2 * scsi_lib.c Copyright (C) 1999 Eric Youngdale
4 * SCSI queueing library.
5 * Initial versions: Eric Youngdale (eric@andante.org).
6 * Based upon conversations with large numbers
7 * of people at Linux Expo.
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/completion.h>
13 #include <linux/kernel.h>
14 #include <linux/mempool.h>
15 #include <linux/slab.h>
16 #include <linux/init.h>
17 #include <linux/pci.h>
19 #include <scsi/scsi_driver.h>
20 #include <scsi/scsi_host.h>
23 #include "scsi_priv.h"
24 #include "scsi_logging.h"
27 #define SG_MEMPOOL_NR (sizeof(scsi_sg_pools)/sizeof(struct scsi_host_sg_pool))
28 #define SG_MEMPOOL_SIZE 32
30 struct scsi_host_sg_pool {
37 #if (SCSI_MAX_PHYS_SEGMENTS < 32)
38 #error SCSI_MAX_PHYS_SEGMENTS is too small
41 #define SP(x) { x, "sgpool-" #x }
42 struct scsi_host_sg_pool scsi_sg_pools[] = {
46 #if (SCSI_MAX_PHYS_SEGMENTS > 32)
48 #if (SCSI_MAX_PHYS_SEGMENTS > 64)
50 #if (SCSI_MAX_PHYS_SEGMENTS > 128)
52 #if (SCSI_MAX_PHYS_SEGMENTS > 256)
53 #error SCSI_MAX_PHYS_SEGMENTS is too large
63 * Function: scsi_insert_special_req()
65 * Purpose: Insert pre-formed request into request queue.
67 * Arguments: sreq - request that is ready to be queued.
68 * at_head - boolean. True if we should insert at head
69 * of queue, false if we should insert at tail.
71 * Lock status: Assumed that lock is not held upon entry.
75 * Notes: This function is called from character device and from
76 * ioctl types of functions where the caller knows exactly
77 * what SCSI command needs to be issued. The idea is that
78 * we merely inject the command into the queue (at the head
79 * for now), and then call the queue request function to actually
82 int scsi_insert_special_req(struct scsi_request *sreq, int at_head)
85 * Because users of this function are apt to reuse requests with no
86 * modification, we have to sanitise the request flags here
88 sreq->sr_request->flags &= ~REQ_DONTPREP;
89 blk_insert_request(sreq->sr_device->request_queue, sreq->sr_request,
95 * Function: scsi_queue_insert()
97 * Purpose: Insert a command in the midlevel queue.
99 * Arguments: cmd - command that we are adding to queue.
100 * reason - why we are inserting command to queue.
102 * Lock status: Assumed that lock is not held upon entry.
106 * Notes: We do this for one of two cases. Either the host is busy
107 * and it cannot accept any more commands for the time being,
108 * or the device returned QUEUE_FULL and can accept no more
110 * Notes: This could be called either from an interrupt context or a
111 * normal process context.
113 int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
115 struct Scsi_Host *host = cmd->device->host;
116 struct scsi_device *device = cmd->device;
119 printk("Inserting command %p into mlqueue\n", cmd));
122 * We are inserting the command into the ml queue. First, we
123 * cancel the timer, so it doesn't time out.
125 scsi_delete_timer(cmd);
128 * Next, set the appropriate busy bit for the device/host.
130 * If the host/device isn't busy, assume that something actually
131 * completed, and that we should be able to queue a command now.
133 * Note that the prior mid-layer assumption that any host could
134 * always queue at least one command is now broken. The mid-layer
135 * will implement a user specifiable stall (see
136 * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
137 * if a command is requeued with no other commands outstanding
138 * either for the device or for the host.
140 if (reason == SCSI_MLQUEUE_HOST_BUSY)
141 host->host_blocked = host->max_host_blocked;
142 else if (reason == SCSI_MLQUEUE_DEVICE_BUSY)
143 device->device_blocked = device->max_device_blocked;
146 * Register the fact that we own the thing for now.
148 cmd->state = SCSI_STATE_MLQUEUE;
149 cmd->owner = SCSI_OWNER_MIDLEVEL;
152 * Decrement the counters, since these commands are no longer
153 * active on the host/device.
155 scsi_device_unbusy(device);
158 * Insert this command at the head of the queue for it's device.
159 * It will go before all other commands that are already in the queue.
161 * NOTE: there is magic here about the way the queue is plugged if
162 * we have no outstanding commands.
