Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 1999 Eric Youngdale
4 * Copyright (C) 2014 Christoph Hellwig
5 *
6 * SCSI queueing library.
7 * Initial versions: Eric Youngdale (eric@andante.org).
8 * Based upon conversations with large numbers
9 * of people at Linux Expo.
10 */
11
12#include <linux/bio.h>
13#include <linux/bitops.h>
14#include <linux/blkdev.h>
15#include <linux/completion.h>
16#include <linux/kernel.h>
17#include <linux/export.h>
18#include <linux/init.h>
19#include <linux/pci.h>
20#include <linux/delay.h>
21#include <linux/hardirq.h>
22#include <linux/scatterlist.h>
23#include <linux/blk-mq.h>
24#include <linux/blk-integrity.h>
25#include <linux/ratelimit.h>
26#include <asm/unaligned.h>
27
28#include <scsi/scsi.h>
29#include <scsi/scsi_cmnd.h>
30#include <scsi/scsi_dbg.h>
31#include <scsi/scsi_device.h>
32#include <scsi/scsi_driver.h>
33#include <scsi/scsi_eh.h>
34#include <scsi/scsi_host.h>
35#include <scsi/scsi_transport.h> /* __scsi_init_queue() */
36#include <scsi/scsi_dh.h>
37
38#include <trace/events/scsi.h>
39
40#include "scsi_debugfs.h"
41#include "scsi_priv.h"
42#include "scsi_logging.h"
43
44/*
45 * Size of integrity metadata is usually small, 1 inline sg should
46 * cover normal cases.
47 */
48#ifdef CONFIG_ARCH_NO_SG_CHAIN
49#define SCSI_INLINE_PROT_SG_CNT 0
50#define SCSI_INLINE_SG_CNT 0
51#else
52#define SCSI_INLINE_PROT_SG_CNT 1
53#define SCSI_INLINE_SG_CNT 2
54#endif
55
56static struct kmem_cache *scsi_sense_cache;
57static DEFINE_MUTEX(scsi_sense_cache_mutex);
58
59static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd);
60
61int scsi_init_sense_cache(struct Scsi_Host *shost)
62{
63 int ret = 0;
64
65 mutex_lock(&scsi_sense_cache_mutex);
66 if (!scsi_sense_cache) {
67 scsi_sense_cache =
68 kmem_cache_create_usercopy("scsi_sense_cache",
69 SCSI_SENSE_BUFFERSIZE, 0, SLAB_HWCACHE_ALIGN,
70 0, SCSI_SENSE_BUFFERSIZE, NULL);
71 if (!scsi_sense_cache)
72 ret = -ENOMEM;
73 }
74 mutex_unlock(&scsi_sense_cache_mutex);
75 return ret;
76}
77
78static void
79scsi_set_blocked(struct scsi_cmnd *cmd, int reason)
80{
81 struct Scsi_Host *host = cmd->device->host;
82 struct scsi_device *device = cmd->device;
83 struct scsi_target *starget = scsi_target(device);
84
85 /*
86 * Set the appropriate busy bit for the device/host.
87 *
88 * If the host/device isn't busy, assume that something actually
89 * completed, and that we should be able to queue a command now.
90 *
91 * Note that the prior mid-layer assumption that any host could
92 * always queue at least one command is now broken. The mid-layer
93 * will implement a user specifiable stall (see
94 * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
95 * if a command is requeued with no other commands outstanding
96 * either for the device or for the host.
97 */
98 switch (reason) {
99 case SCSI_MLQUEUE_HOST_BUSY:
100 atomic_set(&host->host_blocked, host->max_host_blocked);
101 break;
102 case SCSI_MLQUEUE_DEVICE_BUSY:
103 case SCSI_MLQUEUE_EH_RETRY:
104 atomic_set(&device->device_blocked,
105 device->max_device_blocked);
106 break;
107 case SCSI_MLQUEUE_TARGET_BUSY:
108 atomic_set(&starget->target_blocked,
109 starget->max_target_blocked);
110 break;
111 }
112}
113
114static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd, unsigned long msecs)
115{
116 struct request *rq = scsi_cmd_to_rq(cmd);
117
118 if (rq->rq_flags & RQF_DONTPREP) {
119 rq->rq_flags &= ~RQF_DONTPREP;
120 scsi_mq_uninit_cmd(cmd);
121 } else {
122 WARN_ON_ONCE(true);
123 }
124
125 blk_mq_requeue_request(rq, false);
126 if (!scsi_host_in_recovery(cmd->device->host))
127 blk_mq_delay_kick_requeue_list(rq->q, msecs);
128}
129
130/**
131 * __scsi_queue_insert - private queue insertion
132 * @cmd: The SCSI command being requeued
133 * @reason: The reason for the requeue
134 * @unbusy: Whether the queue should be unbusied
135 *
136 * This is a private queue insertion. The public interface
137 * scsi_queue_insert() always assumes the queue should be unbusied
138 * because it's always called before the completion. This function is
139 * for a requeue after completion, which should only occur in this
140 * file.
141 */
142static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, bool unbusy)
143{
144 struct scsi_device *device = cmd->device;
145
146 SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
147 "Inserting command %p into mlqueue\n", cmd));
148
149 scsi_set_blocked(cmd, reason);
150
151 /*
152 * Decrement the counters, since these commands are no longer
153 * active on the host/device.
154 */
155 if (unbusy)
156 scsi_device_unbusy(device, cmd);
157
158 /*
159 * Requeue this command. It will go before all other commands
160 * that are already in the queue. Schedule requeue work under
161 * lock such that the kblockd_schedule_work() call happens
162 * before blk_mq_destroy_queue() finishes.
163 */
164 cmd->result = 0;
165
166 blk_mq_requeue_request(scsi_cmd_to_rq(cmd),
167 !scsi_host_in_recovery(cmd->device->host));
168}
169
170/**
171 * scsi_queue_insert - Reinsert a command in the queue.
172 * @cmd: command that we are adding to queue.
173 * @reason: why we are inserting command to queue.
174 *
175 * We do this for one of two cases. Either the host is busy and it cannot accept
176 * any more commands for the time being, or the device returned QUEUE_FULL and
177 * can accept no more commands.
178 *
179 * Context: This could be called either from an interrupt context or a normal
180 * process context.
181 */
182void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
183{
184 __scsi_queue_insert(cmd, reason, true);
185}
186
187/**
188 * scsi_execute_cmd - insert request and wait for the result
189 * @sdev: scsi_device
190 * @cmd: scsi command
191 * @opf: block layer request cmd_flags
192 * @buffer: data buffer
193 * @bufflen: len of buffer
194 * @timeout: request timeout in HZ
195 * @retries: number of times to retry request
196 * @args: Optional args. See struct definition for field descriptions
197 *
198 * Returns the scsi_cmnd result field if a command was executed, or a negative
199 * Linux error code if we didn't get that far.
200 */
201int scsi_execute_cmd(struct scsi_device *sdev, const unsigned char *cmd,
202 blk_opf_t opf, void *buffer, unsigned int bufflen,
203 int timeout, int retries,
204 const struct scsi_exec_args *args)
205{
206 static const struct scsi_exec_args default_args;
207 struct request *req;
208 struct scsi_cmnd *scmd;
209 int ret;
210
211 if (!args)
212 args = &default_args;
213 else if (WARN_ON_ONCE(args->sense &&
214 args->sense_len != SCSI_SENSE_BUFFERSIZE))
215 return -EINVAL;
216
217 req = scsi_alloc_request(sdev->request_queue, opf, args->req_flags);
218 if (IS_ERR(req))
219 return PTR_ERR(req);
220
221 if (bufflen) {
222 ret = blk_rq_map_kern(sdev->request_queue, req,
223 buffer, bufflen, GFP_NOIO);
224 if (ret)
225 goto out;
226 }
227 scmd = blk_mq_rq_to_pdu(req);
228 scmd->cmd_len = COMMAND_SIZE(cmd[0]);
229 memcpy(scmd->cmnd, cmd, scmd->cmd_len);
230 scmd->allowed = retries;
231 scmd->flags |= args->scmd_flags;
232 req->timeout = timeout;
233 req->rq_flags |= RQF_QUIET;
234
235 /*
236 * head injection *required* here otherwise quiesce won't work
237 */
238 blk_execute_rq(req, true);
239
240 /*
241 * Some devices (USB mass-storage in particular) may transfer
242 * garbage data together with a residue indicating that the data
243 * is invalid. Prevent the garbage from being misinterpreted
244 * and prevent security leaks by zeroing out the excess data.
245 */
246 if (unlikely(scmd->resid_len > 0 && scmd->resid_len <= bufflen))
247 memset(buffer + bufflen - scmd->resid_len, 0, scmd->resid_len);
248
249 if (args->resid)
250 *args->resid = scmd->resid_len;
251 if (args->sense)
252 memcpy(args->sense, scmd->sense_buffer, SCSI_SENSE_BUFFERSIZE);
253 if (args->sshdr)
254 scsi_normalize_sense(scmd->sense_buffer, scmd->sense_len,
255 args->sshdr);
256
257 ret = scmd->result;
258 out:
259 blk_mq_free_request(req);
260
261 return ret;
262}
263EXPORT_SYMBOL(scsi_execute_cmd);
264
265/*
266 * Wake up the error handler if necessary. Avoid as follows that the error
267 * handler is not woken up if host in-flight requests number ==
268 * shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
269 * with an RCU read lock in this function to ensure that this function in
270 * its entirety either finishes before scsi_eh_scmd_add() increases the
271 * host_failed counter or that it notices the shost state change made by
272 * scsi_eh_scmd_add().
273 */
274static void scsi_dec_host_busy(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
275{
276 unsigned long flags;
277
278 rcu_read_lock();
279 __clear_bit(SCMD_STATE_INFLIGHT, &cmd->state);
280 if (unlikely(scsi_host_in_recovery(shost))) {
281 unsigned int busy = scsi_host_busy(shost);
282
283 spin_lock_irqsave(shost->host_lock, flags);
284 if (shost->host_failed || shost->host_eh_scheduled)
285 scsi_eh_wakeup(shost, busy);
286 spin_unlock_irqrestore(shost->host_lock, flags);
287 }
288 rcu_read_unlock();
289}
290
291void scsi_device_unbusy(struct scsi_device *sdev, struct scsi_cmnd *cmd)
292{
293 struct Scsi_Host *shost = sdev->host;
294 struct scsi_target *starget = scsi_target(sdev);
295
296 scsi_dec_host_busy(shost, cmd);
297
298 if (starget->can_queue > 0)
299 atomic_dec(&starget->target_busy);
300
301 sbitmap_put(&sdev->budget_map, cmd->budget_token);
302 cmd->budget_token = -1;
303}
304
305/*
306 * Kick the queue of SCSI device @sdev if @sdev != current_sdev. Called with
307 * interrupts disabled.
308 */
309static void scsi_kick_sdev_queue(struct scsi_device *sdev, void *data)
310{
311 struct scsi_device *current_sdev = data;
312
313 if (sdev != current_sdev)
314 blk_mq_run_hw_queues(sdev->request_queue, true);
315}
316
317/*
318 * Called for single_lun devices on IO completion. Clear starget_sdev_user,
319 * and call blk_run_queue for all the scsi_devices on the target -
320 * including current_sdev first.
321 *
322 * Called with *no* scsi locks held.
323 */
324static void scsi_single_lun_run(struct scsi_device *current_sdev)
325{
326 struct Scsi_Host *shost = current_sdev->host;
327 struct scsi_target *starget = scsi_target(current_sdev);
328 unsigned long flags;
329
330 spin_lock_irqsave(shost->host_lock, flags);
331 starget->starget_sdev_user = NULL;
332 spin_unlock_irqrestore(shost->host_lock, flags);
333
334 /*
335 * Call blk_run_queue for all LUNs on the target, starting with
336 * current_sdev. We race with others (to set starget_sdev_user),
337 * but in most cases, we will be first. Ideally, each LU on the
338 * target would get some limited time or requests on the target.
339 */
340 blk_mq_run_hw_queues(current_sdev->request_queue,
341 shost->queuecommand_may_block);
342
343 spin_lock_irqsave(shost->host_lock, flags);
344 if (!starget->starget_sdev_user)
345 __starget_for_each_device(starget, current_sdev,
346 scsi_kick_sdev_queue);
347 spin_unlock_irqrestore(shost->host_lock, flags);
348}
349
350static inline bool scsi_device_is_busy(struct scsi_device *sdev)
351{
352 if (scsi_device_busy(sdev) >= sdev->queue_depth)
353 return true;
354 if (atomic_read(&sdev->device_blocked) > 0)
355 return true;
356 return false;
357}
358
359static inline bool scsi_target_is_busy(struct scsi_target *starget)
360{
361 if (starget->can_queue > 0) {
362 if (atomic_read(&starget->target_busy) >= starget->can_queue)
363 return true;
364 if (atomic_read(&starget->target_blocked) > 0)
365 return true;
366 }
367 return false;
368}
369
370static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
371{
372 if (atomic_read(&shost->host_blocked) > 0)
373 return true;
374 if (shost->host_self_blocked)
375 return true;
376 return false;
377}
378
379static void scsi_starved_list_run(struct Scsi_Host *shost)
380{
381 LIST_HEAD(starved_list);
382 struct scsi_device *sdev;
383 unsigned long flags;
384
385 spin_lock_irqsave(shost->host_lock, flags);
386 list_splice_init(&shost->starved_list, &starved_list);
387
388 while (!list_empty(&starved_list)) {
389 struct request_queue *slq;
390
391 /*
392 * As long as shost is accepting commands and we have
393 * starved queues, call blk_run_queue. scsi_request_fn
394 * drops the queue_lock and can add us back to the
395 * starved_list.
396 *
397 * host_lock protects the starved_list and starved_entry.
398 * scsi_request_fn must get the host_lock before checking
399 * or modifying starved_list or starved_entry.
400 */
401 if (scsi_host_is_busy(shost))
402 break;
403
404 sdev = list_entry(starved_list.next,
405 struct scsi_device, starved_entry);
406 list_del_init(&sdev->starved_entry);
407 if (scsi_target_is_busy(scsi_target(sdev))) {
408 list_move_tail(&sdev->starved_entry,
409 &shost->starved_list);
410 continue;
411 }
412
413 /*
414 * Once we drop the host lock, a racing scsi_remove_device()
415 * call may remove the sdev from the starved list and destroy
416 * it and the queue. Mitigate by taking a reference to the
417 * queue and never touching the sdev again after we drop the
418 * host lock. Note: if __scsi_remove_device() invokes
419 * blk_mq_destroy_queue() before the queue is run from this
420 * function then blk_run_queue() will return immediately since
421 * blk_mq_destroy_queue() marks the queue with QUEUE_FLAG_DYING.
422 */
423 slq = sdev->request_queue;
424 if (!blk_get_queue(slq))
425 continue;
426 spin_unlock_irqrestore(shost->host_lock, flags);
427
428 blk_mq_run_hw_queues(slq, false);
429 blk_put_queue(slq);
430
431 spin_lock_irqsave(shost->host_lock, flags);
432 }
433 /* put any unprocessed entries back */
434 list_splice(&starved_list, &shost->starved_list);
435 spin_unlock_irqrestore(shost->host_lock, flags);
436}
437
438/**
439 * scsi_run_queue - Select a proper request queue to serve next.
440 * @q: last request's queue
441 *
442 * The previous command was completely finished, start a new one if possible.
443 */
444static void scsi_run_queue(struct request_queue *q)
445{
446 struct scsi_device *sdev = q->queuedata;
447
448 if (scsi_target(sdev)->single_lun)
449 scsi_single_lun_run(sdev);
450 if (!list_empty(&sdev->host->starved_list))
451 scsi_starved_list_run(sdev->host);
452
453 /* Note: blk_mq_kick_requeue_list() runs the queue asynchronously. */
454 blk_mq_kick_requeue_list(q);
455}
456
457void scsi_requeue_run_queue(struct work_struct *work)
458{
459 struct scsi_device *sdev;
460 struct request_queue *q;
461
462 sdev = container_of(work, struct scsi_device, requeue_work);
463 q = sdev->request_queue;
464 scsi_run_queue(q);
465}
466
467void scsi_run_host_queues(struct Scsi_Host *shost)
468{
469 struct scsi_device *sdev;
470
471 shost_for_each_device(sdev, shost)
472 scsi_run_queue(sdev->request_queue);
473}
474
475static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
476{
477 if (!blk_rq_is_passthrough(scsi_cmd_to_rq(cmd))) {
478 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
479
480 if (drv->uninit_command)
481 drv->uninit_command(cmd);
482 }
483}
484
485void scsi_free_sgtables(struct scsi_cmnd *cmd)
486{
487 if (cmd->sdb.table.nents)
488 sg_free_table_chained(&cmd->sdb.table,
489 SCSI_INLINE_SG_CNT);
490 if (scsi_prot_sg_count(cmd))
491 sg_free_table_chained(&cmd->prot_sdb->table,
492 SCSI_INLINE_PROT_SG_CNT);
493}
494EXPORT_SYMBOL_GPL(scsi_free_sgtables);
495
496static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
497{
498 scsi_free_sgtables(cmd);
499 scsi_uninit_cmd(cmd);
500}
501
502static void scsi_run_queue_async(struct scsi_device *sdev)
503{
504 if (scsi_host_in_recovery(sdev->host))
505 return;
506
507 if (scsi_target(sdev)->single_lun ||
508 !list_empty(&sdev->host->starved_list)) {
509 kblockd_schedule_work(&sdev->requeue_work);
510 } else {
511 /*
512 * smp_mb() present in sbitmap_queue_clear() or implied in
513 * .end_io is for ordering writing .device_busy in
514 * scsi_device_unbusy() and reading sdev->restarts.
515 */
516 int old = atomic_read(&sdev->restarts);
517
518 /*
519 * ->restarts has to be kept as non-zero if new budget
520 * contention occurs.
521 *
522 * No need to run queue when either another re-run
523 * queue wins in updating ->restarts or a new budget
524 * contention occurs.
525 */
526 if (old && atomic_cmpxchg(&sdev->restarts, old, 0) == old)
527 blk_mq_run_hw_queues(sdev->request_queue, true);
528 }
529}
530
531/* Returns false when no more bytes to process, true if there are more */
532static bool scsi_end_request(struct request *req, blk_status_t error,
533 unsigned int bytes)
534{
535 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
536 struct scsi_device *sdev = cmd->device;
537 struct request_queue *q = sdev->request_queue;
538
539 if (blk_update_request(req, error, bytes))
540 return true;
541
542 // XXX:
543 if (blk_queue_add_random(q))
544 add_disk_randomness(req->q->disk);
545
546 if (!blk_rq_is_passthrough(req)) {
547 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
548 cmd->flags &= ~SCMD_INITIALIZED;
549 }
550
551 /*
552 * Calling rcu_barrier() is not necessary here because the
553 * SCSI error handler guarantees that the function called by
554 * call_rcu() has been called before scsi_end_request() is
555 * called.
556 */
557 destroy_rcu_head(&cmd->rcu);
558
559 /*
560 * In the MQ case the command gets freed by __blk_mq_end_request,
561 * so we have to do all cleanup that depends on it earlier.
562 *
563 * We also can't kick the queues from irq context, so we
564 * will have to defer it to a workqueue.
565 */
566 scsi_mq_uninit_cmd(cmd);
567
568 /*
569 * queue is still alive, so grab the ref for preventing it
570 * from being cleaned up during running queue.
571 */
572 percpu_ref_get(&q->q_usage_counter);
573
574 __blk_mq_end_request(req, error);
575
576 scsi_run_queue_async(sdev);
577
578 percpu_ref_put(&q->q_usage_counter);
579 return false;
580}
581
582/**
583 * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
584 * @result: scsi error code
585 *
586 * Translate a SCSI result code into a blk_status_t value.
587 */
588static blk_status_t scsi_result_to_blk_status(int result)
589{
590 /*
591 * Check the scsi-ml byte first in case we converted a host or status
592 * byte.
593 */
594 switch (scsi_ml_byte(result)) {
595 case SCSIML_STAT_OK:
596 break;
597 case SCSIML_STAT_RESV_CONFLICT:
598 return BLK_STS_RESV_CONFLICT;
599 case SCSIML_STAT_NOSPC:
600 return BLK_STS_NOSPC;
601 case SCSIML_STAT_MED_ERROR:
602 return BLK_STS_MEDIUM;
603 case SCSIML_STAT_TGT_FAILURE:
604 return BLK_STS_TARGET;
605 case SCSIML_STAT_DL_TIMEOUT:
606 return BLK_STS_DURATION_LIMIT;
607 }
608
609 switch (host_byte(result)) {
610 case DID_OK:
611 if (scsi_status_is_good(result))
612 return BLK_STS_OK;
613 return BLK_STS_IOERR;
614 case DID_TRANSPORT_FAILFAST:
615 case DID_TRANSPORT_MARGINAL:
616 return BLK_STS_TRANSPORT;
617 default:
618 return BLK_STS_IOERR;
619 }
620}
621
622/**
623 * scsi_rq_err_bytes - determine number of bytes till the next failure boundary
624 * @rq: request to examine
625 *
626 * Description:
627 * A request could be merge of IOs which require different failure
628 * handling. This function determines the number of bytes which
629 * can be failed from the beginning of the request without
630 * crossing into area which need to be retried further.
631 *
632 * Return:
633 * The number of bytes to fail.
634 */
635static unsigned int scsi_rq_err_bytes(const struct request *rq)
636{
637 blk_opf_t ff = rq->cmd_flags & REQ_FAILFAST_MASK;
638 unsigned int bytes = 0;
639 struct bio *bio;
640
641 if (!(rq->rq_flags & RQF_MIXED_MERGE))
642 return blk_rq_bytes(rq);
643
644 /*
645 * Currently the only 'mixing' which can happen is between
646 * different fastfail types. We can safely fail portions
647 * which have all the failfast bits that the first one has -
648 * the ones which are at least as eager to fail as the first
649 * one.
650 */
651 for (bio = rq->bio; bio; bio = bio->bi_next) {
652 if ((bio->bi_opf & ff) != ff)
653 break;
654 bytes += bio->bi_iter.bi_size;
655 }
656
657 /* this could lead to infinite loop */
658 BUG_ON(blk_rq_bytes(rq) && !bytes);
659 return bytes;
660}
661
662static bool scsi_cmd_runtime_exceeced(struct scsi_cmnd *cmd)
663{
664 struct request *req = scsi_cmd_to_rq(cmd);
665 unsigned long wait_for;
666
667 if (cmd->allowed == SCSI_CMD_RETRIES_NO_LIMIT)
668 return false;
669
670 wait_for = (cmd->allowed + 1) * req->timeout;
671 if (time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
672 scmd_printk(KERN_ERR, cmd, "timing out command, waited %lus\n",
673 wait_for/HZ);
674 return true;
675 }
676 return false;
677}
678
679/*
680 * When ALUA transition state is returned, reprep the cmd to
681 * use the ALUA handler's transition timeout. Delay the reprep
682 * 1 sec to avoid aggressive retries of the target in that
683 * state.
684 */
685#define ALUA_TRANSITION_REPREP_DELAY 1000
686
687/* Helper for scsi_io_completion() when special action required. */
688static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
689{
690 struct request *req = scsi_cmd_to_rq(cmd);
691 int level = 0;
692 enum {ACTION_FAIL, ACTION_REPREP, ACTION_DELAYED_REPREP,
693 ACTION_RETRY, ACTION_DELAYED_RETRY} action;
694 struct scsi_sense_hdr sshdr;
695 bool sense_valid;
696 bool sense_current = true; /* false implies "deferred sense" */
697 blk_status_t blk_stat;
698
699 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
700 if (sense_valid)
701 sense_current = !scsi_sense_is_deferred(&sshdr);
702
703 blk_stat = scsi_result_to_blk_status(result);
704
705 if (host_byte(result) == DID_RESET) {
706 /* Third party bus reset or reset for error recovery
707 * reasons. Just retry the command and see what
708 * happens.
709 */
710 action = ACTION_RETRY;
711 } else if (sense_valid && sense_current) {
712 switch (sshdr.sense_key) {
713 case UNIT_ATTENTION:
714 if (cmd->device->removable) {
715 /* Detected disc change. Set a bit
716 * and quietly refuse further access.
717 */
718 cmd->device->changed = 1;
719 action = ACTION_FAIL;
720 } else {
721 /* Must have been a power glitch, or a
722 * bus reset. Could not have been a
723 * media change, so we just retry the
724 * command and see what happens.
725 */
726 action = ACTION_RETRY;
727 }
728 break;
729 case ILLEGAL_REQUEST:
730 /* If we had an ILLEGAL REQUEST returned, then
731 * we may have performed an unsupported
732 * command. The only thing this should be
733 * would be a ten byte read where only a six
734 * byte read was supported. Also, on a system
735 * where READ CAPACITY failed, we may have
736 * read past the end of the disk.
737 */
738 if ((cmd->device->use_10_for_rw &&
739 sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
740 (cmd->cmnd[0] == READ_10 ||
741 cmd->cmnd[0] == WRITE_10)) {
742 /* This will issue a new 6-byte command. */
743 cmd->device->use_10_for_rw = 0;
744 action = ACTION_REPREP;
745 } else if (sshdr.asc == 0x10) /* DIX */ {
746 action = ACTION_FAIL;
747 blk_stat = BLK_STS_PROTECTION;
748 /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
749 } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
750 action = ACTION_FAIL;
751 blk_stat = BLK_STS_TARGET;
752 } else
753 action = ACTION_FAIL;
754 break;
755 case ABORTED_COMMAND:
756 action = ACTION_FAIL;
757 if (sshdr.asc == 0x10) /* DIF */
758 blk_stat = BLK_STS_PROTECTION;
759 break;
760 case NOT_READY:
761 /* If the device is in the process of becoming
762 * ready, or has a temporary blockage, retry.
