO, cmd, "killing request\n");
1637
1638 sdev = cmd->device;
1639 starget = scsi_target(sdev);
1640 shost = sdev->host;
1641 scsi_init_cmd_errh(cmd);
1642 cmd->result = DID_NO_CONNECT << 16;
1643 atomic_inc(&cmd->device->iorequest_cnt);
1644
1645 /*
1646 * SCSI request completion path will do scsi_device_unbusy(),
1647 * bump busy counts. To bump the counters, we need to dance
1648 * with the locks as normal issue path does.
1649 */
1650 atomic_inc(&sdev->device_busy);
1651 atomic_inc(&shost->host_busy);
1652 if (starget->can_queue > 0)
1653 atomic_inc(&starget->target_busy);
1654
1655 blk_complete_request(req);
1656}
1657
1658static void scsi_softirq_done(struct request *rq)
1659{
1660 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1661 unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1662 int disposition;
1663
1664 INIT_LIST_HEAD(&cmd->eh_entry);
1665
1666 atomic_inc(&cmd->device->iodone_cnt);
1667 if (cmd->result)
1668 atomic_inc(&cmd->device->ioerr_cnt);
1669
1670 disposition = scsi_decide_disposition(cmd);
1671 if (disposition != SUCCESS &&
1672 time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1673 sdev_printk(KERN_ERR, cmd->device,
1674 "timing out command, waited %lus\n",
1675 wait_for/HZ);
1676 disposition = SUCCESS;
1677 }
1678
1679 scsi_log_completion(cmd, disposition);
1680
1681 switch (disposition) {
1682 case SUCCESS:
1683 scsi_finish_command(cmd);
1684 break;
1685 case NEEDS_RETRY:
1686 scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1687 break;
1688 case ADD_TO_MLQUEUE:
1689 scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1690 break;
1691 default:
1692 scsi_eh_scmd_add(cmd);
1693 break;
1694 }
1695}
1696
1697/**
1698 * scsi_dispatch_command - Dispatch a command to the low-level driver.
1699 * @cmd: command block we are dispatching.
1700 *
1701 * Return: nonzero return request was rejected and device's queue needs to be
1702 * plugged.
1703 */
1704static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1705{
1706 struct Scsi_Host *host = cmd->device->host;
1707 int rtn = 0;
1708
1709 atomic_inc(&cmd->device->iorequest_cnt);
1710
1711 /* check if the device is still usable */
1712 if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1713 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1714 * returns an immediate error upwards, and signals
1715 * that the device is no longer present */
1716 cmd->result = DID_NO_CONNECT << 16;
1717 goto done;
1718 }
1719
1720 /* Check to see if the scsi lld made this device blocked. */
1721 if (unlikely(scsi_device_blocked(cmd->device))) {
1722 /*
1723 * in blocked state, the command is just put back on
1724 * the device queue. The suspend state has already
1725 * blocked the queue so future requests should not
1726 * occur until the device transitions out of the
1727 * suspend state.
1728 */
1729 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1730 "queuecommand : device blocked\n"));
1731 return SCSI_MLQUEUE_DEVICE_BUSY;
1732 }
1733
1734 /* Store the LUN value in cmnd, if needed. */
1735 if (cmd->device->lun_in_cdb)
1736 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1737 (cmd->device->lun << 5 & 0xe0);
1738
1739 scsi_log_send(cmd);
1740
1741 /*
1742 * Before we queue this command, check if the command
1743 * length exceeds what the host adapter can handle.
1744 */
1745 if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1746 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1747 "queuecommand : command too long. "
1748 "cdb_size=%d host->max_cmd_len=%d\n",
1749 cmd->cmd_len, cmd->device->host->max_cmd_len));
1750 cmd->result = (DID_ABORT << 16);
1751 goto done;
1752 }
1753
1754 if (unlikely(host->shost_state == SHOST_DEL)) {
1755 cmd->result = (DID_NO_CONNECT << 16);
1756 goto done;
1757
1758 }
1759
1760 trace_scsi_dispatch_cmd_start(cmd);
1761 rtn = host->hostt->queuecommand(host, cmd);
1762 if (rtn) {
1763 trace_scsi_dispatch_cmd_error(cmd, rtn);
1764 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1765 rtn != SCSI_MLQUEUE_TARGET_BUSY)
1766 rtn = SCSI_MLQUEUE_HOST_BUSY;
1767
1768 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1769 "queuecommand : request rejected\n"));
1770 }
1771
1772 return rtn;
1773 done:
1774 cmd->scsi_done(cmd);
1775 return 0;
1776}
1777
1778/**
1779 * scsi_done - Invoke completion on finished SCSI command.
1780 * @cmd: The SCSI Command for which a low-level device driver (LLDD) gives
1781 * ownership back to SCSI Core -- i.e. the LLDD has finished with it.
1782 *
1783 * Description: This function is the mid-level's (SCSI Core) interrupt routine,
1784 * which regains ownership of the SCSI command (de facto) from a LLDD, and
1785 * calls blk_complete_request() for further processing.
1786 *
1787 * This function is interrupt context safe.
1788 */
1789static void scsi_done(struct scsi_cmnd *cmd)
1790{
1791 trace_scsi_dispatch_cmd_done(cmd);
1792 blk_complete_request(cmd->request);
1793}
1794
1795/*
1796 * Function: scsi_request_fn()
1797 *
1798 * Purpose: Main strategy routine for SCSI.
1799 *
1800 * Arguments: q - Pointer to actual queue.
1801 *
1802 * Returns: Nothing
1803 *
1804 * Lock status: request queue lock assumed to be held when called.
1805 *
1806 * Note: See sd_zbc.c sd_zbc_write_lock_zone() for write order
1807 * protection for ZBC disks.
1808 */
1809static void scsi_request_fn(struct request_queue *q)
1810 __releases(q->queue_lock)
1811 __acquires(q->queue_lock)
1812{
1813 struct scsi_device *sdev = q->queuedata;
1814 struct Scsi_Host *shost;
1815 struct scsi_cmnd *cmd;
1816 struct request *req;
1817
1818 /*
1819 * To start with, we keep looping until the queue is empty, or until
1820 * the host is no longer able to accept any more requests.
1821 */
1822 shost = sdev->host;
1823 for (;;) {
1824 int rtn;
1825 /*
1826 * get next queueable request. We do this early to make sure
1827 * that the request is fully prepared even if we cannot
1828 * accept it.
1829 */
1830 req = blk_peek_request(q);
1831 if (!req)
1832 break;
1833
1834 if (unlikely(!scsi_device_online(sdev))) {
1835 sdev_printk(KERN_ERR, sdev,
1836 "rejecting I/O to offline device\n");
1837 scsi_kill_request(req, q);
1838 continue;
1839 }
1840
1841 if (!scsi_dev_queue_ready(q, sdev))
1842 break;
1843
1844 /*
1845 * Remove the request from the request list.
1846 */
1847 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1848 blk_start_request(req);
1849
1850 spin_unlock_irq(q->queue_lock);
1851 cmd = blk_mq_rq_to_pdu(req);
1852 if (cmd != req->special) {
1853 printk(KERN_CRIT "impossible request in %s.\n"
1854 "please mail a stack trace to "
1855 "linux-scsi@vger.kernel.org\n",
1856 __func__);
1857 blk_dump_rq_flags(req, "foo");
1858 BUG();
1859 }
1860
1861 /*
1862 * We hit this when the driver is using a host wide
1863 * tag map. For device level tag maps the queue_depth check
1864 * in the device ready fn would prevent us from trying
1865 * to allocate a tag. Since the map is a shared host resource
1866 * we add the dev to the starved list so it eventually gets
1867 * a run when a tag is freed.
1868 */
1869 if (blk_queue_tagged(q) && !(req->rq_flags & RQF_QUEUED)) {
1870 spin_lock_irq(shost->host_lock);
1871 if (list_empty(&sdev->starved_entry))
1872 list_add_tail(&sdev->starved_entry,
1873 &shost->starved_list);
1874 spin_unlock_irq(shost->host_lock);
1875 goto not_ready;
1876 }
1877
1878 if (!scsi_target_queue_ready(shost, sdev))
1879 goto not_ready;
1880
1881 if (!scsi_host_queue_ready(q, shost, sdev))
1882 goto host_not_ready;
1883
1884 if (sdev->simple_tags)
1885 cmd->flags |= SCMD_TAGGED;
1886 else
1887 cmd->flags &= ~SCMD_TAGGED;
1888
1889 /*
1890 * Finally, initialize any error handling parameters, and set up
1891 * the timers for timeouts.
1892 */
1893 scsi_init_cmd_errh(cmd);
1894
1895 /*
1896 * Dispatch the command to the low-level driver.
1897 */
1898 cmd->scsi_done = scsi_done;
1899 rtn = scsi_dispatch_cmd(cmd);
1900 if (rtn) {
1901 scsi_queue_insert(cmd, rtn);
1902 spin_lock_irq(q->queue_lock);
1903 goto out_delay;
1904 }
1905 spin_lock_irq(q->queue_lock);
1906 }
1907
1908 return;
1909
1910 host_not_ready:
1911 if (scsi_target(sdev)->can_queue > 0)
1912 atomic_dec(&scsi_target(sdev)->target_busy);
1913 not_ready:
1914 /*
1915 * lock q, handle tag, requeue req, and decrement device_busy. We
1916 * must return with queue_lock held.
1917 *
1918 * Decrementing device_busy without checking it is OK, as all such
1919 * cases (host limits or settings) should run the queue at some
1920 * later time.
1921 */
1922 spin_lock_irq(q->queue_lock);
1923 blk_requeue_request(q, req);
1924 atomic_dec(&sdev->device_busy);
1925out_delay:
1926 if (!atomic_read(&sdev->device_busy) && !scsi_device_blocked(sdev))
1927 blk_delay_queue(q, SCSI_QUEUE_DELAY);
1928}
1929
1930static inline blk_status_t prep_to_mq(int ret)
1931{
1932 switch (ret) {
1933 case BLKPREP_OK:
1934 return BLK_STS_OK;
1935 case BLKPREP_DEFER:
1936 return BLK_STS_RESOURCE;
1937 default:
1938 return BLK_STS_IOERR;
1939 }
1940}
1941
1942/* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1943static unsigned int scsi_mq_sgl_size(struct Scsi_Host *shost)
1944{
1945 return min_t(unsigned int, shost->sg_tablesize, SG_CHUNK_SIZE) *
1946 sizeof(struct scatterlist);
1947}
1948
1949static int scsi_mq_prep_fn(struct request *req)
1950{
1951 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1952 struct scsi_device *sdev = req->q->queuedata;
1953 struct Scsi_Host *shost = sdev->host;
1954 struct scatterlist *sg;
1955
1956 scsi_init_command(sdev, cmd);
1957
1958 req->special = cmd;
1959
1960 cmd->request = req;
1961
1962 cmd->tag = req->tag;
1963 cmd->prot_op = SCSI_PROT_NORMAL;
1964
1965 sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1966 cmd->sdb.table.sgl = sg;
1967
1968 if (scsi_host_get_prot(shost)) {
1969 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1970
1971 cmd->prot_sdb->table.sgl =
1972 (struct scatterlist *)(cmd->prot_sdb + 1);
1973 }
1974
1975 if (blk_bidi_rq(req)) {
1976 struct request *next_rq = req->next_rq;
1977 struct scsi_data_buffer *bidi_sdb = blk_mq_rq_to_pdu(next_rq);
1978
1979 memset(bidi_sdb, 0, sizeof(struct scsi_data_buffer));
1980 bidi_sdb->table.sgl =
1981 (struct scatterlist *)(bidi_sdb + 1);
1982
1983 next_rq->special = bidi_sdb;
1984 }
1985
1986 blk_mq_start_request(req);
1987
1988 return scsi_setup_cmnd(sdev, req);
1989}
1990
1991static void scsi_mq_done(struct scsi_cmnd *cmd)
1992{
1993 trace_scsi_dispatch_cmd_done(cmd);
1994 blk_mq_complete_request(cmd->request);
1995}
1996
1997static void scsi_mq_put_budget(struct blk_mq_hw_ctx *hctx)
1998{
1999 struct request_queue *q = hctx->queue;
2000 struct scsi_device *sdev = q->queuedata;
2001
2002 atomic_dec(&sdev->device_busy);
2003 put_device(&sdev->sdev_gendev);
2004}
2005
2006static bool scsi_mq_get_budget(struct blk_mq_hw_ctx *hctx)
2007{
2008 struct request_queue *q = hctx->queue;
2009 struct scsi_device *sdev = q->queuedata;
2010
2011 if (!get_device(&sdev->sdev_gendev))
2012 goto out;
2013 if (!scsi_dev_queue_ready(q, sdev))
2014 goto out_put_device;
2015
2016 return true;
2017
2018out_put_device:
2019 put_device(&sdev->sdev_gendev);
2020out:
2021 if (atomic_read(&sdev->device_busy) == 0 && !scsi_device_blocked(sdev))
2022 blk_mq_delay_run_hw_queue(hctx, SCSI_QUEUE_DELAY);
2023 return false;
2024}
2025
2026static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
2027 const struct blk_mq_queue_data *bd)
2028{
2029 struct request *req = bd->rq;
2030 struct request_queue *q = req->q;
2031 struct scsi_device *sdev = q->queuedata;
2032 struct Scsi_Host *shost = sdev->host;
2033 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
2034 blk_status_t ret;
2035 int reason;
2036
2037 ret = prep_to_mq(scsi_prep_state_check(sdev, req));
2038 if (ret != BLK_STS_OK)
2039 goto out_put_budget;
2040
2041 ret = BLK_STS_RESOURCE;
2042 if (!scsi_target_queue_ready(shost, sdev))
2043 goto out_put_budget;
2044 if (!scsi_host_queue_ready(q, shost, sdev))
2045 goto out_dec_target_busy;
2046
2047 if (!(req->rq_flags & RQF_DONTPREP)) {
2048 ret = prep_to_mq(scsi_mq_prep_fn(req));
2049 if (ret != BLK_STS_OK)
2050 goto out_dec_host_busy;
2051 req->rq_flags |= RQF_DONTPREP;
2052 } else {
2053 blk_mq_start_request(req);
2054 }
2055
2056 if (sdev->simple_tags)
2057 cmd->flags |= SCMD_TAGGED;
2058 else
2059 cmd->flags &= ~SCMD_TAGGED;
2060
2061 scsi_init_cmd_errh(cmd);
2062 cmd->scsi_done = scsi_mq_done;
2063
2064 reason = scsi_dispatch_cmd(cmd);
2065 if (reason) {
2066 scsi_set_blocked(cmd, reason);
2067 ret = BLK_STS_RESOURCE;
2068 goto out_dec_host_busy;
2069 }
2070
2071 return BLK_STS_OK;
2072
2073out_dec_host_busy:
2074 scsi_dec_host_busy(shost);
2075out_dec_target_busy:
2076 if (scsi_target(sdev)->can_queue > 0)
2077 atomic_dec(&scsi_target(sdev)->target_busy);
2078out_put_budget:
2079 scsi_mq_put_budget(hctx);
2080 switch (ret) {
2081 case BLK_STS_OK:
2082 break;
2083 case BLK_STS_RESOURCE:
2084 if (atomic_read(&sdev->device_busy) ||
2085 scsi_device_blocked(sdev))
2086 ret = BLK_STS_DEV_RESOURCE;
2087 break;
2088 default:
2089 /*
2090 * Make sure to release all allocated ressources when
2091 * we hit an error, as we will never see this command
2092 * again.
2093 */
2094 if (req->rq_flags & RQF_DONTPREP)
2095 scsi_mq_uninit_cmd(cmd);
2096 break;
2097 }
2098 return ret;
2099}
2100
2101static enum blk_eh_timer_return scsi_timeout(struct request *req,
2102 bool reserved)
2103{
2104 if (reserved)
2105 return BLK_EH_RESET_TIMER;
2106 return scsi_times_out(req);
2107}
2108
2109static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2110 unsigned int hctx_idx, unsigned int numa_node)
2111{
2112 struct Scsi_Host *shost = set->driver_data;
2113 const bool unchecked_isa_dma = shost->unchecked_isa_dma;
2114 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2115 struct scatterlist *sg;
2116
2117 if (unchecked_isa_dma)
2118 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
2119 cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
2120 GFP_KERNEL, numa_node);
2121 if (!cmd->sense_buffer)
2122 return -ENOMEM;
2123 cmd->req.sense = cmd->sense_buffer;
2124
2125 if (scsi_host_get_prot(shost)) {
2126 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
2127 shost->hostt->cmd_size;
2128 cmd->prot_sdb = (void *)sg + scsi_mq_sgl_size(shost);
2129 }
2130
2131 return 0;
2132}
2133
2134static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
2135 unsigned int hctx_idx)
2136{
2137 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2138
2139 scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
2140 cmd->sense_buffer);
2141}
2142
2143static int scsi_map_queues(struct blk_mq_tag_set *set)
2144{
2145 struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
2146
2147 if (shost->hostt->map_queues)
2148 return shost->hostt->map_queues(shost);
2149 return blk_mq_map_queues(set);
2150}
2151
2152static u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
2153{
2154 struct device *host_dev;
2155 u64 bounce_limit = 0xffffffff;
2156
2157 if (shost->unchecked_isa_dma)
2158 return BLK_BOUNCE_ISA;
2159 /*
2160 * Platforms with virtual-DMA translation
2161 * hardware have no practical limit.
2162 */
2163 if (!PCI_DMA_BUS_IS_PHYS)
2164 return BLK_BOUNCE_ANY;
2165
2166 host_dev = scsi_get_device(shost);
2167 if (host_dev && host_dev->dma_mask)
2168 bounce_limit = (u64)dma_max_pfn(host_dev) << PAGE_SHIFT;
2169
2170 return bounce_limit;
2171}
2172
2173void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
2174{
2175 struct device *dev = shost->dma_dev;
2176
2177 /*
2178 * this limit is imposed by hardware restrictions
2179 */
2180 blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
2181 SG_MAX_SEGMENTS));
2182
2183 if (scsi_host_prot_dma(shost)) {
2184 shost->sg_prot_tablesize =
2185 min_not_zero(shost->sg_prot_tablesize,
2186 (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
2187 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
2188 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
2189 }
2190
2191 blk_queue_max_hw_sectors(q, shost->max_sectors);
2192 blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
2193 blk_queue_segment_boundary(q, shost->dma_boundary);
2194 dma_set_seg_boundary(dev, shost->dma_boundary);
2195
2196 blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
2197
2198 if (!shost->use_clustering)
2199 q->limits.cluster = 0;
2200
2201 /*
2202 * Set a reasonable default alignment: The larger of 32-byte (dword),
2203 * which is a common minimum for HBAs, and the minimum DMA alignment,
2204 * which is set by the platform.
2205 *
2206 * Devices that require a bigger alignment can increase it later.