164 * Although this *doesn't* plug the queue, it does call the request
165 * function. The SCSI request function detects the blocked condition
166 * and plugs the queue appropriately.
168 blk_insert_request(device->request_queue, cmd->request, 1, cmd, 1);
173 * Function: scsi_do_req
175 * Purpose: Queue a SCSI request
177 * Arguments: sreq - command descriptor.
178 * cmnd - actual SCSI command to be performed.
179 * buffer - data buffer.
180 * bufflen - size of data buffer.
181 * done - completion function to be run.
182 * timeout - how long to let it run before timeout.
183 * retries - number of retries we allow.
185 * Lock status: No locks held upon entry.
189 * Notes: This function is only used for queueing requests for things
190 * like ioctls and character device requests - this is because
191 * we essentially just inject a request into the queue for the
194 * In order to support the scsi_device_quiesce function, we
195 * now inject requests on the *head* of the device queue
196 * rather than the tail.
198 void scsi_do_req(struct scsi_request *sreq, const void *cmnd,
199 void *buffer, unsigned bufflen,
200 void (*done)(struct scsi_cmnd *),
201 int timeout, int retries)
204 * If the upper level driver is reusing these things, then
205 * we should release the low-level block now. Another one will
206 * be allocated later when this request is getting queued.
208 __scsi_release_request(sreq);
211 * Our own function scsi_done (which marks the host as not busy,
212 * disables the timeout counter, etc) will be called by us or by the
213 * scsi_hosts[host].queuecommand() function needs to also call
214 * the completion function for the high level driver.
216 memcpy(sreq->sr_cmnd, cmnd, sizeof(sreq->sr_cmnd));
217 sreq->sr_bufflen = bufflen;
218 sreq->sr_buffer = buffer;
219 sreq->sr_allowed = retries;
220 sreq->sr_done = done;
221 sreq->sr_timeout_per_command = timeout;
223 if (sreq->sr_cmd_len == 0)
224 sreq->sr_cmd_len = COMMAND_SIZE(sreq->sr_cmnd[0]);
227 * head injection *required* here otherwise quiesce won't work
229 scsi_insert_special_req(sreq, 1);
232 static void scsi_wait_done(struct scsi_cmnd *cmd)
234 struct request *req = cmd->request;
235 struct request_queue *q = cmd->device->request_queue;
238 req->rq_status = RQ_SCSI_DONE; /* Busy, but indicate request done */
240 spin_lock_irqsave(q->queue_lock, flags);
241 if (blk_rq_tagged(req))
242 blk_queue_end_tag(q, req);
243 spin_unlock_irqrestore(q->queue_lock, flags);
246 complete(req->waiting);
249 void scsi_wait_req(struct scsi_request *sreq, const void *cmnd, void *buffer,
250 unsigned bufflen, int timeout, int retries)
252 DECLARE_COMPLETION(wait);
254 sreq->sr_request->waiting = &wait;
255 sreq->sr_request->rq_status = RQ_SCSI_BUSY;
256 scsi_do_req(sreq, cmnd, buffer, bufflen, scsi_wait_done,
258 wait_for_completion(&wait);
259 sreq->sr_request->waiting = NULL;
260 if (sreq->sr_request->rq_status != RQ_SCSI_DONE)
261 sreq->sr_result |= (DRIVER_ERROR << 24);
263 __scsi_release_request(sreq);
267 * Function: scsi_init_cmd_errh()
269 * Purpose: Initialize cmd fields related to error handling.
271 * Arguments: cmd - command that is ready to be queued.
275 * Notes: This function has the job of initializing a number of
276 * fields related to error handling. Typically this will
277 * be called once for each command, as required.
279 static int scsi_init_cmd_errh(struct scsi_cmnd *cmd)
281 cmd->owner = SCSI_OWNER_MIDLEVEL;
282 cmd->serial_number = 0;
283 cmd->serial_number_at_timeout = 0;
284 cmd->abort_reason = 0;
286 memset(cmd->sense_buffer, 0, sizeof cmd->sense_buffer);
288 if (cmd->cmd_len == 0)
289 cmd->cmd_len = COMMAND_SIZE(cmd->cmnd[0]);
292 * We need saved copies of a number of fields - this is because
293 * error handling may need to overwrite these with different values
294 * to run different commands, and once error handling is complete,
295 * we will need to restore these values prior to running the actual
298 cmd->old_use_sg = cmd->use_sg;
299 cmd->old_cmd_len = cmd->cmd_len;
300 cmd->sc_old_data_direction = cmd->sc_data_direction;
301 cmd->old_underflow = cmd->underflow;
302 memcpy(cmd->data_cmnd, cmd->cmnd, sizeof(cmd->cmnd));
303 cmd->buffer = cmd->request_buffer;
304 cmd->bufflen = cmd->request_bufflen;
305 cmd->internal_timeout = NORMAL_TIMEOUT;
306 cmd->abort_reason = 0;
312 * Function: scsi_setup_cmd_retry()
314 * Purpose: Restore the command state for a retry
316 * Arguments: cmd - command to be restored
320 * Notes: Immediately prior to retrying a command, we need
321 * to restore certain fields that we saved above.