763 */
764 if (sshdr.asc == 0x04) {
765 switch (sshdr.ascq) {
766 case 0x01: /* becoming ready */
767 case 0x04: /* format in progress */
768 case 0x05: /* rebuild in progress */
769 case 0x06: /* recalculation in progress */
770 case 0x07: /* operation in progress */
771 case 0x08: /* Long write in progress */
772 case 0x09: /* self test in progress */
773 case 0x11: /* notify (enable spinup) required */
774 case 0x14: /* space allocation in progress */
775 case 0x1a: /* start stop unit in progress */
776 case 0x1b: /* sanitize in progress */
777 case 0x1d: /* configuration in progress */
778 case 0x24: /* depopulation in progress */
779 case 0x25: /* depopulation restore in progress */
780 action = ACTION_DELAYED_RETRY;
781 break;
782 case 0x0a: /* ALUA state transition */
783 action = ACTION_DELAYED_REPREP;
784 break;
785 default:
786 action = ACTION_FAIL;
787 break;
788 }
789 } else
790 action = ACTION_FAIL;
791 break;
792 case VOLUME_OVERFLOW:
793 /* See SSC3rXX or current. */
794 action = ACTION_FAIL;
795 break;
796 case DATA_PROTECT:
797 action = ACTION_FAIL;
798 if ((sshdr.asc == 0x0C && sshdr.ascq == 0x12) ||
799 (sshdr.asc == 0x55 &&
800 (sshdr.ascq == 0x0E || sshdr.ascq == 0x0F))) {
801 /* Insufficient zone resources */
802 blk_stat = BLK_STS_ZONE_OPEN_RESOURCE;
803 }
804 break;
805 case COMPLETED:
806 fallthrough;
807 default:
808 action = ACTION_FAIL;
809 break;
810 }
811 } else
812 action = ACTION_FAIL;
813
814 if (action != ACTION_FAIL && scsi_cmd_runtime_exceeced(cmd))
815 action = ACTION_FAIL;
816
817 switch (action) {
818 case ACTION_FAIL:
819 /* Give up and fail the remainder of the request */
820 if (!(req->rq_flags & RQF_QUIET)) {
821 static DEFINE_RATELIMIT_STATE(_rs,
822 DEFAULT_RATELIMIT_INTERVAL,
823 DEFAULT_RATELIMIT_BURST);
824
825 if (unlikely(scsi_logging_level))
826 level =
827 SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
828 SCSI_LOG_MLCOMPLETE_BITS);
829
830 /*
831 * if logging is enabled the failure will be printed
832 * in scsi_log_completion(), so avoid duplicate messages
833 */
834 if (!level && __ratelimit(&_rs)) {
835 scsi_print_result(cmd, NULL, FAILED);
836 if (sense_valid)
837 scsi_print_sense(cmd);
838 scsi_print_command(cmd);
839 }
840 }
841 if (!scsi_end_request(req, blk_stat, scsi_rq_err_bytes(req)))
842 return;
843 fallthrough;
844 case ACTION_REPREP:
845 scsi_mq_requeue_cmd(cmd, 0);
846 break;
847 case ACTION_DELAYED_REPREP:
848 scsi_mq_requeue_cmd(cmd, ALUA_TRANSITION_REPREP_DELAY);
849 break;
850 case ACTION_RETRY:
851 /* Retry the same command immediately */
852 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
853 break;
854 case ACTION_DELAYED_RETRY:
855 /* Retry the same command after a delay */
856 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
857 break;
858 }
859}
860
861/*
862 * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
863 * new result that may suppress further error checking. Also modifies
864 * *blk_statp in some cases.
865 */
866static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
867 blk_status_t *blk_statp)
868{
869 bool sense_valid;
870 bool sense_current = true; /* false implies "deferred sense" */
871 struct request *req = scsi_cmd_to_rq(cmd);
872 struct scsi_sense_hdr sshdr;
873
874 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
875 if (sense_valid)
876 sense_current = !scsi_sense_is_deferred(&sshdr);
877
878 if (blk_rq_is_passthrough(req)) {
879 if (sense_valid) {
880 /*
881 * SG_IO wants current and deferred errors
882 */
883 cmd->sense_len = min(8 + cmd->sense_buffer[7],
884 SCSI_SENSE_BUFFERSIZE);
885 }
886 if (sense_current)
887 *blk_statp = scsi_result_to_blk_status(result);
888 } else if (blk_rq_bytes(req) == 0 && sense_current) {
889 /*
890 * Flush commands do not transfers any data, and thus cannot use
891 * good_bytes != blk_rq_bytes(req) as the signal for an error.
892 * This sets *blk_statp explicitly for the problem case.
893 */
894 *blk_statp = scsi_result_to_blk_status(result);
895 }
896 /*
897 * Recovered errors need reporting, but they're always treated as
898 * success, so fiddle the result code here. For passthrough requests
899 * we already took a copy of the original into sreq->result which
900 * is what gets returned to the user
901 */
902 if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
903 bool do_print = true;
904 /*
905 * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
906 * skip print since caller wants ATA registers. Only occurs
907 * on SCSI ATA PASS_THROUGH commands when CK_COND=1
908 */
909 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
910 do_print = false;
911 else if (req->rq_flags & RQF_QUIET)
912 do_print = false;
913 if (do_print)
914 scsi_print_sense(cmd);
915 result = 0;
916 /* for passthrough, *blk_statp may be set */
917 *blk_statp = BLK_STS_OK;
918 }
919 /*
920 * Another corner case: the SCSI status byte is non-zero but 'good'.
921 * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
922 * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
923 * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
924 * intermediate statuses (both obsolete in SAM-4) as good.
925 */
926 if ((result & 0xff) && scsi_status_is_good(result)) {
927 result = 0;
928 *blk_statp = BLK_STS_OK;
929 }
930 return result;
931}
932
933/**
934 * scsi_io_completion - Completion processing for SCSI commands.
935 * @cmd: command that is finished.
936 * @good_bytes: number of processed bytes.
937 *
938 * We will finish off the specified number of sectors. If we are done, the
939 * command block will be released and the queue function will be goosed. If we
940 * are not done then we have to figure out what to do next:
941 *
942 * a) We can call scsi_mq_requeue_cmd(). The request will be
943 * unprepared and put back on the queue. Then a new command will
944 * be created for it. This should be used if we made forward
945 * progress, or if we want to switch from READ(10) to READ(6) for
946 * example.
947 *
948 * b) We can call scsi_io_completion_action(). The request will be
949 * put back on the queue and retried using the same command as
950 * before, possibly after a delay.
951 *
952 * c) We can call scsi_end_request() with blk_stat other than
953 * BLK_STS_OK, to fail the remainder of the request.
954 */
955void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
956{
957 int result = cmd->result;
958 struct request *req = scsi_cmd_to_rq(cmd);
959 blk_status_t blk_stat = BLK_STS_OK;
960
961 if (unlikely(result)) /* a nz result may or may not be an error */
962 result = scsi_io_completion_nz_result(cmd, result, &blk_stat);
963
964 /*
965 * Next deal with any sectors which we were able to correctly
966 * handle.
967 */
968 SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
969 "%u sectors total, %d bytes done.\n",
970 blk_rq_sectors(req), good_bytes));
971
972 /*
973 * Failed, zero length commands always need to drop down
974 * to retry code. Fast path should return in this block.
975 */
976 if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
977 if (likely(!scsi_end_request(req, blk_stat, good_bytes)))
978 return; /* no bytes remaining */
979 }
980
981 /* Kill remainder if no retries. */
982 if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
983 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req)))
984 WARN_ONCE(true,
985 "Bytes remaining after failed, no-retry command");
986 return;
987 }
988
989 /*
990 * If there had been no error, but we have leftover bytes in the
991 * request just queue the command up again.
992 */
993 if (likely(result == 0))
994 scsi_mq_requeue_cmd(cmd, 0);
995 else
996 scsi_io_completion_action(cmd, result);
997}
998
999static inline bool scsi_cmd_needs_dma_drain(struct scsi_device *sdev,
1000 struct request *rq)
1001{
1002 return sdev->dma_drain_len && blk_rq_is_passthrough(rq) &&
1003 !op_is_write(req_op(rq)) &&
1004 sdev->host->hostt->dma_need_drain(rq);
1005}
1006
1007/**
1008 * scsi_alloc_sgtables - Allocate and initialize data and integrity scatterlists
1009 * @cmd: SCSI command data structure to initialize.
1010 *
1011 * Initializes @cmd->sdb and also @cmd->prot_sdb if data integrity is enabled
1012 * for @cmd.
1013 *
1014 * Returns:
1015 * * BLK_STS_OK - on success
1016 * * BLK_STS_RESOURCE - if the failure is retryable
1017 * * BLK_STS_IOERR - if the failure is fatal
1018 */
1019blk_status_t scsi_alloc_sgtables(struct scsi_cmnd *cmd)
1020{
1021 struct scsi_device *sdev = cmd->device;
1022 struct request *rq = scsi_cmd_to_rq(cmd);
1023 unsigned short nr_segs = blk_rq_nr_phys_segments(rq);
1024 struct scatterlist *last_sg = NULL;
1025 blk_status_t ret;
1026 bool need_drain = scsi_cmd_needs_dma_drain(sdev, rq);
1027 int count;
1028
1029 if (WARN_ON_ONCE(!nr_segs))
1030 return BLK_STS_IOERR;
1031
1032 /*
1033 * Make sure there is space for the drain. The driver must adjust
1034 * max_hw_segments to be prepared for this.
1035 */
1036 if (need_drain)
1037 nr_segs++;
1038
1039 /*
1040 * If sg table allocation fails, requeue request later.
1041 */
1042 if (unlikely(sg_alloc_table_chained(&cmd->sdb.table, nr_segs,
1043 cmd->sdb.table.sgl, SCSI_INLINE_SG_CNT)))
1044 return BLK_STS_RESOURCE;
1045
1046 /*
1047 * Next, walk the list, and fill in the addresses and sizes of
1048 * each segment.
1049 */
1050 count = __blk_rq_map_sg(rq->q, rq, cmd->sdb.table.sgl, &last_sg);
1051
1052 if (blk_rq_bytes(rq) & rq->q->dma_pad_mask) {
1053 unsigned int pad_len =
1054 (rq->q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
1055
1056 last_sg->length += pad_len;
1057 cmd->extra_len += pad_len;
1058 }
1059
1060 if (need_drain) {
1061 sg_unmark_end(last_sg);
1062 last_sg = sg_next(last_sg);
1063 sg_set_buf(last_sg, sdev->dma_drain_buf, sdev->dma_drain_len);
1064 sg_mark_end(last_sg);
1065
1066 cmd->extra_len += sdev->dma_drain_len;
1067 count++;
1068 }
1069
1070 BUG_ON(count > cmd->sdb.table.nents);
1071 cmd->sdb.table.nents = count;
1072 cmd->sdb.length = blk_rq_payload_bytes(rq);
1073
1074 if (blk_integrity_rq(rq)) {
1075 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1076 int ivecs;
1077
1078 if (WARN_ON_ONCE(!prot_sdb)) {
1079 /*
1080 * This can happen if someone (e.g. multipath)
1081 * queues a command to a device on an adapter
1082 * that does not support DIX.
1083 */
1084 ret = BLK_STS_IOERR;
1085 goto out_free_sgtables;
1086 }
1087
1088 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1089
1090 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1091 prot_sdb->table.sgl,
1092 SCSI_INLINE_PROT_SG_CNT)) {
1093 ret = BLK_STS_RESOURCE;
1094 goto out_free_sgtables;
1095 }
1096
1097 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1098 prot_sdb->table.sgl);
1099 BUG_ON(count > ivecs);
1100 BUG_ON(count > queue_max_integrity_segments(rq->q));
1101
1102 cmd->prot_sdb = prot_sdb;
1103 cmd->prot_sdb->table.nents = count;
1104 }
1105
1106 return BLK_STS_OK;
1107out_free_sgtables:
1108 scsi_free_sgtables(cmd);
1109 return ret;
1110}
1111EXPORT_SYMBOL(scsi_alloc_sgtables);
1112
1113/**
1114 * scsi_initialize_rq - initialize struct scsi_cmnd partially
1115 * @rq: Request associated with the SCSI command to be initialized.
1116 *
1117 * This function initializes the members of struct scsi_cmnd that must be
1118 * initialized before request processing starts and that won't be
1119 * reinitialized if a SCSI command is requeued.
1120 */
1121static void scsi_initialize_rq(struct request *rq)
1122{
1123 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1124
1125 memset(cmd->cmnd, 0, sizeof(cmd->cmnd));
1126 cmd->cmd_len = MAX_COMMAND_SIZE;
1127 cmd->sense_len = 0;
1128 init_rcu_head(&cmd->rcu);
1129 cmd->jiffies_at_alloc = jiffies;
1130 cmd->retries = 0;
1131}
1132
1133struct request *scsi_alloc_request(struct request_queue *q, blk_opf_t opf,
1134 blk_mq_req_flags_t flags)
1135{
1136 struct request *rq;
1137
1138 rq = blk_mq_alloc_request(q, opf, flags);
1139 if (!IS_ERR(rq))
1140 scsi_initialize_rq(rq);
1141 return rq;
1142}
1143EXPORT_SYMBOL_GPL(scsi_alloc_request);
1144
1145/*
1146 * Only called when the request isn't completed by SCSI, and not freed by
1147 * SCSI
1148 */
1149static void scsi_cleanup_rq(struct request *rq)
1150{
1151 if (rq->rq_flags & RQF_DONTPREP) {
1152 scsi_mq_uninit_cmd(blk_mq_rq_to_pdu(rq));
1153 rq->rq_flags &= ~RQF_DONTPREP;
1154 }
1155}
1156
1157/* Called before a request is prepared. See also scsi_mq_prep_fn(). */
1158void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1159{
1160 struct request *rq = scsi_cmd_to_rq(cmd);
1161
1162 if (!blk_rq_is_passthrough(rq) && !(cmd->flags & SCMD_INITIALIZED)) {
1163 cmd->flags |= SCMD_INITIALIZED;
1164 scsi_initialize_rq(rq);
1165 }
1166
1167 cmd->device = dev;
1168 INIT_LIST_HEAD(&cmd->eh_entry);
1169 INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1170}
1171
1172static blk_status_t scsi_setup_scsi_cmnd(struct scsi_device *sdev,
1173 struct request *req)
1174{
1175 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1176
1177 /*
1178 * Passthrough requests may transfer data, in which case they must
1179 * a bio attached to them. Or they might contain a SCSI command
1180 * that does not transfer data, in which case they may optionally
1181 * submit a request without an attached bio.
1182 */
1183 if (req->bio) {
1184 blk_status_t ret = scsi_alloc_sgtables(cmd);
1185 if (unlikely(ret != BLK_STS_OK))
1186 return ret;
1187 } else {
1188 BUG_ON(blk_rq_bytes(req));
1189
1190 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1191 }
1192
1193 cmd->transfersize = blk_rq_bytes(req);
1194 return BLK_STS_OK;
1195}
1196
1197static blk_status_t
1198scsi_device_state_check(struct scsi_device *sdev, struct request *req)
1199{
1200 switch (sdev->sdev_state) {
1201 case SDEV_CREATED:
1202 return BLK_STS_OK;
1203 case SDEV_OFFLINE:
1204 case SDEV_TRANSPORT_OFFLINE:
1205 /*
1206 * If the device is offline we refuse to process any
1207 * commands. The device must be brought online
1208 * before trying any recovery commands.
1209 */
1210 if (!sdev->offline_already) {
1211 sdev->offline_already = true;
1212 sdev_printk(KERN_ERR, sdev,
1213 "rejecting I/O to offline device\n");
1214 }
1215 return BLK_STS_IOERR;
1216 case SDEV_DEL:
1217 /*
1218 * If the device is fully deleted, we refuse to
1219 * process any commands as well.
1220 */
1221 sdev_printk(KERN_ERR, sdev,
1222 "rejecting I/O to dead device\n");
1223 return BLK_STS_IOERR;
1224 case SDEV_BLOCK:
1225 case SDEV_CREATED_BLOCK:
1226 return BLK_STS_RESOURCE;
1227 case SDEV_QUIESCE:
1228 /*
1229 * If the device is blocked we only accept power management
1230 * commands.
1231 */
1232 if (req && WARN_ON_ONCE(!(req->rq_flags & RQF_PM)))
1233 return BLK_STS_RESOURCE;
1234 return BLK_STS_OK;
1235 default:
1236 /*
1237 * For any other not fully online state we only allow
1238 * power management commands.
1239 */
1240 if (req && !(req->rq_flags & RQF_PM))
1241 return BLK_STS_OFFLINE;
1242 return BLK_STS_OK;
1243 }
1244}
1245
1246/*
1247 * scsi_dev_queue_ready: if we can send requests to sdev, assign one token
1248 * and return the token else return -1.
1249 */
1250static inline int scsi_dev_queue_ready(struct request_queue *q,
1251 struct scsi_device *sdev)
1252{
1253 int token;
1254
1255 token = sbitmap_get(&sdev->budget_map);
1256 if (token < 0)
1257 return -1;
1258
1259 if (!atomic_read(&sdev->device_blocked))
1260 return token;
1261
1262 /*
1263 * Only unblock if no other commands are pending and
1264 * if device_blocked has decreased to zero
1265 */
1266 if (scsi_device_busy(sdev) > 1 ||
1267 atomic_dec_return(&sdev->device_blocked) > 0) {
1268 sbitmap_put(&sdev->budget_map, token);
1269 return -1;
1270 }
1271
1272 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1273 "unblocking device at zero depth\n"));
1274
1275 return token;
1276}
1277
1278/*
1279 * scsi_target_queue_ready: checks if there we can send commands to target
1280 * @sdev: scsi device on starget to check.
1281 */
1282static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1283 struct scsi_device *sdev)
1284{
1285 struct scsi_target *starget = scsi_target(sdev);
1286 unsigned int busy;
1287
1288 if (starget->single_lun) {
1289 spin_lock_irq(shost->host_lock);
1290 if (starget->starget_sdev_user &&
1291 starget->starget_sdev_user != sdev) {
1292 spin_unlock_irq(shost->host_lock);
1293 return 0;
1294 }
1295 starget->starget_sdev_user = sdev;
1296 spin_unlock_irq(shost->host_lock);
1297 }
1298
1299 if (starget->can_queue <= 0)
1300 return 1;
1301
1302 busy = atomic_inc_return(&starget->target_busy) - 1;
1303 if (atomic_read(&starget->target_blocked) > 0) {
1304 if (busy)
1305 goto starved;
1306
1307 /*
1308 * unblock after target_blocked iterates to zero
1309 */
1310 if (atomic_dec_return(&starget->target_blocked) > 0)
1311 goto out_dec;
1312
1313 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1314 "unblocking target at zero depth\n"));
1315 }
1316
1317 if (busy >= starget->can_queue)
1318 goto starved;
1319
1320 return 1;
1321
1322starved:
1323 spin_lock_irq(shost->host_lock);
1324 list_move_tail(&sdev->starved_entry, &shost->starved_list);
1325 spin_unlock_irq(shost->host_lock);
1326out_dec:
1327 if (starget->can_queue > 0)
1328 atomic_dec(&starget->target_busy);
1329 return 0;
1330}
1331
1332/*
1333 * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1334 * return 0. We must end up running the queue again whenever 0 is
1335 * returned, else IO can hang.
1336 */
1337static inline int scsi_host_queue_ready(struct request_queue *q,
1338 struct Scsi_Host *shost,
1339 struct scsi_device *sdev,
1340 struct scsi_cmnd *cmd)
1341{
1342 if (atomic_read(&shost->host_blocked) > 0) {
1343 if (scsi_host_busy(shost) > 0)
1344 goto starved;
1345
1346 /*
1347 * unblock after host_blocked iterates to zero
1348 */
1349 if (atomic_dec_return(&shost->host_blocked) > 0)
1350 goto out_dec;
1351
1352 SCSI_LOG_MLQUEUE(3,
1353 shost_printk(KERN_INFO, shost,
1354 "unblocking host at zero depth\n"));
1355 }
1356
1357 if (shost->host_self_blocked)
1358 goto starved;
1359
1360 /* We're OK to process the command, so we can't be starved */
1361 if (!list_empty(&sdev->starved_entry)) {
1362 spin_lock_irq(shost->host_lock);
1363 if (!list_empty(&sdev->starved_entry))
1364 list_del_init(&sdev->starved_entry);
1365 spin_unlock_irq(shost->host_lock);
1366 }
1367
1368 __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1369
1370 return 1;
1371
1372starved:
1373 spin_lock_irq(shost->host_lock);
1374 if (list_empty(&sdev->starved_entry))
1375 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1376 spin_unlock_irq(shost->host_lock);
1377out_dec:
1378 scsi_dec_host_busy(shost, cmd);
1379 return 0;
1380}
1381
1382/*
1383 * Busy state exporting function for request stacking drivers.
1384 *
1385 * For efficiency, no lock is taken to check the busy state of
1386 * shost/starget/sdev, since the returned value is not guaranteed and
1387 * may be changed after request stacking drivers call the function,
1388 * regardless of taking lock or not.
1389 *
1390 * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1391 * needs to return 'not busy'. Otherwise, request stacking drivers
1392 * may hold requests forever.
1393 */
1394static bool scsi_mq_lld_busy(struct request_queue *q)
1395{
1396 struct scsi_device *sdev = q->queuedata;
1397 struct Scsi_Host *shost;
1398
1399 if (blk_queue_dying(q))
1400 return false;
1401
1402 shost = sdev->host;
1403
1404 /*
1405 * Ignore host/starget busy state.
1406 * Since block layer does not have a concept of fairness across
1407 * multiple queues, congestion of host/starget needs to be handled
1408 * in SCSI layer.
1409 */
1410 if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1411 return true;
1412
1413 return false;
1414}
1415
1416/*
1417 * Block layer request completion callback. May be called from interrupt
1418 * context.
1419 */
1420static void scsi_complete(struct request *rq)
1421{
1422 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1423 enum scsi_disposition disposition;
1424
1425 INIT_LIST_HEAD(&cmd->eh_entry);
1426
1427 atomic_inc(&cmd->device->iodone_cnt);
1428 if (cmd->result)
1429 atomic_inc(&cmd->device->ioerr_cnt);
1430
1431 disposition = scsi_decide_disposition(cmd);
1432 if (disposition != SUCCESS && scsi_cmd_runtime_exceeced(cmd))
1433 disposition = SUCCESS;
1434
1435 scsi_log_completion(cmd, disposition);
1436
1437 switch (disposition) {
1438 case SUCCESS:
1439 scsi_finish_command(cmd);
1440 break;
1441 case NEEDS_RETRY:
1442 scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1443 break;
1444 case ADD_TO_MLQUEUE:
1445 scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1446 break;
1447 default:
1448 scsi_eh_scmd_add(cmd);
1449 break;
1450 }
1451}
1452
1453/**
1454 * scsi_dispatch_cmd - Dispatch a command to the low-level driver.
1455 * @cmd: command block we are dispatching.
1456 *
1457 * Return: nonzero return request was rejected and device's queue needs to be
1458 * plugged.
1459 */
1460static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1461{
1462 struct Scsi_Host *host = cmd->device->host;
1463 int rtn = 0;
1464
1465 atomic_inc(&cmd->device->iorequest_cnt);
1466
1467 /* check if the device is still usable */
1468 if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1469 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1470 * returns an immediate error upwards, and signals
1471 * that the device is no longer present */
1472 cmd->result = DID_NO_CONNECT << 16;
1473 goto done;
1474 }
1475
1476 /* Check to see if the scsi lld made this device blocked. */
1477 if (unlikely(scsi_device_blocked(cmd->device))) {
1478 /*
1479 * in blocked state, the command is just put back on
1480 * the device queue. The suspend state has already
1481 * blocked the queue so future requests should not
1482 * occur until the device transitions out of the
1483 * suspend state.
1484 */
1485 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1486 "queuecommand : device blocked\n"));
1487 atomic_dec(&cmd->device->iorequest_cnt);
1488 return SCSI_MLQUEUE_DEVICE_BUSY;
1489 }
1490
1491 /* Store the LUN value in cmnd, if needed. */
1492 if (cmd->device->lun_in_cdb)
1493 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1494 (cmd->device->lun << 5 & 0xe0);
1495
1496 scsi_log_send(cmd);
1497
1498 /*
1499 * Before we queue this command, check if the command
1500 * length exceeds what the host adapter can handle.
1501 */
1502 if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1503 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1504 "queuecommand : command too long. "
1505 "cdb_size=%d host->max_cmd_len=%d\n",
1506 cmd->cmd_len, cmd->device->host->max_cmd_len));
1507 cmd->result = (DID_ABORT << 16);
1508 goto done;
1509 }
1510
1511 if (unlikely(host->shost_state == SHOST_DEL)) {
1512 cmd->result = (DID_NO_CONNECT << 16);
1513 goto done;
1514
1515 }
1516
1517 trace_scsi_dispatch_cmd_start(cmd);
1518 rtn = host->hostt->queuecommand(host, cmd);
1519 if (rtn) {
1520 atomic_dec(&cmd->device->iorequest_cnt);
1521 trace_scsi_dispatch_cmd_error(cmd, rtn);
1522 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1523 rtn != SCSI_MLQUEUE_TARGET_BUSY)
1524 rtn = SCSI_MLQUEUE_HOST_BUSY;
1525
1526 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1527 "queuecommand : request rejected\n"));
1528 }
1529
1530 return rtn;
1531 done:
1532 scsi_done(cmd);
1533 return 0;
1534}
1535
1536/* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1537static unsigned int scsi_mq_inline_sgl_size(struct Scsi_Host *shost)
1538{
1539 return min_t(unsigned int, shost->sg_tablesize, SCSI_INLINE_SG_CNT) *
1540 sizeof(struct scatterlist);
1541}
1542
1543static blk_status_t scsi_prepare_cmd(struct request *req)
1544{
1545 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1546 struct scsi_device *sdev = req->q->queuedata;
1547 struct Scsi_Host *shost = sdev->host;
1548 bool in_flight = test_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1549 struct scatterlist *sg;
1550
1551 scsi_init_command(sdev, cmd);
1552
1553 cmd->eh_eflags = 0;
1554 cmd->prot_type = 0;
1555 cmd->prot_flags = 0;
1556 cmd->submitter = 0;
1557 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1558 cmd->underflow = 0;
1559 cmd->transfersize = 0;
1560 cmd->host_scribble = NULL;
1561 cmd->result = 0;
1562 cmd->extra_len = 0;
1563 cmd->state = 0;
1564 if (in_flight)
1565 __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1566
1567 /*
1568 * Only clear the driver-private command data if the LLD does not supply
1569 * a function to initialize that data.