2207 */
2208 blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
2209}
2210EXPORT_SYMBOL_GPL(__scsi_init_queue);
2211
2212static int scsi_old_init_rq(struct request_queue *q, struct request *rq,
2213 gfp_t gfp)
2214{
2215 struct Scsi_Host *shost = q->rq_alloc_data;
2216 const bool unchecked_isa_dma = shost->unchecked_isa_dma;
2217 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2218
2219 memset(cmd, 0, sizeof(*cmd));
2220
2221 if (unchecked_isa_dma)
2222 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
2223 cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma, gfp,
2224 NUMA_NO_NODE);
2225 if (!cmd->sense_buffer)
2226 goto fail;
2227 cmd->req.sense = cmd->sense_buffer;
2228
2229 if (scsi_host_get_prot(shost) >= SHOST_DIX_TYPE0_PROTECTION) {
2230 cmd->prot_sdb = kmem_cache_zalloc(scsi_sdb_cache, gfp);
2231 if (!cmd->prot_sdb)
2232 goto fail_free_sense;
2233 }
2234
2235 return 0;
2236
2237fail_free_sense:
2238 scsi_free_sense_buffer(unchecked_isa_dma, cmd->sense_buffer);
2239fail:
2240 return -ENOMEM;
2241}
2242
2243static void scsi_old_exit_rq(struct request_queue *q, struct request *rq)
2244{
2245 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2246
2247 if (cmd->prot_sdb)
2248 kmem_cache_free(scsi_sdb_cache, cmd->prot_sdb);
2249 scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
2250 cmd->sense_buffer);
2251}
2252
2253struct request_queue *scsi_old_alloc_queue(struct scsi_device *sdev)
2254{
2255 struct Scsi_Host *shost = sdev->host;
2256 struct request_queue *q;
2257
2258 q = blk_alloc_queue_node(GFP_KERNEL, NUMA_NO_NODE, NULL);
2259 if (!q)
2260 return NULL;
2261 q->cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
2262 q->rq_alloc_data = shost;
2263 q->request_fn = scsi_request_fn;
2264 q->init_rq_fn = scsi_old_init_rq;
2265 q->exit_rq_fn = scsi_old_exit_rq;
2266 q->initialize_rq_fn = scsi_initialize_rq;
2267
2268 if (blk_init_allocated_queue(q) < 0) {
2269 blk_cleanup_queue(q);
2270 return NULL;
2271 }
2272
2273 __scsi_init_queue(shost, q);
2274 blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, q);
2275 blk_queue_prep_rq(q, scsi_prep_fn);
2276 blk_queue_unprep_rq(q, scsi_unprep_fn);
2277 blk_queue_softirq_done(q, scsi_softirq_done);
2278 blk_queue_rq_timed_out(q, scsi_times_out);
2279 blk_queue_lld_busy(q, scsi_lld_busy);
2280 return q;
2281}
2282
2283static const struct blk_mq_ops scsi_mq_ops = {
2284 .get_budget = scsi_mq_get_budget,
2285 .put_budget = scsi_mq_put_budget,
2286 .queue_rq = scsi_queue_rq,
2287 .complete = scsi_softirq_done,
2288 .timeout = scsi_timeout,
2289#ifdef CONFIG_BLK_DEBUG_FS
2290 .show_rq = scsi_show_rq,
2291#endif
2292 .init_request = scsi_mq_init_request,
2293 .exit_request = scsi_mq_exit_request,
2294 .initialize_rq_fn = scsi_initialize_rq,
2295 .map_queues = scsi_map_queues,
2296};
2297
2298struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
2299{
2300 sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
2301 if (IS_ERR(sdev->request_queue))
2302 return NULL;
2303
2304 sdev->request_queue->queuedata = sdev;
2305 __scsi_init_queue(sdev->host, sdev->request_queue);
2306 blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, sdev->request_queue);
2307 return sdev->request_queue;
2308}
2309
2310int scsi_mq_setup_tags(struct Scsi_Host *shost)
2311{
2312 unsigned int cmd_size, sgl_size;
2313
2314 sgl_size = scsi_mq_sgl_size(shost);
2315 cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
2316 if (scsi_host_get_prot(shost))
2317 cmd_size += sizeof(struct scsi_data_buffer) + sgl_size;
2318
2319 memset(&shost->tag_set, 0, sizeof(shost->tag_set));
2320 shost->tag_set.ops = &scsi_mq_ops;
2321 shost->tag_set.nr_hw_queues = shost->nr_hw_queues ? : 1;
2322 shost->tag_set.queue_depth = shost->can_queue;
2323 shost->tag_set.cmd_size = cmd_size;
2324 shost->tag_set.numa_node = NUMA_NO_NODE;
2325 shost->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
2326 shost->tag_set.flags |=
2327 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
2328 shost->tag_set.driver_data = shost;
2329
2330 return blk_mq_alloc_tag_set(&shost->tag_set);
2331}
2332
2333void scsi_mq_destroy_tags(struct Scsi_Host *shost)
2334{
2335 blk_mq_free_tag_set(&shost->tag_set);
2336}
2337
2338/**
2339 * scsi_device_from_queue - return sdev associated with a request_queue
2340 * @q: The request queue to return the sdev from
2341 *
2342 * Return the sdev associated with a request queue or NULL if the
2343 * request_queue does not reference a SCSI device.
2344 */
2345struct scsi_device *scsi_device_from_queue(struct request_queue *q)
2346{
2347 struct scsi_device *sdev = NULL;
2348
2349 if (q->mq_ops) {
2350 if (q->mq_ops == &scsi_mq_ops)
2351 sdev = q->queuedata;
2352 } else if (q->request_fn == scsi_request_fn)
2353 sdev = q->queuedata;
2354 if (!sdev || !get_device(&sdev->sdev_gendev))
2355 sdev = NULL;
2356
2357 return sdev;
2358}
2359EXPORT_SYMBOL_GPL(scsi_device_from_queue);
2360
2361/*
2362 * Function: scsi_block_requests()
2363 *
2364 * Purpose: Utility function used by low-level drivers to prevent further
2365 * commands from being queued to the device.
2366 *
2367 * Arguments: shost - Host in question
2368 *
2369 * Returns: Nothing
2370 *
2371 * Lock status: No locks are assumed held.
2372 *
2373 * Notes: There is no timer nor any other means by which the requests
2374 * get unblocked other than the low-level driver calling
2375 * scsi_unblock_requests().
2376 */
2377void scsi_block_requests(struct Scsi_Host *shost)
2378{
2379 shost->host_self_blocked = 1;
2380}
2381EXPORT_SYMBOL(scsi_block_requests);
2382
2383/*
2384 * Function: scsi_unblock_requests()
2385 *
2386 * Purpose: Utility function used by low-level drivers to allow further
2387 * commands from being queued to the device.
2388 *
2389 * Arguments: shost - Host in question
2390 *
2391 * Returns: Nothing
2392 *
2393 * Lock status: No locks are assumed held.
2394 *
2395 * Notes: There is no timer nor any other means by which the requests
2396 * get unblocked other than the low-level driver calling
2397 * scsi_unblock_requests().
2398 *
2399 * This is done as an API function so that changes to the
2400 * internals of the scsi mid-layer won't require wholesale
2401 * changes to drivers that use this feature.
2402 */
2403void scsi_unblock_requests(struct Scsi_Host *shost)
2404{
2405 shost->host_self_blocked = 0;
2406 scsi_run_host_queues(shost);
2407}
2408EXPORT_SYMBOL(scsi_unblock_requests);
2409
2410int __init scsi_init_queue(void)
2411{
2412 scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
2413 sizeof(struct scsi_data_buffer),
2414 0, 0, NULL);
2415 if (!scsi_sdb_cache) {
2416 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
2417 return -ENOMEM;
2418 }
2419
2420 return 0;
2421}
2422
2423void scsi_exit_queue(void)
2424{
2425 kmem_cache_destroy(scsi_sense_cache);
2426 kmem_cache_destroy(scsi_sense_isadma_cache);
2427 kmem_cache_destroy(scsi_sdb_cache);
2428}
2429
2430/**
2431 * scsi_mode_select - issue a mode select
2432 * @sdev: SCSI device to be queried
2433 * @pf: Page format bit (1 == standard, 0 == vendor specific)
2434 * @sp: Save page bit (0 == don't save, 1 == save)
2435 * @modepage: mode page being requested
2436 * @buffer: request buffer (may not be smaller than eight bytes)
2437 * @len: length of request buffer.
2438 * @timeout: command timeout
2439 * @retries: number of retries before failing
2440 * @data: returns a structure abstracting the mode header data
2441 * @sshdr: place to put sense data (or NULL if no sense to be collected).
2442 * must be SCSI_SENSE_BUFFERSIZE big.
2443 *
2444 * Returns zero if successful; negative error number or scsi
2445 * status on error
2446 *
2447 */
2448int
2449scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
2450 unsigned char *buffer, int len, int timeout, int retries,
2451 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2452{
2453 unsigned char cmd[10];
2454 unsigned char *real_buffer;
2455 int ret;
2456
2457 memset(cmd, 0, sizeof(cmd));
2458 cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2459
2460 if (sdev->use_10_for_ms) {
2461 if (len > 65535)
2462 return -EINVAL;
2463 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2464 if (!real_buffer)
2465 return -ENOMEM;
2466 memcpy(real_buffer + 8, buffer, len);
2467 len += 8;
2468 real_buffer[0] = 0;
2469 real_buffer[1] = 0;
2470 real_buffer[2] = data->medium_type;
2471 real_buffer[3] = data->device_specific;
2472 real_buffer[4] = data->longlba ? 0x01 : 0;
2473 real_buffer[5] = 0;
2474 real_buffer[6] = data->block_descriptor_length >> 8;
2475 real_buffer[7] = data->block_descriptor_length;
2476
2477 cmd[0] = MODE_SELECT_10;
2478 cmd[7] = len >> 8;
2479 cmd[8] = len;
2480 } else {
2481 if (len > 255 || data->block_descriptor_length > 255 ||
2482 data->longlba)
2483 return -EINVAL;
2484
2485 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2486 if (!real_buffer)
2487 return -ENOMEM;
2488 memcpy(real_buffer + 4, buffer, len);
2489 len += 4;
2490 real_buffer[0] = 0;
2491 real_buffer[1] = data->medium_type;
2492 real_buffer[2] = data->device_specific;
2493 real_buffer[3] = data->block_descriptor_length;
2494
2495
2496 cmd[0] = MODE_SELECT;
2497 cmd[4] = len;
2498 }
2499
2500 ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2501 sshdr, timeout, retries, NULL);
2502 kfree(real_buffer);
2503 return ret;
2504}
2505EXPORT_SYMBOL_GPL(scsi_mode_select);
2506
2507/**
2508 * scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2509 * @sdev: SCSI device to be queried
2510 * @dbd: set if mode sense will allow block descriptors to be returned
2511 * @modepage: mode page being requested
2512 * @buffer: request buffer (may not be smaller than eight bytes)
2513 * @len: length of request buffer.
2514 * @timeout: command timeout
2515 * @retries: number of retries before failing
2516 * @data: returns a structure abstracting the mode header data
2517 * @sshdr: place to put sense data (or NULL if no sense to be collected).
2518 * must be SCSI_SENSE_BUFFERSIZE big.
2519 *
2520 * Returns zero if unsuccessful, or the header offset (either 4
2521 * or 8 depending on whether a six or ten byte command was
2522 * issued) if successful.
2523 */
2524int
2525scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2526 unsigned char *buffer, int len, int timeout, int retries,
2527 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2528{
2529 unsigned char cmd[12];
2530 int use_10_for_ms;
2531 int header_length;
2532 int result, retry_count = retries;
2533 struct scsi_sense_hdr my_sshdr;
2534
2535 memset(data, 0, sizeof(*data));
2536 memset(&cmd[0], 0, 12);
2537 cmd[1] = dbd & 0x18; /* allows DBD and LLBA bits */
2538 cmd[2] = modepage;
2539
2540 /* caller might not be interested in sense, but we need it */
2541 if (!sshdr)
2542 sshdr = &my_sshdr;
2543
2544 retry:
2545 use_10_for_ms = sdev->use_10_for_ms;
2546
2547 if (use_10_for_ms) {
2548 if (len < 8)
2549 len = 8;
2550
2551 cmd[0] = MODE_SENSE_10;
2552 cmd[8] = len;
2553 header_length = 8;
2554 } else {
2555 if (len < 4)
2556 len = 4;
2557
2558 cmd[0] = MODE_SENSE;
2559 cmd[4] = len;
2560 header_length = 4;
2561 }
2562
2563 memset(buffer, 0, len);
2564
2565 result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2566 sshdr, timeout, retries, NULL);
2567
2568 /* This code looks awful: what it's doing is making sure an
2569 * ILLEGAL REQUEST sense return identifies the actual command
2570 * byte as the problem. MODE_SENSE commands can return
2571 * ILLEGAL REQUEST if the code page isn't supported */
2572
2573 if (use_10_for_ms && !scsi_status_is_good(result) &&
2574 (driver_byte(result) & DRIVER_SENSE)) {
2575 if (scsi_sense_valid(sshdr)) {
2576 if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2577 (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2578 /*
2579 * Invalid command operation code
2580 */
2581 sdev->use_10_for_ms = 0;
2582 goto retry;
2583 }
2584 }
2585 }
2586
2587 if(scsi_status_is_good(result)) {
2588 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2589 (modepage == 6 || modepage == 8))) {
2590 /* Initio breakage? */
2591 header_length = 0;
2592 data->length = 13;
2593 data->medium_type = 0;
2594 data->device_specific = 0;
2595 data->longlba = 0;
2596 data->block_descriptor_length = 0;
2597 } else if(use_10_for_ms) {
2598 data->length = buffer[0]*256 + buffer[1] + 2;
2599 data->medium_type = buffer[2];
2600 data->device_specific = buffer[3];
2601 data->longlba = buffer[4] & 0x01;
2602 data->block_descriptor_length = buffer[6]*256
2603 + buffer[7];
2604 } else {
2605 data->length = buffer[0] + 1;
2606 data->medium_type = buffer[1];
2607 data->device_specific = buffer[2];
2608 data->block_descriptor_length = buffer[3];
2609 }
2610 data->header_length = header_length;
2611 } else if ((status_byte(result) == CHECK_CONDITION) &&
2612 scsi_sense_valid(sshdr) &&
2613 sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2614 retry_count--;
2615 goto retry;
2616 }
2617
2618 return result;
2619}
2620EXPORT_SYMBOL(scsi_mode_sense);
2621
2622/**
2623 * scsi_test_unit_ready - test if unit is ready
2624 * @sdev: scsi device to change the state of.
2625 * @timeout: command timeout
2626 * @retries: number of retries before failing
2627 * @sshdr: outpout pointer for decoded sense information.
2628 *
2629 * Returns zero if unsuccessful or an error if TUR failed. For
2630 * removable media, UNIT_ATTENTION sets ->changed flag.
2631 **/
2632int
2633scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2634 struct scsi_sense_hdr *sshdr)
2635{
2636 char cmd[] = {
2637 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2638 };
2639 int result;
2640
2641 /* try to eat the UNIT_ATTENTION if there are enough retries */
2642 do {
2643 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2644 timeout, 1, NULL);
2645 if (sdev->removable && scsi_sense_valid(sshdr) &&
2646 sshdr->sense_key == UNIT_ATTENTION)
2647 sdev->changed = 1;
2648 } while (scsi_sense_valid(sshdr) &&
2649 sshdr->sense_key == UNIT_ATTENTION && --retries);
2650
2651 return result;
2652}
2653EXPORT_SYMBOL(scsi_test_unit_ready);
2654
2655/**
2656 * scsi_device_set_state - Take the given device through the device state model.
2657 * @sdev: scsi device to change the state of.
2658 * @state: state to change to.
2659 *
2660 * Returns zero if successful or an error if the requested
2661 * transition is illegal.
2662 */
2663int
2664scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2665{
2666 enum scsi_device_state oldstate = sdev->sdev_state;
2667
2668 if (state == oldstate)
2669 return 0;
2670
2671 switch (state) {
2672 case SDEV_CREATED:
2673 switch (oldstate) {
2674 case SDEV_CREATED_BLOCK:
2675 break;
2676 default:
2677 goto illegal;
2678 }
2679 break;
2680
2681 case SDEV_RUNNING:
2682 switch (oldstate) {
2683 case SDEV_CREATED:
2684 case SDEV_OFFLINE:
2685 case SDEV_TRANSPORT_OFFLINE:
2686 case SDEV_QUIESCE:
2687 case SDEV_BLOCK:
2688 break;
2689 default:
2690 goto illegal;
2691 }
2692 break;
2693
2694 case SDEV_QUIESCE:
2695 switch (oldstate) {
2696 case SDEV_RUNNING:
2697 case SDEV_OFFLINE:
2698 case SDEV_TRANSPORT_OFFLINE:
2699 break;
2700 default:
2701 goto illegal;
2702 }
2703 break;
2704
2705 case SDEV_OFFLINE:
2706 case SDEV_TRANSPORT_OFFLINE:
2707 switch (oldstate) {
2708 case SDEV_CREATED:
2709 case SDEV_RUNNING:
2710 case SDEV_QUIESCE:
2711 case SDEV_BLOCK:
2712 break;
2713 default:
2714 goto illegal;
2715 }
2716 break;
2717
2718 case SDEV_BLOCK:
2719 switch (oldstate) {
2720 case SDEV_RUNNING:
2721 case SDEV_CREATED_BLOCK:
2722 break;
2723 default:
2724 goto illegal;
2725 }
2726 break;
2727
2728 case SDEV_CREATED_BLOCK:
2729 switch (oldstate) {
2730 case SDEV_CREATED:
2731 break;
2732 default:
2733 goto illegal;
2734 }
2735 break;
2736
2737 case SDEV_CANCEL:
2738 switch (oldstate) {
2739 case SDEV_CREATED:
2740 case SDEV_RUNNING:
2741 case SDEV_QUIESCE:
2742 case SDEV_OFFLINE:
2743 case SDEV_TRANSPORT_OFFLINE:
2744 break;
2745 default:
2746 goto illegal;
2747 }
2748 break;
2749
2750 case SDEV_DEL:
2751 switch (oldstate) {
2752 case SDEV_CREATED:
2753 case SDEV_RUNNING:
2754 case SDEV_OFFLINE:
2755 case SDEV_TRANSPORT_OFFLINE:
2756 case SDEV_CANCEL:
2757 case SDEV_BLOCK:
2758 case SDEV_CREATED_BLOCK:
2759 break;
2760 default:
2761 goto illegal;
2762 }
2763 break;
2764
2765 }
2766 sdev->sdev_state = state;
2767 return 0;
2768
2769 illegal:
2770 SCSI_LOG_ERROR_RECOVERY(1,
2771 sdev_printk(KERN_ERR, sdev,
2772 "Illegal state transition %s->%s",
2773 scsi_device_state_name(oldstate),
2774 scsi_device_state_name(state))
2775 );
2776 return -EINVAL;
2777}
2778EXPORT_SYMBOL(scsi_device_set_state);
2779
2780/**
2781 * sdev_evt_emit - emit a single SCSI device uevent
2782 * @sdev: associated SCSI device
2783 * @evt: event to emit
2784 *
2785 * Send a single uevent (scsi_event) to the associated scsi_device.
2786 */
2787static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2788{
2789 int idx = 0;
2790 char *envp[3];
2791
2792 switch (evt->evt_type) {
2793 case SDEV_EVT_MEDIA_CHANGE:
2794 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2795 break;
2796 case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2797 scsi_rescan_device(&sdev->sdev_gendev);
2798 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2799 break;
2800 case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2801 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2802 break;
2803 case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2804 envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2805 break;
2806 case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2807 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2808 break;
2809 case SDEV_EVT_LUN_CHANGE_REPORTED:
2810 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2811 break;
2812 case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2813 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2814 break;
2815 case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2816 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2817 break;
2818 default:
2819 /* do nothing */
2820 break;
2821 }
2822
2823 envp[idx++] = NULL;
2824
2825 kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2826}
2827
2828/**
2829 * sdev_evt_thread - send a uevent for each scsi event
2830 * @work: work struct for scsi_device
2831 *
2832 * Dispatch queued events to their associated scsi_device kobjects
2833 * as uevents.
2834 */
2835void scsi_evt_thread(struct work_struct *work)
2836{
2837 struct scsi_device *sdev;
2838 enum scsi_device_event evt_type;
2839 LIST_HEAD(event_list);
2840
2841 sdev = container_of(work, struct scsi_device, event_work);
2842
2843 for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2844 if (test_and_clear_bit(evt_type, sdev->pending_events))
2845 sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2846
2847 while (1) {
2848 struct scsi_event *evt;
2849 struct list_head *this, *tmp;
2850 unsigned long flags;
2851
2852 spin_lock_irqsave(&sdev->list_lock, flags);
2853 list_splice_init(&sdev->event_list, &event_list);
2854 spin_unlock_irqrestore(&sdev->list_lock, flags);
2855
2856 if (list_empty(&event_list))
2857 break;
2858
2859 list_for_each_safe(this, tmp, &event_list) {
2860 evt = list_entry(this, struct scsi_event, node);
2861 list_del(&evt->node);
2862 scsi_evt_emit(sdev, evt);
2863 kfree(evt);
2864 }
2865 }
2866}
2867
2868/**
2869 * sdev_evt_send - send asserted event to uevent thread
2870 * @sdev: scsi_device event occurred on
2871 * @evt: event to send
2872 *
2873 * Assert scsi device event asynchronously.
2874 */
2875void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2876{
2877 unsigned long flags;
2878
2879#if 0
2880 /* FIXME: currently this check eliminates all media change events
2881 * for polled devices. Need to update to discriminate between AN
2882 * and polled events */
2883 if (!test_bit(evt->evt_type, sdev->supported_events)) {
2884 kfree(evt);
2885 return;
2886 }
2887#endif
2888
2889 spin_lock_irqsave(&sdev->list_lock, flags);
2890 list_add_tail(&evt->node, &sdev->event_list);
2891 schedule_work(&sdev->event_work);
2892 spin_unlock_irqrestore(&sdev->list_lock, flags);
2893}
2894EXPORT_SYMBOL_GPL(sdev_evt_send);
2895
2896/**
2897 * sdev_evt_alloc - allocate a new scsi event
2898 * @evt_type: type of event to allocate
2899 * @gfpflags: GFP flags for allocation
2900 *
2901 * Allocates and returns a new scsi_event.