323 void scsi_setup_cmd_retry(struct scsi_cmnd *cmd)
325 memcpy(cmd->cmnd, cmd->data_cmnd, sizeof(cmd->data_cmnd));
326 cmd->request_buffer = cmd->buffer;
327 cmd->request_bufflen = cmd->bufflen;
328 cmd->use_sg = cmd->old_use_sg;
329 cmd->cmd_len = cmd->old_cmd_len;
330 cmd->sc_data_direction = cmd->sc_old_data_direction;
331 cmd->underflow = cmd->old_underflow;
334 void scsi_device_unbusy(struct scsi_device *sdev)
336 struct Scsi_Host *shost = sdev->host;
339 spin_lock_irqsave(shost->host_lock, flags);
341 if (unlikely(test_bit(SHOST_RECOVERY, &shost->shost_state) &&
343 scsi_eh_wakeup(shost);
344 spin_unlock(shost->host_lock);
345 spin_lock(&sdev->sdev_lock);
347 spin_unlock_irqrestore(&sdev->sdev_lock, flags);
351 * Called for single_lun devices on IO completion. Clear starget_sdev_user,
352 * and call blk_run_queue for all the scsi_devices on the target -
353 * including current_sdev first.
355 * Called with *no* scsi locks held.
357 static void scsi_single_lun_run(struct scsi_device *current_sdev)
359 struct Scsi_Host *shost = current_sdev->host;
360 struct scsi_device *sdev, *tmp;
363 spin_lock_irqsave(shost->host_lock, flags);
364 current_sdev->sdev_target->starget_sdev_user = NULL;
365 spin_unlock_irqrestore(shost->host_lock, flags);
368 * Call blk_run_queue for all LUNs on the target, starting with
369 * current_sdev. We race with others (to set starget_sdev_user),
370 * but in most cases, we will be first. Ideally, each LU on the
371 * target would get some limited time or requests on the target.
373 blk_run_queue(current_sdev->request_queue);
375 spin_lock_irqsave(shost->host_lock, flags);
376 if (current_sdev->sdev_target->starget_sdev_user)
378 list_for_each_entry_safe(sdev, tmp, ¤t_sdev->same_target_siblings,
379 same_target_siblings) {
380 if (scsi_device_get(sdev))
383 spin_unlock_irqrestore(shost->host_lock, flags);
384 blk_run_queue(sdev->request_queue);
385 spin_lock_irqsave(shost->host_lock, flags);
387 scsi_device_put(sdev);
390 spin_unlock_irqrestore(shost->host_lock, flags);
394 * Function: scsi_run_queue()
396 * Purpose: Select a proper request queue to serve next
398 * Arguments: q - last request's queue
402 * Notes: The previous command was completely finished, start
403 * a new one if possible.
405 static void scsi_run_queue(struct request_queue *q)
407 struct scsi_device *sdev = q->queuedata;
408 struct Scsi_Host *shost = sdev->host;
411 if (sdev->single_lun)
412 scsi_single_lun_run(sdev);
414 spin_lock_irqsave(shost->host_lock, flags);
415 while (!list_empty(&shost->starved_list) &&
416 !shost->host_blocked && !shost->host_self_blocked &&
417 !((shost->can_queue > 0) &&
418 (shost->host_busy >= shost->can_queue))) {
420 * As long as shost is accepting commands and we have
421 * starved queues, call blk_run_queue. scsi_request_fn
422 * drops the queue_lock and can add us back to the
425 * host_lock protects the starved_list and starved_entry.
426 * scsi_request_fn must get the host_lock before checking
427 * or modifying starved_list or starved_entry.