1570 */
1571 if (!shost->hostt->init_cmd_priv)
1572 memset(cmd + 1, 0, shost->hostt->cmd_size);
1573
1574 cmd->prot_op = SCSI_PROT_NORMAL;
1575 if (blk_rq_bytes(req))
1576 cmd->sc_data_direction = rq_dma_dir(req);
1577 else
1578 cmd->sc_data_direction = DMA_NONE;
1579
1580 sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1581 cmd->sdb.table.sgl = sg;
1582
1583 if (scsi_host_get_prot(shost)) {
1584 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1585
1586 cmd->prot_sdb->table.sgl =
1587 (struct scatterlist *)(cmd->prot_sdb + 1);
1588 }
1589
1590 /*
1591 * Special handling for passthrough commands, which don't go to the ULP
1592 * at all:
1593 */
1594 if (blk_rq_is_passthrough(req))
1595 return scsi_setup_scsi_cmnd(sdev, req);
1596
1597 if (sdev->handler && sdev->handler->prep_fn) {
1598 blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1599
1600 if (ret != BLK_STS_OK)
1601 return ret;
1602 }
1603
1604 /* Usually overridden by the ULP */
1605 cmd->allowed = 0;
1606 memset(cmd->cmnd, 0, sizeof(cmd->cmnd));
1607 return scsi_cmd_to_driver(cmd)->init_command(cmd);
1608}
1609
1610static void scsi_done_internal(struct scsi_cmnd *cmd, bool complete_directly)
1611{
1612 struct request *req = scsi_cmd_to_rq(cmd);
1613
1614 switch (cmd->submitter) {
1615 case SUBMITTED_BY_BLOCK_LAYER:
1616 break;
1617 case SUBMITTED_BY_SCSI_ERROR_HANDLER:
1618 return scsi_eh_done(cmd);
1619 case SUBMITTED_BY_SCSI_RESET_IOCTL:
1620 return;
1621 }
1622
1623 if (unlikely(blk_should_fake_timeout(scsi_cmd_to_rq(cmd)->q)))
1624 return;
1625 if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1626 return;
1627 trace_scsi_dispatch_cmd_done(cmd);
1628
1629 if (complete_directly)
1630 blk_mq_complete_request_direct(req, scsi_complete);
1631 else
1632 blk_mq_complete_request(req);
1633}
1634
1635void scsi_done(struct scsi_cmnd *cmd)
1636{
1637 scsi_done_internal(cmd, false);
1638}
1639EXPORT_SYMBOL(scsi_done);
1640
1641void scsi_done_direct(struct scsi_cmnd *cmd)
1642{
1643 scsi_done_internal(cmd, true);
1644}
1645EXPORT_SYMBOL(scsi_done_direct);
1646
1647static void scsi_mq_put_budget(struct request_queue *q, int budget_token)
1648{
1649 struct scsi_device *sdev = q->queuedata;
1650
1651 sbitmap_put(&sdev->budget_map, budget_token);
1652}
1653
1654/*
1655 * When to reinvoke queueing after a resource shortage. It's 3 msecs to
1656 * not change behaviour from the previous unplug mechanism, experimentation
1657 * may prove this needs changing.
1658 */
1659#define SCSI_QUEUE_DELAY 3
1660
1661static int scsi_mq_get_budget(struct request_queue *q)
1662{
1663 struct scsi_device *sdev = q->queuedata;
1664 int token = scsi_dev_queue_ready(q, sdev);
1665
1666 if (token >= 0)
1667 return token;
1668
1669 atomic_inc(&sdev->restarts);
1670
1671 /*
1672 * Orders atomic_inc(&sdev->restarts) and atomic_read(&sdev->device_busy).
1673 * .restarts must be incremented before .device_busy is read because the
1674 * code in scsi_run_queue_async() depends on the order of these operations.
1675 */
1676 smp_mb__after_atomic();
1677
1678 /*
1679 * If all in-flight requests originated from this LUN are completed
1680 * before reading .device_busy, sdev->device_busy will be observed as
1681 * zero, then blk_mq_delay_run_hw_queues() will dispatch this request
1682 * soon. Otherwise, completion of one of these requests will observe
1683 * the .restarts flag, and the request queue will be run for handling
1684 * this request, see scsi_end_request().
1685 */
1686 if (unlikely(scsi_device_busy(sdev) == 0 &&
1687 !scsi_device_blocked(sdev)))
1688 blk_mq_delay_run_hw_queues(sdev->request_queue, SCSI_QUEUE_DELAY);
1689 return -1;
1690}
1691
1692static void scsi_mq_set_rq_budget_token(struct request *req, int token)
1693{
1694 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1695
1696 cmd->budget_token = token;
1697}
1698
1699static int scsi_mq_get_rq_budget_token(struct request *req)
1700{
1701 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1702
1703 return cmd->budget_token;
1704}
1705
1706static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1707 const struct blk_mq_queue_data *bd)
1708{
1709 struct request *req = bd->rq;
1710 struct request_queue *q = req->q;
1711 struct scsi_device *sdev = q->queuedata;
1712 struct Scsi_Host *shost = sdev->host;
1713 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1714 blk_status_t ret;
1715 int reason;
1716
1717 WARN_ON_ONCE(cmd->budget_token < 0);
1718
1719 /*
1720 * If the device is not in running state we will reject some or all
1721 * commands.
1722 */
1723 if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1724 ret = scsi_device_state_check(sdev, req);
1725 if (ret != BLK_STS_OK)
1726 goto out_put_budget;
1727 }
1728
1729 ret = BLK_STS_RESOURCE;
1730 if (!scsi_target_queue_ready(shost, sdev))
1731 goto out_put_budget;
1732 if (unlikely(scsi_host_in_recovery(shost))) {
1733 if (cmd->flags & SCMD_FAIL_IF_RECOVERING)
1734 ret = BLK_STS_OFFLINE;
1735 goto out_dec_target_busy;
1736 }
1737 if (!scsi_host_queue_ready(q, shost, sdev, cmd))
1738 goto out_dec_target_busy;
1739
1740 if (!(req->rq_flags & RQF_DONTPREP)) {
1741 ret = scsi_prepare_cmd(req);
1742 if (ret != BLK_STS_OK)
1743 goto out_dec_host_busy;
1744 req->rq_flags |= RQF_DONTPREP;
1745 } else {
1746 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1747 }
1748
1749 cmd->flags &= SCMD_PRESERVED_FLAGS;
1750 if (sdev->simple_tags)
1751 cmd->flags |= SCMD_TAGGED;
1752 if (bd->last)
1753 cmd->flags |= SCMD_LAST;
1754
1755 scsi_set_resid(cmd, 0);
1756 memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
1757 cmd->submitter = SUBMITTED_BY_BLOCK_LAYER;
1758
1759 blk_mq_start_request(req);
1760 reason = scsi_dispatch_cmd(cmd);
1761 if (reason) {
1762 scsi_set_blocked(cmd, reason);
1763 ret = BLK_STS_RESOURCE;
1764 goto out_dec_host_busy;
1765 }
1766
1767 return BLK_STS_OK;
1768
1769out_dec_host_busy:
1770 scsi_dec_host_busy(shost, cmd);
1771out_dec_target_busy:
1772 if (scsi_target(sdev)->can_queue > 0)
1773 atomic_dec(&scsi_target(sdev)->target_busy);
1774out_put_budget:
1775 scsi_mq_put_budget(q, cmd->budget_token);
1776 cmd->budget_token = -1;
1777 switch (ret) {
1778 case BLK_STS_OK:
1779 break;
1780 case BLK_STS_RESOURCE:
1781 case BLK_STS_ZONE_RESOURCE:
1782 if (scsi_device_blocked(sdev))
1783 ret = BLK_STS_DEV_RESOURCE;
1784 break;
1785 case BLK_STS_AGAIN:
1786 cmd->result = DID_BUS_BUSY << 16;
1787 if (req->rq_flags & RQF_DONTPREP)
1788 scsi_mq_uninit_cmd(cmd);
1789 break;
1790 default:
1791 if (unlikely(!scsi_device_online(sdev)))
1792 cmd->result = DID_NO_CONNECT << 16;
1793 else
1794 cmd->result = DID_ERROR << 16;
1795 /*
1796 * Make sure to release all allocated resources when
1797 * we hit an error, as we will never see this command
1798 * again.
1799 */
1800 if (req->rq_flags & RQF_DONTPREP)
1801 scsi_mq_uninit_cmd(cmd);
1802 scsi_run_queue_async(sdev);
1803 break;
1804 }
1805 return ret;
1806}
1807
1808static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1809 unsigned int hctx_idx, unsigned int numa_node)
1810{
1811 struct Scsi_Host *shost = set->driver_data;
1812 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1813 struct scatterlist *sg;
1814 int ret = 0;
1815
1816 cmd->sense_buffer =
1817 kmem_cache_alloc_node(scsi_sense_cache, GFP_KERNEL, numa_node);
1818 if (!cmd->sense_buffer)
1819 return -ENOMEM;
1820
1821 if (scsi_host_get_prot(shost)) {
1822 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1823 shost->hostt->cmd_size;
1824 cmd->prot_sdb = (void *)sg + scsi_mq_inline_sgl_size(shost);
1825 }
1826
1827 if (shost->hostt->init_cmd_priv) {
1828 ret = shost->hostt->init_cmd_priv(shost, cmd);
1829 if (ret < 0)
1830 kmem_cache_free(scsi_sense_cache, cmd->sense_buffer);
1831 }
1832
1833 return ret;
1834}
1835
1836static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1837 unsigned int hctx_idx)
1838{
1839 struct Scsi_Host *shost = set->driver_data;
1840 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1841
1842 if (shost->hostt->exit_cmd_priv)
1843 shost->hostt->exit_cmd_priv(shost, cmd);
1844 kmem_cache_free(scsi_sense_cache, cmd->sense_buffer);
1845}
1846
1847
1848static int scsi_mq_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
1849{
1850 struct Scsi_Host *shost = hctx->driver_data;
1851
1852 if (shost->hostt->mq_poll)
1853 return shost->hostt->mq_poll(shost, hctx->queue_num);
1854
1855 return 0;
1856}
1857
1858static int scsi_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
1859 unsigned int hctx_idx)
1860{
1861 struct Scsi_Host *shost = data;
1862
1863 hctx->driver_data = shost;
1864 return 0;
1865}
1866
1867static void scsi_map_queues(struct blk_mq_tag_set *set)
1868{
1869 struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
1870
1871 if (shost->hostt->map_queues)
1872 return shost->hostt->map_queues(shost);
1873 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
1874}
1875
1876void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
1877{
1878 struct device *dev = shost->dma_dev;
1879
1880 /*
1881 * this limit is imposed by hardware restrictions
1882 */
1883 blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1884 SG_MAX_SEGMENTS));
1885
1886 if (scsi_host_prot_dma(shost)) {
1887 shost->sg_prot_tablesize =
1888 min_not_zero(shost->sg_prot_tablesize,
1889 (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1890 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1891 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1892 }
1893
1894 blk_queue_max_hw_sectors(q, shost->max_sectors);
1895 blk_queue_segment_boundary(q, shost->dma_boundary);
1896 dma_set_seg_boundary(dev, shost->dma_boundary);
1897
1898 blk_queue_max_segment_size(q, shost->max_segment_size);
1899 blk_queue_virt_boundary(q, shost->virt_boundary_mask);
1900 dma_set_max_seg_size(dev, queue_max_segment_size(q));
1901
1902 /*
1903 * Set a reasonable default alignment: The larger of 32-byte (dword),
1904 * which is a common minimum for HBAs, and the minimum DMA alignment,
1905 * which is set by the platform.
1906 *
1907 * Devices that require a bigger alignment can increase it later.
1908 */
1909 blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
1910}
1911EXPORT_SYMBOL_GPL(__scsi_init_queue);
1912
1913static const struct blk_mq_ops scsi_mq_ops_no_commit = {
1914 .get_budget = scsi_mq_get_budget,
1915 .put_budget = scsi_mq_put_budget,
1916 .queue_rq = scsi_queue_rq,
1917 .complete = scsi_complete,
1918 .timeout = scsi_timeout,
1919#ifdef CONFIG_BLK_DEBUG_FS
1920 .show_rq = scsi_show_rq,
1921#endif
1922 .init_request = scsi_mq_init_request,
1923 .exit_request = scsi_mq_exit_request,
1924 .cleanup_rq = scsi_cleanup_rq,
1925 .busy = scsi_mq_lld_busy,
1926 .map_queues = scsi_map_queues,
1927 .init_hctx = scsi_init_hctx,
1928 .poll = scsi_mq_poll,
1929 .set_rq_budget_token = scsi_mq_set_rq_budget_token,
1930 .get_rq_budget_token = scsi_mq_get_rq_budget_token,
1931};
1932
1933
1934static void scsi_commit_rqs(struct blk_mq_hw_ctx *hctx)
1935{
1936 struct Scsi_Host *shost = hctx->driver_data;
1937
1938 shost->hostt->commit_rqs(shost, hctx->queue_num);
1939}
1940
1941static const struct blk_mq_ops scsi_mq_ops = {
1942 .get_budget = scsi_mq_get_budget,
1943 .put_budget = scsi_mq_put_budget,
1944 .queue_rq = scsi_queue_rq,
1945 .commit_rqs = scsi_commit_rqs,
1946 .complete = scsi_complete,
1947 .timeout = scsi_timeout,
1948#ifdef CONFIG_BLK_DEBUG_FS
1949 .show_rq = scsi_show_rq,
1950#endif
1951 .init_request = scsi_mq_init_request,
1952 .exit_request = scsi_mq_exit_request,
1953 .cleanup_rq = scsi_cleanup_rq,
1954 .busy = scsi_mq_lld_busy,
1955 .map_queues = scsi_map_queues,
1956 .init_hctx = scsi_init_hctx,
1957 .poll = scsi_mq_poll,
1958 .set_rq_budget_token = scsi_mq_set_rq_budget_token,
1959 .get_rq_budget_token = scsi_mq_get_rq_budget_token,
1960};
1961
1962int scsi_mq_setup_tags(struct Scsi_Host *shost)
1963{
1964 unsigned int cmd_size, sgl_size;
1965 struct blk_mq_tag_set *tag_set = &shost->tag_set;
1966
1967 sgl_size = max_t(unsigned int, sizeof(struct scatterlist),
1968 scsi_mq_inline_sgl_size(shost));
1969 cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
1970 if (scsi_host_get_prot(shost))
1971 cmd_size += sizeof(struct scsi_data_buffer) +
1972 sizeof(struct scatterlist) * SCSI_INLINE_PROT_SG_CNT;
1973
1974 memset(tag_set, 0, sizeof(*tag_set));
1975 if (shost->hostt->commit_rqs)
1976 tag_set->ops = &scsi_mq_ops;
1977 else
1978 tag_set->ops = &scsi_mq_ops_no_commit;
1979 tag_set->nr_hw_queues = shost->nr_hw_queues ? : 1;
1980 tag_set->nr_maps = shost->nr_maps ? : 1;
1981 tag_set->queue_depth = shost->can_queue;
1982 tag_set->cmd_size = cmd_size;
1983 tag_set->numa_node = dev_to_node(shost->dma_dev);
1984 tag_set->flags = BLK_MQ_F_SHOULD_MERGE;
1985 tag_set->flags |=
1986 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
1987 if (shost->queuecommand_may_block)
1988 tag_set->flags |= BLK_MQ_F_BLOCKING;
1989 tag_set->driver_data = shost;
1990 if (shost->host_tagset)
1991 tag_set->flags |= BLK_MQ_F_TAG_HCTX_SHARED;
1992
1993 return blk_mq_alloc_tag_set(tag_set);
1994}
1995
1996void scsi_mq_free_tags(struct kref *kref)
1997{
1998 struct Scsi_Host *shost = container_of(kref, typeof(*shost),
1999 tagset_refcnt);
2000
2001 blk_mq_free_tag_set(&shost->tag_set);
2002 complete(&shost->tagset_freed);
2003}
2004
2005/**
2006 * scsi_device_from_queue - return sdev associated with a request_queue
2007 * @q: The request queue to return the sdev from
2008 *
2009 * Return the sdev associated with a request queue or NULL if the
2010 * request_queue does not reference a SCSI device.
2011 */
2012struct scsi_device *scsi_device_from_queue(struct request_queue *q)
2013{
2014 struct scsi_device *sdev = NULL;
2015
2016 if (q->mq_ops == &scsi_mq_ops_no_commit ||
2017 q->mq_ops == &scsi_mq_ops)
2018 sdev = q->queuedata;
2019 if (!sdev || !get_device(&sdev->sdev_gendev))
2020 sdev = NULL;
2021
2022 return sdev;
2023}
2024/*
2025 * pktcdvd should have been integrated into the SCSI layers, but for historical
2026 * reasons like the old IDE driver it isn't. This export allows it to safely
2027 * probe if a given device is a SCSI one and only attach to that.
2028 */
2029#ifdef CONFIG_CDROM_PKTCDVD_MODULE
2030EXPORT_SYMBOL_GPL(scsi_device_from_queue);
2031#endif
2032
2033/**
2034 * scsi_block_requests - Utility function used by low-level drivers to prevent
2035 * further commands from being queued to the device.
2036 * @shost: host in question
2037 *
2038 * There is no timer nor any other means by which the requests get unblocked
2039 * other than the low-level driver calling scsi_unblock_requests().
2040 */
2041void scsi_block_requests(struct Scsi_Host *shost)
2042{
2043 shost->host_self_blocked = 1;
2044}
2045EXPORT_SYMBOL(scsi_block_requests);
2046
2047/**
2048 * scsi_unblock_requests - Utility function used by low-level drivers to allow
2049 * further commands to be queued to the device.
2050 * @shost: host in question
2051 *
2052 * There is no timer nor any other means by which the requests get unblocked
2053 * other than the low-level driver calling scsi_unblock_requests(). This is done
2054 * as an API function so that changes to the internals of the scsi mid-layer
2055 * won't require wholesale changes to drivers that use this feature.
2056 */
2057void scsi_unblock_requests(struct Scsi_Host *shost)
2058{
2059 shost->host_self_blocked = 0;
2060 scsi_run_host_queues(shost);
2061}
2062EXPORT_SYMBOL(scsi_unblock_requests);
2063
2064void scsi_exit_queue(void)
2065{
2066 kmem_cache_destroy(scsi_sense_cache);
2067}
2068
2069/**
2070 * scsi_mode_select - issue a mode select
2071 * @sdev: SCSI device to be queried
2072 * @pf: Page format bit (1 == standard, 0 == vendor specific)
2073 * @sp: Save page bit (0 == don't save, 1 == save)
2074 * @buffer: request buffer (may not be smaller than eight bytes)
2075 * @len: length of request buffer.
2076 * @timeout: command timeout
2077 * @retries: number of retries before failing
2078 * @data: returns a structure abstracting the mode header data
2079 * @sshdr: place to put sense data (or NULL if no sense to be collected).
2080 * must be SCSI_SENSE_BUFFERSIZE big.
2081 *
2082 * Returns zero if successful; negative error number or scsi
2083 * status on error
2084 *
2085 */
2086int scsi_mode_select(struct scsi_device *sdev, int pf, int sp,
2087 unsigned char *buffer, int len, int timeout, int retries,
2088 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2089{
2090 unsigned char cmd[10];
2091 unsigned char *real_buffer;
2092 const struct scsi_exec_args exec_args = {
2093 .sshdr = sshdr,
2094 };
2095 int ret;
2096
2097 memset(cmd, 0, sizeof(cmd));
2098 cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2099
2100 /*
2101 * Use MODE SELECT(10) if the device asked for it or if the mode page
2102 * and the mode select header cannot fit within the maximumm 255 bytes
2103 * of the MODE SELECT(6) command.
2104 */
2105 if (sdev->use_10_for_ms ||
2106 len + 4 > 255 ||
2107 data->block_descriptor_length > 255) {
2108 if (len > 65535 - 8)
2109 return -EINVAL;
2110 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2111 if (!real_buffer)
2112 return -ENOMEM;
2113 memcpy(real_buffer + 8, buffer, len);
2114 len += 8;
2115 real_buffer[0] = 0;
2116 real_buffer[1] = 0;
2117 real_buffer[2] = data->medium_type;
2118 real_buffer[3] = data->device_specific;
2119 real_buffer[4] = data->longlba ? 0x01 : 0;
2120 real_buffer[5] = 0;
2121 put_unaligned_be16(data->block_descriptor_length,
2122 &real_buffer[6]);
2123
2124 cmd[0] = MODE_SELECT_10;
2125 put_unaligned_be16(len, &cmd[7]);
2126 } else {
2127 if (data->longlba)
2128 return -EINVAL;
2129
2130 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2131 if (!real_buffer)
2132 return -ENOMEM;
2133 memcpy(real_buffer + 4, buffer, len);
2134 len += 4;
2135 real_buffer[0] = 0;
2136 real_buffer[1] = data->medium_type;
2137 real_buffer[2] = data->device_specific;
2138 real_buffer[3] = data->block_descriptor_length;
2139
2140 cmd[0] = MODE_SELECT;
2141 cmd[4] = len;
2142 }
2143
2144 ret = scsi_execute_cmd(sdev, cmd, REQ_OP_DRV_OUT, real_buffer, len,
2145 timeout, retries, &exec_args);
2146 kfree(real_buffer);
2147 return ret;
2148}
2149EXPORT_SYMBOL_GPL(scsi_mode_select);
2150
2151/**
2152 * scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2153 * @sdev: SCSI device to be queried
2154 * @dbd: set to prevent mode sense from returning block descriptors
2155 * @modepage: mode page being requested
2156 * @subpage: sub-page of the mode page being requested
2157 * @buffer: request buffer (may not be smaller than eight bytes)
2158 * @len: length of request buffer.
2159 * @timeout: command timeout
2160 * @retries: number of retries before failing
2161 * @data: returns a structure abstracting the mode header data
2162 * @sshdr: place to put sense data (or NULL if no sense to be collected).
2163 * must be SCSI_SENSE_BUFFERSIZE big.
2164 *
2165 * Returns zero if successful, or a negative error number on failure
2166 */
2167int
2168scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage, int subpage,
2169 unsigned char *buffer, int len, int timeout, int retries,
2170 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2171{
2172 unsigned char cmd[12];
2173 int use_10_for_ms;
2174 int header_length;
2175 int result, retry_count = retries;
2176 struct scsi_sense_hdr my_sshdr;
2177 const struct scsi_exec_args exec_args = {
2178 /* caller might not be interested in sense, but we need it */
2179 .sshdr = sshdr ? : &my_sshdr,
2180 };
2181
2182 memset(data, 0, sizeof(*data));
2183 memset(&cmd[0], 0, 12);
2184
2185 dbd = sdev->set_dbd_for_ms ? 8 : dbd;
2186 cmd[1] = dbd & 0x18; /* allows DBD and LLBA bits */
2187 cmd[2] = modepage;
2188 cmd[3] = subpage;
2189
2190 sshdr = exec_args.sshdr;
2191
2192 retry:
2193 use_10_for_ms = sdev->use_10_for_ms || len > 255;
2194
2195 if (use_10_for_ms) {
2196 if (len < 8 || len > 65535)
2197 return -EINVAL;
2198
2199 cmd[0] = MODE_SENSE_10;
2200 put_unaligned_be16(len, &cmd[7]);
2201 header_length = 8;
2202 } else {
2203 if (len < 4)
2204 return -EINVAL;
2205
2206 cmd[0] = MODE_SENSE;
2207 cmd[4] = len;
2208 header_length = 4;
2209 }
2210
2211 memset(buffer, 0, len);
2212
2213 result = scsi_execute_cmd(sdev, cmd, REQ_OP_DRV_IN, buffer, len,
2214 timeout, retries, &exec_args);
2215 if (result < 0)
2216 return result;
2217
2218 /* This code looks awful: what it's doing is making sure an
2219 * ILLEGAL REQUEST sense return identifies the actual command
2220 * byte as the problem. MODE_SENSE commands can return
2221 * ILLEGAL REQUEST if the code page isn't supported */
2222
2223 if (!scsi_status_is_good(result)) {
2224 if (scsi_sense_valid(sshdr)) {
2225 if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2226 (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2227 /*
2228 * Invalid command operation code: retry using
2229 * MODE SENSE(6) if this was a MODE SENSE(10)
2230 * request, except if the request mode page is
2231 * too large for MODE SENSE single byte
2232 * allocation length field.
2233 */
2234 if (use_10_for_ms) {
2235 if (len > 255)
2236 return -EIO;
2237 sdev->use_10_for_ms = 0;
2238 goto retry;
2239 }
2240 }
2241 if (scsi_status_is_check_condition(result) &&
2242 sshdr->sense_key == UNIT_ATTENTION &&
2243 retry_count) {
2244 retry_count--;
2245 goto retry;
2246 }
2247 }
2248 return -EIO;
2249 }
2250 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2251 (modepage == 6 || modepage == 8))) {
2252 /* Initio breakage? */
2253 header_length = 0;
2254 data->length = 13;
2255 data->medium_type = 0;
2256 data->device_specific = 0;
2257 data->longlba = 0;
2258 data->block_descriptor_length = 0;
2259 } else if (use_10_for_ms) {
2260 data->length = get_unaligned_be16(&buffer[0]) + 2;
2261 data->medium_type = buffer[2];
2262 data->device_specific = buffer[3];
2263 data->longlba = buffer[4] & 0x01;
2264 data->block_descriptor_length = get_unaligned_be16(&buffer[6]);
2265 } else {
2266 data->length = buffer[0] + 1;
2267 data->medium_type = buffer[1];
2268 data->device_specific = buffer[2];
2269 data->block_descriptor_length = buffer[3];
2270 }
2271 data->header_length = header_length;
2272
2273 return 0;
2274}
2275EXPORT_SYMBOL(scsi_mode_sense);
2276
2277/**
2278 * scsi_test_unit_ready - test if unit is ready
2279 * @sdev: scsi device to change the state of.
2280 * @timeout: command timeout
2281 * @retries: number of retries before failing
2282 * @sshdr: outpout pointer for decoded sense information.
2283 *
2284 * Returns zero if unsuccessful or an error if TUR failed. For
2285 * removable media, UNIT_ATTENTION sets ->changed flag.