2902 */
2903struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2904 gfp_t gfpflags)
2905{
2906 struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2907 if (!evt)
2908 return NULL;
2909
2910 evt->evt_type = evt_type;
2911 INIT_LIST_HEAD(&evt->node);
2912
2913 /* evt_type-specific initialization, if any */
2914 switch (evt_type) {
2915 case SDEV_EVT_MEDIA_CHANGE:
2916 case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2917 case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2918 case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2919 case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2920 case SDEV_EVT_LUN_CHANGE_REPORTED:
2921 case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2922 case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2923 default:
2924 /* do nothing */
2925 break;
2926 }
2927
2928 return evt;
2929}
2930EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2931
2932/**
2933 * sdev_evt_send_simple - send asserted event to uevent thread
2934 * @sdev: scsi_device event occurred on
2935 * @evt_type: type of event to send
2936 * @gfpflags: GFP flags for allocation
2937 *
2938 * Assert scsi device event asynchronously, given an event type.
2939 */
2940void sdev_evt_send_simple(struct scsi_device *sdev,
2941 enum scsi_device_event evt_type, gfp_t gfpflags)
2942{
2943 struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2944 if (!evt) {
2945 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2946 evt_type);
2947 return;
2948 }
2949
2950 sdev_evt_send(sdev, evt);
2951}
2952EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2953
2954/**
2955 * scsi_request_fn_active() - number of kernel threads inside scsi_request_fn()
2956 * @sdev: SCSI device to count the number of scsi_request_fn() callers for.
2957 */
2958static int scsi_request_fn_active(struct scsi_device *sdev)
2959{
2960 struct request_queue *q = sdev->request_queue;
2961 int request_fn_active;
2962
2963 WARN_ON_ONCE(sdev->host->use_blk_mq);
2964
2965 spin_lock_irq(q->queue_lock);
2966 request_fn_active = q->request_fn_active;
2967 spin_unlock_irq(q->queue_lock);
2968
2969 return request_fn_active;
2970}
2971
2972/**
2973 * scsi_wait_for_queuecommand() - wait for ongoing queuecommand() calls
2974 * @sdev: SCSI device pointer.
2975 *
2976 * Wait until the ongoing shost->hostt->queuecommand() calls that are
2977 * invoked from scsi_request_fn() have finished.
2978 */
2979static void scsi_wait_for_queuecommand(struct scsi_device *sdev)
2980{
2981 WARN_ON_ONCE(sdev->host->use_blk_mq);
2982
2983 while (scsi_request_fn_active(sdev))
2984 msleep(20);
2985}
2986
2987/**
2988 * scsi_device_quiesce - Block user issued commands.
2989 * @sdev: scsi device to quiesce.
2990 *
2991 * This works by trying to transition to the SDEV_QUIESCE state
2992 * (which must be a legal transition). When the device is in this
2993 * state, only special requests will be accepted, all others will
2994 * be deferred. Since special requests may also be requeued requests,
2995 * a successful return doesn't guarantee the device will be
2996 * totally quiescent.
2997 *
2998 * Must be called with user context, may sleep.
2999 *
3000 * Returns zero if unsuccessful or an error if not.
3001 */
3002int
3003scsi_device_quiesce(struct scsi_device *sdev)
3004{
3005 struct request_queue *q = sdev->request_queue;
3006 int err;
3007
3008 /*
3009 * It is allowed to call scsi_device_quiesce() multiple times from
3010 * the same context but concurrent scsi_device_quiesce() calls are
3011 * not allowed.
3012 */
3013 WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
3014
3015 blk_set_preempt_only(q);
3016
3017 blk_mq_freeze_queue(q);
3018 /*
3019 * Ensure that the effect of blk_set_preempt_only() will be visible
3020 * for percpu_ref_tryget() callers that occur after the queue
3021 * unfreeze even if the queue was already frozen before this function
3022 * was called. See also https://lwn.net/Articles/573497/.
3023 */
3024 synchronize_rcu();
3025 blk_mq_unfreeze_queue(q);
3026
3027 mutex_lock(&sdev->state_mutex);
3028 err = scsi_device_set_state(sdev, SDEV_QUIESCE);
3029 if (err == 0)
3030 sdev->quiesced_by = current;
3031 else
3032 blk_clear_preempt_only(q);
3033 mutex_unlock(&sdev->state_mutex);
3034
3035 return err;
3036}
3037EXPORT_SYMBOL(scsi_device_quiesce);
3038
3039/**
3040 * scsi_device_resume - Restart user issued commands to a quiesced device.
3041 * @sdev: scsi device to resume.
3042 *
3043 * Moves the device from quiesced back to running and restarts the
3044 * queues.
3045 *
3046 * Must be called with user context, may sleep.
3047 */
3048void scsi_device_resume(struct scsi_device *sdev)
3049{
3050 /* check if the device state was mutated prior to resume, and if
3051 * so assume the state is being managed elsewhere (for example
3052 * device deleted during suspend)
3053 */
3054 mutex_lock(&sdev->state_mutex);
3055 WARN_ON_ONCE(!sdev->quiesced_by);
3056 sdev->quiesced_by = NULL;
3057 blk_clear_preempt_only(sdev->request_queue);
3058 if (sdev->sdev_state == SDEV_QUIESCE)
3059 scsi_device_set_state(sdev, SDEV_RUNNING);
3060 mutex_unlock(&sdev->state_mutex);
3061}
3062EXPORT_SYMBOL(scsi_device_resume);
3063
3064static void
3065device_quiesce_fn(struct scsi_device *sdev, void *data)
3066{
3067 scsi_device_quiesce(sdev);
3068}
3069
3070void
3071scsi_target_quiesce(struct scsi_target *starget)
3072{
3073 starget_for_each_device(starget, NULL, device_quiesce_fn);
3074}
3075EXPORT_SYMBOL(scsi_target_quiesce);
3076
3077static void
3078device_resume_fn(struct scsi_device *sdev, void *data)
3079{
3080 scsi_device_resume(sdev);
3081}
3082
3083void
3084scsi_target_resume(struct scsi_target *starget)
3085{
3086 starget_for_each_device(starget, NULL, device_resume_fn);
3087}
3088EXPORT_SYMBOL(scsi_target_resume);
3089
3090/**
3091 * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
3092 * @sdev: device to block
3093 *
3094 * Pause SCSI command processing on the specified device. Does not sleep.
3095 *
3096 * Returns zero if successful or a negative error code upon failure.
3097 *
3098 * Notes:
3099 * This routine transitions the device to the SDEV_BLOCK state (which must be
3100 * a legal transition). When the device is in this state, command processing
3101 * is paused until the device leaves the SDEV_BLOCK state. See also
3102 * scsi_internal_device_unblock_nowait().
3103 */
3104int scsi_internal_device_block_nowait(struct scsi_device *sdev)
3105{
3106 struct request_queue *q = sdev->request_queue;
3107 unsigned long flags;
3108 int err = 0;
3109
3110 err = scsi_device_set_state(sdev, SDEV_BLOCK);
3111 if (err) {
3112 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
3113
3114 if (err)
3115 return err;
3116 }
3117
3118 /*
3119 * The device has transitioned to SDEV_BLOCK. Stop the
3120 * block layer from calling the midlayer with this device's
3121 * request queue.
3122 */
3123 if (q->mq_ops) {
3124 blk_mq_quiesce_queue_nowait(q);
3125 } else {
3126 spin_lock_irqsave(q->queue_lock, flags);
3127 blk_stop_queue(q);
3128 spin_unlock_irqrestore(q->queue_lock, flags);
3129 }
3130
3131 return 0;
3132}
3133EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
3134
3135/**
3136 * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
3137 * @sdev: device to block
3138 *
3139 * Pause SCSI command processing on the specified device and wait until all
3140 * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
3141 *
3142 * Returns zero if successful or a negative error code upon failure.
3143 *
3144 * Note:
3145 * This routine transitions the device to the SDEV_BLOCK state (which must be
3146 * a legal transition). When the device is in this state, command processing
3147 * is paused until the device leaves the SDEV_BLOCK state. See also
3148 * scsi_internal_device_unblock().
3149 *
3150 * To do: avoid that scsi_send_eh_cmnd() calls queuecommand() after
3151 * scsi_internal_device_block() has blocked a SCSI device and also
3152 * remove the rport mutex lock and unlock calls from srp_queuecommand().
3153 */
3154static int scsi_internal_device_block(struct scsi_device *sdev)
3155{
3156 struct request_queue *q = sdev->request_queue;
3157 int err;
3158
3159 mutex_lock(&sdev->state_mutex);
3160 err = scsi_internal_device_block_nowait(sdev);
3161 if (err == 0) {
3162 if (q->mq_ops)
3163 blk_mq_quiesce_queue(q);
3164 else
3165 scsi_wait_for_queuecommand(sdev);
3166 }
3167 mutex_unlock(&sdev->state_mutex);
3168
3169 return err;
3170}
3171
3172void scsi_start_queue(struct scsi_device *sdev)
3173{
3174 struct request_queue *q = sdev->request_queue;
3175 unsigned long flags;
3176
3177 if (q->mq_ops) {
3178 blk_mq_unquiesce_queue(q);
3179 } else {
3180 spin_lock_irqsave(q->queue_lock, flags);
3181 blk_start_queue(q);
3182 spin_unlock_irqrestore(q->queue_lock, flags);
3183 }
3184}
3185
3186/**
3187 * scsi_internal_device_unblock_nowait - resume a device after a block request
3188 * @sdev: device to resume
3189 * @new_state: state to set the device to after unblocking
3190 *
3191 * Restart the device queue for a previously suspended SCSI device. Does not
3192 * sleep.
3193 *
3194 * Returns zero if successful or a negative error code upon failure.
3195 *
3196 * Notes:
3197 * This routine transitions the device to the SDEV_RUNNING state or to one of
3198 * the offline states (which must be a legal transition) allowing the midlayer
3199 * to goose the queue for this device.
3200 */
3201int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
3202 enum scsi_device_state new_state)
3203{
3204 /*
3205 * Try to transition the scsi device to SDEV_RUNNING or one of the
3206 * offlined states and goose the device queue if successful.
3207 */
3208 switch (sdev->sdev_state) {
3209 case SDEV_BLOCK:
3210 case SDEV_TRANSPORT_OFFLINE:
3211 sdev->sdev_state = new_state;
3212 break;
3213 case SDEV_CREATED_BLOCK:
3214 if (new_state == SDEV_TRANSPORT_OFFLINE ||
3215 new_state == SDEV_OFFLINE)
3216 sdev->sdev_state = new_state;
3217 else
3218 sdev->sdev_state = SDEV_CREATED;
3219 break;
3220 case SDEV_CANCEL:
3221 case SDEV_OFFLINE:
3222 break;
3223 default:
3224 return -EINVAL;
3225 }
3226 scsi_start_queue(sdev);
3227
3228 return 0;
3229}
3230EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
3231
3232/**
3233 * scsi_internal_device_unblock - resume a device after a block request
3234 * @sdev: device to resume
3235 * @new_state: state to set the device to after unblocking
3236 *
3237 * Restart the device queue for a previously suspended SCSI device. May sleep.
3238 *
3239 * Returns zero if successful or a negative error code upon failure.
3240 *
3241 * Notes:
3242 * This routine transitions the device to the SDEV_RUNNING state or to one of
3243 * the offline states (which must be a legal transition) allowing the midlayer
3244 * to goose the queue for this device.
3245 */
3246static int scsi_internal_device_unblock(struct scsi_device *sdev,
3247 enum scsi_device_state new_state)
3248{
3249 int ret;
3250
3251 mutex_lock(&sdev->state_mutex);
3252 ret = scsi_internal_device_unblock_nowait(sdev, new_state);
3253 mutex_unlock(&sdev->state_mutex);
3254
3255 return ret;
3256}
3257
3258static void
3259device_block(struct scsi_device *sdev, void *data)
3260{
3261 scsi_internal_device_block(sdev);
3262}
3263
3264static int
3265target_block(struct device *dev, void *data)
3266{
3267 if (scsi_is_target_device(dev))
3268 starget_for_each_device(to_scsi_target(dev), NULL,
3269 device_block);
3270 return 0;
3271}
3272
3273void
3274scsi_target_block(struct device *dev)
3275{
3276 if (scsi_is_target_device(dev))
3277 starget_for_each_device(to_scsi_target(dev), NULL,
3278 device_block);
3279 else
3280 device_for_each_child(dev, NULL, target_block);
3281}
3282EXPORT_SYMBOL_GPL(scsi_target_block);
3283
3284static void
3285device_unblock(struct scsi_device *sdev, void *data)
3286{
3287 scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
3288}
3289
3290static int
3291target_unblock(struct device *dev, void *data)
3292{
3293 if (scsi_is_target_device(dev))
3294 starget_for_each_device(to_scsi_target(dev), data,
3295 device_unblock);
3296 return 0;
3297}
3298
3299void
3300scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
3301{
3302 if (scsi_is_target_device(dev))
3303 starget_for_each_device(to_scsi_target(dev), &new_state,
3304 device_unblock);
3305 else
3306 device_for_each_child(dev, &new_state, target_unblock);
3307}
3308EXPORT_SYMBOL_GPL(scsi_target_unblock);
3309
3310/**
3311 * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
3312 * @sgl: scatter-gather list
3313 * @sg_count: number of segments in sg
3314 * @offset: offset in bytes into sg, on return offset into the mapped area
3315 * @len: bytes to map, on return number of bytes mapped
3316 *
3317 * Returns virtual address of the start of the mapped page
3318 */
3319void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
3320 size_t *offset, size_t *len)
3321{
3322 int i;
3323 size_t sg_len = 0, len_complete = 0;
3324 struct scatterlist *sg;
3325 struct page *page;
3326
3327 WARN_ON(!irqs_disabled());
3328
3329 for_each_sg(sgl, sg, sg_count, i) {
3330 len_complete = sg_len; /* Complete sg-entries */
3331 sg_len += sg->length;
3332 if (sg_len > *offset)
3333 break;
3334 }
3335
3336 if (unlikely(i == sg_count)) {
3337 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
3338 "elements %d\n",
3339 __func__, sg_len, *offset, sg_count);
3340 WARN_ON(1);
3341 return NULL;
3342 }
3343
3344 /* Offset starting from the beginning of first page in this sg-entry */
3345 *offset = *offset - len_complete + sg->offset;
3346
3347 /* Assumption: contiguous pages can be accessed as "page + i" */
3348 page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
3349 *offset &= ~PAGE_MASK;
3350
3351 /* Bytes in this sg-entry from *offset to the end of the page */
3352 sg_len = PAGE_SIZE - *offset;
3353 if (*len > sg_len)
3354 *len = sg_len;
3355
3356 return kmap_atomic(page);
3357}
3358EXPORT_SYMBOL(scsi_kmap_atomic_sg);
3359
3360/**
3361 * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
3362 * @virt: virtual address to be unmapped
3363 */
3364void scsi_kunmap_atomic_sg(void *virt)
3365{
3366 kunmap_atomic(virt);
3367}
3368EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
3369
3370void sdev_disable_disk_events(struct scsi_device *sdev)
3371{
3372 atomic_inc(&sdev->disk_events_disable_depth);
3373}
3374EXPORT_SYMBOL(sdev_disable_disk_events);
3375
3376void sdev_enable_disk_events(struct scsi_device *sdev)
3377{
3378 if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
3379 return;
3380 atomic_dec(&sdev->disk_events_disable_depth);
3381}
3382EXPORT_SYMBOL(sdev_enable_disk_events);
3383
3384/**
3385 * scsi_vpd_lun_id - return a unique device identification
3386 * @sdev: SCSI device
3387 * @id: buffer for the identification
3388 * @id_len: length of the buffer
3389 *
3390 * Copies a unique device identification into @id based
3391 * on the information in the VPD page 0x83 of the device.
3392 * The string will be formatted as a SCSI name string.
3393 *
3394 * Returns the length of the identification or error on failure.
3395 * If the identifier is longer than the supplied buffer the actual
3396 * identifier length is returned and the buffer is not zero-padded.
3397 */
3398int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
3399{
3400 u8 cur_id_type = 0xff;
3401 u8 cur_id_size = 0;
3402 const unsigned char *d, *cur_id_str;
3403 const struct scsi_vpd *vpd_pg83;
3404 int id_size = -EINVAL;
3405
3406 rcu_read_lock();
3407 vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3408 if (!vpd_pg83) {
3409 rcu_read_unlock();
3410 return -ENXIO;
3411 }
3412
3413 /*
3414 * Look for the correct descriptor.
3415 * Order of preference for lun descriptor:
3416 * - SCSI name string
3417 * - NAA IEEE Registered Extended
3418 * - EUI-64 based 16-byte
3419 * - EUI-64 based 12-byte
3420 * - NAA IEEE Registered
3421 * - NAA IEEE Extended
3422 * - T10 Vendor ID
3423 * as longer descriptors reduce the likelyhood
3424 * of identification clashes.
3425 */
3426
3427 /* The id string must be at least 20 bytes + terminating NULL byte */
3428 if (id_len < 21) {
3429 rcu_read_unlock();
3430 return -EINVAL;
3431 }
3432
3433 memset(id, 0, id_len);
3434 d = vpd_pg83->data + 4;
3435 while (d < vpd_pg83->data + vpd_pg83->len) {
3436 /* Skip designators not referring to the LUN */
3437 if ((d[1] & 0x30) != 0x00)
3438 goto next_desig;
3439
3440 switch (d[1] & 0xf) {
3441 case 0x1:
3442 /* T10 Vendor ID */
3443 if (cur_id_size > d[3])
3444 break;
3445 /* Prefer anything */
3446 if (cur_id_type > 0x01 && cur_id_type != 0xff)
3447 break;
3448 cur_id_size = d[3];
3449 if (cur_id_size + 4 > id_len)
3450 cur_id_size = id_len - 4;
3451 cur_id_str = d + 4;
3452 cur_id_type = d[1] & 0xf;
3453 id_size = snprintf(id, id_len, "t10.%*pE",
3454 cur_id_size, cur_id_str);
3455 break;
3456 case 0x2:
3457 /* EUI-64 */
3458 if (cur_id_size > d[3])
3459 break;
3460 /* Prefer NAA IEEE Registered Extended */
3461 if (cur_id_type == 0x3 &&
3462 cur_id_size == d[3])
3463 break;
3464 cur_id_size = d[3];
3465 cur_id_str = d + 4;
3466 cur_id_type = d[1] & 0xf;
3467 switch (cur_id_size) {
3468 case 8:
3469 id_size = snprintf(id, id_len,
3470 "eui.%8phN",
3471 cur_id_str);
3472 break;
3473 case 12:
3474 id_size = snprintf(id, id_len,
3475 "eui.%12phN",
3476 cur_id_str);
3477 break;
3478 case 16:
3479 id_size = snprintf(id, id_len,
3480 "eui.%16phN",
3481 cur_id_str);
3482 break;
3483 default:
3484 cur_id_size = 0;
3485 break;
3486 }
3487 break;
3488 case 0x3:
3489 /* NAA */
3490 if (cur_id_size > d[3])
3491 break;
3492 cur_id_size = d[3];
3493 cur_id_str = d + 4;
3494 cur_id_type = d[1] & 0xf;
3495 switch (cur_id_size) {
3496 case 8:
3497 id_size = snprintf(id, id_len,
3498 "naa.%8phN",
3499 cur_id_str);
3500 break;
3501 case 16:
3502 id_size = snprintf(id, id_len,
3503 "naa.%16phN",
3504 cur_id_str);
3505 break;
3506 default:
3507 cur_id_size = 0;
3508 break;
3509 }
3510 break;
3511 case 0x8:
3512 /* SCSI name string */
3513 if (cur_id_size + 4 > d[3])
3514 break;
3515 /* Prefer others for truncated descriptor */
3516 if (cur_id_size && d[3] > id_len)
3517 break;
3518 cur_id_size = id_size = d[3];
3519 cur_id_str = d + 4;
3520 cur_id_type = d[1] & 0xf;
3521 if (cur_id_size >= id_len)
3522 cur_id_size = id_len - 1;
3523 memcpy(id, cur_id_str, cur_id_size);
3524 /* Decrease priority for truncated descriptor */
3525 if (cur_id_size != id_size)
3526 cur_id_size = 6;
3527 break;
3528 default:
3529 break;
3530 }
3531next_desig:
3532 d += d[3] + 4;
3533 }
3534 rcu_read_unlock();
3535
3536 return id_size;
3537}
3538EXPORT_SYMBOL(scsi_vpd_lun_id);
3539
3540/*
3541 * scsi_vpd_tpg_id - return a target port group identifier
3542 * @sdev: SCSI device
3543 *
3544 * Returns the Target Port Group identifier from the information
3545 * froom VPD page 0x83 of the device.