429 sdev = list_entry(shost->starved_list.next,
430 struct scsi_device, starved_entry);
431 list_del_init(&sdev->starved_entry);
432 spin_unlock_irqrestore(shost->host_lock, flags);
434 blk_run_queue(sdev->request_queue);
436 spin_lock_irqsave(shost->host_lock, flags);
437 if (unlikely(!list_empty(&sdev->starved_entry)))
439 * sdev lost a race, and was put back on the
440 * starved list. This is unlikely but without this
441 * in theory we could loop forever.
445 spin_unlock_irqrestore(shost->host_lock, flags);
451 * Function: scsi_requeue_command()
453 * Purpose: Handle post-processing of completed commands.
455 * Arguments: q - queue to operate on
456 * cmd - command that may need to be requeued.
460 * Notes: After command completion, there may be blocks left
461 * over which weren't finished by the previous command
462 * this can be for a number of reasons - the main one is
463 * I/O errors in the middle of the request, in which case
464 * we need to request the blocks that come after the bad
467 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
469 cmd->request->flags &= ~REQ_DONTPREP;
470 blk_insert_request(q, cmd->request, 1, cmd, 1);
475 void scsi_next_command(struct scsi_cmnd *cmd)
477 struct request_queue *q = cmd->device->request_queue;
479 scsi_put_command(cmd);
483 void scsi_run_host_queues(struct Scsi_Host *shost)
485 struct scsi_device *sdev;
487 shost_for_each_device(sdev, shost)
488 scsi_run_queue(sdev->request_queue);
492 * Function: scsi_end_request()
494 * Purpose: Post-processing of completed commands called from interrupt
495 * handler or a bottom-half handler.
497 * Arguments: cmd - command that is complete.
498 * uptodate - 1 if I/O indicates success, 0 for I/O error.
499 * sectors - number of sectors we want to mark.
500 * requeue - indicates whether we should requeue leftovers.
501 * frequeue - indicates that if we release the command block
502 * that the queue request function should be called.
504 * Lock status: Assumed that lock is not held upon entry.
508 * Notes: This is called for block device requests in order to
509 * mark some number of sectors as complete.
511 * We are guaranteeing that the request queue will be goosed
512 * at some point during this call.
514 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
515 int bytes, int requeue)
517 request_queue_t *q = cmd->device->request_queue;
518 struct request *req = cmd->request;
522 * If there are blocks left over at the end, set up the command
523 * to queue the remainder of them.
525 if (end_that_request_chunk(req, uptodate, bytes)) {
526 int leftover = (req->hard_nr_sectors << 9);
528 if (blk_pc_request(req))
529 leftover = req->data_len;
531 /* kill remainder if no retrys */
532 if (!uptodate && blk_noretry_request(req))
533 end_that_request_chunk(req, 0, leftover);
537 * Bleah. Leftovers again. Stick the
538 * leftovers in the front of the
539 * queue, and goose the queue again.
541 scsi_requeue_command(q, cmd);
547 add_disk_randomness(req->rq_disk);
549 spin_lock_irqsave(q->queue_lock, flags);
550 if (blk_rq_tagged(req))
551 blk_queue_end_tag(q, req);
552 end_that_request_last(req);
553 spin_unlock_irqrestore(q->queue_lock, flags);
556 * This will goose the queue request function at the end, so we don't
557 * need to worry about launching another command.
559 scsi_next_command(cmd);
563 static struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, int gfp_mask)
565 struct scsi_host_sg_pool *sgp;
566 struct scatterlist *sgl;
568 BUG_ON(!cmd->use_sg);
570 switch (cmd->use_sg) {
580 #if (SCSI_MAX_PHYS_SEGMENTS > 32)
584 #if (SCSI_MAX_PHYS_SEGMENTS > 64)
588 #if (SCSI_MAX_PHYS_SEGMENTS > 128)
599 sgp = scsi_sg_pools + cmd->sglist_len;
600 sgl = mempool_alloc(sgp->pool, gfp_mask);
602 memset(sgl, 0, sgp->size);
606 static void scsi_free_sgtable(struct scatterlist *sgl, int index)
608 struct scsi_host_sg_pool *sgp;
610 BUG_ON(index > SG_MEMPOOL_NR);
612 sgp = scsi_sg_pools + index;
613 mempool_free(sgl, sgp->pool);
617 * Function: scsi_release_buffers()
619 * Purpose: Completion processing for block device I/O requests.
621 * Arguments: cmd - command that we are bailing.
623 * Lock status: Assumed that no lock is held upon entry.