2286 **/
2287int
2288scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2289 struct scsi_sense_hdr *sshdr)
2290{
2291 char cmd[] = {
2292 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2293 };
2294 const struct scsi_exec_args exec_args = {
2295 .sshdr = sshdr,
2296 };
2297 int result;
2298
2299 /* try to eat the UNIT_ATTENTION if there are enough retries */
2300 do {
2301 result = scsi_execute_cmd(sdev, cmd, REQ_OP_DRV_IN, NULL, 0,
2302 timeout, 1, &exec_args);
2303 if (sdev->removable && result > 0 && scsi_sense_valid(sshdr) &&
2304 sshdr->sense_key == UNIT_ATTENTION)
2305 sdev->changed = 1;
2306 } while (result > 0 && scsi_sense_valid(sshdr) &&
2307 sshdr->sense_key == UNIT_ATTENTION && --retries);
2308
2309 return result;
2310}
2311EXPORT_SYMBOL(scsi_test_unit_ready);
2312
2313/**
2314 * scsi_device_set_state - Take the given device through the device state model.
2315 * @sdev: scsi device to change the state of.
2316 * @state: state to change to.
2317 *
2318 * Returns zero if successful or an error if the requested
2319 * transition is illegal.
2320 */
2321int
2322scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2323{
2324 enum scsi_device_state oldstate = sdev->sdev_state;
2325
2326 if (state == oldstate)
2327 return 0;
2328
2329 switch (state) {
2330 case SDEV_CREATED:
2331 switch (oldstate) {
2332 case SDEV_CREATED_BLOCK:
2333 break;
2334 default:
2335 goto illegal;
2336 }
2337 break;
2338
2339 case SDEV_RUNNING:
2340 switch (oldstate) {
2341 case SDEV_CREATED:
2342 case SDEV_OFFLINE:
2343 case SDEV_TRANSPORT_OFFLINE:
2344 case SDEV_QUIESCE:
2345 case SDEV_BLOCK:
2346 break;
2347 default:
2348 goto illegal;
2349 }
2350 break;
2351
2352 case SDEV_QUIESCE:
2353 switch (oldstate) {
2354 case SDEV_RUNNING:
2355 case SDEV_OFFLINE:
2356 case SDEV_TRANSPORT_OFFLINE:
2357 break;
2358 default:
2359 goto illegal;
2360 }
2361 break;
2362
2363 case SDEV_OFFLINE:
2364 case SDEV_TRANSPORT_OFFLINE:
2365 switch (oldstate) {
2366 case SDEV_CREATED:
2367 case SDEV_RUNNING:
2368 case SDEV_QUIESCE:
2369 case SDEV_BLOCK:
2370 break;
2371 default:
2372 goto illegal;
2373 }
2374 break;
2375
2376 case SDEV_BLOCK:
2377 switch (oldstate) {
2378 case SDEV_RUNNING:
2379 case SDEV_CREATED_BLOCK:
2380 case SDEV_QUIESCE:
2381 case SDEV_OFFLINE:
2382 break;
2383 default:
2384 goto illegal;
2385 }
2386 break;
2387
2388 case SDEV_CREATED_BLOCK:
2389 switch (oldstate) {
2390 case SDEV_CREATED:
2391 break;
2392 default:
2393 goto illegal;
2394 }
2395 break;
2396
2397 case SDEV_CANCEL:
2398 switch (oldstate) {
2399 case SDEV_CREATED:
2400 case SDEV_RUNNING:
2401 case SDEV_QUIESCE:
2402 case SDEV_OFFLINE:
2403 case SDEV_TRANSPORT_OFFLINE:
2404 break;
2405 default:
2406 goto illegal;
2407 }
2408 break;
2409
2410 case SDEV_DEL:
2411 switch (oldstate) {
2412 case SDEV_CREATED:
2413 case SDEV_RUNNING:
2414 case SDEV_OFFLINE:
2415 case SDEV_TRANSPORT_OFFLINE:
2416 case SDEV_CANCEL:
2417 case SDEV_BLOCK:
2418 case SDEV_CREATED_BLOCK:
2419 break;
2420 default:
2421 goto illegal;
2422 }
2423 break;
2424
2425 }
2426 sdev->offline_already = false;
2427 sdev->sdev_state = state;
2428 return 0;
2429
2430 illegal:
2431 SCSI_LOG_ERROR_RECOVERY(1,
2432 sdev_printk(KERN_ERR, sdev,
2433 "Illegal state transition %s->%s",
2434 scsi_device_state_name(oldstate),
2435 scsi_device_state_name(state))
2436 );
2437 return -EINVAL;
2438}
2439EXPORT_SYMBOL(scsi_device_set_state);
2440
2441/**
2442 * scsi_evt_emit - emit a single SCSI device uevent
2443 * @sdev: associated SCSI device
2444 * @evt: event to emit
2445 *
2446 * Send a single uevent (scsi_event) to the associated scsi_device.
2447 */
2448static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2449{
2450 int idx = 0;
2451 char *envp[3];
2452
2453 switch (evt->evt_type) {
2454 case SDEV_EVT_MEDIA_CHANGE:
2455 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2456 break;
2457 case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2458 scsi_rescan_device(sdev);
2459 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2460 break;
2461 case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2462 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2463 break;
2464 case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2465 envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2466 break;
2467 case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2468 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2469 break;
2470 case SDEV_EVT_LUN_CHANGE_REPORTED:
2471 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2472 break;
2473 case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2474 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2475 break;
2476 case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2477 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2478 break;
2479 default:
2480 /* do nothing */
2481 break;
2482 }
2483
2484 envp[idx++] = NULL;
2485
2486 kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2487}
2488
2489/**
2490 * scsi_evt_thread - send a uevent for each scsi event
2491 * @work: work struct for scsi_device
2492 *
2493 * Dispatch queued events to their associated scsi_device kobjects
2494 * as uevents.
2495 */
2496void scsi_evt_thread(struct work_struct *work)
2497{
2498 struct scsi_device *sdev;
2499 enum scsi_device_event evt_type;
2500 LIST_HEAD(event_list);
2501
2502 sdev = container_of(work, struct scsi_device, event_work);
2503
2504 for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2505 if (test_and_clear_bit(evt_type, sdev->pending_events))
2506 sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2507
2508 while (1) {
2509 struct scsi_event *evt;
2510 struct list_head *this, *tmp;
2511 unsigned long flags;
2512
2513 spin_lock_irqsave(&sdev->list_lock, flags);
2514 list_splice_init(&sdev->event_list, &event_list);
2515 spin_unlock_irqrestore(&sdev->list_lock, flags);
2516
2517 if (list_empty(&event_list))
2518 break;
2519
2520 list_for_each_safe(this, tmp, &event_list) {
2521 evt = list_entry(this, struct scsi_event, node);
2522 list_del(&evt->node);
2523 scsi_evt_emit(sdev, evt);
2524 kfree(evt);
2525 }
2526 }
2527}
2528
2529/**
2530 * sdev_evt_send - send asserted event to uevent thread
2531 * @sdev: scsi_device event occurred on
2532 * @evt: event to send
2533 *
2534 * Assert scsi device event asynchronously.
2535 */
2536void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2537{
2538 unsigned long flags;
2539
2540#if 0
2541 /* FIXME: currently this check eliminates all media change events
2542 * for polled devices. Need to update to discriminate between AN
2543 * and polled events */
2544 if (!test_bit(evt->evt_type, sdev->supported_events)) {
2545 kfree(evt);
2546 return;
2547 }
2548#endif
2549
2550 spin_lock_irqsave(&sdev->list_lock, flags);
2551 list_add_tail(&evt->node, &sdev->event_list);
2552 schedule_work(&sdev->event_work);
2553 spin_unlock_irqrestore(&sdev->list_lock, flags);
2554}
2555EXPORT_SYMBOL_GPL(sdev_evt_send);
2556
2557/**
2558 * sdev_evt_alloc - allocate a new scsi event
2559 * @evt_type: type of event to allocate
2560 * @gfpflags: GFP flags for allocation
2561 *
2562 * Allocates and returns a new scsi_event.
2563 */
2564struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2565 gfp_t gfpflags)
2566{
2567 struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2568 if (!evt)
2569 return NULL;
2570
2571 evt->evt_type = evt_type;
2572 INIT_LIST_HEAD(&evt->node);
2573
2574 /* evt_type-specific initialization, if any */
2575 switch (evt_type) {
2576 case SDEV_EVT_MEDIA_CHANGE:
2577 case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2578 case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2579 case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2580 case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2581 case SDEV_EVT_LUN_CHANGE_REPORTED:
2582 case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2583 case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2584 default:
2585 /* do nothing */
2586 break;
2587 }
2588
2589 return evt;
2590}
2591EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2592
2593/**
2594 * sdev_evt_send_simple - send asserted event to uevent thread
2595 * @sdev: scsi_device event occurred on
2596 * @evt_type: type of event to send
2597 * @gfpflags: GFP flags for allocation
2598 *
2599 * Assert scsi device event asynchronously, given an event type.
2600 */
2601void sdev_evt_send_simple(struct scsi_device *sdev,
2602 enum scsi_device_event evt_type, gfp_t gfpflags)
2603{
2604 struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2605 if (!evt) {
2606 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2607 evt_type);
2608 return;
2609 }
2610
2611 sdev_evt_send(sdev, evt);
2612}
2613EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2614
2615/**
2616 * scsi_device_quiesce - Block all commands except power management.
2617 * @sdev: scsi device to quiesce.
2618 *
2619 * This works by trying to transition to the SDEV_QUIESCE state
2620 * (which must be a legal transition). When the device is in this
2621 * state, only power management requests will be accepted, all others will
2622 * be deferred.
2623 *
2624 * Must be called with user context, may sleep.
2625 *
2626 * Returns zero if unsuccessful or an error if not.
2627 */
2628int
2629scsi_device_quiesce(struct scsi_device *sdev)
2630{
2631 struct request_queue *q = sdev->request_queue;
2632 int err;
2633
2634 /*
2635 * It is allowed to call scsi_device_quiesce() multiple times from
2636 * the same context but concurrent scsi_device_quiesce() calls are
2637 * not allowed.
2638 */
2639 WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2640
2641 if (sdev->quiesced_by == current)
2642 return 0;
2643
2644 blk_set_pm_only(q);
2645
2646 blk_mq_freeze_queue(q);
2647 /*
2648 * Ensure that the effect of blk_set_pm_only() will be visible
2649 * for percpu_ref_tryget() callers that occur after the queue
2650 * unfreeze even if the queue was already frozen before this function
2651 * was called. See also https://lwn.net/Articles/573497/.
2652 */
2653 synchronize_rcu();
2654 blk_mq_unfreeze_queue(q);
2655
2656 mutex_lock(&sdev->state_mutex);
2657 err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2658 if (err == 0)
2659 sdev->quiesced_by = current;
2660 else
2661 blk_clear_pm_only(q);
2662 mutex_unlock(&sdev->state_mutex);
2663
2664 return err;
2665}
2666EXPORT_SYMBOL(scsi_device_quiesce);
2667
2668/**
2669 * scsi_device_resume - Restart user issued commands to a quiesced device.
2670 * @sdev: scsi device to resume.
2671 *
2672 * Moves the device from quiesced back to running and restarts the
2673 * queues.
2674 *
2675 * Must be called with user context, may sleep.
2676 */
2677void scsi_device_resume(struct scsi_device *sdev)
2678{
2679 /* check if the device state was mutated prior to resume, and if
2680 * so assume the state is being managed elsewhere (for example
2681 * device deleted during suspend)
2682 */
2683 mutex_lock(&sdev->state_mutex);
2684 if (sdev->sdev_state == SDEV_QUIESCE)
2685 scsi_device_set_state(sdev, SDEV_RUNNING);
2686 if (sdev->quiesced_by) {
2687 sdev->quiesced_by = NULL;
2688 blk_clear_pm_only(sdev->request_queue);
2689 }
2690 mutex_unlock(&sdev->state_mutex);
2691}
2692EXPORT_SYMBOL(scsi_device_resume);
2693
2694static void
2695device_quiesce_fn(struct scsi_device *sdev, void *data)
2696{
2697 scsi_device_quiesce(sdev);
2698}
2699
2700void
2701scsi_target_quiesce(struct scsi_target *starget)
2702{
2703 starget_for_each_device(starget, NULL, device_quiesce_fn);
2704}
2705EXPORT_SYMBOL(scsi_target_quiesce);
2706
2707static void
2708device_resume_fn(struct scsi_device *sdev, void *data)
2709{
2710 scsi_device_resume(sdev);
2711}
2712
2713void
2714scsi_target_resume(struct scsi_target *starget)
2715{
2716 starget_for_each_device(starget, NULL, device_resume_fn);
2717}
2718EXPORT_SYMBOL(scsi_target_resume);
2719
2720static int __scsi_internal_device_block_nowait(struct scsi_device *sdev)
2721{
2722 if (scsi_device_set_state(sdev, SDEV_BLOCK))
2723 return scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2724
2725 return 0;
2726}
2727
2728void scsi_start_queue(struct scsi_device *sdev)
2729{
2730 if (cmpxchg(&sdev->queue_stopped, 1, 0))
2731 blk_mq_unquiesce_queue(sdev->request_queue);
2732}
2733
2734static void scsi_stop_queue(struct scsi_device *sdev)
2735{
2736 /*
2737 * The atomic variable of ->queue_stopped covers that
2738 * blk_mq_quiesce_queue* is balanced with blk_mq_unquiesce_queue.
2739 *
2740 * The caller needs to wait until quiesce is done.
2741 */
2742 if (!cmpxchg(&sdev->queue_stopped, 0, 1))
2743 blk_mq_quiesce_queue_nowait(sdev->request_queue);
2744}
2745
2746/**
2747 * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2748 * @sdev: device to block
2749 *
2750 * Pause SCSI command processing on the specified device. Does not sleep.
2751 *
2752 * Returns zero if successful or a negative error code upon failure.
2753 *
2754 * Notes:
2755 * This routine transitions the device to the SDEV_BLOCK state (which must be
2756 * a legal transition). When the device is in this state, command processing
2757 * is paused until the device leaves the SDEV_BLOCK state. See also
2758 * scsi_internal_device_unblock_nowait().
2759 */
2760int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2761{
2762 int ret = __scsi_internal_device_block_nowait(sdev);
2763
2764 /*
2765 * The device has transitioned to SDEV_BLOCK. Stop the
2766 * block layer from calling the midlayer with this device's
2767 * request queue.
2768 */
2769 if (!ret)
2770 scsi_stop_queue(sdev);
2771 return ret;
2772}
2773EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2774
2775/**
2776 * scsi_device_block - try to transition to the SDEV_BLOCK state
2777 * @sdev: device to block
2778 * @data: dummy argument, ignored
2779 *
2780 * Pause SCSI command processing on the specified device. Callers must wait
2781 * until all ongoing scsi_queue_rq() calls have finished after this function
2782 * returns.
2783 *
2784 * Note:
2785 * This routine transitions the device to the SDEV_BLOCK state (which must be
2786 * a legal transition). When the device is in this state, command processing
2787 * is paused until the device leaves the SDEV_BLOCK state. See also
2788 * scsi_internal_device_unblock().
2789 */
2790static void scsi_device_block(struct scsi_device *sdev, void *data)
2791{
2792 int err;
2793 enum scsi_device_state state;
2794
2795 mutex_lock(&sdev->state_mutex);
2796 err = __scsi_internal_device_block_nowait(sdev);
2797 state = sdev->sdev_state;
2798 if (err == 0)
2799 /*
2800 * scsi_stop_queue() must be called with the state_mutex
2801 * held. Otherwise a simultaneous scsi_start_queue() call
2802 * might unquiesce the queue before we quiesce it.
2803 */
2804 scsi_stop_queue(sdev);
2805
2806 mutex_unlock(&sdev->state_mutex);
2807
2808 WARN_ONCE(err, "%s: failed to block %s in state %d\n",
2809 __func__, dev_name(&sdev->sdev_gendev), state);
2810}
2811
2812/**
2813 * scsi_internal_device_unblock_nowait - resume a device after a block request
2814 * @sdev: device to resume
2815 * @new_state: state to set the device to after unblocking
2816 *
2817 * Restart the device queue for a previously suspended SCSI device. Does not
2818 * sleep.
2819 *
2820 * Returns zero if successful or a negative error code upon failure.
2821 *
2822 * Notes:
2823 * This routine transitions the device to the SDEV_RUNNING state or to one of
2824 * the offline states (which must be a legal transition) allowing the midlayer
2825 * to goose the queue for this device.
2826 */
2827int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
2828 enum scsi_device_state new_state)
2829{
2830 switch (new_state) {
2831 case SDEV_RUNNING:
2832 case SDEV_TRANSPORT_OFFLINE:
2833 break;
2834 default:
2835 return -EINVAL;
2836 }
2837
2838 /*
2839 * Try to transition the scsi device to SDEV_RUNNING or one of the
2840 * offlined states and goose the device queue if successful.
2841 */
2842 switch (sdev->sdev_state) {
2843 case SDEV_BLOCK:
2844 case SDEV_TRANSPORT_OFFLINE:
2845 sdev->sdev_state = new_state;
2846 break;
2847 case SDEV_CREATED_BLOCK:
2848 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2849 new_state == SDEV_OFFLINE)
2850 sdev->sdev_state = new_state;
2851 else
2852 sdev->sdev_state = SDEV_CREATED;
2853 break;
2854 case SDEV_CANCEL:
2855 case SDEV_OFFLINE:
2856 break;
2857 default:
2858 return -EINVAL;
2859 }
2860 scsi_start_queue(sdev);
2861
2862 return 0;
2863}
2864EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
2865
2866/**
2867 * scsi_internal_device_unblock - resume a device after a block request
2868 * @sdev: device to resume
2869 * @new_state: state to set the device to after unblocking
2870 *
2871 * Restart the device queue for a previously suspended SCSI device. May sleep.
2872 *
2873 * Returns zero if successful or a negative error code upon failure.
2874 *
2875 * Notes:
2876 * This routine transitions the device to the SDEV_RUNNING state or to one of
2877 * the offline states (which must be a legal transition) allowing the midlayer
2878 * to goose the queue for this device.
2879 */
2880static int scsi_internal_device_unblock(struct scsi_device *sdev,
2881 enum scsi_device_state new_state)
2882{
2883 int ret;
2884
2885 mutex_lock(&sdev->state_mutex);
2886 ret = scsi_internal_device_unblock_nowait(sdev, new_state);
2887 mutex_unlock(&sdev->state_mutex);
2888
2889 return ret;
2890}
2891
2892static int
2893target_block(struct device *dev, void *data)
2894{
2895 if (scsi_is_target_device(dev))
2896 starget_for_each_device(to_scsi_target(dev), NULL,
2897 scsi_device_block);
2898 return 0;
2899}
2900
2901/**
2902 * scsi_block_targets - transition all SCSI child devices to SDEV_BLOCK state
2903 * @dev: a parent device of one or more scsi_target devices
2904 * @shost: the Scsi_Host to which this device belongs
2905 *
2906 * Iterate over all children of @dev, which should be scsi_target devices,
2907 * and switch all subordinate scsi devices to SDEV_BLOCK state. Wait for
2908 * ongoing scsi_queue_rq() calls to finish. May sleep.
2909 *
2910 * Note:
2911 * @dev must not itself be a scsi_target device.
2912 */
2913void
2914scsi_block_targets(struct Scsi_Host *shost, struct device *dev)
2915{
2916 WARN_ON_ONCE(scsi_is_target_device(dev));
2917 device_for_each_child(dev, NULL, target_block);
2918 blk_mq_wait_quiesce_done(&shost->tag_set);
2919}
2920EXPORT_SYMBOL_GPL(scsi_block_targets);
2921
2922static void
2923device_unblock(struct scsi_device *sdev, void *data)
2924{
2925 scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2926}
2927
2928static int
2929target_unblock(struct device *dev, void *data)
2930{
2931 if (scsi_is_target_device(dev))
2932 starget_for_each_device(to_scsi_target(dev), data,
2933 device_unblock);
2934 return 0;
2935}
2936
2937void
2938scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2939{
2940 if (scsi_is_target_device(dev))
2941 starget_for_each_device(to_scsi_target(dev), &new_state,
2942 device_unblock);
2943 else
2944 device_for_each_child(dev, &new_state, target_unblock);
2945}
2946EXPORT_SYMBOL_GPL(scsi_target_unblock);
2947
2948/**
2949 * scsi_host_block - Try to transition all logical units to the SDEV_BLOCK state
2950 * @shost: device to block
2951 *
2952 * Pause SCSI command processing for all logical units associated with the SCSI
2953 * host and wait until pending scsi_queue_rq() calls have finished.
2954 *
2955 * Returns zero if successful or a negative error code upon failure.
2956 */
2957int
2958scsi_host_block(struct Scsi_Host *shost)
2959{
2960 struct scsi_device *sdev;
2961 int ret;
2962
2963 /*
2964 * Call scsi_internal_device_block_nowait so we can avoid
2965 * calling synchronize_rcu() for each LUN.
2966 */
2967 shost_for_each_device(sdev, shost) {
2968 mutex_lock(&sdev->state_mutex);
2969 ret = scsi_internal_device_block_nowait(sdev);
2970 mutex_unlock(&sdev->state_mutex);
2971 if (ret) {
2972 scsi_device_put(sdev);
2973 return ret;
2974 }
2975 }
2976
2977 /* Wait for ongoing scsi_queue_rq() calls to finish. */
2978 blk_mq_wait_quiesce_done(&shost->tag_set);
2979
2980 return 0;
2981}
2982EXPORT_SYMBOL_GPL(scsi_host_block);
2983
2984int
2985scsi_host_unblock(struct Scsi_Host *shost, int new_state)
2986{
2987 struct scsi_device *sdev;
2988 int ret = 0;
2989
2990 shost_for_each_device(sdev, shost) {
2991 ret = scsi_internal_device_unblock(sdev, new_state);
2992 if (ret) {
2993 scsi_device_put(sdev);
2994 break;
2995 }
2996 }
2997 return ret;
2998}
2999EXPORT_SYMBOL_GPL(scsi_host_unblock);
3000
3001/**
3002 * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
3003 * @sgl: scatter-gather list
3004 * @sg_count: number of segments in sg
3005 * @offset: offset in bytes into sg, on return offset into the mapped area
3006 * @len: bytes to map, on return number of bytes mapped
3007 *
3008 * Returns virtual address of the start of the mapped page
3009 */
3010void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
3011 size_t *offset, size_t *len)
3012{
3013 int i;
3014 size_t sg_len = 0, len_complete = 0;
3015 struct scatterlist *sg;
3016 struct page *page;
3017
3018 WARN_ON(!irqs_disabled());
3019
3020 for_each_sg(sgl, sg, sg_count, i) {
3021 len_complete = sg_len; /* Complete sg-entries */
3022 sg_len += sg->length;
3023 if (sg_len > *offset)
3024 break;
3025 }
3026
3027 if (unlikely(i == sg_count)) {
3028 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
3029 "elements %d\n",
3030 __func__, sg_len, *offset, sg_count);
3031 WARN_ON(1);
3032 return NULL;
3033 }
3034
3035 /* Offset starting from the beginning of first page in this sg-entry */
3036 *offset = *offset - len_complete + sg->offset;
3037
3038 /* Assumption: contiguous pages can be accessed as "page + i" */
3039 page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
3040 *offset &= ~PAGE_MASK;
3041
3042 /* Bytes in this sg-entry from *offset to the end of the page */
3043 sg_len = PAGE_SIZE - *offset;
3044 if (*len > sg_len)
3045 *len = sg_len;
3046
3047 return kmap_atomic(page);
3048}
3049EXPORT_SYMBOL(scsi_kmap_atomic_sg);
3050
3051/**
3052 * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
3053 * @virt: virtual address to be unmapped
3054 */
3055void scsi_kunmap_atomic_sg(void *virt)
3056{
3057 kunmap_atomic(virt);
3058}
3059EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
3060
3061void sdev_disable_disk_events(struct scsi_device *sdev)
3062{
3063 atomic_inc(&sdev->disk_events_disable_depth);
3064}
3065EXPORT_SYMBOL(sdev_disable_disk_events);
3066
3067void sdev_enable_disk_events(struct scsi_device *sdev)
3068{
3069 if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
3070 return;
3071 atomic_dec(&sdev->disk_events_disable_depth);
3072}
3073EXPORT_SYMBOL(sdev_enable_disk_events);
3074
3075static unsigned char designator_prio(const unsigned char *d)
3076{
3077 if (d[1] & 0x30)
3078 /* not associated with LUN */
3079 return 0;
3080
3081 if (d[3] == 0)
3082 /* invalid length */
3083 return 0;
3084
3085 /*
3086 * Order of preference for lun descriptor:
3087 * - SCSI name string
3088 * - NAA IEEE Registered Extended
3089 * - EUI-64 based 16-byte
3090 * - EUI-64 based 12-byte
3091 * - NAA IEEE Registered
3092 * - NAA IEEE Extended
3093 * - EUI-64 based 8-byte
3094 * - SCSI name string (truncated)
3095 * - T10 Vendor ID
3096 * as longer descriptors reduce the likelyhood
3097 * of identification clashes.
3098 */
3099
3100 switch (d[1] & 0xf) {
3101 case 8:
3102 /* SCSI name string, variable-length UTF-8 */
3103 return 9;
3104 case 3:
3105 switch (d[4] >> 4) {
3106 case 6:
3107 /* NAA registered extended */
3108 return 8;
3109 case 5:
3110 /* NAA registered */
3111 return 5;
3112 case 4:
3113 /* NAA extended */
3114 return 4;
3115 case 3:
3116 /* NAA locally assigned */
3117 return 1;
3118 default:
3119 break;
3120 }
3121 break;
3122 case 2:
3123 switch (d[3]) {
3124 case 16:
3125 /* EUI64-based, 16 byte */
3126 return 7;
3127 case 12:
3128 /* EUI64-based, 12 byte */
3129 return 6;
3130 case 8:
3131 /* EUI64-based, 8 byte */
3132 return 3;
3133 default:
3134 break;
3135 }
3136 break;
3137 case 1:
3138 /* T10 vendor ID */
3139 return 1;
3140 default:
3141 break;
3142 }
3143
3144 return 0;
3145}
3146
3147/**
3148 * scsi_vpd_lun_id - return a unique device identification
3149 * @sdev: SCSI device
3150 * @id: buffer for the identification
3151 * @id_len: length of the buffer
3152 *
3153 * Copies a unique device identification into @id based
3154 * on the information in the VPD page 0x83 of the device.
3155 * The string will be formatted as a SCSI name string.