3546 *
3547 * Returns the identifier or error on failure.
3548 */
3549int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3550{
3551 const unsigned char *d;
3552 const struct scsi_vpd *vpd_pg83;
3553 int group_id = -EAGAIN, rel_port = -1;
3554
3555 rcu_read_lock();
3556 vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3557 if (!vpd_pg83) {
3558 rcu_read_unlock();
3559 return -ENXIO;
3560 }
3561
3562 d = vpd_pg83->data + 4;
3563 while (d < vpd_pg83->data + vpd_pg83->len) {
3564 switch (d[1] & 0xf) {
3565 case 0x4:
3566 /* Relative target port */
3567 rel_port = get_unaligned_be16(&d[6]);
3568 break;
3569 case 0x5:
3570 /* Target port group */
3571 group_id = get_unaligned_be16(&d[6]);
3572 break;
3573 default:
3574 break;
3575 }
3576 d += d[3] + 4;
3577 }
3578 rcu_read_unlock();
3579
3580 if (group_id >= 0 && rel_id && rel_port != -1)
3581 *rel_id = rel_port;
3582
3583 return group_id;
3584}
3585EXPORT_SYMBOL(scsi_vpd_tpg_id);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 1999 Eric Youngdale
4 * Copyright (C) 2014 Christoph Hellwig
5 *
6 * SCSI queueing library.
7 * Initial versions: Eric Youngdale (eric@andante.org).
8 * Based upon conversations with large numbers
9 * of people at Linux Expo.
10 */
11
12#include <linux/bio.h>
13#include <linux/bitops.h>
14#include <linux/blkdev.h>
15#include <linux/completion.h>
16#include <linux/kernel.h>
17#include <linux/export.h>
18#include <linux/init.h>
19#include <linux/pci.h>
20#include <linux/delay.h>
21#include <linux/hardirq.h>
22#include <linux/scatterlist.h>
23#include <linux/blk-mq.h>
24#include <linux/ratelimit.h>
25#include <asm/unaligned.h>
26
27#include <scsi/scsi.h>
28#include <scsi/scsi_cmnd.h>
29#include <scsi/scsi_dbg.h>
30#include <scsi/scsi_device.h>
31#include <scsi/scsi_driver.h>
32#include <scsi/scsi_eh.h>
33#include <scsi/scsi_host.h>
34#include <scsi/scsi_transport.h> /* __scsi_init_queue() */
35#include <scsi/scsi_dh.h>
36
37#include <trace/events/scsi.h>
38
39#include "scsi_debugfs.h"
40#include "scsi_priv.h"
41#include "scsi_logging.h"
42
43/*
44 * Size of integrity metadata is usually small, 1 inline sg should
45 * cover normal cases.
46 */
47#ifdef CONFIG_ARCH_NO_SG_CHAIN
48#define SCSI_INLINE_PROT_SG_CNT 0
49#define SCSI_INLINE_SG_CNT 0
50#else
51#define SCSI_INLINE_PROT_SG_CNT 1
52#define SCSI_INLINE_SG_CNT 2
53#endif
54
55static struct kmem_cache *scsi_sdb_cache;
56static struct kmem_cache *scsi_sense_cache;
57static struct kmem_cache *scsi_sense_isadma_cache;
58static DEFINE_MUTEX(scsi_sense_cache_mutex);
59
60static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd);
61
62static inline struct kmem_cache *
63scsi_select_sense_cache(bool unchecked_isa_dma)
64{
65 return unchecked_isa_dma ? scsi_sense_isadma_cache : scsi_sense_cache;
66}
67
68static void scsi_free_sense_buffer(bool unchecked_isa_dma,
69 unsigned char *sense_buffer)
70{
71 kmem_cache_free(scsi_select_sense_cache(unchecked_isa_dma),
72 sense_buffer);
73}
74
75static unsigned char *scsi_alloc_sense_buffer(bool unchecked_isa_dma,
76 gfp_t gfp_mask, int numa_node)
77{
78 return kmem_cache_alloc_node(scsi_select_sense_cache(unchecked_isa_dma),
79 gfp_mask, numa_node);
80}
81
82int scsi_init_sense_cache(struct Scsi_Host *shost)
83{
84 struct kmem_cache *cache;
85 int ret = 0;
86
87 mutex_lock(&scsi_sense_cache_mutex);
88 cache = scsi_select_sense_cache(shost->unchecked_isa_dma);
89 if (cache)
90 goto exit;
91
92 if (shost->unchecked_isa_dma) {
93 scsi_sense_isadma_cache =
94 kmem_cache_create("scsi_sense_cache(DMA)",
95 SCSI_SENSE_BUFFERSIZE, 0,
96 SLAB_HWCACHE_ALIGN | SLAB_CACHE_DMA, NULL);
97 if (!scsi_sense_isadma_cache)
98 ret = -ENOMEM;
99 } else {
100 scsi_sense_cache =
101 kmem_cache_create_usercopy("scsi_sense_cache",
102 SCSI_SENSE_BUFFERSIZE, 0, SLAB_HWCACHE_ALIGN,
103 0, SCSI_SENSE_BUFFERSIZE, NULL);
104 if (!scsi_sense_cache)
105 ret = -ENOMEM;
106 }
107 exit:
108 mutex_unlock(&scsi_sense_cache_mutex);
109 return ret;
110}
111
112/*
113 * When to reinvoke queueing after a resource shortage. It's 3 msecs to
114 * not change behaviour from the previous unplug mechanism, experimentation
115 * may prove this needs changing.
116 */
117#define SCSI_QUEUE_DELAY 3
118
119static void
120scsi_set_blocked(struct scsi_cmnd *cmd, int reason)
121{
122 struct Scsi_Host *host = cmd->device->host;
123 struct scsi_device *device = cmd->device;
124 struct scsi_target *starget = scsi_target(device);
125
126 /*
127 * Set the appropriate busy bit for the device/host.
128 *
129 * If the host/device isn't busy, assume that something actually
130 * completed, and that we should be able to queue a command now.
131 *
132 * Note that the prior mid-layer assumption that any host could
133 * always queue at least one command is now broken. The mid-layer
134 * will implement a user specifiable stall (see
135 * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
136 * if a command is requeued with no other commands outstanding
137 * either for the device or for the host.
138 */
139 switch (reason) {
140 case SCSI_MLQUEUE_HOST_BUSY:
141 atomic_set(&host->host_blocked, host->max_host_blocked);
142 break;
143 case SCSI_MLQUEUE_DEVICE_BUSY:
144 case SCSI_MLQUEUE_EH_RETRY:
145 atomic_set(&device->device_blocked,
146 device->max_device_blocked);
147 break;
148 case SCSI_MLQUEUE_TARGET_BUSY:
149 atomic_set(&starget->target_blocked,
150 starget->max_target_blocked);
151 break;
152 }
153}
154
155static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd)
156{
157 if (cmd->request->rq_flags & RQF_DONTPREP) {
158 cmd->request->rq_flags &= ~RQF_DONTPREP;
159 scsi_mq_uninit_cmd(cmd);
160 } else {
161 WARN_ON_ONCE(true);
162 }
163 blk_mq_requeue_request(cmd->request, true);
164}
165
166/**
167 * __scsi_queue_insert - private queue insertion
168 * @cmd: The SCSI command being requeued
169 * @reason: The reason for the requeue
170 * @unbusy: Whether the queue should be unbusied
171 *
172 * This is a private queue insertion. The public interface
173 * scsi_queue_insert() always assumes the queue should be unbusied
174 * because it's always called before the completion. This function is
175 * for a requeue after completion, which should only occur in this
176 * file.
177 */
178static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, bool unbusy)
179{
180 struct scsi_device *device = cmd->device;
181
182 SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
183 "Inserting command %p into mlqueue\n", cmd));
184
185 scsi_set_blocked(cmd, reason);
186
187 /*
188 * Decrement the counters, since these commands are no longer
189 * active on the host/device.
190 */
191 if (unbusy)
192 scsi_device_unbusy(device);
193
194 /*
195 * Requeue this command. It will go before all other commands
196 * that are already in the queue. Schedule requeue work under
197 * lock such that the kblockd_schedule_work() call happens
198 * before blk_cleanup_queue() finishes.
199 */
200 cmd->result = 0;
201
202 blk_mq_requeue_request(cmd->request, true);
203}
204
205/*
206 * Function: scsi_queue_insert()
207 *
208 * Purpose: Insert a command in the midlevel queue.
209 *
210 * Arguments: cmd - command that we are adding to queue.
211 * reason - why we are inserting command to queue.
212 *
213 * Lock status: Assumed that lock is not held upon entry.
214 *
215 * Returns: Nothing.
216 *
217 * Notes: We do this for one of two cases. Either the host is busy
218 * and it cannot accept any more commands for the time being,
219 * or the device returned QUEUE_FULL and can accept no more
220 * commands.
221 * Notes: This could be called either from an interrupt context or a
222 * normal process context.
223 */
224void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
225{
226 __scsi_queue_insert(cmd, reason, true);
227}
228
229
230/**
231 * __scsi_execute - insert request and wait for the result
232 * @sdev: scsi device
233 * @cmd: scsi command
234 * @data_direction: data direction
235 * @buffer: data buffer
236 * @bufflen: len of buffer
237 * @sense: optional sense buffer
238 * @sshdr: optional decoded sense header
239 * @timeout: request timeout in seconds
240 * @retries: number of times to retry request
241 * @flags: flags for ->cmd_flags
242 * @rq_flags: flags for ->rq_flags
243 * @resid: optional residual length
244 *
245 * Returns the scsi_cmnd result field if a command was executed, or a negative
246 * Linux error code if we didn't get that far.
247 */
248int __scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
249 int data_direction, void *buffer, unsigned bufflen,
250 unsigned char *sense, struct scsi_sense_hdr *sshdr,
251 int timeout, int retries, u64 flags, req_flags_t rq_flags,
252 int *resid)
253{
254 struct request *req;
255 struct scsi_request *rq;
256 int ret = DRIVER_ERROR << 24;
257
258 req = blk_get_request(sdev->request_queue,
259 data_direction == DMA_TO_DEVICE ?
260 REQ_OP_SCSI_OUT : REQ_OP_SCSI_IN, BLK_MQ_REQ_PREEMPT);
261 if (IS_ERR(req))
262 return ret;
263 rq = scsi_req(req);
264
265 if (bufflen && blk_rq_map_kern(sdev->request_queue, req,
266 buffer, bufflen, GFP_NOIO))
267 goto out;
268
269 rq->cmd_len = COMMAND_SIZE(cmd[0]);
270 memcpy(rq->cmd, cmd, rq->cmd_len);
271 rq->retries = retries;
272 req->timeout = timeout;
273 req->cmd_flags |= flags;
274 req->rq_flags |= rq_flags | RQF_QUIET;
275
276 /*
277 * head injection *required* here otherwise quiesce won't work
278 */
279 blk_execute_rq(req->q, NULL, req, 1);
280
281 /*
282 * Some devices (USB mass-storage in particular) may transfer
283 * garbage data together with a residue indicating that the data
284 * is invalid. Prevent the garbage from being misinterpreted
285 * and prevent security leaks by zeroing out the excess data.
286 */
287 if (unlikely(rq->resid_len > 0 && rq->resid_len <= bufflen))
288 memset(buffer + (bufflen - rq->resid_len), 0, rq->resid_len);
289
290 if (resid)
291 *resid = rq->resid_len;
292 if (sense && rq->sense_len)
293 memcpy(sense, rq->sense, SCSI_SENSE_BUFFERSIZE);
294 if (sshdr)
295 scsi_normalize_sense(rq->sense, rq->sense_len, sshdr);
296 ret = rq->result;
297 out:
298 blk_put_request(req);
299
300 return ret;
301}
302EXPORT_SYMBOL(__scsi_execute);
303
304/*
305 * Function: scsi_init_cmd_errh()
306 *
307 * Purpose: Initialize cmd fields related to error handling.
308 *
309 * Arguments: cmd - command that is ready to be queued.
310 *
311 * Notes: This function has the job of initializing a number of
312 * fields related to error handling. Typically this will
313 * be called once for each command, as required.
314 */
315static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
316{
317 scsi_set_resid(cmd, 0);
318 memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
319 if (cmd->cmd_len == 0)
320 cmd->cmd_len = scsi_command_size(cmd->cmnd);
321}
322
323/*
324 * Decrement the host_busy counter and wake up the error handler if necessary.
325 * Avoid as follows that the error handler is not woken up if shost->host_busy
326 * == shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
327 * with an RCU read lock in this function to ensure that this function in its
328 * entirety either finishes before scsi_eh_scmd_add() increases the
329 * host_failed counter or that it notices the shost state change made by
330 * scsi_eh_scmd_add().
331 */
332static void scsi_dec_host_busy(struct Scsi_Host *shost)
333{
334 unsigned long flags;
335
336 rcu_read_lock();
337 atomic_dec(&shost->host_busy);
338 if (unlikely(scsi_host_in_recovery(shost))) {
339 spin_lock_irqsave(shost->host_lock, flags);
340 if (shost->host_failed || shost->host_eh_scheduled)
341 scsi_eh_wakeup(shost);
342 spin_unlock_irqrestore(shost->host_lock, flags);
343 }
344 rcu_read_unlock();
345}
346
347void scsi_device_unbusy(struct scsi_device *sdev)
348{
349 struct Scsi_Host *shost = sdev->host;
350 struct scsi_target *starget = scsi_target(sdev);
351
352 scsi_dec_host_busy(shost);
353
354 if (starget->can_queue > 0)
355 atomic_dec(&starget->target_busy);
356
357 atomic_dec(&sdev->device_busy);
358}
359
360static void scsi_kick_queue(struct request_queue *q)
361{
362 blk_mq_run_hw_queues(q, false);
363}
364
365/*
366 * Called for single_lun devices on IO completion. Clear starget_sdev_user,
367 * and call blk_run_queue for all the scsi_devices on the target -
368 * including current_sdev first.
369 *
370 * Called with *no* scsi locks held.
371 */
372static void scsi_single_lun_run(struct scsi_device *current_sdev)
373{
374 struct Scsi_Host *shost = current_sdev->host;
375 struct scsi_device *sdev, *tmp;
376 struct scsi_target *starget = scsi_target(current_sdev);
377 unsigned long flags;
378
379 spin_lock_irqsave(shost->host_lock, flags);
380 starget->starget_sdev_user = NULL;
381 spin_unlock_irqrestore(shost->host_lock, flags);
382
383 /*
384 * Call blk_run_queue for all LUNs on the target, starting with
385 * current_sdev. We race with others (to set starget_sdev_user),
386 * but in most cases, we will be first. Ideally, each LU on the
387 * target would get some limited time or requests on the target.
388 */
389 scsi_kick_queue(current_sdev->request_queue);
390
391 spin_lock_irqsave(shost->host_lock, flags);
392 if (starget->starget_sdev_user)
393 goto out;
394 list_for_each_entry_safe(sdev, tmp, &starget->devices,
395 same_target_siblings) {
396 if (sdev == current_sdev)
397 continue;
398 if (scsi_device_get(sdev))
399 continue;
400
401 spin_unlock_irqrestore(shost->host_lock, flags);
402 scsi_kick_queue(sdev->request_queue);
403 spin_lock_irqsave(shost->host_lock, flags);
404
405 scsi_device_put(sdev);
406 }
407 out:
408 spin_unlock_irqrestore(shost->host_lock, flags);
409}
410
411static inline bool scsi_device_is_busy(struct scsi_device *sdev)
412{
413 if (atomic_read(&sdev->device_busy) >= sdev->queue_depth)
414 return true;
415 if (atomic_read(&sdev->device_blocked) > 0)
416 return true;
417 return false;
418}
419
420static inline bool scsi_target_is_busy(struct scsi_target *starget)
421{
422 if (starget->can_queue > 0) {
423 if (atomic_read(&starget->target_busy) >= starget->can_queue)
424 return true;
425 if (atomic_read(&starget->target_blocked) > 0)
426 return true;
427 }
428 return false;
429}
430
431static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
432{
433 if (shost->can_queue > 0 &&
434 atomic_read(&shost->host_busy) >= shost->can_queue)
435 return true;
436 if (atomic_read(&shost->host_blocked) > 0)
437 return true;
438 if (shost->host_self_blocked)
439 return true;
440 return false;
441}
442
443static void scsi_starved_list_run(struct Scsi_Host *shost)
444{
445 LIST_HEAD(starved_list);
446 struct scsi_device *sdev;
447 unsigned long flags;
448
449 spin_lock_irqsave(shost->host_lock, flags);
450 list_splice_init(&shost->starved_list, &starved_list);
451
452 while (!list_empty(&starved_list)) {
453 struct request_queue *slq;
454
455 /*
456 * As long as shost is accepting commands and we have
457 * starved queues, call blk_run_queue. scsi_request_fn
458 * drops the queue_lock and can add us back to the
459 * starved_list.
460 *
461 * host_lock protects the starved_list and starved_entry.
462 * scsi_request_fn must get the host_lock before checking
463 * or modifying starved_list or starved_entry.
464 */
465 if (scsi_host_is_busy(shost))
466 break;
467
468 sdev = list_entry(starved_list.next,
469 struct scsi_device, starved_entry);
470 list_del_init(&sdev->starved_entry);
471 if (scsi_target_is_busy(scsi_target(sdev))) {
472 list_move_tail(&sdev->starved_entry,
473 &shost->starved_list);
474 continue;
475 }
476
477 /*
478 * Once we drop the host lock, a racing scsi_remove_device()
479 * call may remove the sdev from the starved list and destroy
480 * it and the queue. Mitigate by taking a reference to the
481 * queue and never touching the sdev again after we drop the
482 * host lock. Note: if __scsi_remove_device() invokes
483 * blk_cleanup_queue() before the queue is run from this
484 * function then blk_run_queue() will return immediately since
485 * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
486 */
487 slq = sdev->request_queue;
488 if (!blk_get_queue(slq))
489 continue;
490 spin_unlock_irqrestore(shost->host_lock, flags);
491
492 scsi_kick_queue(slq);
493 blk_put_queue(slq);
494
495 spin_lock_irqsave(shost->host_lock, flags);
496 }
497 /* put any unprocessed entries back */
498 list_splice(&starved_list, &shost->starved_list);
499 spin_unlock_irqrestore(shost->host_lock, flags);
500}
501
502/*
503 * Function: scsi_run_queue()
504 *
505 * Purpose: Select a proper request queue to serve next
506 *
507 * Arguments: q - last request's queue
508 *
509 * Returns: Nothing
510 *
511 * Notes: The previous command was completely finished, start
512 * a new one if possible.
513 */
514static void scsi_run_queue(struct request_queue *q)
515{
516 struct scsi_device *sdev = q->queuedata;
517
518 if (scsi_target(sdev)->single_lun)
519 scsi_single_lun_run(sdev);
520 if (!list_empty(&sdev->host->starved_list))
521 scsi_starved_list_run(sdev->host);
522
523 blk_mq_run_hw_queues(q, false);
524}
525
526void scsi_requeue_run_queue(struct work_struct *work)
527{
528 struct scsi_device *sdev;
529 struct request_queue *q;
530
531 sdev = container_of(work, struct scsi_device, requeue_work);
532 q = sdev->request_queue;
533 scsi_run_queue(q);
534}
535
536void scsi_run_host_queues(struct Scsi_Host *shost)
537{
538 struct scsi_device *sdev;
539
540 shost_for_each_device(sdev, shost)
541 scsi_run_queue(sdev->request_queue);
542}
543
544static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
545{
546 if (!blk_rq_is_passthrough(cmd->request)) {
547 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
548
549 if (drv->uninit_command)
550 drv->uninit_command(cmd);
551 }
552}
553
554static void scsi_mq_free_sgtables(struct scsi_cmnd *cmd)
555{
556 if (cmd->sdb.table.nents)
557 sg_free_table_chained(&cmd->sdb.table,
558 SCSI_INLINE_SG_CNT);
559 if (scsi_prot_sg_count(cmd))
560 sg_free_table_chained(&cmd->prot_sdb->table,
561 SCSI_INLINE_PROT_SG_CNT);
562}
563
564static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
565{
566 scsi_mq_free_sgtables(cmd);
567 scsi_uninit_cmd(cmd);
568 scsi_del_cmd_from_list(cmd);
569}
570
571/* Returns false when no more bytes to process, true if there are more */
572static bool scsi_end_request(struct request *req, blk_status_t error,
573 unsigned int bytes)
574{
575 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
576 struct scsi_device *sdev = cmd->device;
577 struct request_queue *q = sdev->request_queue;
578
579 if (blk_update_request(req, error, bytes))
580 return true;
581
582 if (blk_queue_add_random(q))
583 add_disk_randomness(req->rq_disk);
584
585 if (!blk_rq_is_scsi(req)) {
586 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
587 cmd->flags &= ~SCMD_INITIALIZED;
588 }
589
590 /*
591 * Calling rcu_barrier() is not necessary here because the
592 * SCSI error handler guarantees that the function called by
593 * call_rcu() has been called before scsi_end_request() is
594 * called.