627 * Notes: In the event that an upper level driver rejects a
628 * command, we must release resources allocated during
629 * the __init_io() function. Primarily this would involve
630 * the scatter-gather table, and potentially any bounce
633 static void scsi_release_buffers(struct scsi_cmnd *cmd)
635 struct request *req = cmd->request;
638 * Free up any indirection buffers we allocated for DMA purposes.
641 scsi_free_sgtable(cmd->request_buffer, cmd->sglist_len);
642 else if (cmd->request_buffer != req->buffer)
643 kfree(cmd->request_buffer);
646 * Zero these out. They now point to freed memory, and it is
647 * dangerous to hang onto the pointers.
651 cmd->request_buffer = NULL;
652 cmd->request_bufflen = 0;
656 * Function: scsi_io_completion()
658 * Purpose: Completion processing for block device I/O requests.
660 * Arguments: cmd - command that is finished.
662 * Lock status: Assumed that no lock is held upon entry.
666 * Notes: This function is matched in terms of capabilities to
667 * the function that created the scatter-gather list.
668 * In other words, if there are no bounce buffers
669 * (the normal case for most drivers), we don't need
670 * the logic to deal with cleaning up afterwards.
672 * We must do one of several things here:
674 * a) Call scsi_end_request. This will finish off the
675 * specified number of sectors. If we are done, the
676 * command block will be released, and the queue
677 * function will be goosed. If we are not done, then
678 * scsi_end_request will directly goose the queue.
680 * b) We can just use scsi_requeue_command() here. This would
681 * be used if we just wanted to retry, for example.
683 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes,
684 unsigned int block_bytes)
686 int result = cmd->result;
687 int this_count = cmd->bufflen;
688 request_queue_t *q = cmd->device->request_queue;
689 struct request *req = cmd->request;
690 int clear_errors = 1;
693 * Free up any indirection buffers we allocated for DMA purposes.
694 * For the case of a READ, we need to copy the data out of the
695 * bounce buffer and into the real buffer.
698 scsi_free_sgtable(cmd->buffer, cmd->sglist_len);
699 else if (cmd->buffer != req->buffer) {
700 if (rq_data_dir(req) == READ) {
702 char *to = bio_kmap_irq(req->bio, &flags);
703 memcpy(to, cmd->buffer, cmd->bufflen);
704 bio_kunmap_irq(to, &flags);
709 if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
710 req->errors = (driver_byte(result) & DRIVER_SENSE) ?
711 (CHECK_CONDITION << 1) : (result & 0xff);
714 if (cmd->sense_buffer[0] & 0x70) {
715 int len = 8 + cmd->sense_buffer[7];
717 if (len > SCSI_SENSE_BUFFERSIZE)
718 len = SCSI_SENSE_BUFFERSIZE;
719 memcpy(req->sense, cmd->sense_buffer, len);
720 req->sense_len = len;
723 req->data_len -= cmd->bufflen;
727 * Zero these out. They now point to freed memory, and it is
728 * dangerous to hang onto the pointers.
732 cmd->request_buffer = NULL;
733 cmd->request_bufflen = 0;
736 * Next deal with any sectors which we were able to correctly
739 if (good_bytes >= 0) {
740 SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, %d bytes done.\n",
741 req->nr_sectors, good_bytes));
742 SCSI_LOG_HLCOMPLETE(1, printk("use_sg is %d\n", cmd->use_sg));
747 * If multiple sectors are requested in one buffer, then
748 * they will have been finished off by the first command.
749 * If not, then we have a multi-buffer command.
751 * If block_bytes != 0, it means we had a medium error
752 * of some sort, and that we want to mark some number of
753 * sectors as not uptodate. Thus we want to inhibit
754 * requeueing right here - we will requeue down below
755 * when we handle the bad sectors.
757 cmd = scsi_end_request(cmd, 1, good_bytes, result == 0);
760 * If the command completed without error, then either finish off the
761 * rest of the command, or start a new one.
763 if (result == 0 || cmd == NULL ) {
768 * Now, if we were good little boys and girls, Santa left us a request
769 * sense buffer. We can extract information from this, so we
770 * can choose a block to remap, etc.
772 if (driver_byte(result) != 0) {
773 if ((cmd->sense_buffer[0] & 0x7f) == 0x70) {
775 * If the device is in the process of becoming ready,
778 if (cmd->sense_buffer[12] == 0x04 &&
779 cmd->sense_buffer[13] == 0x01) {
780 scsi_requeue_command(q, cmd);
783 if ((cmd->sense_buffer[2] & 0xf) == UNIT_ATTENTION) {
784 if (cmd->device->removable) {
785 /* detected disc change. set a bit
786 * and quietly refuse further access.