3156 *
3157 * Returns the length of the identification or error on failure.
3158 * If the identifier is longer than the supplied buffer the actual
3159 * identifier length is returned and the buffer is not zero-padded.
3160 */
3161int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
3162{
3163 u8 cur_id_prio = 0;
3164 u8 cur_id_size = 0;
3165 const unsigned char *d, *cur_id_str;
3166 const struct scsi_vpd *vpd_pg83;
3167 int id_size = -EINVAL;
3168
3169 rcu_read_lock();
3170 vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3171 if (!vpd_pg83) {
3172 rcu_read_unlock();
3173 return -ENXIO;
3174 }
3175
3176 /* The id string must be at least 20 bytes + terminating NULL byte */
3177 if (id_len < 21) {
3178 rcu_read_unlock();
3179 return -EINVAL;
3180 }
3181
3182 memset(id, 0, id_len);
3183 for (d = vpd_pg83->data + 4;
3184 d < vpd_pg83->data + vpd_pg83->len;
3185 d += d[3] + 4) {
3186 u8 prio = designator_prio(d);
3187
3188 if (prio == 0 || cur_id_prio > prio)
3189 continue;
3190
3191 switch (d[1] & 0xf) {
3192 case 0x1:
3193 /* T10 Vendor ID */
3194 if (cur_id_size > d[3])
3195 break;
3196 cur_id_prio = prio;
3197 cur_id_size = d[3];
3198 if (cur_id_size + 4 > id_len)
3199 cur_id_size = id_len - 4;
3200 cur_id_str = d + 4;
3201 id_size = snprintf(id, id_len, "t10.%*pE",
3202 cur_id_size, cur_id_str);
3203 break;
3204 case 0x2:
3205 /* EUI-64 */
3206 cur_id_prio = prio;
3207 cur_id_size = d[3];
3208 cur_id_str = d + 4;
3209 switch (cur_id_size) {
3210 case 8:
3211 id_size = snprintf(id, id_len,
3212 "eui.%8phN",
3213 cur_id_str);
3214 break;
3215 case 12:
3216 id_size = snprintf(id, id_len,
3217 "eui.%12phN",
3218 cur_id_str);
3219 break;
3220 case 16:
3221 id_size = snprintf(id, id_len,
3222 "eui.%16phN",
3223 cur_id_str);
3224 break;
3225 default:
3226 break;
3227 }
3228 break;
3229 case 0x3:
3230 /* NAA */
3231 cur_id_prio = prio;
3232 cur_id_size = d[3];
3233 cur_id_str = d + 4;
3234 switch (cur_id_size) {
3235 case 8:
3236 id_size = snprintf(id, id_len,
3237 "naa.%8phN",
3238 cur_id_str);
3239 break;
3240 case 16:
3241 id_size = snprintf(id, id_len,
3242 "naa.%16phN",
3243 cur_id_str);
3244 break;
3245 default:
3246 break;
3247 }
3248 break;
3249 case 0x8:
3250 /* SCSI name string */
3251 if (cur_id_size > d[3])
3252 break;
3253 /* Prefer others for truncated descriptor */
3254 if (d[3] > id_len) {
3255 prio = 2;
3256 if (cur_id_prio > prio)
3257 break;
3258 }
3259 cur_id_prio = prio;
3260 cur_id_size = id_size = d[3];
3261 cur_id_str = d + 4;
3262 if (cur_id_size >= id_len)
3263 cur_id_size = id_len - 1;
3264 memcpy(id, cur_id_str, cur_id_size);
3265 break;
3266 default:
3267 break;
3268 }
3269 }
3270 rcu_read_unlock();
3271
3272 return id_size;
3273}
3274EXPORT_SYMBOL(scsi_vpd_lun_id);
3275
3276/*
3277 * scsi_vpd_tpg_id - return a target port group identifier
3278 * @sdev: SCSI device
3279 *
3280 * Returns the Target Port Group identifier from the information
3281 * froom VPD page 0x83 of the device.
3282 *
3283 * Returns the identifier or error on failure.
3284 */
3285int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3286{
3287 const unsigned char *d;
3288 const struct scsi_vpd *vpd_pg83;
3289 int group_id = -EAGAIN, rel_port = -1;
3290
3291 rcu_read_lock();
3292 vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3293 if (!vpd_pg83) {
3294 rcu_read_unlock();
3295 return -ENXIO;
3296 }
3297
3298 d = vpd_pg83->data + 4;
3299 while (d < vpd_pg83->data + vpd_pg83->len) {
3300 switch (d[1] & 0xf) {
3301 case 0x4:
3302 /* Relative target port */
3303 rel_port = get_unaligned_be16(&d[6]);
3304 break;
3305 case 0x5:
3306 /* Target port group */
3307 group_id = get_unaligned_be16(&d[6]);
3308 break;
3309 default:
3310 break;
3311 }
3312 d += d[3] + 4;
3313 }
3314 rcu_read_unlock();
3315
3316 if (group_id >= 0 && rel_id && rel_port != -1)
3317 *rel_id = rel_port;
3318
3319 return group_id;
3320}
3321EXPORT_SYMBOL(scsi_vpd_tpg_id);
3322
3323/**
3324 * scsi_build_sense - build sense data for a command
3325 * @scmd: scsi command for which the sense should be formatted
3326 * @desc: Sense format (non-zero == descriptor format,
3327 * 0 == fixed format)
3328 * @key: Sense key
3329 * @asc: Additional sense code
3330 * @ascq: Additional sense code qualifier
3331 *
3332 **/
3333void scsi_build_sense(struct scsi_cmnd *scmd, int desc, u8 key, u8 asc, u8 ascq)
3334{
3335 scsi_build_sense_buffer(desc, scmd->sense_buffer, key, asc, ascq);
3336 scmd->result = SAM_STAT_CHECK_CONDITION;
3337}
3338EXPORT_SYMBOL_GPL(scsi_build_sense);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 1999 Eric Youngdale
4 * Copyright (C) 2014 Christoph Hellwig
5 *
6 * SCSI queueing library.
7 * Initial versions: Eric Youngdale (eric@andante.org).
8 * Based upon conversations with large numbers
9 * of people at Linux Expo.
10 */
11
12#include <linux/bio.h>
13#include <linux/bitops.h>
14#include <linux/blkdev.h>
15#include <linux/completion.h>
16#include <linux/kernel.h>
17#include <linux/export.h>
18#include <linux/init.h>
19#include <linux/pci.h>
20#include <linux/delay.h>
21#include <linux/hardirq.h>
22#include <linux/scatterlist.h>
23#include <linux/blk-mq.h>
24#include <linux/ratelimit.h>
25#include <asm/unaligned.h>
26
27#include <scsi/scsi.h>
28#include <scsi/scsi_cmnd.h>
29#include <scsi/scsi_dbg.h>
30#include <scsi/scsi_device.h>
31#include <scsi/scsi_driver.h>
32#include <scsi/scsi_eh.h>
33#include <scsi/scsi_host.h>
34#include <scsi/scsi_transport.h> /* __scsi_init_queue() */
35#include <scsi/scsi_dh.h>
36
37#include <trace/events/scsi.h>
38
39#include "scsi_debugfs.h"
40#include "scsi_priv.h"
41#include "scsi_logging.h"
42
43/*
44 * Size of integrity metadata is usually small, 1 inline sg should
45 * cover normal cases.
46 */
47#ifdef CONFIG_ARCH_NO_SG_CHAIN
48#define SCSI_INLINE_PROT_SG_CNT 0
49#define SCSI_INLINE_SG_CNT 0
50#else
51#define SCSI_INLINE_PROT_SG_CNT 1
52#define SCSI_INLINE_SG_CNT 2
53#endif
54
55static struct kmem_cache *scsi_sense_cache;
56static struct kmem_cache *scsi_sense_isadma_cache;
57static DEFINE_MUTEX(scsi_sense_cache_mutex);
58
59static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd);
60
61static inline struct kmem_cache *
62scsi_select_sense_cache(bool unchecked_isa_dma)
63{
64 return unchecked_isa_dma ? scsi_sense_isadma_cache : scsi_sense_cache;
65}
66
67static void scsi_free_sense_buffer(bool unchecked_isa_dma,
68 unsigned char *sense_buffer)
69{
70 kmem_cache_free(scsi_select_sense_cache(unchecked_isa_dma),
71 sense_buffer);
72}
73
74static unsigned char *scsi_alloc_sense_buffer(bool unchecked_isa_dma,
75 gfp_t gfp_mask, int numa_node)
76{
77 return kmem_cache_alloc_node(scsi_select_sense_cache(unchecked_isa_dma),
78 gfp_mask, numa_node);
79}
80
81int scsi_init_sense_cache(struct Scsi_Host *shost)
82{
83 struct kmem_cache *cache;
84 int ret = 0;
85
86 mutex_lock(&scsi_sense_cache_mutex);
87 cache = scsi_select_sense_cache(shost->unchecked_isa_dma);
88 if (cache)
89 goto exit;
90
91 if (shost->unchecked_isa_dma) {
92 scsi_sense_isadma_cache =
93 kmem_cache_create("scsi_sense_cache(DMA)",
94 SCSI_SENSE_BUFFERSIZE, 0,
95 SLAB_HWCACHE_ALIGN | SLAB_CACHE_DMA, NULL);
96 if (!scsi_sense_isadma_cache)
97 ret = -ENOMEM;
98 } else {
99 scsi_sense_cache =
100 kmem_cache_create_usercopy("scsi_sense_cache",
101 SCSI_SENSE_BUFFERSIZE, 0, SLAB_HWCACHE_ALIGN,
102 0, SCSI_SENSE_BUFFERSIZE, NULL);
103 if (!scsi_sense_cache)
104 ret = -ENOMEM;
105 }
106 exit:
107 mutex_unlock(&scsi_sense_cache_mutex);
108 return ret;
109}
110
111/*
112 * When to reinvoke queueing after a resource shortage. It's 3 msecs to
113 * not change behaviour from the previous unplug mechanism, experimentation
114 * may prove this needs changing.
115 */
116#define SCSI_QUEUE_DELAY 3
117
118static void
119scsi_set_blocked(struct scsi_cmnd *cmd, int reason)
120{
121 struct Scsi_Host *host = cmd->device->host;
122 struct scsi_device *device = cmd->device;
123 struct scsi_target *starget = scsi_target(device);
124
125 /*
126 * Set the appropriate busy bit for the device/host.
127 *
128 * If the host/device isn't busy, assume that something actually
129 * completed, and that we should be able to queue a command now.
130 *
131 * Note that the prior mid-layer assumption that any host could
132 * always queue at least one command is now broken. The mid-layer
133 * will implement a user specifiable stall (see
134 * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
135 * if a command is requeued with no other commands outstanding
136 * either for the device or for the host.
137 */
138 switch (reason) {
139 case SCSI_MLQUEUE_HOST_BUSY:
140 atomic_set(&host->host_blocked, host->max_host_blocked);
141 break;
142 case SCSI_MLQUEUE_DEVICE_BUSY:
143 case SCSI_MLQUEUE_EH_RETRY:
144 atomic_set(&device->device_blocked,
145 device->max_device_blocked);
146 break;
147 case SCSI_MLQUEUE_TARGET_BUSY:
148 atomic_set(&starget->target_blocked,
149 starget->max_target_blocked);
150 break;
151 }
152}
153
154static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd)
155{
156 if (cmd->request->rq_flags & RQF_DONTPREP) {
157 cmd->request->rq_flags &= ~RQF_DONTPREP;
158 scsi_mq_uninit_cmd(cmd);
159 } else {
160 WARN_ON_ONCE(true);
161 }
162 blk_mq_requeue_request(cmd->request, true);
163}
164
165/**
166 * __scsi_queue_insert - private queue insertion
167 * @cmd: The SCSI command being requeued
168 * @reason: The reason for the requeue
169 * @unbusy: Whether the queue should be unbusied
170 *
171 * This is a private queue insertion. The public interface
172 * scsi_queue_insert() always assumes the queue should be unbusied
173 * because it's always called before the completion. This function is
174 * for a requeue after completion, which should only occur in this
175 * file.
176 */
177static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, bool unbusy)
178{
179 struct scsi_device *device = cmd->device;
180
181 SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
182 "Inserting command %p into mlqueue\n", cmd));
183
184 scsi_set_blocked(cmd, reason);
185
186 /*
187 * Decrement the counters, since these commands are no longer
188 * active on the host/device.
189 */
190 if (unbusy)
191 scsi_device_unbusy(device, cmd);
192
193 /*
194 * Requeue this command. It will go before all other commands
195 * that are already in the queue. Schedule requeue work under
196 * lock such that the kblockd_schedule_work() call happens
197 * before blk_cleanup_queue() finishes.
198 */
199 cmd->result = 0;
200
201 blk_mq_requeue_request(cmd->request, true);
202}
203
204/**
205 * scsi_queue_insert - Reinsert a command in the queue.
206 * @cmd: command that we are adding to queue.
207 * @reason: why we are inserting command to queue.
208 *
209 * We do this for one of two cases. Either the host is busy and it cannot accept
210 * any more commands for the time being, or the device returned QUEUE_FULL and
211 * can accept no more commands.
212 *
213 * Context: This could be called either from an interrupt context or a normal
214 * process context.
215 */
216void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
217{
218 __scsi_queue_insert(cmd, reason, true);
219}
220
221
222/**
223 * __scsi_execute - insert request and wait for the result
224 * @sdev: scsi device
225 * @cmd: scsi command
226 * @data_direction: data direction
227 * @buffer: data buffer
228 * @bufflen: len of buffer
229 * @sense: optional sense buffer
230 * @sshdr: optional decoded sense header
231 * @timeout: request timeout in seconds
232 * @retries: number of times to retry request
233 * @flags: flags for ->cmd_flags
234 * @rq_flags: flags for ->rq_flags
235 * @resid: optional residual length
236 *
237 * Returns the scsi_cmnd result field if a command was executed, or a negative
238 * Linux error code if we didn't get that far.
239 */
240int __scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
241 int data_direction, void *buffer, unsigned bufflen,
242 unsigned char *sense, struct scsi_sense_hdr *sshdr,
243 int timeout, int retries, u64 flags, req_flags_t rq_flags,
244 int *resid)
245{
246 struct request *req;
247 struct scsi_request *rq;
248 int ret = DRIVER_ERROR << 24;
249
250 req = blk_get_request(sdev->request_queue,
251 data_direction == DMA_TO_DEVICE ?
252 REQ_OP_SCSI_OUT : REQ_OP_SCSI_IN, BLK_MQ_REQ_PREEMPT);
253 if (IS_ERR(req))
254 return ret;
255 rq = scsi_req(req);
256
257 if (bufflen && blk_rq_map_kern(sdev->request_queue, req,
258 buffer, bufflen, GFP_NOIO))
259 goto out;
260
261 rq->cmd_len = COMMAND_SIZE(cmd[0]);
262 memcpy(rq->cmd, cmd, rq->cmd_len);
263 rq->retries = retries;
264 req->timeout = timeout;
265 req->cmd_flags |= flags;
266 req->rq_flags |= rq_flags | RQF_QUIET;
267
268 /*
269 * head injection *required* here otherwise quiesce won't work
270 */
271 blk_execute_rq(req->q, NULL, req, 1);
272
273 /*
274 * Some devices (USB mass-storage in particular) may transfer
275 * garbage data together with a residue indicating that the data
276 * is invalid. Prevent the garbage from being misinterpreted
277 * and prevent security leaks by zeroing out the excess data.
278 */
279 if (unlikely(rq->resid_len > 0 && rq->resid_len <= bufflen))
280 memset(buffer + (bufflen - rq->resid_len), 0, rq->resid_len);
281
282 if (resid)
283 *resid = rq->resid_len;
284 if (sense && rq->sense_len)
285 memcpy(sense, rq->sense, SCSI_SENSE_BUFFERSIZE);
286 if (sshdr)
287 scsi_normalize_sense(rq->sense, rq->sense_len, sshdr);
288 ret = rq->result;
289 out:
290 blk_put_request(req);
291
292 return ret;
293}
294EXPORT_SYMBOL(__scsi_execute);
295
296/**
297 * scsi_init_cmd_errh - Initialize cmd fields related to error handling.
298 * @cmd: command that is ready to be queued.
299 *
300 * This function has the job of initializing a number of fields related to error
301 * handling. Typically this will be called once for each command, as required.
302 */
303static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
304{
305 scsi_set_resid(cmd, 0);
306 memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
307 if (cmd->cmd_len == 0)
308 cmd->cmd_len = scsi_command_size(cmd->cmnd);
309}
310
311/*
312 * Wake up the error handler if necessary. Avoid as follows that the error
313 * handler is not woken up if host in-flight requests number ==
314 * shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
315 * with an RCU read lock in this function to ensure that this function in
316 * its entirety either finishes before scsi_eh_scmd_add() increases the
317 * host_failed counter or that it notices the shost state change made by
318 * scsi_eh_scmd_add().
319 */
320static void scsi_dec_host_busy(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
321{
322 unsigned long flags;
323
324 rcu_read_lock();
325 __clear_bit(SCMD_STATE_INFLIGHT, &cmd->state);
326 if (unlikely(scsi_host_in_recovery(shost))) {
327 spin_lock_irqsave(shost->host_lock, flags);
328 if (shost->host_failed || shost->host_eh_scheduled)
329 scsi_eh_wakeup(shost);
330 spin_unlock_irqrestore(shost->host_lock, flags);
331 }
332 rcu_read_unlock();
333}
334
335void scsi_device_unbusy(struct scsi_device *sdev, struct scsi_cmnd *cmd)
336{
337 struct Scsi_Host *shost = sdev->host;
338 struct scsi_target *starget = scsi_target(sdev);
339
340 scsi_dec_host_busy(shost, cmd);
341
342 if (starget->can_queue > 0)
343 atomic_dec(&starget->target_busy);
344
345 atomic_dec(&sdev->device_busy);
346}
347
348static void scsi_kick_queue(struct request_queue *q)
349{
350 blk_mq_run_hw_queues(q, false);
351}
352
353/*
354 * Called for single_lun devices on IO completion. Clear starget_sdev_user,
355 * and call blk_run_queue for all the scsi_devices on the target -
356 * including current_sdev first.
357 *
358 * Called with *no* scsi locks held.
359 */
360static void scsi_single_lun_run(struct scsi_device *current_sdev)
361{
362 struct Scsi_Host *shost = current_sdev->host;
363 struct scsi_device *sdev, *tmp;
364 struct scsi_target *starget = scsi_target(current_sdev);
365 unsigned long flags;
366
367 spin_lock_irqsave(shost->host_lock, flags);
368 starget->starget_sdev_user = NULL;
369 spin_unlock_irqrestore(shost->host_lock, flags);
370
371 /*
372 * Call blk_run_queue for all LUNs on the target, starting with
373 * current_sdev. We race with others (to set starget_sdev_user),
374 * but in most cases, we will be first. Ideally, each LU on the
375 * target would get some limited time or requests on the target.
376 */
377 scsi_kick_queue(current_sdev->request_queue);
378
379 spin_lock_irqsave(shost->host_lock, flags);
380 if (starget->starget_sdev_user)
381 goto out;
382 list_for_each_entry_safe(sdev, tmp, &starget->devices,
383 same_target_siblings) {
384 if (sdev == current_sdev)
385 continue;
386 if (scsi_device_get(sdev))
387 continue;
388
389 spin_unlock_irqrestore(shost->host_lock, flags);
390 scsi_kick_queue(sdev->request_queue);
391 spin_lock_irqsave(shost->host_lock, flags);
392
393 scsi_device_put(sdev);
394 }
395 out:
396 spin_unlock_irqrestore(shost->host_lock, flags);
397}
398
399static inline bool scsi_device_is_busy(struct scsi_device *sdev)
400{
401 if (atomic_read(&sdev->device_busy) >= sdev->queue_depth)
402 return true;
403 if (atomic_read(&sdev->device_blocked) > 0)
404 return true;
405 return false;
406}
407
408static inline bool scsi_target_is_busy(struct scsi_target *starget)
409{
410 if (starget->can_queue > 0) {
411 if (atomic_read(&starget->target_busy) >= starget->can_queue)
412 return true;
413 if (atomic_read(&starget->target_blocked) > 0)
414 return true;
415 }
416 return false;
417}
418
419static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
420{
421 if (atomic_read(&shost->host_blocked) > 0)
422 return true;
423 if (shost->host_self_blocked)
424 return true;
425 return false;
426}
427
428static void scsi_starved_list_run(struct Scsi_Host *shost)
429{
430 LIST_HEAD(starved_list);
431 struct scsi_device *sdev;
432 unsigned long flags;
433
434 spin_lock_irqsave(shost->host_lock, flags);
435 list_splice_init(&shost->starved_list, &starved_list);
436
437 while (!list_empty(&starved_list)) {
438 struct request_queue *slq;
439
440 /*
441 * As long as shost is accepting commands and we have
442 * starved queues, call blk_run_queue. scsi_request_fn
443 * drops the queue_lock and can add us back to the
444 * starved_list.
445 *
446 * host_lock protects the starved_list and starved_entry.
447 * scsi_request_fn must get the host_lock before checking
448 * or modifying starved_list or starved_entry.
449 */
450 if (scsi_host_is_busy(shost))
451 break;
452
453 sdev = list_entry(starved_list.next,
454 struct scsi_device, starved_entry);
455 list_del_init(&sdev->starved_entry);
456 if (scsi_target_is_busy(scsi_target(sdev))) {
457 list_move_tail(&sdev->starved_entry,
458 &shost->starved_list);
459 continue;
460 }
461
462 /*
463 * Once we drop the host lock, a racing scsi_remove_device()
464 * call may remove the sdev from the starved list and destroy
465 * it and the queue. Mitigate by taking a reference to the
466 * queue and never touching the sdev again after we drop the
467 * host lock. Note: if __scsi_remove_device() invokes
468 * blk_cleanup_queue() before the queue is run from this
469 * function then blk_run_queue() will return immediately since
470 * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
471 */
472 slq = sdev->request_queue;
473 if (!blk_get_queue(slq))
474 continue;
475 spin_unlock_irqrestore(shost->host_lock, flags);
476
477 scsi_kick_queue(slq);
478 blk_put_queue(slq);
479
480 spin_lock_irqsave(shost->host_lock, flags);
481 }
482 /* put any unprocessed entries back */
483 list_splice(&starved_list, &shost->starved_list);
484 spin_unlock_irqrestore(shost->host_lock, flags);
485}
486
487/**
488 * scsi_run_queue - Select a proper request queue to serve next.
489 * @q: last request's queue
490 *
491 * The previous command was completely finished, start a new one if possible.
492 */
493static void scsi_run_queue(struct request_queue *q)
494{
495 struct scsi_device *sdev = q->queuedata;
496
497 if (scsi_target(sdev)->single_lun)
498 scsi_single_lun_run(sdev);
499 if (!list_empty(&sdev->host->starved_list))
500 scsi_starved_list_run(sdev->host);
501
502 blk_mq_run_hw_queues(q, false);
503}
504
505void scsi_requeue_run_queue(struct work_struct *work)
506{
507 struct scsi_device *sdev;
508 struct request_queue *q;
509
510 sdev = container_of(work, struct scsi_device, requeue_work);
511 q = sdev->request_queue;
512 scsi_run_queue(q);
513}
514
515void scsi_run_host_queues(struct Scsi_Host *shost)
516{
517 struct scsi_device *sdev;
518
519 shost_for_each_device(sdev, shost)
520 scsi_run_queue(sdev->request_queue);
521}
522
523static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
524{
525 if (!blk_rq_is_passthrough(cmd->request)) {
526 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
527
528 if (drv->uninit_command)
529 drv->uninit_command(cmd);
530 }
531}
532
533static void scsi_free_sgtables(struct scsi_cmnd *cmd)
534{
535 if (cmd->sdb.table.nents)
536 sg_free_table_chained(&cmd->sdb.table,
537 SCSI_INLINE_SG_CNT);
538 if (scsi_prot_sg_count(cmd))
539 sg_free_table_chained(&cmd->prot_sdb->table,
540 SCSI_INLINE_PROT_SG_CNT);
541}
542
543static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
544{
545 scsi_free_sgtables(cmd);
546 scsi_uninit_cmd(cmd);
547}
548
549static void scsi_run_queue_async(struct scsi_device *sdev)
550{
551 if (scsi_target(sdev)->single_lun ||
552 !list_empty(&sdev->host->starved_list))
553 kblockd_schedule_work(&sdev->requeue_work);
554 else
555 blk_mq_run_hw_queues(sdev->request_queue, true);
556}
557
558/* Returns false when no more bytes to process, true if there are more */
559static bool scsi_end_request(struct request *req, blk_status_t error,
560 unsigned int bytes)
561{
562 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
563 struct scsi_device *sdev = cmd->device;
564 struct request_queue *q = sdev->request_queue;
565
566 if (blk_update_request(req, error, bytes))
567 return true;
568
569 if (blk_queue_add_random(q))
570 add_disk_randomness(req->rq_disk);
571
572 if (!blk_rq_is_scsi(req)) {
573 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
574 cmd->flags &= ~SCMD_INITIALIZED;
575 }
576
577 /*
578 * Calling rcu_barrier() is not necessary here because the
579 * SCSI error handler guarantees that the function called by
580 * call_rcu() has been called before scsi_end_request() is
581 * called.
582 */
583 destroy_rcu_head(&cmd->rcu);
584
585 /*
586 * In the MQ case the command gets freed by __blk_mq_end_request,
587 * so we have to do all cleanup that depends on it earlier.
588 *
589 * We also can't kick the queues from irq context, so we
590 * will have to defer it to a workqueue.
591 */
592 scsi_mq_uninit_cmd(cmd);
593
594 /*
595 * queue is still alive, so grab the ref for preventing it
596 * from being cleaned up during running queue.
597 */
598 percpu_ref_get(&q->q_usage_counter);
599
600 __blk_mq_end_request(req, error);
601
602 scsi_run_queue_async(sdev);
603
604 percpu_ref_put(&q->q_usage_counter);
605 return false;
606}
607
608/**
609 * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
610 * @cmd: SCSI command
611 * @result: scsi error code
612 *
613 * Translate a SCSI result code into a blk_status_t value. May reset the host
614 * byte of @cmd->result.