595 */
596 destroy_rcu_head(&cmd->rcu);
597
598 /*
599 * In the MQ case the command gets freed by __blk_mq_end_request,
600 * so we have to do all cleanup that depends on it earlier.
601 *
602 * We also can't kick the queues from irq context, so we
603 * will have to defer it to a workqueue.
604 */
605 scsi_mq_uninit_cmd(cmd);
606
607 /*
608 * queue is still alive, so grab the ref for preventing it
609 * from being cleaned up during running queue.
610 */
611 percpu_ref_get(&q->q_usage_counter);
612
613 __blk_mq_end_request(req, error);
614
615 if (scsi_target(sdev)->single_lun ||
616 !list_empty(&sdev->host->starved_list))
617 kblockd_schedule_work(&sdev->requeue_work);
618 else
619 blk_mq_run_hw_queues(q, true);
620
621 percpu_ref_put(&q->q_usage_counter);
622 return false;
623}
624
625/**
626 * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
627 * @cmd: SCSI command
628 * @result: scsi error code
629 *
630 * Translate a SCSI result code into a blk_status_t value. May reset the host
631 * byte of @cmd->result.
632 */
633static blk_status_t scsi_result_to_blk_status(struct scsi_cmnd *cmd, int result)
634{
635 switch (host_byte(result)) {
636 case DID_OK:
637 /*
638 * Also check the other bytes than the status byte in result
639 * to handle the case when a SCSI LLD sets result to
640 * DRIVER_SENSE << 24 without setting SAM_STAT_CHECK_CONDITION.
641 */
642 if (scsi_status_is_good(result) && (result & ~0xff) == 0)
643 return BLK_STS_OK;
644 return BLK_STS_IOERR;
645 case DID_TRANSPORT_FAILFAST:
646 return BLK_STS_TRANSPORT;
647 case DID_TARGET_FAILURE:
648 set_host_byte(cmd, DID_OK);
649 return BLK_STS_TARGET;
650 case DID_NEXUS_FAILURE:
651 set_host_byte(cmd, DID_OK);
652 return BLK_STS_NEXUS;
653 case DID_ALLOC_FAILURE:
654 set_host_byte(cmd, DID_OK);
655 return BLK_STS_NOSPC;
656 case DID_MEDIUM_ERROR:
657 set_host_byte(cmd, DID_OK);
658 return BLK_STS_MEDIUM;
659 default:
660 return BLK_STS_IOERR;
661 }
662}
663
664/* Helper for scsi_io_completion() when "reprep" action required. */
665static void scsi_io_completion_reprep(struct scsi_cmnd *cmd,
666 struct request_queue *q)
667{
668 /* A new command will be prepared and issued. */
669 scsi_mq_requeue_cmd(cmd);
670}
671
672/* Helper for scsi_io_completion() when special action required. */
673static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
674{
675 struct request_queue *q = cmd->device->request_queue;
676 struct request *req = cmd->request;
677 int level = 0;
678 enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
679 ACTION_DELAYED_RETRY} action;
680 unsigned long wait_for = (cmd->allowed + 1) * req->timeout;
681 struct scsi_sense_hdr sshdr;
682 bool sense_valid;
683 bool sense_current = true; /* false implies "deferred sense" */
684 blk_status_t blk_stat;
685
686 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
687 if (sense_valid)
688 sense_current = !scsi_sense_is_deferred(&sshdr);
689
690 blk_stat = scsi_result_to_blk_status(cmd, result);
691
692 if (host_byte(result) == DID_RESET) {
693 /* Third party bus reset or reset for error recovery
694 * reasons. Just retry the command and see what
695 * happens.
696 */
697 action = ACTION_RETRY;
698 } else if (sense_valid && sense_current) {
699 switch (sshdr.sense_key) {
700 case UNIT_ATTENTION:
701 if (cmd->device->removable) {
702 /* Detected disc change. Set a bit
703 * and quietly refuse further access.
704 */
705 cmd->device->changed = 1;
706 action = ACTION_FAIL;
707 } else {
708 /* Must have been a power glitch, or a
709 * bus reset. Could not have been a
710 * media change, so we just retry the
711 * command and see what happens.
712 */
713 action = ACTION_RETRY;
714 }
715 break;
716 case ILLEGAL_REQUEST:
717 /* If we had an ILLEGAL REQUEST returned, then
718 * we may have performed an unsupported
719 * command. The only thing this should be
720 * would be a ten byte read where only a six
721 * byte read was supported. Also, on a system
722 * where READ CAPACITY failed, we may have
723 * read past the end of the disk.
724 */
725 if ((cmd->device->use_10_for_rw &&
726 sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
727 (cmd->cmnd[0] == READ_10 ||
728 cmd->cmnd[0] == WRITE_10)) {
729 /* This will issue a new 6-byte command. */
730 cmd->device->use_10_for_rw = 0;
731 action = ACTION_REPREP;
732 } else if (sshdr.asc == 0x10) /* DIX */ {
733 action = ACTION_FAIL;
734 blk_stat = BLK_STS_PROTECTION;
735 /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
736 } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
737 action = ACTION_FAIL;
738 blk_stat = BLK_STS_TARGET;
739 } else
740 action = ACTION_FAIL;
741 break;
742 case ABORTED_COMMAND:
743 action = ACTION_FAIL;
744 if (sshdr.asc == 0x10) /* DIF */
745 blk_stat = BLK_STS_PROTECTION;
746 break;
747 case NOT_READY:
748 /* If the device is in the process of becoming
749 * ready, or has a temporary blockage, retry.
750 */
751 if (sshdr.asc == 0x04) {
752 switch (sshdr.ascq) {
753 case 0x01: /* becoming ready */
754 case 0x04: /* format in progress */
755 case 0x05: /* rebuild in progress */
756 case 0x06: /* recalculation in progress */
757 case 0x07: /* operation in progress */
758 case 0x08: /* Long write in progress */
759 case 0x09: /* self test in progress */
760 case 0x14: /* space allocation in progress */
761 case 0x1a: /* start stop unit in progress */
762 case 0x1b: /* sanitize in progress */
763 case 0x1d: /* configuration in progress */
764 case 0x24: /* depopulation in progress */
765 action = ACTION_DELAYED_RETRY;
766 break;
767 default:
768 action = ACTION_FAIL;
769 break;
770 }
771 } else
772 action = ACTION_FAIL;
773 break;
774 case VOLUME_OVERFLOW:
775 /* See SSC3rXX or current. */
776 action = ACTION_FAIL;
777 break;
778 default:
779 action = ACTION_FAIL;
780 break;
781 }
782 } else
783 action = ACTION_FAIL;
784
785 if (action != ACTION_FAIL &&
786 time_before(cmd->jiffies_at_alloc + wait_for, jiffies))
787 action = ACTION_FAIL;
788
789 switch (action) {
790 case ACTION_FAIL:
791 /* Give up and fail the remainder of the request */
792 if (!(req->rq_flags & RQF_QUIET)) {
793 static DEFINE_RATELIMIT_STATE(_rs,
794 DEFAULT_RATELIMIT_INTERVAL,
795 DEFAULT_RATELIMIT_BURST);
796
797 if (unlikely(scsi_logging_level))
798 level =
799 SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
800 SCSI_LOG_MLCOMPLETE_BITS);
801
802 /*
803 * if logging is enabled the failure will be printed
804 * in scsi_log_completion(), so avoid duplicate messages
805 */
806 if (!level && __ratelimit(&_rs)) {
807 scsi_print_result(cmd, NULL, FAILED);
808 if (driver_byte(result) == DRIVER_SENSE)
809 scsi_print_sense(cmd);
810 scsi_print_command(cmd);
811 }
812 }
813 if (!scsi_end_request(req, blk_stat, blk_rq_err_bytes(req)))
814 return;
815 /*FALLTHRU*/
816 case ACTION_REPREP:
817 scsi_io_completion_reprep(cmd, q);
818 break;
819 case ACTION_RETRY:
820 /* Retry the same command immediately */
821 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
822 break;
823 case ACTION_DELAYED_RETRY:
824 /* Retry the same command after a delay */
825 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
826 break;
827 }
828}
829
830/*
831 * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
832 * new result that may suppress further error checking. Also modifies
833 * *blk_statp in some cases.
834 */
835static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
836 blk_status_t *blk_statp)
837{
838 bool sense_valid;
839 bool sense_current = true; /* false implies "deferred sense" */
840 struct request *req = cmd->request;
841 struct scsi_sense_hdr sshdr;
842
843 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
844 if (sense_valid)
845 sense_current = !scsi_sense_is_deferred(&sshdr);
846
847 if (blk_rq_is_passthrough(req)) {
848 if (sense_valid) {
849 /*
850 * SG_IO wants current and deferred errors
851 */
852 scsi_req(req)->sense_len =
853 min(8 + cmd->sense_buffer[7],
854 SCSI_SENSE_BUFFERSIZE);
855 }
856 if (sense_current)
857 *blk_statp = scsi_result_to_blk_status(cmd, result);
858 } else if (blk_rq_bytes(req) == 0 && sense_current) {
859 /*
860 * Flush commands do not transfers any data, and thus cannot use
861 * good_bytes != blk_rq_bytes(req) as the signal for an error.
862 * This sets *blk_statp explicitly for the problem case.
863 */
864 *blk_statp = scsi_result_to_blk_status(cmd, result);
865 }
866 /*
867 * Recovered errors need reporting, but they're always treated as
868 * success, so fiddle the result code here. For passthrough requests
869 * we already took a copy of the original into sreq->result which
870 * is what gets returned to the user
871 */
872 if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
873 bool do_print = true;
874 /*
875 * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
876 * skip print since caller wants ATA registers. Only occurs
877 * on SCSI ATA PASS_THROUGH commands when CK_COND=1
878 */
879 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
880 do_print = false;
881 else if (req->rq_flags & RQF_QUIET)
882 do_print = false;
883 if (do_print)
884 scsi_print_sense(cmd);
885 result = 0;
886 /* for passthrough, *blk_statp may be set */
887 *blk_statp = BLK_STS_OK;
888 }
889 /*
890 * Another corner case: the SCSI status byte is non-zero but 'good'.
891 * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
892 * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
893 * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
894 * intermediate statuses (both obsolete in SAM-4) as good.
895 */
896 if (status_byte(result) && scsi_status_is_good(result)) {
897 result = 0;
898 *blk_statp = BLK_STS_OK;
899 }
900 return result;
901}
902
903/*
904 * Function: scsi_io_completion()
905 *
906 * Purpose: Completion processing for block device I/O requests.
907 *
908 * Arguments: cmd - command that is finished.
909 *
910 * Lock status: Assumed that no lock is held upon entry.
911 *
912 * Returns: Nothing
913 *
914 * Notes: We will finish off the specified number of sectors. If we
915 * are done, the command block will be released and the queue
916 * function will be goosed. If we are not done then we have to
917 * figure out what to do next:
918 *
919 * a) We can call scsi_requeue_command(). The request
920 * will be unprepared and put back on the queue. Then
921 * a new command will be created for it. This should
922 * be used if we made forward progress, or if we want
923 * to switch from READ(10) to READ(6) for example.
924 *
925 * b) We can call __scsi_queue_insert(). The request will
926 * be put back on the queue and retried using the same
927 * command as before, possibly after a delay.
928 *
929 * c) We can call scsi_end_request() with blk_stat other than
930 * BLK_STS_OK, to fail the remainder of the request.
931 */
932void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
933{
934 int result = cmd->result;
935 struct request_queue *q = cmd->device->request_queue;
936 struct request *req = cmd->request;
937 blk_status_t blk_stat = BLK_STS_OK;
938
939 if (unlikely(result)) /* a nz result may or may not be an error */
940 result = scsi_io_completion_nz_result(cmd, result, &blk_stat);
941
942 if (unlikely(blk_rq_is_passthrough(req))) {
943 /*
944 * scsi_result_to_blk_status may have reset the host_byte
945 */
946 scsi_req(req)->result = cmd->result;
947 }
948
949 /*
950 * Next deal with any sectors which we were able to correctly
951 * handle.
952 */
953 SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
954 "%u sectors total, %d bytes done.\n",
955 blk_rq_sectors(req), good_bytes));
956
957 /*
958 * Next deal with any sectors which we were able to correctly
959 * handle. Failed, zero length commands always need to drop down
960 * to retry code. Fast path should return in this block.
961 */
962 if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
963 if (likely(!scsi_end_request(req, blk_stat, good_bytes)))
964 return; /* no bytes remaining */
965 }
966
967 /* Kill remainder if no retries. */
968 if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
969 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req)))
970 WARN_ONCE(true,
971 "Bytes remaining after failed, no-retry command");
972 return;
973 }
974
975 /*
976 * If there had been no error, but we have leftover bytes in the
977 * requeues just queue the command up again.
978 */
979 if (likely(result == 0))
980 scsi_io_completion_reprep(cmd, q);
981 else
982 scsi_io_completion_action(cmd, result);
983}
984
985static blk_status_t scsi_init_sgtable(struct request *req,
986 struct scsi_data_buffer *sdb)
987{
988 int count;
989
990 /*
991 * If sg table allocation fails, requeue request later.
992 */
993 if (unlikely(sg_alloc_table_chained(&sdb->table,
994 blk_rq_nr_phys_segments(req), sdb->table.sgl,
995 SCSI_INLINE_SG_CNT)))
996 return BLK_STS_RESOURCE;
997
998 /*
999 * Next, walk the list, and fill in the addresses and sizes of
1000 * each segment.
1001 */
1002 count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
1003 BUG_ON(count > sdb->table.nents);
1004 sdb->table.nents = count;
1005 sdb->length = blk_rq_payload_bytes(req);
1006 return BLK_STS_OK;
1007}
1008
1009/*
1010 * Function: scsi_init_io()
1011 *
1012 * Purpose: SCSI I/O initialize function.
1013 *
1014 * Arguments: cmd - Command descriptor we wish to initialize
1015 *
1016 * Returns: BLK_STS_OK on success
1017 * BLK_STS_RESOURCE if the failure is retryable
1018 * BLK_STS_IOERR if the failure is fatal
1019 */
1020blk_status_t scsi_init_io(struct scsi_cmnd *cmd)
1021{
1022 struct request *rq = cmd->request;
1023 blk_status_t ret;
1024
1025 if (WARN_ON_ONCE(!blk_rq_nr_phys_segments(rq)))
1026 return BLK_STS_IOERR;
1027
1028 ret = scsi_init_sgtable(rq, &cmd->sdb);
1029 if (ret)
1030 return ret;
1031
1032 if (blk_integrity_rq(rq)) {
1033 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1034 int ivecs, count;
1035
1036 if (WARN_ON_ONCE(!prot_sdb)) {
1037 /*
1038 * This can happen if someone (e.g. multipath)
1039 * queues a command to a device on an adapter
1040 * that does not support DIX.
1041 */
1042 ret = BLK_STS_IOERR;
1043 goto out_free_sgtables;
1044 }
1045
1046 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1047
1048 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1049 prot_sdb->table.sgl,
1050 SCSI_INLINE_PROT_SG_CNT)) {
1051 ret = BLK_STS_RESOURCE;
1052 goto out_free_sgtables;
1053 }
1054
1055 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1056 prot_sdb->table.sgl);
1057 BUG_ON(count > ivecs);
1058 BUG_ON(count > queue_max_integrity_segments(rq->q));
1059
1060 cmd->prot_sdb = prot_sdb;
1061 cmd->prot_sdb->table.nents = count;
1062 }
1063
1064 return BLK_STS_OK;
1065out_free_sgtables:
1066 scsi_mq_free_sgtables(cmd);
1067 return ret;
1068}
1069EXPORT_SYMBOL(scsi_init_io);
1070
1071/**
1072 * scsi_initialize_rq - initialize struct scsi_cmnd partially
1073 * @rq: Request associated with the SCSI command to be initialized.
1074 *
1075 * This function initializes the members of struct scsi_cmnd that must be
1076 * initialized before request processing starts and that won't be
1077 * reinitialized if a SCSI command is requeued.
1078 *
1079 * Called from inside blk_get_request() for pass-through requests and from
1080 * inside scsi_init_command() for filesystem requests.
1081 */
1082static void scsi_initialize_rq(struct request *rq)
1083{
1084 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1085
1086 scsi_req_init(&cmd->req);
1087 init_rcu_head(&cmd->rcu);
1088 cmd->jiffies_at_alloc = jiffies;
1089 cmd->retries = 0;
1090}
1091
1092/*
1093 * Only called when the request isn't completed by SCSI, and not freed by
1094 * SCSI
1095 */
1096static void scsi_cleanup_rq(struct request *rq)
1097{
1098 if (rq->rq_flags & RQF_DONTPREP) {
1099 scsi_mq_uninit_cmd(blk_mq_rq_to_pdu(rq));
1100 rq->rq_flags &= ~RQF_DONTPREP;
1101 }
1102}
1103
1104/* Add a command to the list used by the aacraid and dpt_i2o drivers */
1105void scsi_add_cmd_to_list(struct scsi_cmnd *cmd)
1106{
1107 struct scsi_device *sdev = cmd->device;
1108 struct Scsi_Host *shost = sdev->host;
1109 unsigned long flags;
1110
1111 if (shost->use_cmd_list) {
1112 spin_lock_irqsave(&sdev->list_lock, flags);
1113 list_add_tail(&cmd->list, &sdev->cmd_list);
1114 spin_unlock_irqrestore(&sdev->list_lock, flags);
1115 }
1116}
1117
1118/* Remove a command from the list used by the aacraid and dpt_i2o drivers */
1119void scsi_del_cmd_from_list(struct scsi_cmnd *cmd)
1120{
1121 struct scsi_device *sdev = cmd->device;
1122 struct Scsi_Host *shost = sdev->host;
1123 unsigned long flags;
1124
1125 if (shost->use_cmd_list) {
1126 spin_lock_irqsave(&sdev->list_lock, flags);
1127 BUG_ON(list_empty(&cmd->list));
1128 list_del_init(&cmd->list);
1129 spin_unlock_irqrestore(&sdev->list_lock, flags);
1130 }
1131}
1132
1133/* Called after a request has been started. */
1134void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1135{
1136 void *buf = cmd->sense_buffer;
1137 void *prot = cmd->prot_sdb;
1138 struct request *rq = blk_mq_rq_from_pdu(cmd);
1139 unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1140 unsigned long jiffies_at_alloc;
1141 int retries;
1142
1143 if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1144 flags |= SCMD_INITIALIZED;
1145 scsi_initialize_rq(rq);
1146 }
1147
1148 jiffies_at_alloc = cmd->jiffies_at_alloc;
1149 retries = cmd->retries;
1150 /* zero out the cmd, except for the embedded scsi_request */
1151 memset((char *)cmd + sizeof(cmd->req), 0,
1152 sizeof(*cmd) - sizeof(cmd->req) + dev->host->hostt->cmd_size);
1153
1154 cmd->device = dev;
1155 cmd->sense_buffer = buf;
1156 cmd->prot_sdb = prot;
1157 cmd->flags = flags;
1158 INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1159 cmd->jiffies_at_alloc = jiffies_at_alloc;
1160 cmd->retries = retries;
1161
1162 scsi_add_cmd_to_list(cmd);
1163}
1164
1165static blk_status_t scsi_setup_scsi_cmnd(struct scsi_device *sdev,
1166 struct request *req)
1167{
1168 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1169
1170 /*
1171 * Passthrough requests may transfer data, in which case they must
1172 * a bio attached to them. Or they might contain a SCSI command
1173 * that does not transfer data, in which case they may optionally
1174 * submit a request without an attached bio.