615 */
616static blk_status_t scsi_result_to_blk_status(struct scsi_cmnd *cmd, int result)
617{
618 switch (host_byte(result)) {
619 case DID_OK:
620 /*
621 * Also check the other bytes than the status byte in result
622 * to handle the case when a SCSI LLD sets result to
623 * DRIVER_SENSE << 24 without setting SAM_STAT_CHECK_CONDITION.
624 */
625 if (scsi_status_is_good(result) && (result & ~0xff) == 0)
626 return BLK_STS_OK;
627 return BLK_STS_IOERR;
628 case DID_TRANSPORT_FAILFAST:
629 return BLK_STS_TRANSPORT;
630 case DID_TARGET_FAILURE:
631 set_host_byte(cmd, DID_OK);
632 return BLK_STS_TARGET;
633 case DID_NEXUS_FAILURE:
634 set_host_byte(cmd, DID_OK);
635 return BLK_STS_NEXUS;
636 case DID_ALLOC_FAILURE:
637 set_host_byte(cmd, DID_OK);
638 return BLK_STS_NOSPC;
639 case DID_MEDIUM_ERROR:
640 set_host_byte(cmd, DID_OK);
641 return BLK_STS_MEDIUM;
642 default:
643 return BLK_STS_IOERR;
644 }
645}
646
647/* Helper for scsi_io_completion() when "reprep" action required. */
648static void scsi_io_completion_reprep(struct scsi_cmnd *cmd,
649 struct request_queue *q)
650{
651 /* A new command will be prepared and issued. */
652 scsi_mq_requeue_cmd(cmd);
653}
654
655/* Helper for scsi_io_completion() when special action required. */
656static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
657{
658 struct request_queue *q = cmd->device->request_queue;
659 struct request *req = cmd->request;
660 int level = 0;
661 enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
662 ACTION_DELAYED_RETRY} action;
663 unsigned long wait_for = (cmd->allowed + 1) * req->timeout;
664 struct scsi_sense_hdr sshdr;
665 bool sense_valid;
666 bool sense_current = true; /* false implies "deferred sense" */
667 blk_status_t blk_stat;
668
669 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
670 if (sense_valid)
671 sense_current = !scsi_sense_is_deferred(&sshdr);
672
673 blk_stat = scsi_result_to_blk_status(cmd, result);
674
675 if (host_byte(result) == DID_RESET) {
676 /* Third party bus reset or reset for error recovery
677 * reasons. Just retry the command and see what
678 * happens.
679 */
680 action = ACTION_RETRY;
681 } else if (sense_valid && sense_current) {
682 switch (sshdr.sense_key) {
683 case UNIT_ATTENTION:
684 if (cmd->device->removable) {
685 /* Detected disc change. Set a bit
686 * and quietly refuse further access.
687 */
688 cmd->device->changed = 1;
689 action = ACTION_FAIL;
690 } else {
691 /* Must have been a power glitch, or a
692 * bus reset. Could not have been a
693 * media change, so we just retry the
694 * command and see what happens.
695 */
696 action = ACTION_RETRY;
697 }
698 break;
699 case ILLEGAL_REQUEST:
700 /* If we had an ILLEGAL REQUEST returned, then
701 * we may have performed an unsupported
702 * command. The only thing this should be
703 * would be a ten byte read where only a six
704 * byte read was supported. Also, on a system
705 * where READ CAPACITY failed, we may have
706 * read past the end of the disk.
707 */
708 if ((cmd->device->use_10_for_rw &&
709 sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
710 (cmd->cmnd[0] == READ_10 ||
711 cmd->cmnd[0] == WRITE_10)) {
712 /* This will issue a new 6-byte command. */
713 cmd->device->use_10_for_rw = 0;
714 action = ACTION_REPREP;
715 } else if (sshdr.asc == 0x10) /* DIX */ {
716 action = ACTION_FAIL;
717 blk_stat = BLK_STS_PROTECTION;
718 /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
719 } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
720 action = ACTION_FAIL;
721 blk_stat = BLK_STS_TARGET;
722 } else
723 action = ACTION_FAIL;
724 break;
725 case ABORTED_COMMAND:
726 action = ACTION_FAIL;
727 if (sshdr.asc == 0x10) /* DIF */
728 blk_stat = BLK_STS_PROTECTION;
729 break;
730 case NOT_READY:
731 /* If the device is in the process of becoming
732 * ready, or has a temporary blockage, retry.
733 */
734 if (sshdr.asc == 0x04) {
735 switch (sshdr.ascq) {
736 case 0x01: /* becoming ready */
737 case 0x04: /* format in progress */
738 case 0x05: /* rebuild in progress */
739 case 0x06: /* recalculation in progress */
740 case 0x07: /* operation in progress */
741 case 0x08: /* Long write in progress */
742 case 0x09: /* self test in progress */
743 case 0x14: /* space allocation in progress */
744 case 0x1a: /* start stop unit in progress */
745 case 0x1b: /* sanitize in progress */
746 case 0x1d: /* configuration in progress */
747 case 0x24: /* depopulation in progress */
748 action = ACTION_DELAYED_RETRY;
749 break;
750 default:
751 action = ACTION_FAIL;
752 break;
753 }
754 } else
755 action = ACTION_FAIL;
756 break;
757 case VOLUME_OVERFLOW:
758 /* See SSC3rXX or current. */
759 action = ACTION_FAIL;
760 break;
761 default:
762 action = ACTION_FAIL;
763 break;
764 }
765 } else
766 action = ACTION_FAIL;
767
768 if (action != ACTION_FAIL &&
769 time_before(cmd->jiffies_at_alloc + wait_for, jiffies))
770 action = ACTION_FAIL;
771
772 switch (action) {
773 case ACTION_FAIL:
774 /* Give up and fail the remainder of the request */
775 if (!(req->rq_flags & RQF_QUIET)) {
776 static DEFINE_RATELIMIT_STATE(_rs,
777 DEFAULT_RATELIMIT_INTERVAL,
778 DEFAULT_RATELIMIT_BURST);
779
780 if (unlikely(scsi_logging_level))
781 level =
782 SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
783 SCSI_LOG_MLCOMPLETE_BITS);
784
785 /*
786 * if logging is enabled the failure will be printed
787 * in scsi_log_completion(), so avoid duplicate messages
788 */
789 if (!level && __ratelimit(&_rs)) {
790 scsi_print_result(cmd, NULL, FAILED);
791 if (driver_byte(result) == DRIVER_SENSE)
792 scsi_print_sense(cmd);
793 scsi_print_command(cmd);
794 }
795 }
796 if (!scsi_end_request(req, blk_stat, blk_rq_err_bytes(req)))
797 return;
798 fallthrough;
799 case ACTION_REPREP:
800 scsi_io_completion_reprep(cmd, q);
801 break;
802 case ACTION_RETRY:
803 /* Retry the same command immediately */
804 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
805 break;
806 case ACTION_DELAYED_RETRY:
807 /* Retry the same command after a delay */
808 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
809 break;
810 }
811}
812
813/*
814 * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
815 * new result that may suppress further error checking. Also modifies
816 * *blk_statp in some cases.
817 */
818static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
819 blk_status_t *blk_statp)
820{
821 bool sense_valid;
822 bool sense_current = true; /* false implies "deferred sense" */
823 struct request *req = cmd->request;
824 struct scsi_sense_hdr sshdr;
825
826 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
827 if (sense_valid)
828 sense_current = !scsi_sense_is_deferred(&sshdr);
829
830 if (blk_rq_is_passthrough(req)) {
831 if (sense_valid) {
832 /*
833 * SG_IO wants current and deferred errors
834 */
835 scsi_req(req)->sense_len =
836 min(8 + cmd->sense_buffer[7],
837 SCSI_SENSE_BUFFERSIZE);
838 }
839 if (sense_current)
840 *blk_statp = scsi_result_to_blk_status(cmd, result);
841 } else if (blk_rq_bytes(req) == 0 && sense_current) {
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 *blk_statp explicitly for the problem case.
846 */
847 *blk_statp = scsi_result_to_blk_status(cmd, result);
848 }
849 /*
850 * Recovered errors need reporting, but they're always treated as
851 * success, so fiddle the result code here. For passthrough requests
852 * we already took a copy of the original into sreq->result which
853 * is what gets returned to the user
854 */
855 if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
856 bool do_print = true;
857 /*
858 * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
859 * skip print since caller wants ATA registers. Only occurs
860 * on SCSI ATA PASS_THROUGH commands when CK_COND=1
861 */
862 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
863 do_print = false;
864 else if (req->rq_flags & RQF_QUIET)
865 do_print = false;
866 if (do_print)
867 scsi_print_sense(cmd);
868 result = 0;
869 /* for passthrough, *blk_statp may be set */
870 *blk_statp = BLK_STS_OK;
871 }
872 /*
873 * Another corner case: the SCSI status byte is non-zero but 'good'.
874 * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
875 * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
876 * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
877 * intermediate statuses (both obsolete in SAM-4) as good.
878 */
879 if (status_byte(result) && scsi_status_is_good(result)) {
880 result = 0;
881 *blk_statp = BLK_STS_OK;
882 }
883 return result;
884}
885
886/**
887 * scsi_io_completion - Completion processing for SCSI commands.
888 * @cmd: command that is finished.
889 * @good_bytes: number of processed bytes.
890 *
891 * We will finish off the specified number of sectors. If we are done, the
892 * command block will be released and the queue function will be goosed. If we
893 * are not done then we have to figure out what to do next:
894 *
895 * a) We can call scsi_io_completion_reprep(). The request will be
896 * unprepared and put back on the queue. Then a new command will
897 * be created for it. This should be used if we made forward
898 * progress, or if we want to switch from READ(10) to READ(6) for
899 * example.
900 *
901 * b) We can call scsi_io_completion_action(). The request will be
902 * put back on the queue and retried using the same command as
903 * before, possibly after a delay.
904 *
905 * c) We can call scsi_end_request() with blk_stat other than
906 * BLK_STS_OK, to fail the remainder of the request.
907 */
908void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
909{
910 int result = cmd->result;
911 struct request_queue *q = cmd->device->request_queue;
912 struct request *req = cmd->request;
913 blk_status_t blk_stat = BLK_STS_OK;
914
915 if (unlikely(result)) /* a nz result may or may not be an error */
916 result = scsi_io_completion_nz_result(cmd, result, &blk_stat);
917
918 if (unlikely(blk_rq_is_passthrough(req))) {
919 /*
920 * scsi_result_to_blk_status may have reset the host_byte
921 */
922 scsi_req(req)->result = cmd->result;
923 }
924
925 /*
926 * Next deal with any sectors which we were able to correctly
927 * handle.
928 */
929 SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
930 "%u sectors total, %d bytes done.\n",
931 blk_rq_sectors(req), good_bytes));
932
933 /*
934 * Failed, zero length commands always need to drop down
935 * to retry code. Fast path should return in this block.
936 */
937 if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
938 if (likely(!scsi_end_request(req, blk_stat, good_bytes)))
939 return; /* no bytes remaining */
940 }
941
942 /* Kill remainder if no retries. */
943 if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
944 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req)))
945 WARN_ONCE(true,
946 "Bytes remaining after failed, no-retry command");
947 return;
948 }
949
950 /*
951 * If there had been no error, but we have leftover bytes in the
952 * requeues just queue the command up again.
953 */
954 if (likely(result == 0))
955 scsi_io_completion_reprep(cmd, q);
956 else
957 scsi_io_completion_action(cmd, result);
958}
959
960static inline bool scsi_cmd_needs_dma_drain(struct scsi_device *sdev,
961 struct request *rq)
962{
963 return sdev->dma_drain_len && blk_rq_is_passthrough(rq) &&
964 !op_is_write(req_op(rq)) &&
965 sdev->host->hostt->dma_need_drain(rq);
966}
967
968/**
969 * scsi_init_io - SCSI I/O initialization function.
970 * @cmd: command descriptor we wish to initialize
971 *
972 * Returns:
973 * * BLK_STS_OK - on success
974 * * BLK_STS_RESOURCE - if the failure is retryable
975 * * BLK_STS_IOERR - if the failure is fatal
976 */
977blk_status_t scsi_init_io(struct scsi_cmnd *cmd)
978{
979 struct scsi_device *sdev = cmd->device;
980 struct request *rq = cmd->request;
981 unsigned short nr_segs = blk_rq_nr_phys_segments(rq);
982 struct scatterlist *last_sg = NULL;
983 blk_status_t ret;
984 bool need_drain = scsi_cmd_needs_dma_drain(sdev, rq);
985 int count;
986
987 if (WARN_ON_ONCE(!nr_segs))
988 return BLK_STS_IOERR;
989
990 /*
991 * Make sure there is space for the drain. The driver must adjust
992 * max_hw_segments to be prepared for this.
993 */
994 if (need_drain)
995 nr_segs++;
996
997 /*
998 * If sg table allocation fails, requeue request later.
999 */
1000 if (unlikely(sg_alloc_table_chained(&cmd->sdb.table, nr_segs,
1001 cmd->sdb.table.sgl, SCSI_INLINE_SG_CNT)))
1002 return BLK_STS_RESOURCE;
1003
1004 /*
1005 * Next, walk the list, and fill in the addresses and sizes of
1006 * each segment.
1007 */
1008 count = __blk_rq_map_sg(rq->q, rq, cmd->sdb.table.sgl, &last_sg);
1009
1010 if (blk_rq_bytes(rq) & rq->q->dma_pad_mask) {
1011 unsigned int pad_len =
1012 (rq->q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
1013
1014 last_sg->length += pad_len;
1015 cmd->extra_len += pad_len;
1016 }
1017
1018 if (need_drain) {
1019 sg_unmark_end(last_sg);
1020 last_sg = sg_next(last_sg);
1021 sg_set_buf(last_sg, sdev->dma_drain_buf, sdev->dma_drain_len);
1022 sg_mark_end(last_sg);
1023
1024 cmd->extra_len += sdev->dma_drain_len;
1025 count++;
1026 }
1027
1028 BUG_ON(count > cmd->sdb.table.nents);
1029 cmd->sdb.table.nents = count;
1030 cmd->sdb.length = blk_rq_payload_bytes(rq);
1031
1032 if (blk_integrity_rq(rq)) {
1033 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1034 int ivecs;
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_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/* Called before a request is prepared. See also scsi_mq_prep_fn(). */
1105void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1106{
1107 void *buf = cmd->sense_buffer;
1108 void *prot = cmd->prot_sdb;
1109 struct request *rq = blk_mq_rq_from_pdu(cmd);
1110 unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1111 unsigned long jiffies_at_alloc;
1112 int retries, to_clear;
1113 bool in_flight;
1114
1115 if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1116 flags |= SCMD_INITIALIZED;
1117 scsi_initialize_rq(rq);
1118 }
1119
1120 jiffies_at_alloc = cmd->jiffies_at_alloc;
1121 retries = cmd->retries;
1122 in_flight = test_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1123 /*
1124 * Zero out the cmd, except for the embedded scsi_request. Only clear
1125 * the driver-private command data if the LLD does not supply a
1126 * function to initialize that data.
1127 */
1128 to_clear = sizeof(*cmd) - sizeof(cmd->req);
1129 if (!dev->host->hostt->init_cmd_priv)
1130 to_clear += dev->host->hostt->cmd_size;
1131 memset((char *)cmd + sizeof(cmd->req), 0, to_clear);
1132
1133 cmd->device = dev;
1134 cmd->sense_buffer = buf;
1135 cmd->prot_sdb = prot;
1136 cmd->flags = flags;
1137 INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1138 cmd->jiffies_at_alloc = jiffies_at_alloc;
1139 cmd->retries = retries;
1140 if (in_flight)
1141 __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1142
1143}
1144
1145static blk_status_t scsi_setup_scsi_cmnd(struct scsi_device *sdev,
1146 struct request *req)
1147{
1148 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1149
1150 /*
1151 * Passthrough requests may transfer data, in which case they must
1152 * a bio attached to them. Or they might contain a SCSI command
1153 * that does not transfer data, in which case they may optionally
1154 * submit a request without an attached bio.
1155 */
1156 if (req->bio) {
1157 blk_status_t ret = scsi_init_io(cmd);
1158 if (unlikely(ret != BLK_STS_OK))
1159 return ret;
1160 } else {
1161 BUG_ON(blk_rq_bytes(req));
1162
1163 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1164 }
1165
1166 cmd->cmd_len = scsi_req(req)->cmd_len;
1167 cmd->cmnd = scsi_req(req)->cmd;
1168 cmd->transfersize = blk_rq_bytes(req);
1169 cmd->allowed = scsi_req(req)->retries;
1170 return BLK_STS_OK;
1171}
1172
1173/*
1174 * Setup a normal block command. These are simple request from filesystems
1175 * that still need to be translated to SCSI CDBs from the ULD.
1176 */
1177static blk_status_t scsi_setup_fs_cmnd(struct scsi_device *sdev,
1178 struct request *req)
1179{
1180 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1181
1182 if (unlikely(sdev->handler && sdev->handler->prep_fn)) {
1183 blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1184 if (ret != BLK_STS_OK)
1185 return ret;
1186 }
1187
1188 cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1189 memset(cmd->cmnd, 0, BLK_MAX_CDB);
1190 return scsi_cmd_to_driver(cmd)->init_command(cmd);
1191}
1192
1193static blk_status_t scsi_setup_cmnd(struct scsi_device *sdev,
1194 struct request *req)
1195{
1196 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1197 blk_status_t ret;
1198
1199 if (!blk_rq_bytes(req))
1200 cmd->sc_data_direction = DMA_NONE;
1201 else if (rq_data_dir(req) == WRITE)
1202 cmd->sc_data_direction = DMA_TO_DEVICE;
1203 else
1204 cmd->sc_data_direction = DMA_FROM_DEVICE;
1205
1206 if (blk_rq_is_scsi(req))
1207 ret = scsi_setup_scsi_cmnd(sdev, req);
1208 else
1209 ret = scsi_setup_fs_cmnd(sdev, req);
1210
1211 if (ret != BLK_STS_OK)
1212 scsi_free_sgtables(cmd);
1213
1214 return ret;
1215}
1216
1217static blk_status_t
1218scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1219{
1220 switch (sdev->sdev_state) {
1221 case SDEV_OFFLINE:
1222 case SDEV_TRANSPORT_OFFLINE:
1223 /*
1224 * If the device is offline we refuse to process any
1225 * commands. The device must be brought online
1226 * before trying any recovery commands.
1227 */
1228 if (!sdev->offline_already) {
1229 sdev->offline_already = true;
1230 sdev_printk(KERN_ERR, sdev,
1231 "rejecting I/O to offline device\n");
1232 }
1233 return BLK_STS_IOERR;
1234 case SDEV_DEL:
1235 /*
1236 * If the device is fully deleted, we refuse to
1237 * process any commands as well.
1238 */
1239 sdev_printk(KERN_ERR, sdev,
1240 "rejecting I/O to dead device\n");
1241 return BLK_STS_IOERR;
1242 case SDEV_BLOCK:
1243 case SDEV_CREATED_BLOCK:
1244 return BLK_STS_RESOURCE;
1245 case SDEV_QUIESCE:
1246 /*
1247 * If the devices is blocked we defer normal commands.
1248 */
1249 if (req && !(req->rq_flags & RQF_PREEMPT))
1250 return BLK_STS_RESOURCE;
1251 return BLK_STS_OK;
1252 default:
1253 /*
1254 * For any other not fully online state we only allow
1255 * special commands. In particular any user initiated
1256 * command is not allowed.
1257 */
1258 if (req && !(req->rq_flags & RQF_PREEMPT))
1259 return BLK_STS_IOERR;
1260 return BLK_STS_OK;
1261 }
1262}
1263
1264/*
1265 * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1266 * return 0.
1267 *
1268 * Called with the queue_lock held.
1269 */
1270static inline int scsi_dev_queue_ready(struct request_queue *q,
1271 struct scsi_device *sdev)
1272{
1273 unsigned int busy;
1274
1275 busy = atomic_inc_return(&sdev->device_busy) - 1;
1276 if (atomic_read(&sdev->device_blocked)) {
1277 if (busy)
1278 goto out_dec;
1279
1280 /*
1281 * unblock after device_blocked iterates to zero
1282 */
1283 if (atomic_dec_return(&sdev->device_blocked) > 0)
1284 goto out_dec;
1285 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1286 "unblocking device at zero depth\n"));
1287 }
1288
1289 if (busy >= sdev->queue_depth)
1290 goto out_dec;
1291
1292 return 1;
1293out_dec:
1294 atomic_dec(&sdev->device_busy);
1295 return 0;
1296}
1297
1298/*
1299 * scsi_target_queue_ready: checks if there we can send commands to target
1300 * @sdev: scsi device on starget to check.
1301 */
1302static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1303 struct scsi_device *sdev)
1304{
1305 struct scsi_target *starget = scsi_target(sdev);
1306 unsigned int busy;
1307
1308 if (starget->single_lun) {
1309 spin_lock_irq(shost->host_lock);
1310 if (starget->starget_sdev_user &&
1311 starget->starget_sdev_user != sdev) {
1312 spin_unlock_irq(shost->host_lock);
1313 return 0;
1314 }
1315 starget->starget_sdev_user = sdev;
1316 spin_unlock_irq(shost->host_lock);
1317 }
1318
1319 if (starget->can_queue <= 0)
1320 return 1;
1321
1322 busy = atomic_inc_return(&starget->target_busy) - 1;
1323 if (atomic_read(&starget->target_blocked) > 0) {
1324 if (busy)
1325 goto starved;
1326
1327 /*
1328 * unblock after target_blocked iterates to zero
1329 */
1330 if (atomic_dec_return(&starget->target_blocked) > 0)
1331 goto out_dec;
1332
1333 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1334 "unblocking target at zero depth\n"));
1335 }
1336
1337 if (busy >= starget->can_queue)
1338 goto starved;
1339
1340 return 1;
1341
1342starved:
1343 spin_lock_irq(shost->host_lock);
1344 list_move_tail(&sdev->starved_entry, &shost->starved_list);
1345 spin_unlock_irq(shost->host_lock);
1346out_dec:
1347 if (starget->can_queue > 0)
1348 atomic_dec(&starget->target_busy);
1349 return 0;
1350}
1351
1352/*
1353 * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1354 * return 0. We must end up running the queue again whenever 0 is
1355 * returned, else IO can hang.
1356 */
1357static inline int scsi_host_queue_ready(struct request_queue *q,
1358 struct Scsi_Host *shost,
1359 struct scsi_device *sdev,
1360 struct scsi_cmnd *cmd)
1361{
1362 if (scsi_host_in_recovery(shost))
1363 return 0;
1364
1365 if (atomic_read(&shost->host_blocked) > 0) {
1366 if (scsi_host_busy(shost) > 0)
1367 goto starved;
1368
1369 /*
1370 * unblock after host_blocked iterates to zero
1371 */
1372 if (atomic_dec_return(&shost->host_blocked) > 0)
1373 goto out_dec;
1374
1375 SCSI_LOG_MLQUEUE(3,
1376 shost_printk(KERN_INFO, shost,
1377 "unblocking host at zero depth\n"));
1378 }
1379
1380 if (shost->host_self_blocked)
1381 goto starved;
1382
1383 /* We're OK to process the command, so we can't be starved */
1384 if (!list_empty(&sdev->starved_entry)) {
1385 spin_lock_irq(shost->host_lock);
1386 if (!list_empty(&sdev->starved_entry))
1387 list_del_init(&sdev->starved_entry);
1388 spin_unlock_irq(shost->host_lock);
1389 }
1390
1391 __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1392
1393 return 1;
1394
1395starved:
1396 spin_lock_irq(shost->host_lock);
1397 if (list_empty(&sdev->starved_entry))
1398 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1399 spin_unlock_irq(shost->host_lock);
1400out_dec:
1401 scsi_dec_host_busy(shost, cmd);
1402 return 0;
1403}
1404
1405/*
1406 * Busy state exporting function for request stacking drivers.
1407 *
1408 * For efficiency, no lock is taken to check the busy state of
1409 * shost/starget/sdev, since the returned value is not guaranteed and
1410 * may be changed after request stacking drivers call the function,
1411 * regardless of taking lock or not.
1412 *
1413 * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1414 * needs to return 'not busy'. Otherwise, request stacking drivers
1415 * may hold requests forever.
1416 */
1417static bool scsi_mq_lld_busy(struct request_queue *q)
1418{
1419 struct scsi_device *sdev = q->queuedata;
1420 struct Scsi_Host *shost;
1421
1422 if (blk_queue_dying(q))
1423 return false;
1424
1425 shost = sdev->host;
1426
1427 /*
1428 * Ignore host/starget busy state.
1429 * Since block layer does not have a concept of fairness across
1430 * multiple queues, congestion of host/starget needs to be handled
1431 * in SCSI layer.
1432 */
1433 if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1434 return true;
1435
1436 return false;
1437}
1438
1439static void scsi_softirq_done(struct request *rq)
1440{
1441 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1442 unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1443 int disposition;
1444
1445 INIT_LIST_HEAD(&cmd->eh_entry);
1446
1447 atomic_inc(&cmd->device->iodone_cnt);
1448 if (cmd->result)
1449 atomic_inc(&cmd->device->ioerr_cnt);
1450
1451 disposition = scsi_decide_disposition(cmd);
1452 if (disposition != SUCCESS &&
1453 time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1454 scmd_printk(KERN_ERR, cmd,
1455 "timing out command, waited %lus\n",
1456 wait_for/HZ);
1457 disposition = SUCCESS;
1458 }
1459
1460 scsi_log_completion(cmd, disposition);
1461
1462 switch (disposition) {
1463 case SUCCESS:
1464 scsi_finish_command(cmd);
1465 break;
1466 case NEEDS_RETRY:
1467 scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1468 break;
1469 case ADD_TO_MLQUEUE:
1470 scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1471 break;
1472 default:
1473 scsi_eh_scmd_add(cmd);
1474 break;
1475 }
1476}
1477
1478/**
1479 * scsi_dispatch_command - Dispatch a command to the low-level driver.
1480 * @cmd: command block we are dispatching.
1481 *
1482 * Return: nonzero return request was rejected and device's queue needs to be
1483 * plugged.