1175 */
1176 if (req->bio) {
1177 blk_status_t ret = scsi_init_io(cmd);
1178 if (unlikely(ret != BLK_STS_OK))
1179 return ret;
1180 } else {
1181 BUG_ON(blk_rq_bytes(req));
1182
1183 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1184 }
1185
1186 cmd->cmd_len = scsi_req(req)->cmd_len;
1187 cmd->cmnd = scsi_req(req)->cmd;
1188 cmd->transfersize = blk_rq_bytes(req);
1189 cmd->allowed = scsi_req(req)->retries;
1190 return BLK_STS_OK;
1191}
1192
1193/*
1194 * Setup a normal block command. These are simple request from filesystems
1195 * that still need to be translated to SCSI CDBs from the ULD.
1196 */
1197static blk_status_t scsi_setup_fs_cmnd(struct scsi_device *sdev,
1198 struct request *req)
1199{
1200 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1201
1202 if (unlikely(sdev->handler && sdev->handler->prep_fn)) {
1203 blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1204 if (ret != BLK_STS_OK)
1205 return ret;
1206 }
1207
1208 cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1209 memset(cmd->cmnd, 0, BLK_MAX_CDB);
1210 return scsi_cmd_to_driver(cmd)->init_command(cmd);
1211}
1212
1213static blk_status_t scsi_setup_cmnd(struct scsi_device *sdev,
1214 struct request *req)
1215{
1216 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1217
1218 if (!blk_rq_bytes(req))
1219 cmd->sc_data_direction = DMA_NONE;
1220 else if (rq_data_dir(req) == WRITE)
1221 cmd->sc_data_direction = DMA_TO_DEVICE;
1222 else
1223 cmd->sc_data_direction = DMA_FROM_DEVICE;
1224
1225 if (blk_rq_is_scsi(req))
1226 return scsi_setup_scsi_cmnd(sdev, req);
1227 else
1228 return scsi_setup_fs_cmnd(sdev, req);
1229}
1230
1231static blk_status_t
1232scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1233{
1234 switch (sdev->sdev_state) {
1235 case SDEV_OFFLINE:
1236 case SDEV_TRANSPORT_OFFLINE:
1237 /*
1238 * If the device is offline we refuse to process any
1239 * commands. The device must be brought online
1240 * before trying any recovery commands.
1241 */
1242 sdev_printk(KERN_ERR, sdev,
1243 "rejecting I/O to offline device\n");
1244 return BLK_STS_IOERR;
1245 case SDEV_DEL:
1246 /*
1247 * If the device is fully deleted, we refuse to
1248 * process any commands as well.
1249 */
1250 sdev_printk(KERN_ERR, sdev,
1251 "rejecting I/O to dead device\n");
1252 return BLK_STS_IOERR;
1253 case SDEV_BLOCK:
1254 case SDEV_CREATED_BLOCK:
1255 return BLK_STS_RESOURCE;
1256 case SDEV_QUIESCE:
1257 /*
1258 * If the devices is blocked we defer normal commands.
1259 */
1260 if (req && !(req->rq_flags & RQF_PREEMPT))
1261 return BLK_STS_RESOURCE;
1262 return BLK_STS_OK;
1263 default:
1264 /*
1265 * For any other not fully online state we only allow
1266 * special commands. In particular any user initiated
1267 * command is not allowed.
1268 */
1269 if (req && !(req->rq_flags & RQF_PREEMPT))
1270 return BLK_STS_IOERR;
1271 return BLK_STS_OK;
1272 }
1273}
1274
1275/*
1276 * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1277 * return 0.
1278 *
1279 * Called with the queue_lock held.
1280 */
1281static inline int scsi_dev_queue_ready(struct request_queue *q,
1282 struct scsi_device *sdev)
1283{
1284 unsigned int busy;
1285
1286 busy = atomic_inc_return(&sdev->device_busy) - 1;
1287 if (atomic_read(&sdev->device_blocked)) {
1288 if (busy)
1289 goto out_dec;
1290
1291 /*
1292 * unblock after device_blocked iterates to zero
1293 */
1294 if (atomic_dec_return(&sdev->device_blocked) > 0)
1295 goto out_dec;
1296 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1297 "unblocking device at zero depth\n"));
1298 }
1299
1300 if (busy >= sdev->queue_depth)
1301 goto out_dec;
1302
1303 return 1;
1304out_dec:
1305 atomic_dec(&sdev->device_busy);
1306 return 0;
1307}
1308
1309/*
1310 * scsi_target_queue_ready: checks if there we can send commands to target
1311 * @sdev: scsi device on starget to check.
1312 */
1313static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1314 struct scsi_device *sdev)
1315{
1316 struct scsi_target *starget = scsi_target(sdev);
1317 unsigned int busy;
1318
1319 if (starget->single_lun) {
1320 spin_lock_irq(shost->host_lock);
1321 if (starget->starget_sdev_user &&
1322 starget->starget_sdev_user != sdev) {
1323 spin_unlock_irq(shost->host_lock);
1324 return 0;
1325 }
1326 starget->starget_sdev_user = sdev;
1327 spin_unlock_irq(shost->host_lock);
1328 }
1329
1330 if (starget->can_queue <= 0)
1331 return 1;
1332
1333 busy = atomic_inc_return(&starget->target_busy) - 1;
1334 if (atomic_read(&starget->target_blocked) > 0) {
1335 if (busy)
1336 goto starved;
1337
1338 /*
1339 * unblock after target_blocked iterates to zero
1340 */
1341 if (atomic_dec_return(&starget->target_blocked) > 0)
1342 goto out_dec;
1343
1344 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1345 "unblocking target at zero depth\n"));
1346 }
1347
1348 if (busy >= starget->can_queue)
1349 goto starved;
1350
1351 return 1;
1352
1353starved:
1354 spin_lock_irq(shost->host_lock);
1355 list_move_tail(&sdev->starved_entry, &shost->starved_list);
1356 spin_unlock_irq(shost->host_lock);
1357out_dec:
1358 if (starget->can_queue > 0)
1359 atomic_dec(&starget->target_busy);
1360 return 0;
1361}
1362
1363/*
1364 * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1365 * return 0. We must end up running the queue again whenever 0 is
1366 * returned, else IO can hang.
1367 */
1368static inline int scsi_host_queue_ready(struct request_queue *q,
1369 struct Scsi_Host *shost,
1370 struct scsi_device *sdev)
1371{
1372 unsigned int busy;
1373
1374 if (scsi_host_in_recovery(shost))
1375 return 0;
1376
1377 busy = atomic_inc_return(&shost->host_busy) - 1;
1378 if (atomic_read(&shost->host_blocked) > 0) {
1379 if (busy)
1380 goto starved;
1381
1382 /*
1383 * unblock after host_blocked iterates to zero
1384 */
1385 if (atomic_dec_return(&shost->host_blocked) > 0)
1386 goto out_dec;
1387
1388 SCSI_LOG_MLQUEUE(3,
1389 shost_printk(KERN_INFO, shost,
1390 "unblocking host at zero depth\n"));
1391 }
1392
1393 if (shost->can_queue > 0 && busy >= shost->can_queue)
1394 goto starved;
1395 if (shost->host_self_blocked)
1396 goto starved;
1397
1398 /* We're OK to process the command, so we can't be starved */
1399 if (!list_empty(&sdev->starved_entry)) {
1400 spin_lock_irq(shost->host_lock);
1401 if (!list_empty(&sdev->starved_entry))
1402 list_del_init(&sdev->starved_entry);
1403 spin_unlock_irq(shost->host_lock);
1404 }
1405
1406 return 1;
1407
1408starved:
1409 spin_lock_irq(shost->host_lock);
1410 if (list_empty(&sdev->starved_entry))
1411 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1412 spin_unlock_irq(shost->host_lock);
1413out_dec:
1414 scsi_dec_host_busy(shost);
1415 return 0;
1416}
1417
1418/*
1419 * Busy state exporting function for request stacking drivers.
1420 *
1421 * For efficiency, no lock is taken to check the busy state of
1422 * shost/starget/sdev, since the returned value is not guaranteed and
1423 * may be changed after request stacking drivers call the function,
1424 * regardless of taking lock or not.
1425 *
1426 * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1427 * needs to return 'not busy'. Otherwise, request stacking drivers
1428 * may hold requests forever.
1429 */
1430static bool scsi_mq_lld_busy(struct request_queue *q)
1431{
1432 struct scsi_device *sdev = q->queuedata;
1433 struct Scsi_Host *shost;
1434
1435 if (blk_queue_dying(q))
1436 return false;
1437
1438 shost = sdev->host;
1439
1440 /*
1441 * Ignore host/starget busy state.
1442 * Since block layer does not have a concept of fairness across
1443 * multiple queues, congestion of host/starget needs to be handled
1444 * in SCSI layer.
1445 */
1446 if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1447 return true;
1448
1449 return false;
1450}
1451
1452static void scsi_softirq_done(struct request *rq)
1453{
1454 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1455 unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1456 int disposition;
1457
1458 INIT_LIST_HEAD(&cmd->eh_entry);
1459
1460 atomic_inc(&cmd->device->iodone_cnt);
1461 if (cmd->result)
1462 atomic_inc(&cmd->device->ioerr_cnt);
1463
1464 disposition = scsi_decide_disposition(cmd);
1465 if (disposition != SUCCESS &&
1466 time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1467 scmd_printk(KERN_ERR, cmd,
1468 "timing out command, waited %lus\n",
1469 wait_for/HZ);
1470 disposition = SUCCESS;
1471 }
1472
1473 scsi_log_completion(cmd, disposition);
1474
1475 switch (disposition) {
1476 case SUCCESS:
1477 scsi_finish_command(cmd);
1478 break;
1479 case NEEDS_RETRY:
1480 scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1481 break;
1482 case ADD_TO_MLQUEUE:
1483 scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1484 break;
1485 default:
1486 scsi_eh_scmd_add(cmd);
1487 break;
1488 }
1489}
1490
1491/**
1492 * scsi_dispatch_command - Dispatch a command to the low-level driver.
1493 * @cmd: command block we are dispatching.
1494 *
1495 * Return: nonzero return request was rejected and device's queue needs to be
1496 * plugged.
1497 */
1498static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1499{
1500 struct Scsi_Host *host = cmd->device->host;
1501 int rtn = 0;
1502
1503 atomic_inc(&cmd->device->iorequest_cnt);
1504
1505 /* check if the device is still usable */
1506 if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1507 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1508 * returns an immediate error upwards, and signals
1509 * that the device is no longer present */
1510 cmd->result = DID_NO_CONNECT << 16;
1511 goto done;
1512 }
1513
1514 /* Check to see if the scsi lld made this device blocked. */
1515 if (unlikely(scsi_device_blocked(cmd->device))) {
1516 /*
1517 * in blocked state, the command is just put back on
1518 * the device queue. The suspend state has already
1519 * blocked the queue so future requests should not
1520 * occur until the device transitions out of the
1521 * suspend state.
1522 */
1523 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1524 "queuecommand : device blocked\n"));
1525 return SCSI_MLQUEUE_DEVICE_BUSY;
1526 }
1527
1528 /* Store the LUN value in cmnd, if needed. */
1529 if (cmd->device->lun_in_cdb)
1530 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1531 (cmd->device->lun << 5 & 0xe0);
1532
1533 scsi_log_send(cmd);
1534
1535 /*
1536 * Before we queue this command, check if the command
1537 * length exceeds what the host adapter can handle.
1538 */
1539 if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1540 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1541 "queuecommand : command too long. "
1542 "cdb_size=%d host->max_cmd_len=%d\n",
1543 cmd->cmd_len, cmd->device->host->max_cmd_len));
1544 cmd->result = (DID_ABORT << 16);
1545 goto done;
1546 }
1547
1548 if (unlikely(host->shost_state == SHOST_DEL)) {
1549 cmd->result = (DID_NO_CONNECT << 16);
1550 goto done;
1551
1552 }
1553
1554 trace_scsi_dispatch_cmd_start(cmd);
1555 rtn = host->hostt->queuecommand(host, cmd);
1556 if (rtn) {
1557 trace_scsi_dispatch_cmd_error(cmd, rtn);
1558 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1559 rtn != SCSI_MLQUEUE_TARGET_BUSY)
1560 rtn = SCSI_MLQUEUE_HOST_BUSY;
1561
1562 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1563 "queuecommand : request rejected\n"));
1564 }
1565
1566 return rtn;
1567 done:
1568 cmd->scsi_done(cmd);
1569 return 0;
1570}
1571
1572/* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1573static unsigned int scsi_mq_inline_sgl_size(struct Scsi_Host *shost)
1574{
1575 return min_t(unsigned int, shost->sg_tablesize, SCSI_INLINE_SG_CNT) *
1576 sizeof(struct scatterlist);
1577}
1578
1579static blk_status_t scsi_mq_prep_fn(struct request *req)
1580{
1581 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1582 struct scsi_device *sdev = req->q->queuedata;
1583 struct Scsi_Host *shost = sdev->host;
1584 struct scatterlist *sg;
1585
1586 scsi_init_command(sdev, cmd);
1587
1588 cmd->request = req;
1589 cmd->tag = req->tag;
1590 cmd->prot_op = SCSI_PROT_NORMAL;
1591
1592 sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1593 cmd->sdb.table.sgl = sg;
1594
1595 if (scsi_host_get_prot(shost)) {
1596 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1597
1598 cmd->prot_sdb->table.sgl =
1599 (struct scatterlist *)(cmd->prot_sdb + 1);
1600 }
1601
1602 blk_mq_start_request(req);
1603
1604 return scsi_setup_cmnd(sdev, req);
1605}
1606
1607static void scsi_mq_done(struct scsi_cmnd *cmd)
1608{
1609 if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1610 return;
1611 trace_scsi_dispatch_cmd_done(cmd);
1612
1613 /*
1614 * If the block layer didn't complete the request due to a timeout
1615 * injection, scsi must clear its internal completed state so that the
1616 * timeout handler will see it needs to escalate its own error
1617 * recovery.
1618 */
1619 if (unlikely(!blk_mq_complete_request(cmd->request)))
1620 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1621}
1622
1623static void scsi_mq_put_budget(struct blk_mq_hw_ctx *hctx)
1624{
1625 struct request_queue *q = hctx->queue;
1626 struct scsi_device *sdev = q->queuedata;
1627
1628 atomic_dec(&sdev->device_busy);
1629}
1630
1631static bool scsi_mq_get_budget(struct blk_mq_hw_ctx *hctx)
1632{
1633 struct request_queue *q = hctx->queue;
1634 struct scsi_device *sdev = q->queuedata;
1635
1636 if (scsi_dev_queue_ready(q, sdev))
1637 return true;
1638
1639 if (atomic_read(&sdev->device_busy) == 0 && !scsi_device_blocked(sdev))
1640 blk_mq_delay_run_hw_queue(hctx, SCSI_QUEUE_DELAY);
1641 return false;
1642}
1643
1644static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1645 const struct blk_mq_queue_data *bd)
1646{
1647 struct request *req = bd->rq;
1648 struct request_queue *q = req->q;
1649 struct scsi_device *sdev = q->queuedata;
1650 struct Scsi_Host *shost = sdev->host;
1651 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1652 blk_status_t ret;
1653 int reason;
1654
1655 /*
1656 * If the device is not in running state we will reject some or all
1657 * commands.
1658 */
1659 if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1660 ret = scsi_prep_state_check(sdev, req);
1661 if (ret != BLK_STS_OK)
1662 goto out_put_budget;
1663 }
1664
1665 ret = BLK_STS_RESOURCE;
1666 if (!scsi_target_queue_ready(shost, sdev))
1667 goto out_put_budget;
1668 if (!scsi_host_queue_ready(q, shost, sdev))
1669 goto out_dec_target_busy;
1670
1671 if (!(req->rq_flags & RQF_DONTPREP)) {
1672 ret = scsi_mq_prep_fn(req);
1673 if (ret != BLK_STS_OK)
1674 goto out_dec_host_busy;
1675 req->rq_flags |= RQF_DONTPREP;
1676 } else {
1677 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1678 blk_mq_start_request(req);
1679 }
1680
1681 cmd->flags &= SCMD_PRESERVED_FLAGS;
1682 if (sdev->simple_tags)
1683 cmd->flags |= SCMD_TAGGED;
1684 if (bd->last)
1685 cmd->flags |= SCMD_LAST;
1686
1687 scsi_init_cmd_errh(cmd);
1688 cmd->scsi_done = scsi_mq_done;
1689
1690 reason = scsi_dispatch_cmd(cmd);
1691 if (reason) {
1692 scsi_set_blocked(cmd, reason);
1693 ret = BLK_STS_RESOURCE;
1694 goto out_dec_host_busy;
1695 }
1696
1697 return BLK_STS_OK;
1698
1699out_dec_host_busy:
1700 scsi_dec_host_busy(shost);
1701out_dec_target_busy:
1702 if (scsi_target(sdev)->can_queue > 0)
1703 atomic_dec(&scsi_target(sdev)->target_busy);
1704out_put_budget:
1705 scsi_mq_put_budget(hctx);
1706 switch (ret) {
1707 case BLK_STS_OK:
1708 break;
1709 case BLK_STS_RESOURCE:
1710 if (atomic_read(&sdev->device_busy) ||
1711 scsi_device_blocked(sdev))
1712 ret = BLK_STS_DEV_RESOURCE;
1713 break;
1714 default:
1715 if (unlikely(!scsi_device_online(sdev)))
1716 scsi_req(req)->result = DID_NO_CONNECT << 16;
1717 else
1718 scsi_req(req)->result = DID_ERROR << 16;
1719 /*
1720 * Make sure to release all allocated resources when
1721 * we hit an error, as we will never see this command
1722 * again.
1723 */
1724 if (req->rq_flags & RQF_DONTPREP)
1725 scsi_mq_uninit_cmd(cmd);
1726 break;
1727 }
1728 return ret;
1729}
1730
1731static enum blk_eh_timer_return scsi_timeout(struct request *req,
1732 bool reserved)
1733{
1734 if (reserved)
1735 return BLK_EH_RESET_TIMER;
1736 return scsi_times_out(req);
1737}
1738
1739static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1740 unsigned int hctx_idx, unsigned int numa_node)
1741{
1742 struct Scsi_Host *shost = set->driver_data;
1743 const bool unchecked_isa_dma = shost->unchecked_isa_dma;
1744 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1745 struct scatterlist *sg;
1746
1747 if (unchecked_isa_dma)
1748 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
1749 cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
1750 GFP_KERNEL, numa_node);
1751 if (!cmd->sense_buffer)
1752 return -ENOMEM;
1753 cmd->req.sense = cmd->sense_buffer;
1754
1755 if (scsi_host_get_prot(shost)) {
1756 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1757 shost->hostt->cmd_size;
1758 cmd->prot_sdb = (void *)sg + scsi_mq_inline_sgl_size(shost);
1759 }
1760
1761 return 0;
1762}
1763
1764static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1765 unsigned int hctx_idx)
1766{
1767 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1768
1769 scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
1770 cmd->sense_buffer);
1771}
1772
1773static int scsi_map_queues(struct blk_mq_tag_set *set)
1774{
1775 struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
1776
1777 if (shost->hostt->map_queues)
1778 return shost->hostt->map_queues(shost);
1779 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
1780}
1781
1782void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
1783{
1784 struct device *dev = shost->dma_dev;
1785
1786 /*
1787 * this limit is imposed by hardware restrictions
1788 */
1789 blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1790 SG_MAX_SEGMENTS));
1791
1792 if (scsi_host_prot_dma(shost)) {
1793 shost->sg_prot_tablesize =
1794 min_not_zero(shost->sg_prot_tablesize,
1795 (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1796 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1797 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1798 }
1799
1800 if (dev->dma_mask) {
1801 shost->max_sectors = min_t(unsigned int, shost->max_sectors,
1802 dma_max_mapping_size(dev) >> SECTOR_SHIFT);
1803 }
1804 blk_queue_max_hw_sectors(q, shost->max_sectors);
1805 if (shost->unchecked_isa_dma)
1806 blk_queue_bounce_limit(q, BLK_BOUNCE_ISA);
1807 blk_queue_segment_boundary(q, shost->dma_boundary);
1808 dma_set_seg_boundary(dev, shost->dma_boundary);
1809
1810 blk_queue_max_segment_size(q, shost->max_segment_size);
1811 blk_queue_virt_boundary(q, shost->virt_boundary_mask);
1812 dma_set_max_seg_size(dev, queue_max_segment_size(q));
1813
1814 /*
1815 * Set a reasonable default alignment: The larger of 32-byte (dword),
1816 * which is a common minimum for HBAs, and the minimum DMA alignment,
1817 * which is set by the platform.
1818 *
1819 * Devices that require a bigger alignment can increase it later.
1820 */
1821 blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
1822}
1823EXPORT_SYMBOL_GPL(__scsi_init_queue);
1824
1825static const struct blk_mq_ops scsi_mq_ops_no_commit = {
1826 .get_budget = scsi_mq_get_budget,
1827 .put_budget = scsi_mq_put_budget,
1828 .queue_rq = scsi_queue_rq,
1829 .complete = scsi_softirq_done,
1830 .timeout = scsi_timeout,
1831#ifdef CONFIG_BLK_DEBUG_FS
1832 .show_rq = scsi_show_rq,
1833#endif
1834 .init_request = scsi_mq_init_request,
1835 .exit_request = scsi_mq_exit_request,
1836 .initialize_rq_fn = scsi_initialize_rq,
1837 .cleanup_rq = scsi_cleanup_rq,
1838 .busy = scsi_mq_lld_busy,
1839 .map_queues = scsi_map_queues,
1840};
1841
1842
1843static void scsi_commit_rqs(struct blk_mq_hw_ctx *hctx)
1844{
1845 struct request_queue *q = hctx->queue;
1846 struct scsi_device *sdev = q->queuedata;
1847 struct Scsi_Host *shost = sdev->host;
1848
1849 shost->hostt->commit_rqs(shost, hctx->queue_num);
1850}
1851
1852static const struct blk_mq_ops scsi_mq_ops = {
1853 .get_budget = scsi_mq_get_budget,
1854 .put_budget = scsi_mq_put_budget,
1855 .queue_rq = scsi_queue_rq,
1856 .commit_rqs = scsi_commit_rqs,
1857 .complete = scsi_softirq_done,
1858 .timeout = scsi_timeout,
1859#ifdef CONFIG_BLK_DEBUG_FS
1860 .show_rq = scsi_show_rq,
1861#endif
1862 .init_request = scsi_mq_init_request,
1863 .exit_request = scsi_mq_exit_request,
1864 .initialize_rq_fn = scsi_initialize_rq,
1865 .cleanup_rq = scsi_cleanup_rq,
1866 .busy = scsi_mq_lld_busy,
1867 .map_queues = scsi_map_queues,
1868};
1869
1870struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
1871{
1872 sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
1873 if (IS_ERR(sdev->request_queue))
1874 return NULL;
1875
1876 sdev->request_queue->queuedata = sdev;
1877 __scsi_init_queue(sdev->host, sdev->request_queue);
1878 blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, sdev->request_queue);
1879 return sdev->request_queue;
1880}
1881
1882int scsi_mq_setup_tags(struct Scsi_Host *shost)
1883{
1884 unsigned int cmd_size, sgl_size;
1885
1886 sgl_size = max_t(unsigned int, sizeof(struct scatterlist),
1887 scsi_mq_inline_sgl_size(shost));
1888 cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
1889 if (scsi_host_get_prot(shost))
1890 cmd_size += sizeof(struct scsi_data_buffer) +
1891 sizeof(struct scatterlist) * SCSI_INLINE_PROT_SG_CNT;
1892
1893 memset(&shost->tag_set, 0, sizeof(shost->tag_set));
1894 if (shost->hostt->commit_rqs)
1895 shost->tag_set.ops = &scsi_mq_ops;
1896 else
1897 shost->tag_set.ops = &scsi_mq_ops_no_commit;
1898 shost->tag_set.nr_hw_queues = shost->nr_hw_queues ? : 1;
1899 shost->tag_set.queue_depth = shost->can_queue;
1900 shost->tag_set.cmd_size = cmd_size;
1901 shost->tag_set.numa_node = NUMA_NO_NODE;
1902 shost->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1903 shost->tag_set.flags |=
1904 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
1905 shost->tag_set.driver_data = shost;
1906
1907 return blk_mq_alloc_tag_set(&shost->tag_set);
1908}
1909
1910void scsi_mq_destroy_tags(struct Scsi_Host *shost)
1911{
1912 blk_mq_free_tag_set(&shost->tag_set);
1913}
1914
1915/**
1916 * scsi_device_from_queue - return sdev associated with a request_queue
1917 * @q: The request queue to return the sdev from
1918 *
1919 * Return the sdev associated with a request queue or NULL if the
1920 * request_queue does not reference a SCSI device.
1921 */
1922struct scsi_device *scsi_device_from_queue(struct request_queue *q)
1923{
1924 struct scsi_device *sdev = NULL;
1925
1926 if (q->mq_ops == &scsi_mq_ops_no_commit ||
1927 q->mq_ops == &scsi_mq_ops)
1928 sdev = q->queuedata;
1929 if (!sdev || !get_device(&sdev->sdev_gendev))
1930 sdev = NULL;
1931
1932 return sdev;
1933}
1934EXPORT_SYMBOL_GPL(scsi_device_from_queue);
1935
1936/*
1937 * Function: scsi_block_requests()
1938 *
1939 * Purpose: Utility function used by low-level drivers to prevent further
1940 * commands from being queued to the device.
1941 *
1942 * Arguments: shost - Host in question
1943 *
1944 * Returns: Nothing
1945 *
1946 * Lock status: No locks are assumed held.
1947 *
1948 * Notes: There is no timer nor any other means by which the requests
1949 * get unblocked other than the low-level driver calling
1950 * scsi_unblock_requests().
1951 */
1952void scsi_block_requests(struct Scsi_Host *shost)
1953{
1954 shost->host_self_blocked = 1;
1955}
1956EXPORT_SYMBOL(scsi_block_requests);
1957
1958/*
1959 * Function: scsi_unblock_requests()
1960 *
1961 * Purpose: Utility function used by low-level drivers to allow further
1962 * commands from being queued to the device.
1963 *
1964 * Arguments: shost - Host in question
1965 *
1966 * Returns: Nothing
1967 *
1968 * Lock status: No locks are assumed held.
1969 *
1970 * Notes: There is no timer nor any other means by which the requests
1971 * get unblocked other than the low-level driver calling
1972 * scsi_unblock_requests().
1973 *
1974 * This is done as an API function so that changes to the
1975 * internals of the scsi mid-layer won't require wholesale
1976 * changes to drivers that use this feature.
1977 */
1978void scsi_unblock_requests(struct Scsi_Host *shost)
1979{
1980 shost->host_self_blocked = 0;
1981 scsi_run_host_queues(shost);
1982}
1983EXPORT_SYMBOL(scsi_unblock_requests);
1984
1985int __init scsi_init_queue(void)
1986{
1987 scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1988 sizeof(struct scsi_data_buffer),
1989 0, 0, NULL);
1990 if (!scsi_sdb_cache) {
1991 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1992 return -ENOMEM;
1993 }
1994
1995 return 0;
1996}
1997
1998void scsi_exit_queue(void)
1999{
2000 kmem_cache_destroy(scsi_sense_cache);
2001 kmem_cache_destroy(scsi_sense_isadma_cache);
2002 kmem_cache_destroy(scsi_sdb_cache);
2003}
2004
2005/**
2006 * scsi_mode_select - issue a mode select
2007 * @sdev: SCSI device to be queried
2008 * @pf: Page format bit (1 == standard, 0 == vendor specific)
2009 * @sp: Save page bit (0 == don't save, 1 == save)
2010 * @modepage: mode page being requested
2011 * @buffer: request buffer (may not be smaller than eight bytes)
2012 * @len: length of request buffer.
2013 * @timeout: command timeout
2014 * @retries: number of retries before failing
2015 * @data: returns a structure abstracting the mode header data
2016 * @sshdr: place to put sense data (or NULL if no sense to be collected).
2017 * must be SCSI_SENSE_BUFFERSIZE big.
2018 *
2019 * Returns zero if successful; negative error number or scsi
2020 * status on error
2021 *
2022 */
2023int
2024scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
2025 unsigned char *buffer, int len, int timeout, int retries,
2026 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2027{
2028 unsigned char cmd[10];
2029 unsigned char *real_buffer;
2030 int ret;
2031
2032 memset(cmd, 0, sizeof(cmd));
2033 cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2034
2035 if (sdev->use_10_for_ms) {
2036 if (len > 65535)
2037 return -EINVAL;
2038 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2039 if (!real_buffer)
2040 return -ENOMEM;
2041 memcpy(real_buffer + 8, buffer, len);
2042 len += 8;
2043 real_buffer[0] = 0;
2044 real_buffer[1] = 0;
2045 real_buffer[2] = data->medium_type;
2046 real_buffer[3] = data->device_specific;
2047 real_buffer[4] = data->longlba ? 0x01 : 0;
2048 real_buffer[5] = 0;
2049 real_buffer[6] = data->block_descriptor_length >> 8;
2050 real_buffer[7] = data->block_descriptor_length;
2051
2052 cmd[0] = MODE_SELECT_10;
2053 cmd[7] = len >> 8;
2054 cmd[8] = len;
2055 } else {
2056 if (len > 255 || data->block_descriptor_length > 255 ||
2057 data->longlba)
2058 return -EINVAL;
2059
2060 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2061 if (!real_buffer)
2062 return -ENOMEM;
2063 memcpy(real_buffer + 4, buffer, len);
2064 len += 4;
2065 real_buffer[0] = 0;
2066 real_buffer[1] = data->medium_type;
2067 real_buffer[2] = data->device_specific;
2068 real_buffer[3] = data->block_descriptor_length;
2069
2070
2071 cmd[0] = MODE_SELECT;
2072 cmd[4] = len;
2073 }
2074
2075 ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2076 sshdr, timeout, retries, NULL);
2077 kfree(real_buffer);
2078 return ret;
2079}
2080EXPORT_SYMBOL_GPL(scsi_mode_select);
2081
2082/**
2083 * scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2084 * @sdev: SCSI device to be queried
2085 * @dbd: set if mode sense will allow block descriptors to be returned
2086 * @modepage: mode page being requested
2087 * @buffer: request buffer (may not be smaller than eight bytes)
2088 * @len: length of request buffer.
2089 * @timeout: command timeout
2090 * @retries: number of retries before failing
2091 * @data: returns a structure abstracting the mode header data
2092 * @sshdr: place to put sense data (or NULL if no sense to be collected).
2093 * must be SCSI_SENSE_BUFFERSIZE big.
2094 *
2095 * Returns zero if unsuccessful, or the header offset (either 4
2096 * or 8 depending on whether a six or ten byte command was
2097 * issued) if successful.
2098 */
2099int
2100scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2101 unsigned char *buffer, int len, int timeout, int retries,
2102 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2103{
2104 unsigned char cmd[12];
2105 int use_10_for_ms;
2106 int header_length;
2107 int result, retry_count = retries;
2108 struct scsi_sense_hdr my_sshdr;
2109
2110 memset(data, 0, sizeof(*data));
2111 memset(&cmd[0], 0, 12);
2112 cmd[1] = dbd & 0x18; /* allows DBD and LLBA bits */
2113 cmd[2] = modepage;
2114
2115 /* caller might not be interested in sense, but we need it */
2116 if (!sshdr)
2117 sshdr = &my_sshdr;
2118
2119 retry:
2120 use_10_for_ms = sdev->use_10_for_ms;
2121
2122 if (use_10_for_ms) {
2123 if (len < 8)
2124 len = 8;
2125
2126 cmd[0] = MODE_SENSE_10;
2127 cmd[8] = len;
2128 header_length = 8;
2129 } else {
2130 if (len < 4)
2131 len = 4;
2132
2133 cmd[0] = MODE_SENSE;
2134 cmd[4] = len;
2135 header_length = 4;
2136 }
2137
2138 memset(buffer, 0, len);
2139
2140 result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2141 sshdr, timeout, retries, NULL);
2142
2143 /* This code looks awful: what it's doing is making sure an
2144 * ILLEGAL REQUEST sense return identifies the actual command
2145 * byte as the problem. MODE_SENSE commands can return
2146 * ILLEGAL REQUEST if the code page isn't supported */
2147
2148 if (use_10_for_ms && !scsi_status_is_good(result) &&
2149 driver_byte(result) == DRIVER_SENSE) {
2150 if (scsi_sense_valid(sshdr)) {
2151 if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2152 (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2153 /*
2154 * Invalid command operation code
2155 */
2156 sdev->use_10_for_ms = 0;
2157 goto retry;
2158 }
2159 }
2160 }
2161
2162 if(scsi_status_is_good(result)) {
2163 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2164 (modepage == 6 || modepage == 8))) {
2165 /* Initio breakage? */
2166 header_length = 0;
2167 data->length = 13;
2168 data->medium_type = 0;
2169 data->device_specific = 0;
2170 data->longlba = 0;
2171 data->block_descriptor_length = 0;
2172 } else if(use_10_for_ms) {
2173 data->length = buffer[0]*256 + buffer[1] + 2;
2174 data->medium_type = buffer[2];
2175 data->device_specific = buffer[3];
2176 data->longlba = buffer[4] & 0x01;
2177 data->block_descriptor_length = buffer[6]*256
2178 + buffer[7];
2179 } else {
2180 data->length = buffer[0] + 1;
2181 data->medium_type = buffer[1];
2182 data->device_specific = buffer[2];
2183 data->block_descriptor_length = buffer[3];
2184 }
2185 data->header_length = header_length;
2186 } else if ((status_byte(result) == CHECK_CONDITION) &&
2187 scsi_sense_valid(sshdr) &&
2188 sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2189 retry_count--;
2190 goto retry;
2191 }
2192
2193 return result;
2194}
2195EXPORT_SYMBOL(scsi_mode_sense);
2196
2197/**
2198 * scsi_test_unit_ready - test if unit is ready
2199 * @sdev: scsi device to change the state of.
2200 * @timeout: command timeout
2201 * @retries: number of retries before failing
2202 * @sshdr: outpout pointer for decoded sense information.
2203 *
2204 * Returns zero if unsuccessful or an error if TUR failed. For
2205 * removable media, UNIT_ATTENTION sets ->changed flag.
2206 **/
2207int
2208scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2209 struct scsi_sense_hdr *sshdr)
2210{
2211 char cmd[] = {
2212 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2213 };
2214 int result;
2215
2216 /* try to eat the UNIT_ATTENTION if there are enough retries */
2217 do {
2218 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2219 timeout, 1, NULL);
2220 if (sdev->removable && scsi_sense_valid(sshdr) &&
2221 sshdr->sense_key == UNIT_ATTENTION)
2222 sdev->changed = 1;
2223 } while (scsi_sense_valid(sshdr) &&
2224 sshdr->sense_key == UNIT_ATTENTION && --retries);
2225
2226 return result;
2227}
2228EXPORT_SYMBOL(scsi_test_unit_ready);
2229
2230/**
2231 * scsi_device_set_state - Take the given device through the device state model.
2232 * @sdev: scsi device to change the state of.
2233 * @state: state to change to.
2234 *
2235 * Returns zero if successful or an error if the requested
2236 * transition is illegal.
2237 */
2238int
2239scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2240{
2241 enum scsi_device_state oldstate = sdev->sdev_state;
2242
2243 if (state == oldstate)
2244 return 0;
2245
2246 switch (state) {
2247 case SDEV_CREATED:
2248 switch (oldstate) {
2249 case SDEV_CREATED_BLOCK:
2250 break;
2251 default:
2252 goto illegal;
2253 }
2254 break;
2255
2256 case SDEV_RUNNING:
2257 switch (oldstate) {
2258 case SDEV_CREATED:
2259 case SDEV_OFFLINE:
2260 case SDEV_TRANSPORT_OFFLINE:
2261 case SDEV_QUIESCE:
2262 case SDEV_BLOCK:
2263 break;
2264 default:
2265 goto illegal;
2266 }
2267 break;
2268
2269 case SDEV_QUIESCE:
2270 switch (oldstate) {
2271 case SDEV_RUNNING:
2272 case SDEV_OFFLINE:
2273 case SDEV_TRANSPORT_OFFLINE:
2274 break;
2275 default:
2276 goto illegal;
2277 }
2278 break;
2279
2280 case SDEV_OFFLINE:
2281 case SDEV_TRANSPORT_OFFLINE:
2282 switch (oldstate) {
2283 case SDEV_CREATED:
2284 case SDEV_RUNNING:
2285 case SDEV_QUIESCE:
2286 case SDEV_BLOCK:
2287 break;
2288 default:
2289 goto illegal;
2290 }
2291 break;
2292
2293 case SDEV_BLOCK:
2294 switch (oldstate) {
2295 case SDEV_RUNNING:
2296 case SDEV_CREATED_BLOCK:
2297 case SDEV_OFFLINE:
2298 break;
2299 default:
2300 goto illegal;
2301 }
2302 break;
2303
2304 case SDEV_CREATED_BLOCK:
2305 switch (oldstate) {
2306 case SDEV_CREATED:
2307 break;
2308 default:
2309 goto illegal;
2310 }
2311 break;
2312
2313 case SDEV_CANCEL:
2314 switch (oldstate) {
2315 case SDEV_CREATED:
2316 case SDEV_RUNNING:
2317 case SDEV_QUIESCE:
2318 case SDEV_OFFLINE:
2319 case SDEV_TRANSPORT_OFFLINE:
2320 break;
2321 default:
2322 goto illegal;
2323 }
2324 break;
2325
2326 case SDEV_DEL:
2327 switch (oldstate) {
2328 case SDEV_CREATED:
2329 case SDEV_RUNNING:
2330 case SDEV_OFFLINE:
2331 case SDEV_TRANSPORT_OFFLINE:
2332 case SDEV_CANCEL:
2333 case SDEV_BLOCK:
2334 case SDEV_CREATED_BLOCK:
2335 break;
2336 default:
2337 goto illegal;
2338 }
2339 break;
2340
2341 }
2342 sdev->sdev_state = state;
2343 return 0;
2344
2345 illegal:
2346 SCSI_LOG_ERROR_RECOVERY(1,
2347 sdev_printk(KERN_ERR, sdev,
2348 "Illegal state transition %s->%s",
2349 scsi_device_state_name(oldstate),
2350 scsi_device_state_name(state))
2351 );
2352 return -EINVAL;
2353}
2354EXPORT_SYMBOL(scsi_device_set_state);
2355
2356/**
2357 * sdev_evt_emit - emit a single SCSI device uevent
2358 * @sdev: associated SCSI device
2359 * @evt: event to emit
2360 *
2361 * Send a single uevent (scsi_event) to the associated scsi_device.
2362 */
2363static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2364{
2365 int idx = 0;
2366 char *envp[3];
2367
2368 switch (evt->evt_type) {
2369 case SDEV_EVT_MEDIA_CHANGE:
2370 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2371 break;
2372 case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2373 scsi_rescan_device(&sdev->sdev_gendev);
2374 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2375 break;
2376 case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2377 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2378 break;
2379 case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2380 envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2381 break;
2382 case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2383 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2384 break;
2385 case SDEV_EVT_LUN_CHANGE_REPORTED:
2386 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2387 break;
2388 case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2389 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2390 break;
2391 case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2392 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2393 break;
2394 default:
2395 /* do nothing */
2396 break;
2397 }
2398
2399 envp[idx++] = NULL;
2400
2401 kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2402}
2403
2404/**
2405 * sdev_evt_thread - send a uevent for each scsi event
2406 * @work: work struct for scsi_device
2407 *
2408 * Dispatch queued events to their associated scsi_device kobjects
2409 * as uevents.
2410 */
2411void scsi_evt_thread(struct work_struct *work)
2412{
2413 struct scsi_device *sdev;
2414 enum scsi_device_event evt_type;
2415 LIST_HEAD(event_list);
2416
2417 sdev = container_of(work, struct scsi_device, event_work);
2418
2419 for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2420 if (test_and_clear_bit(evt_type, sdev->pending_events))
2421 sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2422
2423 while (1) {
2424 struct scsi_event *evt;
2425 struct list_head *this, *tmp;
2426 unsigned long flags;
2427
2428 spin_lock_irqsave(&sdev->list_lock, flags);
2429 list_splice_init(&sdev->event_list, &event_list);
2430 spin_unlock_irqrestore(&sdev->list_lock, flags);
2431
2432 if (list_empty(&event_list))
2433 break;
2434
2435 list_for_each_safe(this, tmp, &event_list) {
2436 evt = list_entry(this, struct scsi_event, node);
2437 list_del(&evt->node);
2438 scsi_evt_emit(sdev, evt);
2439 kfree(evt);
2440 }
2441 }
2442}
2443
2444/**
2445 * sdev_evt_send - send asserted event to uevent thread
2446 * @sdev: scsi_device event occurred on
2447 * @evt: event to send
2448 *
2449 * Assert scsi device event asynchronously.