1484 */
1485static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1486{
1487 struct Scsi_Host *host = cmd->device->host;
1488 int rtn = 0;
1489
1490 atomic_inc(&cmd->device->iorequest_cnt);
1491
1492 /* check if the device is still usable */
1493 if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1494 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1495 * returns an immediate error upwards, and signals
1496 * that the device is no longer present */
1497 cmd->result = DID_NO_CONNECT << 16;
1498 goto done;
1499 }
1500
1501 /* Check to see if the scsi lld made this device blocked. */
1502 if (unlikely(scsi_device_blocked(cmd->device))) {
1503 /*
1504 * in blocked state, the command is just put back on
1505 * the device queue. The suspend state has already
1506 * blocked the queue so future requests should not
1507 * occur until the device transitions out of the
1508 * suspend state.
1509 */
1510 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1511 "queuecommand : device blocked\n"));
1512 return SCSI_MLQUEUE_DEVICE_BUSY;
1513 }
1514
1515 /* Store the LUN value in cmnd, if needed. */
1516 if (cmd->device->lun_in_cdb)
1517 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1518 (cmd->device->lun << 5 & 0xe0);
1519
1520 scsi_log_send(cmd);
1521
1522 /*
1523 * Before we queue this command, check if the command
1524 * length exceeds what the host adapter can handle.
1525 */
1526 if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1527 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1528 "queuecommand : command too long. "
1529 "cdb_size=%d host->max_cmd_len=%d\n",
1530 cmd->cmd_len, cmd->device->host->max_cmd_len));
1531 cmd->result = (DID_ABORT << 16);
1532 goto done;
1533 }
1534
1535 if (unlikely(host->shost_state == SHOST_DEL)) {
1536 cmd->result = (DID_NO_CONNECT << 16);
1537 goto done;
1538
1539 }
1540
1541 trace_scsi_dispatch_cmd_start(cmd);
1542 rtn = host->hostt->queuecommand(host, cmd);
1543 if (rtn) {
1544 trace_scsi_dispatch_cmd_error(cmd, rtn);
1545 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1546 rtn != SCSI_MLQUEUE_TARGET_BUSY)
1547 rtn = SCSI_MLQUEUE_HOST_BUSY;
1548
1549 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1550 "queuecommand : request rejected\n"));
1551 }
1552
1553 return rtn;
1554 done:
1555 cmd->scsi_done(cmd);
1556 return 0;
1557}
1558
1559/* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1560static unsigned int scsi_mq_inline_sgl_size(struct Scsi_Host *shost)
1561{
1562 return min_t(unsigned int, shost->sg_tablesize, SCSI_INLINE_SG_CNT) *
1563 sizeof(struct scatterlist);
1564}
1565
1566static blk_status_t scsi_mq_prep_fn(struct request *req)
1567{
1568 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1569 struct scsi_device *sdev = req->q->queuedata;
1570 struct Scsi_Host *shost = sdev->host;
1571 struct scatterlist *sg;
1572
1573 scsi_init_command(sdev, cmd);
1574
1575 cmd->request = req;
1576 cmd->tag = req->tag;
1577 cmd->prot_op = SCSI_PROT_NORMAL;
1578
1579 sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1580 cmd->sdb.table.sgl = sg;
1581
1582 if (scsi_host_get_prot(shost)) {
1583 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1584
1585 cmd->prot_sdb->table.sgl =
1586 (struct scatterlist *)(cmd->prot_sdb + 1);
1587 }
1588
1589 blk_mq_start_request(req);
1590
1591 return scsi_setup_cmnd(sdev, req);
1592}
1593
1594static void scsi_mq_done(struct scsi_cmnd *cmd)
1595{
1596 if (unlikely(blk_should_fake_timeout(cmd->request->q)))
1597 return;
1598 if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1599 return;
1600 trace_scsi_dispatch_cmd_done(cmd);
1601 blk_mq_complete_request(cmd->request);
1602}
1603
1604static void scsi_mq_put_budget(struct request_queue *q)
1605{
1606 struct scsi_device *sdev = q->queuedata;
1607
1608 atomic_dec(&sdev->device_busy);
1609}
1610
1611static bool scsi_mq_get_budget(struct request_queue *q)
1612{
1613 struct scsi_device *sdev = q->queuedata;
1614
1615 return scsi_dev_queue_ready(q, sdev);
1616}
1617
1618static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1619 const struct blk_mq_queue_data *bd)
1620{
1621 struct request *req = bd->rq;
1622 struct request_queue *q = req->q;
1623 struct scsi_device *sdev = q->queuedata;
1624 struct Scsi_Host *shost = sdev->host;
1625 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1626 blk_status_t ret;
1627 int reason;
1628
1629 /*
1630 * If the device is not in running state we will reject some or all
1631 * commands.
1632 */
1633 if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1634 ret = scsi_prep_state_check(sdev, req);
1635 if (ret != BLK_STS_OK)
1636 goto out_put_budget;
1637 }
1638
1639 ret = BLK_STS_RESOURCE;
1640 if (!scsi_target_queue_ready(shost, sdev))
1641 goto out_put_budget;
1642 if (!scsi_host_queue_ready(q, shost, sdev, cmd))
1643 goto out_dec_target_busy;
1644
1645 if (!(req->rq_flags & RQF_DONTPREP)) {
1646 ret = scsi_mq_prep_fn(req);
1647 if (ret != BLK_STS_OK)
1648 goto out_dec_host_busy;
1649 req->rq_flags |= RQF_DONTPREP;
1650 } else {
1651 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1652 blk_mq_start_request(req);
1653 }
1654
1655 cmd->flags &= SCMD_PRESERVED_FLAGS;
1656 if (sdev->simple_tags)
1657 cmd->flags |= SCMD_TAGGED;
1658 if (bd->last)
1659 cmd->flags |= SCMD_LAST;
1660
1661 scsi_init_cmd_errh(cmd);
1662 cmd->scsi_done = scsi_mq_done;
1663
1664 reason = scsi_dispatch_cmd(cmd);
1665 if (reason) {
1666 scsi_set_blocked(cmd, reason);
1667 ret = BLK_STS_RESOURCE;
1668 goto out_dec_host_busy;
1669 }
1670
1671 return BLK_STS_OK;
1672
1673out_dec_host_busy:
1674 scsi_dec_host_busy(shost, cmd);
1675out_dec_target_busy:
1676 if (scsi_target(sdev)->can_queue > 0)
1677 atomic_dec(&scsi_target(sdev)->target_busy);
1678out_put_budget:
1679 scsi_mq_put_budget(q);
1680 switch (ret) {
1681 case BLK_STS_OK:
1682 break;
1683 case BLK_STS_RESOURCE:
1684 case BLK_STS_ZONE_RESOURCE:
1685 if (atomic_read(&sdev->device_busy) ||
1686 scsi_device_blocked(sdev))
1687 ret = BLK_STS_DEV_RESOURCE;
1688 break;
1689 default:
1690 if (unlikely(!scsi_device_online(sdev)))
1691 scsi_req(req)->result = DID_NO_CONNECT << 16;
1692 else
1693 scsi_req(req)->result = DID_ERROR << 16;
1694 /*
1695 * Make sure to release all allocated resources when
1696 * we hit an error, as we will never see this command
1697 * again.
1698 */
1699 if (req->rq_flags & RQF_DONTPREP)
1700 scsi_mq_uninit_cmd(cmd);
1701 scsi_run_queue_async(sdev);
1702 break;
1703 }
1704 return ret;
1705}
1706
1707static enum blk_eh_timer_return scsi_timeout(struct request *req,
1708 bool reserved)
1709{
1710 if (reserved)
1711 return BLK_EH_RESET_TIMER;
1712 return scsi_times_out(req);
1713}
1714
1715static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1716 unsigned int hctx_idx, unsigned int numa_node)
1717{
1718 struct Scsi_Host *shost = set->driver_data;
1719 const bool unchecked_isa_dma = shost->unchecked_isa_dma;
1720 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1721 struct scatterlist *sg;
1722 int ret = 0;
1723
1724 if (unchecked_isa_dma)
1725 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
1726 cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
1727 GFP_KERNEL, numa_node);
1728 if (!cmd->sense_buffer)
1729 return -ENOMEM;
1730 cmd->req.sense = cmd->sense_buffer;
1731
1732 if (scsi_host_get_prot(shost)) {
1733 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1734 shost->hostt->cmd_size;
1735 cmd->prot_sdb = (void *)sg + scsi_mq_inline_sgl_size(shost);
1736 }
1737
1738 if (shost->hostt->init_cmd_priv) {
1739 ret = shost->hostt->init_cmd_priv(shost, cmd);
1740 if (ret < 0)
1741 scsi_free_sense_buffer(unchecked_isa_dma,
1742 cmd->sense_buffer);
1743 }
1744
1745 return ret;
1746}
1747
1748static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1749 unsigned int hctx_idx)
1750{
1751 struct Scsi_Host *shost = set->driver_data;
1752 struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1753
1754 if (shost->hostt->exit_cmd_priv)
1755 shost->hostt->exit_cmd_priv(shost, cmd);
1756 scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
1757 cmd->sense_buffer);
1758}
1759
1760static int scsi_map_queues(struct blk_mq_tag_set *set)
1761{
1762 struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
1763
1764 if (shost->hostt->map_queues)
1765 return shost->hostt->map_queues(shost);
1766 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
1767}
1768
1769void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
1770{
1771 struct device *dev = shost->dma_dev;
1772
1773 /*
1774 * this limit is imposed by hardware restrictions
1775 */
1776 blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1777 SG_MAX_SEGMENTS));
1778
1779 if (scsi_host_prot_dma(shost)) {
1780 shost->sg_prot_tablesize =
1781 min_not_zero(shost->sg_prot_tablesize,
1782 (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1783 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1784 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1785 }
1786
1787 if (dev->dma_mask) {
1788 shost->max_sectors = min_t(unsigned int, shost->max_sectors,
1789 dma_max_mapping_size(dev) >> SECTOR_SHIFT);
1790 }
1791 blk_queue_max_hw_sectors(q, shost->max_sectors);
1792 if (shost->unchecked_isa_dma)
1793 blk_queue_bounce_limit(q, BLK_BOUNCE_ISA);
1794 blk_queue_segment_boundary(q, shost->dma_boundary);
1795 dma_set_seg_boundary(dev, shost->dma_boundary);
1796
1797 blk_queue_max_segment_size(q, shost->max_segment_size);
1798 blk_queue_virt_boundary(q, shost->virt_boundary_mask);
1799 dma_set_max_seg_size(dev, queue_max_segment_size(q));
1800
1801 /*
1802 * Set a reasonable default alignment: The larger of 32-byte (dword),
1803 * which is a common minimum for HBAs, and the minimum DMA alignment,
1804 * which is set by the platform.
1805 *
1806 * Devices that require a bigger alignment can increase it later.
1807 */
1808 blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
1809}
1810EXPORT_SYMBOL_GPL(__scsi_init_queue);
1811
1812static const struct blk_mq_ops scsi_mq_ops_no_commit = {
1813 .get_budget = scsi_mq_get_budget,
1814 .put_budget = scsi_mq_put_budget,
1815 .queue_rq = scsi_queue_rq,
1816 .complete = scsi_softirq_done,
1817 .timeout = scsi_timeout,
1818#ifdef CONFIG_BLK_DEBUG_FS
1819 .show_rq = scsi_show_rq,
1820#endif
1821 .init_request = scsi_mq_init_request,
1822 .exit_request = scsi_mq_exit_request,
1823 .initialize_rq_fn = scsi_initialize_rq,
1824 .cleanup_rq = scsi_cleanup_rq,
1825 .busy = scsi_mq_lld_busy,
1826 .map_queues = scsi_map_queues,
1827};
1828
1829
1830static void scsi_commit_rqs(struct blk_mq_hw_ctx *hctx)
1831{
1832 struct request_queue *q = hctx->queue;
1833 struct scsi_device *sdev = q->queuedata;
1834 struct Scsi_Host *shost = sdev->host;
1835
1836 shost->hostt->commit_rqs(shost, hctx->queue_num);
1837}
1838
1839static const struct blk_mq_ops scsi_mq_ops = {
1840 .get_budget = scsi_mq_get_budget,
1841 .put_budget = scsi_mq_put_budget,
1842 .queue_rq = scsi_queue_rq,
1843 .commit_rqs = scsi_commit_rqs,
1844 .complete = scsi_softirq_done,
1845 .timeout = scsi_timeout,
1846#ifdef CONFIG_BLK_DEBUG_FS
1847 .show_rq = scsi_show_rq,
1848#endif
1849 .init_request = scsi_mq_init_request,
1850 .exit_request = scsi_mq_exit_request,
1851 .initialize_rq_fn = scsi_initialize_rq,
1852 .cleanup_rq = scsi_cleanup_rq,
1853 .busy = scsi_mq_lld_busy,
1854 .map_queues = scsi_map_queues,
1855};
1856
1857struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
1858{
1859 sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
1860 if (IS_ERR(sdev->request_queue))
1861 return NULL;
1862
1863 sdev->request_queue->queuedata = sdev;
1864 __scsi_init_queue(sdev->host, sdev->request_queue);
1865 blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, sdev->request_queue);
1866 return sdev->request_queue;
1867}
1868
1869int scsi_mq_setup_tags(struct Scsi_Host *shost)
1870{
1871 unsigned int cmd_size, sgl_size;
1872 struct blk_mq_tag_set *tag_set = &shost->tag_set;
1873
1874 sgl_size = max_t(unsigned int, sizeof(struct scatterlist),
1875 scsi_mq_inline_sgl_size(shost));
1876 cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
1877 if (scsi_host_get_prot(shost))
1878 cmd_size += sizeof(struct scsi_data_buffer) +
1879 sizeof(struct scatterlist) * SCSI_INLINE_PROT_SG_CNT;
1880
1881 memset(tag_set, 0, sizeof(*tag_set));
1882 if (shost->hostt->commit_rqs)
1883 tag_set->ops = &scsi_mq_ops;
1884 else
1885 tag_set->ops = &scsi_mq_ops_no_commit;
1886 tag_set->nr_hw_queues = shost->nr_hw_queues ? : 1;
1887 tag_set->queue_depth = shost->can_queue;
1888 tag_set->cmd_size = cmd_size;
1889 tag_set->numa_node = NUMA_NO_NODE;
1890 tag_set->flags = BLK_MQ_F_SHOULD_MERGE;
1891 tag_set->flags |=
1892 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
1893 tag_set->driver_data = shost;
1894
1895 return blk_mq_alloc_tag_set(tag_set);
1896}
1897
1898void scsi_mq_destroy_tags(struct Scsi_Host *shost)
1899{
1900 blk_mq_free_tag_set(&shost->tag_set);
1901}
1902
1903/**
1904 * scsi_device_from_queue - return sdev associated with a request_queue
1905 * @q: The request queue to return the sdev from
1906 *
1907 * Return the sdev associated with a request queue or NULL if the
1908 * request_queue does not reference a SCSI device.
1909 */
1910struct scsi_device *scsi_device_from_queue(struct request_queue *q)
1911{
1912 struct scsi_device *sdev = NULL;
1913
1914 if (q->mq_ops == &scsi_mq_ops_no_commit ||
1915 q->mq_ops == &scsi_mq_ops)
1916 sdev = q->queuedata;
1917 if (!sdev || !get_device(&sdev->sdev_gendev))
1918 sdev = NULL;
1919
1920 return sdev;
1921}
1922EXPORT_SYMBOL_GPL(scsi_device_from_queue);
1923
1924/**
1925 * scsi_block_requests - Utility function used by low-level drivers to prevent
1926 * further commands from being queued to the device.
1927 * @shost: host in question
1928 *
1929 * There is no timer nor any other means by which the requests get unblocked
1930 * other than the low-level driver calling scsi_unblock_requests().
1931 */
1932void scsi_block_requests(struct Scsi_Host *shost)
1933{
1934 shost->host_self_blocked = 1;
1935}
1936EXPORT_SYMBOL(scsi_block_requests);
1937
1938/**
1939 * scsi_unblock_requests - Utility function used by low-level drivers to allow
1940 * further commands to be queued to the device.
1941 * @shost: host in question
1942 *
1943 * There is no timer nor any other means by which the requests get unblocked
1944 * other than the low-level driver calling scsi_unblock_requests(). This is done
1945 * as an API function so that changes to the internals of the scsi mid-layer
1946 * won't require wholesale changes to drivers that use this feature.
1947 */
1948void scsi_unblock_requests(struct Scsi_Host *shost)
1949{
1950 shost->host_self_blocked = 0;
1951 scsi_run_host_queues(shost);
1952}
1953EXPORT_SYMBOL(scsi_unblock_requests);
1954
1955void scsi_exit_queue(void)
1956{
1957 kmem_cache_destroy(scsi_sense_cache);
1958 kmem_cache_destroy(scsi_sense_isadma_cache);
1959}
1960
1961/**
1962 * scsi_mode_select - issue a mode select
1963 * @sdev: SCSI device to be queried
1964 * @pf: Page format bit (1 == standard, 0 == vendor specific)
1965 * @sp: Save page bit (0 == don't save, 1 == save)
1966 * @modepage: mode page being requested
1967 * @buffer: request buffer (may not be smaller than eight bytes)
1968 * @len: length of request buffer.
1969 * @timeout: command timeout
1970 * @retries: number of retries before failing
1971 * @data: returns a structure abstracting the mode header data
1972 * @sshdr: place to put sense data (or NULL if no sense to be collected).
1973 * must be SCSI_SENSE_BUFFERSIZE big.
1974 *
1975 * Returns zero if successful; negative error number or scsi
1976 * status on error
1977 *
1978 */
1979int
1980scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1981 unsigned char *buffer, int len, int timeout, int retries,
1982 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1983{
1984 unsigned char cmd[10];
1985 unsigned char *real_buffer;
1986 int ret;
1987
1988 memset(cmd, 0, sizeof(cmd));
1989 cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1990
1991 if (sdev->use_10_for_ms) {
1992 if (len > 65535)
1993 return -EINVAL;
1994 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1995 if (!real_buffer)
1996 return -ENOMEM;
1997 memcpy(real_buffer + 8, buffer, len);
1998 len += 8;
1999 real_buffer[0] = 0;
2000 real_buffer[1] = 0;
2001 real_buffer[2] = data->medium_type;
2002 real_buffer[3] = data->device_specific;
2003 real_buffer[4] = data->longlba ? 0x01 : 0;
2004 real_buffer[5] = 0;
2005 real_buffer[6] = data->block_descriptor_length >> 8;
2006 real_buffer[7] = data->block_descriptor_length;
2007
2008 cmd[0] = MODE_SELECT_10;
2009 cmd[7] = len >> 8;
2010 cmd[8] = len;
2011 } else {
2012 if (len > 255 || data->block_descriptor_length > 255 ||
2013 data->longlba)
2014 return -EINVAL;
2015
2016 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2017 if (!real_buffer)
2018 return -ENOMEM;
2019 memcpy(real_buffer + 4, buffer, len);
2020 len += 4;
2021 real_buffer[0] = 0;
2022 real_buffer[1] = data->medium_type;
2023 real_buffer[2] = data->device_specific;
2024 real_buffer[3] = data->block_descriptor_length;
2025
2026 cmd[0] = MODE_SELECT;
2027 cmd[4] = len;
2028 }
2029
2030 ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2031 sshdr, timeout, retries, NULL);
2032 kfree(real_buffer);
2033 return ret;
2034}
2035EXPORT_SYMBOL_GPL(scsi_mode_select);
2036
2037/**
2038 * scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2039 * @sdev: SCSI device to be queried
2040 * @dbd: set if mode sense will allow block descriptors to be returned
2041 * @modepage: mode page being requested
2042 * @buffer: request buffer (may not be smaller than eight bytes)
2043 * @len: length of request buffer.
2044 * @timeout: command timeout
2045 * @retries: number of retries before failing
2046 * @data: returns a structure abstracting the mode header data
2047 * @sshdr: place to put sense data (or NULL if no sense to be collected).
2048 * must be SCSI_SENSE_BUFFERSIZE big.
2049 *
2050 * Returns zero if unsuccessful, or the header offset (either 4
2051 * or 8 depending on whether a six or ten byte command was
2052 * issued) if successful.
2053 */
2054int
2055scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2056 unsigned char *buffer, int len, int timeout, int retries,
2057 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2058{
2059 unsigned char cmd[12];
2060 int use_10_for_ms;
2061 int header_length;
2062 int result, retry_count = retries;
2063 struct scsi_sense_hdr my_sshdr;
2064
2065 memset(data, 0, sizeof(*data));
2066 memset(&cmd[0], 0, 12);
2067
2068 dbd = sdev->set_dbd_for_ms ? 8 : dbd;
2069 cmd[1] = dbd & 0x18; /* allows DBD and LLBA bits */
2070 cmd[2] = modepage;
2071
2072 /* caller might not be interested in sense, but we need it */
2073 if (!sshdr)
2074 sshdr = &my_sshdr;
2075
2076 retry:
2077 use_10_for_ms = sdev->use_10_for_ms;
2078
2079 if (use_10_for_ms) {
2080 if (len < 8)
2081 len = 8;
2082
2083 cmd[0] = MODE_SENSE_10;
2084 cmd[8] = len;
2085 header_length = 8;
2086 } else {
2087 if (len < 4)
2088 len = 4;
2089
2090 cmd[0] = MODE_SENSE;
2091 cmd[4] = len;
2092 header_length = 4;
2093 }
2094
2095 memset(buffer, 0, len);
2096
2097 result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2098 sshdr, timeout, retries, NULL);
2099
2100 /* This code looks awful: what it's doing is making sure an
2101 * ILLEGAL REQUEST sense return identifies the actual command
2102 * byte as the problem. MODE_SENSE commands can return
2103 * ILLEGAL REQUEST if the code page isn't supported */
2104
2105 if (use_10_for_ms && !scsi_status_is_good(result) &&
2106 driver_byte(result) == DRIVER_SENSE) {
2107 if (scsi_sense_valid(sshdr)) {
2108 if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2109 (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2110 /*
2111 * Invalid command operation code
2112 */
2113 sdev->use_10_for_ms = 0;
2114 goto retry;
2115 }
2116 }
2117 }
2118
2119 if (scsi_status_is_good(result)) {
2120 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2121 (modepage == 6 || modepage == 8))) {
2122 /* Initio breakage? */
2123 header_length = 0;
2124 data->length = 13;
2125 data->medium_type = 0;
2126 data->device_specific = 0;
2127 data->longlba = 0;
2128 data->block_descriptor_length = 0;
2129 } else if (use_10_for_ms) {
2130 data->length = buffer[0]*256 + buffer[1] + 2;
2131 data->medium_type = buffer[2];
2132 data->device_specific = buffer[3];
2133 data->longlba = buffer[4] & 0x01;
2134 data->block_descriptor_length = buffer[6]*256
2135 + buffer[7];
2136 } else {
2137 data->length = buffer[0] + 1;
2138 data->medium_type = buffer[1];
2139 data->device_specific = buffer[2];
2140 data->block_descriptor_length = buffer[3];
2141 }
2142 data->header_length = header_length;
2143 } else if ((status_byte(result) == CHECK_CONDITION) &&
2144 scsi_sense_valid(sshdr) &&
2145 sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2146 retry_count--;
2147 goto retry;
2148 }
2149
2150 return result;
2151}
2152EXPORT_SYMBOL(scsi_mode_sense);
2153
2154/**
2155 * scsi_test_unit_ready - test if unit is ready
2156 * @sdev: scsi device to change the state of.
2157 * @timeout: command timeout
2158 * @retries: number of retries before failing
2159 * @sshdr: outpout pointer for decoded sense information.
2160 *
2161 * Returns zero if unsuccessful or an error if TUR failed. For
2162 * removable media, UNIT_ATTENTION sets ->changed flag.
2163 **/
2164int
2165scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2166 struct scsi_sense_hdr *sshdr)
2167{
2168 char cmd[] = {
2169 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2170 };
2171 int result;
2172
2173 /* try to eat the UNIT_ATTENTION if there are enough retries */
2174 do {
2175 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2176 timeout, 1, NULL);
2177 if (sdev->removable && scsi_sense_valid(sshdr) &&
2178 sshdr->sense_key == UNIT_ATTENTION)
2179 sdev->changed = 1;
2180 } while (scsi_sense_valid(sshdr) &&
2181 sshdr->sense_key == UNIT_ATTENTION && --retries);
2182
2183 return result;
2184}
2185EXPORT_SYMBOL(scsi_test_unit_ready);
2186
2187/**
2188 * scsi_device_set_state - Take the given device through the device state model.
2189 * @sdev: scsi device to change the state of.
2190 * @state: state to change to.
2191 *
2192 * Returns zero if successful or an error if the requested
2193 * transition is illegal.
2194 */
2195int
2196scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2197{
2198 enum scsi_device_state oldstate = sdev->sdev_state;
2199
2200 if (state == oldstate)
2201 return 0;
2202
2203 switch (state) {
2204 case SDEV_CREATED:
2205 switch (oldstate) {
2206 case SDEV_CREATED_BLOCK:
2207 break;
2208 default:
2209 goto illegal;
2210 }
2211 break;
2212
2213 case SDEV_RUNNING:
2214 switch (oldstate) {
2215 case SDEV_CREATED:
2216 case SDEV_OFFLINE:
2217 case SDEV_TRANSPORT_OFFLINE:
2218 case SDEV_QUIESCE:
2219 case SDEV_BLOCK:
2220 break;
2221 default:
2222 goto illegal;
2223 }
2224 break;
2225
2226 case SDEV_QUIESCE:
2227 switch (oldstate) {
2228 case SDEV_RUNNING:
2229 case SDEV_OFFLINE:
2230 case SDEV_TRANSPORT_OFFLINE:
2231 break;
2232 default:
2233 goto illegal;
2234 }
2235 break;
2236
2237 case SDEV_OFFLINE:
2238 case SDEV_TRANSPORT_OFFLINE:
2239 switch (oldstate) {
2240 case SDEV_CREATED:
2241 case SDEV_RUNNING:
2242 case SDEV_QUIESCE:
2243 case SDEV_BLOCK:
2244 break;
2245 default:
2246 goto illegal;
2247 }
2248 break;
2249
2250 case SDEV_BLOCK:
2251 switch (oldstate) {
2252 case SDEV_RUNNING:
2253 case SDEV_CREATED_BLOCK:
2254 case SDEV_QUIESCE:
2255 case SDEV_OFFLINE:
2256 break;
2257 default:
2258 goto illegal;
2259 }
2260 break;
2261
2262 case SDEV_CREATED_BLOCK:
2263 switch (oldstate) {
2264 case SDEV_CREATED:
2265 break;
2266 default:
2267 goto illegal;
2268 }
2269 break;
2270
2271 case SDEV_CANCEL:
2272 switch (oldstate) {
2273 case SDEV_CREATED:
2274 case SDEV_RUNNING:
2275 case SDEV_QUIESCE:
2276 case SDEV_OFFLINE:
2277 case SDEV_TRANSPORT_OFFLINE:
2278 break;
2279 default:
2280 goto illegal;
2281 }
2282 break;
2283
2284 case SDEV_DEL:
2285 switch (oldstate) {
2286 case SDEV_CREATED:
2287 case SDEV_RUNNING:
2288 case SDEV_OFFLINE:
2289 case SDEV_TRANSPORT_OFFLINE:
2290 case SDEV_CANCEL:
2291 case SDEV_BLOCK:
2292 case SDEV_CREATED_BLOCK:
2293 break;
2294 default:
2295 goto illegal;
2296 }
2297 break;
2298
2299 }
2300 sdev->offline_already = false;
2301 sdev->sdev_state = state;
2302 return 0;
2303
2304 illegal:
2305 SCSI_LOG_ERROR_RECOVERY(1,
2306 sdev_printk(KERN_ERR, sdev,
2307 "Illegal state transition %s->%s",
2308 scsi_device_state_name(oldstate),
2309 scsi_device_state_name(state))
2310 );
2311 return -EINVAL;
2312}
2313EXPORT_SYMBOL(scsi_device_set_state);
2314
2315/**
2316 * sdev_evt_emit - emit a single SCSI device uevent
2317 * @sdev: associated SCSI device
2318 * @evt: event to emit
2319 *
2320 * Send a single uevent (scsi_event) to the associated scsi_device.
2321 */
2322static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2323{
2324 int idx = 0;
2325 char *envp[3];
2326
2327 switch (evt->evt_type) {
2328 case SDEV_EVT_MEDIA_CHANGE:
2329 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2330 break;
2331 case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2332 scsi_rescan_device(&sdev->sdev_gendev);
2333 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2334 break;
2335 case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2336 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2337 break;
2338 case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2339 envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2340 break;
2341 case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2342 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2343 break;
2344 case SDEV_EVT_LUN_CHANGE_REPORTED:
2345 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2346 break;
2347 case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2348 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2349 break;
2350 case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2351 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2352 break;
2353 default:
2354 /* do nothing */
2355 break;
2356 }
2357
2358 envp[idx++] = NULL;
2359
2360 kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2361}
2362
2363/**
2364 * sdev_evt_thread - send a uevent for each scsi event
2365 * @work: work struct for scsi_device
2366 *
2367 * Dispatch queued events to their associated scsi_device kobjects
2368 * as uevents.
2369 */
2370void scsi_evt_thread(struct work_struct *work)
2371{
2372 struct scsi_device *sdev;
2373 enum scsi_device_event evt_type;
2374 LIST_HEAD(event_list);
2375
2376 sdev = container_of(work, struct scsi_device, event_work);
2377
2378 for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2379 if (test_and_clear_bit(evt_type, sdev->pending_events))
2380 sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2381
2382 while (1) {
2383 struct scsi_event *evt;
2384 struct list_head *this, *tmp;
2385 unsigned long flags;
2386
2387 spin_lock_irqsave(&sdev->list_lock, flags);
2388 list_splice_init(&sdev->event_list, &event_list);
2389 spin_unlock_irqrestore(&sdev->list_lock, flags);
2390
2391 if (list_empty(&event_list))
2392 break;
2393
2394 list_for_each_safe(this, tmp, &event_list) {
2395 evt = list_entry(this, struct scsi_event, node);
2396 list_del(&evt->node);
2397 scsi_evt_emit(sdev, evt);
2398 kfree(evt);
2399 }
2400 }
2401}
2402
2403/**
2404 * sdev_evt_send - send asserted event to uevent thread
2405 * @sdev: scsi_device event occurred on
2406 * @evt: event to send
2407 *
2408 * Assert scsi device event asynchronously.
2409 */
2410void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2411{
2412 unsigned long flags;
2413
2414#if 0
2415 /* FIXME: currently this check eliminates all media change events
2416 * for polled devices. Need to update to discriminate between AN
2417 * and polled events */
2418 if (!test_bit(evt->evt_type, sdev->supported_events)) {
2419 kfree(evt);
2420 return;
2421 }
2422#endif
2423
2424 spin_lock_irqsave(&sdev->list_lock, flags);
2425 list_add_tail(&evt->node, &sdev->event_list);
2426 schedule_work(&sdev->event_work);
2427 spin_unlock_irqrestore(&sdev->list_lock, flags);
2428}
2429EXPORT_SYMBOL_GPL(sdev_evt_send);
2430
2431/**
2432 * sdev_evt_alloc - allocate a new scsi event
2433 * @evt_type: type of event to allocate
2434 * @gfpflags: GFP flags for allocation
2435 *
2436 * Allocates and returns a new scsi_event.
2437 */
2438struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2439 gfp_t gfpflags)
2440{
2441 struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2442 if (!evt)
2443 return NULL;
2444
2445 evt->evt_type = evt_type;
2446 INIT_LIST_HEAD(&evt->node);
2447
2448 /* evt_type-specific initialization, if any */
2449 switch (evt_type) {
2450 case SDEV_EVT_MEDIA_CHANGE:
2451 case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2452 case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2453 case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2454 case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2455 case SDEV_EVT_LUN_CHANGE_REPORTED:
2456 case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2457 case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2458 default:
2459 /* do nothing */
2460 break;
2461 }
2462
2463 return evt;
2464}
2465EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2466
2467/**
2468 * sdev_evt_send_simple - send asserted event to uevent thread
2469 * @sdev: scsi_device event occurred on
2470 * @evt_type: type of event to send
2471 * @gfpflags: GFP flags for allocation
2472 *
2473 * Assert scsi device event asynchronously, given an event type.
2474 */
2475void sdev_evt_send_simple(struct scsi_device *sdev,
2476 enum scsi_device_event evt_type, gfp_t gfpflags)
2477{
2478 struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2479 if (!evt) {
2480 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2481 evt_type);
2482 return;
2483 }
2484
2485 sdev_evt_send(sdev, evt);
2486}
2487EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2488
2489/**
2490 * scsi_device_quiesce - Block user issued commands.
2491 * @sdev: scsi device to quiesce.
2492 *
2493 * This works by trying to transition to the SDEV_QUIESCE state
2494 * (which must be a legal transition). When the device is in this
2495 * state, only special requests will be accepted, all others will
2496 * be deferred. Since special requests may also be requeued requests,
2497 * a successful return doesn't guarantee the device will be
2498 * totally quiescent.
2499 *
2500 * Must be called with user context, may sleep.
2501 *
2502 * Returns zero if unsuccessful or an error if not.
2503 */
2504int
2505scsi_device_quiesce(struct scsi_device *sdev)
2506{
2507 struct request_queue *q = sdev->request_queue;
2508 int err;
2509
2510 /*
2511 * It is allowed to call scsi_device_quiesce() multiple times from
2512 * the same context but concurrent scsi_device_quiesce() calls are
2513 * not allowed.
2514 */
2515 WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2516
2517 if (sdev->quiesced_by == current)
2518 return 0;
2519
2520 blk_set_pm_only(q);
2521
2522 blk_mq_freeze_queue(q);
2523 /*
2524 * Ensure that the effect of blk_set_pm_only() will be visible
2525 * for percpu_ref_tryget() callers that occur after the queue
2526 * unfreeze even if the queue was already frozen before this function
2527 * was called. See also https://lwn.net/Articles/573497/.
2528 */
2529 synchronize_rcu();
2530 blk_mq_unfreeze_queue(q);
2531
2532 mutex_lock(&sdev->state_mutex);
2533 err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2534 if (err == 0)
2535 sdev->quiesced_by = current;
2536 else
2537 blk_clear_pm_only(q);
2538 mutex_unlock(&sdev->state_mutex);
2539
2540 return err;
2541}
2542EXPORT_SYMBOL(scsi_device_quiesce);
2543
2544/**
2545 * scsi_device_resume - Restart user issued commands to a quiesced device.
2546 * @sdev: scsi device to resume.
2547 *
2548 * Moves the device from quiesced back to running and restarts the
2549 * queues.
2550 *
2551 * Must be called with user context, may sleep.
2552 */
2553void scsi_device_resume(struct scsi_device *sdev)
2554{
2555 /* check if the device state was mutated prior to resume, and if
2556 * so assume the state is being managed elsewhere (for example
2557 * device deleted during suspend)
2558 */
2559 mutex_lock(&sdev->state_mutex);
2560 if (sdev->quiesced_by) {
2561 sdev->quiesced_by = NULL;
2562 blk_clear_pm_only(sdev->request_queue);
2563 }
2564 if (sdev->sdev_state == SDEV_QUIESCE)
2565 scsi_device_set_state(sdev, SDEV_RUNNING);
2566 mutex_unlock(&sdev->state_mutex);
2567}
2568EXPORT_SYMBOL(scsi_device_resume);
2569
2570static void
2571device_quiesce_fn(struct scsi_device *sdev, void *data)
2572{
2573 scsi_device_quiesce(sdev);
2574}
2575
2576void
2577scsi_target_quiesce(struct scsi_target *starget)
2578{
2579 starget_for_each_device(starget, NULL, device_quiesce_fn);
2580}
2581EXPORT_SYMBOL(scsi_target_quiesce);
2582
2583static void
2584device_resume_fn(struct scsi_device *sdev, void *data)
2585{
2586 scsi_device_resume(sdev);
2587}
2588
2589void
2590scsi_target_resume(struct scsi_target *starget)
2591{
2592 starget_for_each_device(starget, NULL, device_resume_fn);
2593}
2594EXPORT_SYMBOL(scsi_target_resume);
2595
2596/**
2597 * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2598 * @sdev: device to block
2599 *
2600 * Pause SCSI command processing on the specified device. Does not sleep.
2601 *
2602 * Returns zero if successful or a negative error code upon failure.
2603 *
2604 * Notes:
2605 * This routine transitions the device to the SDEV_BLOCK state (which must be
2606 * a legal transition). When the device is in this state, command processing
2607 * is paused until the device leaves the SDEV_BLOCK state. See also
2608 * scsi_internal_device_unblock_nowait().
2609 */
2610int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2611{
2612 struct request_queue *q = sdev->request_queue;
2613 int err = 0;
2614
2615 err = scsi_device_set_state(sdev, SDEV_BLOCK);
2616 if (err) {
2617 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2618
2619 if (err)
2620 return err;
2621 }
2622
2623 /*
2624 * The device has transitioned to SDEV_BLOCK. Stop the
2625 * block layer from calling the midlayer with this device's
2626 * request queue.
2627 */
2628 blk_mq_quiesce_queue_nowait(q);
2629 return 0;
2630}
2631EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2632
2633/**
2634 * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
2635 * @sdev: device to block
2636 *
2637 * Pause SCSI command processing on the specified device and wait until all
2638 * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
2639 *
2640 * Returns zero if successful or a negative error code upon failure.
2641 *
2642 * Note:
2643 * This routine transitions the device to the SDEV_BLOCK state (which must be
2644 * a legal transition). When the device is in this state, command processing
2645 * is paused until the device leaves the SDEV_BLOCK state. See also
2646 * scsi_internal_device_unblock().
2647 */
2648static int scsi_internal_device_block(struct scsi_device *sdev)
2649{
2650 struct request_queue *q = sdev->request_queue;
2651 int err;
2652
2653 mutex_lock(&sdev->state_mutex);
2654 err = scsi_internal_device_block_nowait(sdev);
2655 if (err == 0)
2656 blk_mq_quiesce_queue(q);
2657 mutex_unlock(&sdev->state_mutex);
2658
2659 return err;
2660}
2661
2662void scsi_start_queue(struct scsi_device *sdev)
2663{
2664 struct request_queue *q = sdev->request_queue;
2665
2666 blk_mq_unquiesce_queue(q);
2667}
2668
2669/**
2670 * scsi_internal_device_unblock_nowait - resume a device after a block request
2671 * @sdev: device to resume
2672 * @new_state: state to set the device to after unblocking
2673 *
2674 * Restart the device queue for a previously suspended SCSI device. Does not
2675 * sleep.
2676 *
2677 * Returns zero if successful or a negative error code upon failure.
2678 *
2679 * Notes:
2680 * This routine transitions the device to the SDEV_RUNNING state or to one of
2681 * the offline states (which must be a legal transition) allowing the midlayer
2682 * to goose the queue for this device.
2683 */
2684int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
2685 enum scsi_device_state new_state)
2686{
2687 switch (new_state) {
2688 case SDEV_RUNNING:
2689 case SDEV_TRANSPORT_OFFLINE:
2690 break;
2691 default:
2692 return -EINVAL;
2693 }
2694
2695 /*
2696 * Try to transition the scsi device to SDEV_RUNNING or one of the
2697 * offlined states and goose the device queue if successful.
2698 */
2699 switch (sdev->sdev_state) {
2700 case SDEV_BLOCK:
2701 case SDEV_TRANSPORT_OFFLINE:
2702 sdev->sdev_state = new_state;
2703 break;
2704 case SDEV_CREATED_BLOCK:
2705 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2706 new_state == SDEV_OFFLINE)
2707 sdev->sdev_state = new_state;
2708 else
2709 sdev->sdev_state = SDEV_CREATED;
2710 break;
2711 case SDEV_CANCEL:
2712 case SDEV_OFFLINE:
2713 break;
2714 default:
2715 return -EINVAL;
2716 }
2717 scsi_start_queue(sdev);
2718
2719 return 0;
2720}
2721EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
2722
2723/**
2724 * scsi_internal_device_unblock - resume a device after a block request
2725 * @sdev: device to resume
2726 * @new_state: state to set the device to after unblocking
2727 *
2728 * Restart the device queue for a previously suspended SCSI device. May sleep.
2729 *
2730 * Returns zero if successful or a negative error code upon failure.
2731 *
2732 * Notes:
2733 * This routine transitions the device to the SDEV_RUNNING state or to one of
2734 * the offline states (which must be a legal transition) allowing the midlayer
2735 * to goose the queue for this device.
2736 */
2737static int scsi_internal_device_unblock(struct scsi_device *sdev,
2738 enum scsi_device_state new_state)
2739{
2740 int ret;
2741
2742 mutex_lock(&sdev->state_mutex);
2743 ret = scsi_internal_device_unblock_nowait(sdev, new_state);
2744 mutex_unlock(&sdev->state_mutex);
2745
2746 return ret;
2747}
2748
2749static void
2750device_block(struct scsi_device *sdev, void *data)
2751{
2752 int ret;
2753
2754 ret = scsi_internal_device_block(sdev);
2755
2756 WARN_ONCE(ret, "scsi_internal_device_block(%s) failed: ret = %d\n",
2757 dev_name(&sdev->sdev_gendev), ret);
2758}
2759
2760static int
2761target_block(struct device *dev, void *data)
2762{
2763 if (scsi_is_target_device(dev))
2764 starget_for_each_device(to_scsi_target(dev), NULL,
2765 device_block);
2766 return 0;
2767}
2768
2769void
2770scsi_target_block(struct device *dev)
2771{
2772 if (scsi_is_target_device(dev))
2773 starget_for_each_device(to_scsi_target(dev), NULL,
2774 device_block);
2775 else
2776 device_for_each_child(dev, NULL, target_block);
2777}
2778EXPORT_SYMBOL_GPL(scsi_target_block);
2779
2780static void
2781device_unblock(struct scsi_device *sdev, void *data)
2782{
2783 scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2784}
2785
2786static int
2787target_unblock(struct device *dev, void *data)
2788{
2789 if (scsi_is_target_device(dev))
2790 starget_for_each_device(to_scsi_target(dev), data,
2791 device_unblock);
2792 return 0;
2793}
2794
2795void
2796scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2797{
2798 if (scsi_is_target_device(dev))
2799 starget_for_each_device(to_scsi_target(dev), &new_state,
2800 device_unblock);
2801 else
2802 device_for_each_child(dev, &new_state, target_unblock);
2803}
2804EXPORT_SYMBOL_GPL(scsi_target_unblock);
2805
2806int
2807scsi_host_block(struct Scsi_Host *shost)
2808{
2809 struct scsi_device *sdev;
2810 int ret = 0;
2811
2812 /*
2813 * Call scsi_internal_device_block_nowait so we can avoid
2814 * calling synchronize_rcu() for each LUN.
2815 */
2816 shost_for_each_device(sdev, shost) {
2817 mutex_lock(&sdev->state_mutex);
2818 ret = scsi_internal_device_block_nowait(sdev);
2819 mutex_unlock(&sdev->state_mutex);
2820 if (ret) {
2821 scsi_device_put(sdev);
2822 break;
2823 }
2824 }
2825
2826 /*
2827 * SCSI never enables blk-mq's BLK_MQ_F_BLOCKING flag so
2828 * calling synchronize_rcu() once is enough.
2829 */
2830 WARN_ON_ONCE(shost->tag_set.flags & BLK_MQ_F_BLOCKING);
2831
2832 if (!ret)
2833 synchronize_rcu();
2834
2835 return ret;
2836}
2837EXPORT_SYMBOL_GPL(scsi_host_block);
2838
2839int
2840scsi_host_unblock(struct Scsi_Host *shost, int new_state)
2841{
2842 struct scsi_device *sdev;
2843 int ret = 0;
2844
2845 shost_for_each_device(sdev, shost) {
2846 ret = scsi_internal_device_unblock(sdev, new_state);
2847 if (ret) {
2848 scsi_device_put(sdev);
2849 break;
2850 }
2851 }
2852 return ret;
2853}
2854EXPORT_SYMBOL_GPL(scsi_host_unblock);
2855
2856/**
2857 * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2858 * @sgl: scatter-gather list
2859 * @sg_count: number of segments in sg
2860 * @offset: offset in bytes into sg, on return offset into the mapped area
2861 * @len: bytes to map, on return number of bytes mapped
2862 *
2863 * Returns virtual address of the start of the mapped page
2864 */
2865void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2866 size_t *offset, size_t *len)
2867{
2868 int i;
2869 size_t sg_len = 0, len_complete = 0;
2870 struct scatterlist *sg;
2871 struct page *page;
2872
2873 WARN_ON(!irqs_disabled());
2874
2875 for_each_sg(sgl, sg, sg_count, i) {
2876 len_complete = sg_len; /* Complete sg-entries */
2877 sg_len += sg->length;
2878 if (sg_len > *offset)
2879 break;
2880 }
2881
2882 if (unlikely(i == sg_count)) {
2883 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2884 "elements %d\n",
2885 __func__, sg_len, *offset, sg_count);
2886 WARN_ON(1);
2887 return NULL;
2888 }
2889
2890 /* Offset starting from the beginning of first page in this sg-entry */
2891 *offset = *offset - len_complete + sg->offset;
2892
2893 /* Assumption: contiguous pages can be accessed as "page + i" */
2894 page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2895 *offset &= ~PAGE_MASK;
2896
2897 /* Bytes in this sg-entry from *offset to the end of the page */
2898 sg_len = PAGE_SIZE - *offset;
2899 if (*len > sg_len)
2900 *len = sg_len;
2901
2902 return kmap_atomic(page);
2903}
2904EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2905
2906/**
2907 * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2908 * @virt: virtual address to be unmapped
2909 */
2910void scsi_kunmap_atomic_sg(void *virt)
2911{
2912 kunmap_atomic(virt);
2913}
2914EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
2915
2916void sdev_disable_disk_events(struct scsi_device *sdev)
2917{
2918 atomic_inc(&sdev->disk_events_disable_depth);
2919}
2920EXPORT_SYMBOL(sdev_disable_disk_events);
2921
2922void sdev_enable_disk_events(struct scsi_device *sdev)
2923{
2924 if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
2925 return;
2926 atomic_dec(&sdev->disk_events_disable_depth);
2927}
2928EXPORT_SYMBOL(sdev_enable_disk_events);
2929
2930/**
2931 * scsi_vpd_lun_id - return a unique device identification
2932 * @sdev: SCSI device
2933 * @id: buffer for the identification
2934 * @id_len: length of the buffer
2935 *
2936 * Copies a unique device identification into @id based
2937 * on the information in the VPD page 0x83 of the device.
2938 * The string will be formatted as a SCSI name string.
2939 *
2940 * Returns the length of the identification or error on failure.
2941 * If the identifier is longer than the supplied buffer the actual
2942 * identifier length is returned and the buffer is not zero-padded.
2943 */
2944int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
2945{
2946 u8 cur_id_type = 0xff;
2947 u8 cur_id_size = 0;
2948 const unsigned char *d, *cur_id_str;
2949 const struct scsi_vpd *vpd_pg83;
2950 int id_size = -EINVAL;
2951
2952 rcu_read_lock();
2953 vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
2954 if (!vpd_pg83) {
2955 rcu_read_unlock();
2956 return -ENXIO;
2957 }
2958
2959 /*
2960 * Look for the correct descriptor.
2961 * Order of preference for lun descriptor:
2962 * - SCSI name string
2963 * - NAA IEEE Registered Extended
2964 * - EUI-64 based 16-byte
2965 * - EUI-64 based 12-byte
2966 * - NAA IEEE Registered
2967 * - NAA IEEE Extended
2968 * - T10 Vendor ID
2969 * as longer descriptors reduce the likelyhood
2970 * of identification clashes.
2971 */
2972
2973 /* The id string must be at least 20 bytes + terminating NULL byte */
2974 if (id_len < 21) {
2975 rcu_read_unlock();
2976 return -EINVAL;
2977 }
2978
2979 memset(id, 0, id_len);
2980 d = vpd_pg83->data + 4;
2981 while (d < vpd_pg83->data + vpd_pg83->len) {
2982 /* Skip designators not referring to the LUN */
2983 if ((d[1] & 0x30) != 0x00)
2984 goto next_desig;
2985
2986 switch (d[1] & 0xf) {
2987 case 0x1:
2988 /* T10 Vendor ID */
2989 if (cur_id_size > d[3])
2990 break;
2991 /* Prefer anything */
2992 if (cur_id_type > 0x01 && cur_id_type != 0xff)
2993 break;
2994 cur_id_size = d[3];
2995 if (cur_id_size + 4 > id_len)
2996 cur_id_size = id_len - 4;
2997 cur_id_str = d + 4;
2998 cur_id_type = d[1] & 0xf;
2999 id_size = snprintf(id, id_len, "t10.%*pE",
3000 cur_id_size, cur_id_str);
3001 break;
3002 case 0x2:
3003 /* EUI-64 */
3004 if (cur_id_size > d[3])
3005 break;
3006 /* Prefer NAA IEEE Registered Extended */
3007 if (cur_id_type == 0x3 &&
3008 cur_id_size == d[3])
3009 break;
3010 cur_id_size = d[3];
3011 cur_id_str = d + 4;
3012 cur_id_type = d[1] & 0xf;
3013 switch (cur_id_size) {
3014 case 8:
3015 id_size = snprintf(id, id_len,
3016 "eui.%8phN",
3017 cur_id_str);
3018 break;
3019 case 12:
3020 id_size = snprintf(id, id_len,
3021 "eui.%12phN",
3022 cur_id_str);
3023 break;
3024 case 16:
3025 id_size = snprintf(id, id_len,
3026 "eui.%16phN",
3027 cur_id_str);
3028 break;
3029 default:
3030 cur_id_size = 0;
3031 break;
3032 }
3033 break;
3034 case 0x3:
3035 /* NAA */
3036 if (cur_id_size > d[3])
3037 break;
3038 cur_id_size = d[3];
3039 cur_id_str = d + 4;
3040 cur_id_type = d[1] & 0xf;
3041 switch (cur_id_size) {
3042 case 8:
3043 id_size = snprintf(id, id_len,
3044 "naa.%8phN",
3045 cur_id_str);
3046 break;
3047 case 16:
3048 id_size = snprintf(id, id_len,
3049 "naa.%16phN",
3050 cur_id_str);
3051 break;
3052 default:
3053 cur_id_size = 0;
3054 break;
3055 }
3056 break;
3057 case 0x8:
3058 /* SCSI name string */
3059 if (cur_id_size + 4 > d[3])
3060 break;
3061 /* Prefer others for truncated descriptor */
3062 if (cur_id_size && d[3] > id_len)
3063 break;
3064 cur_id_size = id_size = d[3];
3065 cur_id_str = d + 4;
3066 cur_id_type = d[1] & 0xf;
3067 if (cur_id_size >= id_len)
3068 cur_id_size = id_len - 1;
3069 memcpy(id, cur_id_str, cur_id_size);
3070 /* Decrease priority for truncated descriptor */
3071 if (cur_id_size != id_size)
3072 cur_id_size = 6;
3073 break;
3074 default:
3075 break;
3076 }
3077next_desig:
3078 d += d[3] + 4;
3079 }
3080 rcu_read_unlock();
3081
3082 return id_size;
3083}
3084EXPORT_SYMBOL(scsi_vpd_lun_id);
3085
3086/*
3087 * scsi_vpd_tpg_id - return a target port group identifier
3088 * @sdev: SCSI device
3089 *
3090 * Returns the Target Port Group identifier from the information
3091 * froom VPD page 0x83 of the device.
3092 *
3093 * Returns the identifier or error on failure.
3094 */
3095int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3096{
3097 const unsigned char *d;
3098 const struct scsi_vpd *vpd_pg83;
3099 int group_id = -EAGAIN, rel_port = -1;
3100
3101 rcu_read_lock();
3102 vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3103 if (!vpd_pg83) {
3104 rcu_read_unlock();
3105 return -ENXIO;
3106 }
3107
3108 d = vpd_pg83->data + 4;
3109 while (d < vpd_pg83->data + vpd_pg83->len) {
3110 switch (d[1] & 0xf) {
3111 case 0x4:
3112 /* Relative target port */
3113 rel_port = get_unaligned_be16(&d[6]);
3114 break;
3115 case 0x5:
3116 /* Target port group */
3117 group_id = get_unaligned_be16(&d[6]);
3118 break;
3119 default:
3120 break;
3121 }
3122 d += d[3] + 4;
3123 }
3124 rcu_read_unlock();
3125
3126 if (group_id >= 0 && rel_id && rel_port != -1)
3127 *rel_id = rel_port;
3128
3129 return group_id;
3130}
3131EXPORT_SYMBOL(scsi_vpd_tpg_id);