2450 */
2451void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2452{
2453 unsigned long flags;
2454
2455#if 0
2456 /* FIXME: currently this check eliminates all media change events
2457 * for polled devices. Need to update to discriminate between AN
2458 * and polled events */
2459 if (!test_bit(evt->evt_type, sdev->supported_events)) {
2460 kfree(evt);
2461 return;
2462 }
2463#endif
2464
2465 spin_lock_irqsave(&sdev->list_lock, flags);
2466 list_add_tail(&evt->node, &sdev->event_list);
2467 schedule_work(&sdev->event_work);
2468 spin_unlock_irqrestore(&sdev->list_lock, flags);
2469}
2470EXPORT_SYMBOL_GPL(sdev_evt_send);
2471
2472/**
2473 * sdev_evt_alloc - allocate a new scsi event
2474 * @evt_type: type of event to allocate
2475 * @gfpflags: GFP flags for allocation
2476 *
2477 * Allocates and returns a new scsi_event.
2478 */
2479struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2480 gfp_t gfpflags)
2481{
2482 struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2483 if (!evt)
2484 return NULL;
2485
2486 evt->evt_type = evt_type;
2487 INIT_LIST_HEAD(&evt->node);
2488
2489 /* evt_type-specific initialization, if any */
2490 switch (evt_type) {
2491 case SDEV_EVT_MEDIA_CHANGE:
2492 case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2493 case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2494 case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2495 case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2496 case SDEV_EVT_LUN_CHANGE_REPORTED:
2497 case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2498 case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2499 default:
2500 /* do nothing */
2501 break;
2502 }
2503
2504 return evt;
2505}
2506EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2507
2508/**
2509 * sdev_evt_send_simple - send asserted event to uevent thread
2510 * @sdev: scsi_device event occurred on
2511 * @evt_type: type of event to send
2512 * @gfpflags: GFP flags for allocation
2513 *
2514 * Assert scsi device event asynchronously, given an event type.
2515 */
2516void sdev_evt_send_simple(struct scsi_device *sdev,
2517 enum scsi_device_event evt_type, gfp_t gfpflags)
2518{
2519 struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2520 if (!evt) {
2521 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2522 evt_type);
2523 return;
2524 }
2525
2526 sdev_evt_send(sdev, evt);
2527}
2528EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2529
2530/**
2531 * scsi_device_quiesce - Block user issued commands.
2532 * @sdev: scsi device to quiesce.
2533 *
2534 * This works by trying to transition to the SDEV_QUIESCE state
2535 * (which must be a legal transition). When the device is in this
2536 * state, only special requests will be accepted, all others will
2537 * be deferred. Since special requests may also be requeued requests,
2538 * a successful return doesn't guarantee the device will be
2539 * totally quiescent.
2540 *
2541 * Must be called with user context, may sleep.
2542 *
2543 * Returns zero if unsuccessful or an error if not.
2544 */
2545int
2546scsi_device_quiesce(struct scsi_device *sdev)
2547{
2548 struct request_queue *q = sdev->request_queue;
2549 int err;
2550
2551 /*
2552 * It is allowed to call scsi_device_quiesce() multiple times from
2553 * the same context but concurrent scsi_device_quiesce() calls are
2554 * not allowed.
2555 */
2556 WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2557
2558 if (sdev->quiesced_by == current)
2559 return 0;
2560
2561 blk_set_pm_only(q);
2562
2563 blk_mq_freeze_queue(q);
2564 /*
2565 * Ensure that the effect of blk_set_pm_only() will be visible
2566 * for percpu_ref_tryget() callers that occur after the queue
2567 * unfreeze even if the queue was already frozen before this function
2568 * was called. See also https://lwn.net/Articles/573497/.
2569 */
2570 synchronize_rcu();
2571 blk_mq_unfreeze_queue(q);
2572
2573 mutex_lock(&sdev->state_mutex);
2574 err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2575 if (err == 0)
2576 sdev->quiesced_by = current;
2577 else
2578 blk_clear_pm_only(q);
2579 mutex_unlock(&sdev->state_mutex);
2580
2581 return err;
2582}
2583EXPORT_SYMBOL(scsi_device_quiesce);
2584
2585/**
2586 * scsi_device_resume - Restart user issued commands to a quiesced device.
2587 * @sdev: scsi device to resume.
2588 *
2589 * Moves the device from quiesced back to running and restarts the
2590 * queues.
2591 *
2592 * Must be called with user context, may sleep.
2593 */
2594void scsi_device_resume(struct scsi_device *sdev)
2595{
2596 /* check if the device state was mutated prior to resume, and if
2597 * so assume the state is being managed elsewhere (for example
2598 * device deleted during suspend)
2599 */
2600 mutex_lock(&sdev->state_mutex);
2601 if (sdev->quiesced_by) {
2602 sdev->quiesced_by = NULL;
2603 blk_clear_pm_only(sdev->request_queue);
2604 }
2605 if (sdev->sdev_state == SDEV_QUIESCE)
2606 scsi_device_set_state(sdev, SDEV_RUNNING);
2607 mutex_unlock(&sdev->state_mutex);
2608}
2609EXPORT_SYMBOL(scsi_device_resume);
2610
2611static void
2612device_quiesce_fn(struct scsi_device *sdev, void *data)
2613{
2614 scsi_device_quiesce(sdev);
2615}
2616
2617void
2618scsi_target_quiesce(struct scsi_target *starget)
2619{
2620 starget_for_each_device(starget, NULL, device_quiesce_fn);
2621}
2622EXPORT_SYMBOL(scsi_target_quiesce);
2623
2624static void
2625device_resume_fn(struct scsi_device *sdev, void *data)
2626{
2627 scsi_device_resume(sdev);
2628}
2629
2630void
2631scsi_target_resume(struct scsi_target *starget)
2632{
2633 starget_for_each_device(starget, NULL, device_resume_fn);
2634}
2635EXPORT_SYMBOL(scsi_target_resume);
2636
2637/**
2638 * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2639 * @sdev: device to block
2640 *
2641 * Pause SCSI command processing on the specified device. Does not sleep.
2642 *
2643 * Returns zero if successful or a negative error code upon failure.
2644 *
2645 * Notes:
2646 * This routine transitions the device to the SDEV_BLOCK state (which must be
2647 * a legal transition). When the device is in this state, command processing
2648 * is paused until the device leaves the SDEV_BLOCK state. See also
2649 * scsi_internal_device_unblock_nowait().
2650 */
2651int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2652{
2653 struct request_queue *q = sdev->request_queue;
2654 int err = 0;
2655
2656 err = scsi_device_set_state(sdev, SDEV_BLOCK);
2657 if (err) {
2658 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2659
2660 if (err)
2661 return err;
2662 }
2663
2664 /*
2665 * The device has transitioned to SDEV_BLOCK. Stop the
2666 * block layer from calling the midlayer with this device's
2667 * request queue.
2668 */
2669 blk_mq_quiesce_queue_nowait(q);
2670 return 0;
2671}
2672EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2673
2674/**
2675 * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
2676 * @sdev: device to block
2677 *
2678 * Pause SCSI command processing on the specified device and wait until all
2679 * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
2680 *
2681 * Returns zero if successful or a negative error code upon failure.
2682 *
2683 * Note:
2684 * This routine transitions the device to the SDEV_BLOCK state (which must be
2685 * a legal transition). When the device is in this state, command processing
2686 * is paused until the device leaves the SDEV_BLOCK state. See also
2687 * scsi_internal_device_unblock().
2688 */
2689static int scsi_internal_device_block(struct scsi_device *sdev)
2690{
2691 struct request_queue *q = sdev->request_queue;
2692 int err;
2693
2694 mutex_lock(&sdev->state_mutex);
2695 err = scsi_internal_device_block_nowait(sdev);
2696 if (err == 0)
2697 blk_mq_quiesce_queue(q);
2698 mutex_unlock(&sdev->state_mutex);
2699
2700 return err;
2701}
2702
2703void scsi_start_queue(struct scsi_device *sdev)
2704{
2705 struct request_queue *q = sdev->request_queue;
2706
2707 blk_mq_unquiesce_queue(q);
2708}
2709
2710/**
2711 * scsi_internal_device_unblock_nowait - resume a device after a block request
2712 * @sdev: device to resume
2713 * @new_state: state to set the device to after unblocking
2714 *
2715 * Restart the device queue for a previously suspended SCSI device. Does not
2716 * sleep.
2717 *
2718 * Returns zero if successful or a negative error code upon failure.
2719 *
2720 * Notes:
2721 * This routine transitions the device to the SDEV_RUNNING state or to one of
2722 * the offline states (which must be a legal transition) allowing the midlayer
2723 * to goose the queue for this device.
2724 */
2725int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
2726 enum scsi_device_state new_state)
2727{
2728 switch (new_state) {
2729 case SDEV_RUNNING:
2730 case SDEV_TRANSPORT_OFFLINE:
2731 break;
2732 default:
2733 return -EINVAL;
2734 }
2735
2736 /*
2737 * Try to transition the scsi device to SDEV_RUNNING or one of the
2738 * offlined states and goose the device queue if successful.
2739 */
2740 switch (sdev->sdev_state) {
2741 case SDEV_BLOCK:
2742 case SDEV_TRANSPORT_OFFLINE:
2743 sdev->sdev_state = new_state;
2744 break;
2745 case SDEV_CREATED_BLOCK:
2746 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2747 new_state == SDEV_OFFLINE)
2748 sdev->sdev_state = new_state;
2749 else
2750 sdev->sdev_state = SDEV_CREATED;
2751 break;
2752 case SDEV_CANCEL:
2753 case SDEV_OFFLINE:
2754 break;
2755 default:
2756 return -EINVAL;
2757 }
2758 scsi_start_queue(sdev);
2759
2760 return 0;
2761}
2762EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
2763
2764/**
2765 * scsi_internal_device_unblock - resume a device after a block request
2766 * @sdev: device to resume
2767 * @new_state: state to set the device to after unblocking
2768 *
2769 * Restart the device queue for a previously suspended SCSI device. May sleep.
2770 *
2771 * Returns zero if successful or a negative error code upon failure.
2772 *
2773 * Notes:
2774 * This routine transitions the device to the SDEV_RUNNING state or to one of
2775 * the offline states (which must be a legal transition) allowing the midlayer
2776 * to goose the queue for this device.
2777 */
2778static int scsi_internal_device_unblock(struct scsi_device *sdev,
2779 enum scsi_device_state new_state)
2780{
2781 int ret;
2782
2783 mutex_lock(&sdev->state_mutex);
2784 ret = scsi_internal_device_unblock_nowait(sdev, new_state);
2785 mutex_unlock(&sdev->state_mutex);
2786
2787 return ret;
2788}
2789
2790static void
2791device_block(struct scsi_device *sdev, void *data)
2792{
2793 int ret;
2794
2795 ret = scsi_internal_device_block(sdev);
2796
2797 WARN_ONCE(ret, "scsi_internal_device_block(%s) failed: ret = %d\n",
2798 dev_name(&sdev->sdev_gendev), ret);
2799}
2800
2801static int
2802target_block(struct device *dev, void *data)
2803{
2804 if (scsi_is_target_device(dev))
2805 starget_for_each_device(to_scsi_target(dev), NULL,
2806 device_block);
2807 return 0;
2808}
2809
2810void
2811scsi_target_block(struct device *dev)
2812{
2813 if (scsi_is_target_device(dev))
2814 starget_for_each_device(to_scsi_target(dev), NULL,
2815 device_block);
2816 else
2817 device_for_each_child(dev, NULL, target_block);
2818}
2819EXPORT_SYMBOL_GPL(scsi_target_block);
2820
2821static void
2822device_unblock(struct scsi_device *sdev, void *data)
2823{
2824 scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2825}
2826
2827static int
2828target_unblock(struct device *dev, void *data)
2829{
2830 if (scsi_is_target_device(dev))
2831 starget_for_each_device(to_scsi_target(dev), data,
2832 device_unblock);
2833 return 0;
2834}
2835
2836void
2837scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2838{
2839 if (scsi_is_target_device(dev))
2840 starget_for_each_device(to_scsi_target(dev), &new_state,
2841 device_unblock);
2842 else
2843 device_for_each_child(dev, &new_state, target_unblock);
2844}
2845EXPORT_SYMBOL_GPL(scsi_target_unblock);
2846
2847/**
2848 * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2849 * @sgl: scatter-gather list
2850 * @sg_count: number of segments in sg
2851 * @offset: offset in bytes into sg, on return offset into the mapped area
2852 * @len: bytes to map, on return number of bytes mapped
2853 *
2854 * Returns virtual address of the start of the mapped page
2855 */
2856void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2857 size_t *offset, size_t *len)
2858{
2859 int i;
2860 size_t sg_len = 0, len_complete = 0;
2861 struct scatterlist *sg;
2862 struct page *page;
2863
2864 WARN_ON(!irqs_disabled());
2865
2866 for_each_sg(sgl, sg, sg_count, i) {
2867 len_complete = sg_len; /* Complete sg-entries */
2868 sg_len += sg->length;
2869 if (sg_len > *offset)
2870 break;
2871 }
2872
2873 if (unlikely(i == sg_count)) {
2874 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2875 "elements %d\n",
2876 __func__, sg_len, *offset, sg_count);
2877 WARN_ON(1);
2878 return NULL;
2879 }
2880
2881 /* Offset starting from the beginning of first page in this sg-entry */
2882 *offset = *offset - len_complete + sg->offset;
2883
2884 /* Assumption: contiguous pages can be accessed as "page + i" */
2885 page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2886 *offset &= ~PAGE_MASK;
2887
2888 /* Bytes in this sg-entry from *offset to the end of the page */
2889 sg_len = PAGE_SIZE - *offset;
2890 if (*len > sg_len)
2891 *len = sg_len;
2892
2893 return kmap_atomic(page);
2894}
2895EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2896
2897/**
2898 * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2899 * @virt: virtual address to be unmapped
2900 */
2901void scsi_kunmap_atomic_sg(void *virt)
2902{
2903 kunmap_atomic(virt);
2904}
2905EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
2906
2907void sdev_disable_disk_events(struct scsi_device *sdev)
2908{
2909 atomic_inc(&sdev->disk_events_disable_depth);
2910}
2911EXPORT_SYMBOL(sdev_disable_disk_events);
2912
2913void sdev_enable_disk_events(struct scsi_device *sdev)
2914{
2915 if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
2916 return;
2917 atomic_dec(&sdev->disk_events_disable_depth);
2918}
2919EXPORT_SYMBOL(sdev_enable_disk_events);
2920
2921/**
2922 * scsi_vpd_lun_id - return a unique device identification
2923 * @sdev: SCSI device
2924 * @id: buffer for the identification
2925 * @id_len: length of the buffer
2926 *
2927 * Copies a unique device identification into @id based
2928 * on the information in the VPD page 0x83 of the device.
2929 * The string will be formatted as a SCSI name string.
2930 *
2931 * Returns the length of the identification or error on failure.
2932 * If the identifier is longer than the supplied buffer the actual
2933 * identifier length is returned and the buffer is not zero-padded.
2934 */
2935int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
2936{
2937 u8 cur_id_type = 0xff;
2938 u8 cur_id_size = 0;
2939 const unsigned char *d, *cur_id_str;
2940 const struct scsi_vpd *vpd_pg83;
2941 int id_size = -EINVAL;
2942
2943 rcu_read_lock();
2944 vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
2945 if (!vpd_pg83) {
2946 rcu_read_unlock();
2947 return -ENXIO;
2948 }
2949
2950 /*
2951 * Look for the correct descriptor.
2952 * Order of preference for lun descriptor:
2953 * - SCSI name string
2954 * - NAA IEEE Registered Extended
2955 * - EUI-64 based 16-byte
2956 * - EUI-64 based 12-byte
2957 * - NAA IEEE Registered
2958 * - NAA IEEE Extended
2959 * - T10 Vendor ID
2960 * as longer descriptors reduce the likelyhood
2961 * of identification clashes.
2962 */
2963
2964 /* The id string must be at least 20 bytes + terminating NULL byte */
2965 if (id_len < 21) {
2966 rcu_read_unlock();
2967 return -EINVAL;
2968 }
2969
2970 memset(id, 0, id_len);
2971 d = vpd_pg83->data + 4;
2972 while (d < vpd_pg83->data + vpd_pg83->len) {
2973 /* Skip designators not referring to the LUN */
2974 if ((d[1] & 0x30) != 0x00)
2975 goto next_desig;
2976
2977 switch (d[1] & 0xf) {
2978 case 0x1:
2979 /* T10 Vendor ID */
2980 if (cur_id_size > d[3])
2981 break;
2982 /* Prefer anything */
2983 if (cur_id_type > 0x01 && cur_id_type != 0xff)
2984 break;
2985 cur_id_size = d[3];
2986 if (cur_id_size + 4 > id_len)
2987 cur_id_size = id_len - 4;
2988 cur_id_str = d + 4;
2989 cur_id_type = d[1] & 0xf;
2990 id_size = snprintf(id, id_len, "t10.%*pE",
2991 cur_id_size, cur_id_str);
2992 break;
2993 case 0x2:
2994 /* EUI-64 */
2995 if (cur_id_size > d[3])
2996 break;
2997 /* Prefer NAA IEEE Registered Extended */
2998 if (cur_id_type == 0x3 &&
2999 cur_id_size == d[3])
3000 break;
3001 cur_id_size = d[3];
3002 cur_id_str = d + 4;
3003 cur_id_type = d[1] & 0xf;
3004 switch (cur_id_size) {
3005 case 8:
3006 id_size = snprintf(id, id_len,
3007 "eui.%8phN",
3008 cur_id_str);
3009 break;
3010 case 12:
3011 id_size = snprintf(id, id_len,
3012 "eui.%12phN",
3013 cur_id_str);
3014 break;
3015 case 16:
3016 id_size = snprintf(id, id_len,
3017 "eui.%16phN",
3018 cur_id_str);
3019 break;
3020 default:
3021 cur_id_size = 0;
3022 break;
3023 }
3024 break;
3025 case 0x3:
3026 /* NAA */
3027 if (cur_id_size > d[3])
3028 break;
3029 cur_id_size = d[3];
3030 cur_id_str = d + 4;
3031 cur_id_type = d[1] & 0xf;
3032 switch (cur_id_size) {
3033 case 8:
3034 id_size = snprintf(id, id_len,
3035 "naa.%8phN",
3036 cur_id_str);
3037 break;
3038 case 16:
3039 id_size = snprintf(id, id_len,
3040 "naa.%16phN",
3041 cur_id_str);
3042 break;
3043 default:
3044 cur_id_size = 0;
3045 break;
3046 }
3047 break;
3048 case 0x8:
3049 /* SCSI name string */
3050 if (cur_id_size + 4 > d[3])
3051 break;
3052 /* Prefer others for truncated descriptor */
3053 if (cur_id_size && d[3] > id_len)
3054 break;
3055 cur_id_size = id_size = d[3];
3056 cur_id_str = d + 4;
3057 cur_id_type = d[1] & 0xf;
3058 if (cur_id_size >= id_len)
3059 cur_id_size = id_len - 1;
3060 memcpy(id, cur_id_str, cur_id_size);
3061 /* Decrease priority for truncated descriptor */
3062 if (cur_id_size != id_size)
3063 cur_id_size = 6;
3064 break;
3065 default:
3066 break;
3067 }
3068next_desig:
3069 d += d[3] + 4;
3070 }
3071 rcu_read_unlock();
3072
3073 return id_size;
3074}
3075EXPORT_SYMBOL(scsi_vpd_lun_id);
3076
3077/*
3078 * scsi_vpd_tpg_id - return a target port group identifier
3079 * @sdev: SCSI device
3080 *
3081 * Returns the Target Port Group identifier from the information
3082 * froom VPD page 0x83 of the device.
3083 *
3084 * Returns the identifier or error on failure.
3085 */
3086int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3087{
3088 const unsigned char *d;
3089 const struct scsi_vpd *vpd_pg83;
3090 int group_id = -EAGAIN, rel_port = -1;
3091
3092 rcu_read_lock();
3093 vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3094 if (!vpd_pg83) {
3095 rcu_read_unlock();
3096 return -ENXIO;
3097 }
3098
3099 d = vpd_pg83->data + 4;
3100 while (d < vpd_pg83->data + vpd_pg83->len) {
3101 switch (d[1] & 0xf) {
3102 case 0x4:
3103 /* Relative target port */
3104 rel_port = get_unaligned_be16(&d[6]);
3105 break;
3106 case 0x5:
3107 /* Target port group */
3108 group_id = get_unaligned_be16(&d[6]);
3109 break;
3110 default:
3111 break;
3112 }
3113 d += d[3] + 4;
3114 }
3115 rcu_read_unlock();
3116
3117 if (group_id >= 0 && rel_id && rel_port != -1)
3118 *rel_id = rel_port;
3119
3120 return group_id;
3121}
3122EXPORT_SYMBOL(scsi_vpd_tpg_id);