788 cmd->device->changed = 1;
789 cmd = scsi_end_request(cmd, 0,
794 * Must have been a power glitch, or a
795 * bus reset. Could not have been a
796 * media change, so we just retry the
797 * request and see what happens.
799 scsi_requeue_command(q, cmd);
805 * If we had an ILLEGAL REQUEST returned, then we may have
806 * performed an unsupported command. The only thing this
807 * should be would be a ten byte read where only a six byte
808 * read was supported. Also, on a system where READ CAPACITY
809 * failed, we may have read past the end of the disk.
812 switch (cmd->sense_buffer[2]) {
813 case ILLEGAL_REQUEST:
814 if (cmd->device->use_10_for_rw &&
815 (cmd->cmnd[0] == READ_10 ||
816 cmd->cmnd[0] == WRITE_10)) {
817 cmd->device->use_10_for_rw = 0;
819 * This will cause a retry with a 6-byte
822 scsi_requeue_command(q, cmd);
825 cmd = scsi_end_request(cmd, 0, this_count, 1);
830 printk(KERN_INFO "Device %s not ready.\n",
831 req->rq_disk ? req->rq_disk->disk_name : "");
832 cmd = scsi_end_request(cmd, 0, this_count, 1);
836 case VOLUME_OVERFLOW:
837 printk("scsi%d: ERROR on channel %d, id %d, lun %d, CDB: ",
838 cmd->device->host->host_no, (int) cmd->device->channel,
839 (int) cmd->device->id, (int) cmd->device->lun);
840 print_command(cmd->data_cmnd);
841 print_sense("", cmd);
842 cmd = scsi_end_request(cmd, 0, block_bytes, 1);
847 } /* driver byte != 0 */
848 if (host_byte(result) == DID_RESET) {
850 * Third party bus reset or reset for error
851 * recovery reasons. Just retry the request
852 * and see what happens.
854 scsi_requeue_command(q, cmd);
858 printk("SCSI error : <%d %d %d %d> return code = 0x%x\n",
859 cmd->device->host->host_no,
860 cmd->device->channel,
862 cmd->device->lun, result);
864 if (driver_byte(result) & DRIVER_SENSE)
865 print_sense("", cmd);
867 * Mark a single buffer as not uptodate. Queue the remainder.
868 * We sometimes get this cruft in the event that a medium error
869 * isn't properly reported.
871 block_bytes = req->hard_cur_sectors << 9;
873 block_bytes = req->data_len;
874 cmd = scsi_end_request(cmd, 0, block_bytes, 1);
879 * Function: scsi_init_io()
881 * Purpose: SCSI I/O initialize function.
883 * Arguments: cmd - Command descriptor we wish to initialize
885 * Returns: 0 on success
886 * BLKPREP_DEFER if the failure is retryable
887 * BLKPREP_KILL if the failure is fatal
889 static int scsi_init_io(struct scsi_cmnd *cmd)
891 struct request *req = cmd->request;
892 struct scatterlist *sgpnt;
896 * if this is a rq->data based REQ_BLOCK_PC, setup for a non-sg xfer
898 if ((req->flags & REQ_BLOCK_PC) && !req->bio) {
899 cmd->request_bufflen = req->data_len;
900 cmd->request_buffer = req->data;
901 req->buffer = req->data;
907 * we used to not use scatter-gather for single segment request,
908 * but now we do (it makes highmem I/O easier to support without
911 cmd->use_sg = req->nr_phys_segments;
914 * if sg table allocation fails, requeue request later.
916 sgpnt = scsi_alloc_sgtable(cmd, GFP_ATOMIC);
917 if (unlikely(!sgpnt)) {
918 req->flags |= REQ_SPECIAL;
919 return BLKPREP_DEFER;
922 cmd->request_buffer = (char *) sgpnt;
923 cmd->request_bufflen = req->nr_sectors << 9;
924 if (blk_pc_request(req))
925 cmd->request_bufflen = req->data_len;
929 * Next, walk the list, and fill in the addresses and sizes of
932 count = blk_rq_map_sg(req->q, req, cmd->request_buffer);
935 * mapped well, send it off
937 if (likely(count <= cmd->use_sg)) {
942 printk(KERN_ERR "Incorrect number of segments after building list\n");
943 printk(KERN_ERR "counted %d, received %d\n", count, cmd->use_sg);
944 printk(KERN_ERR "req nr_sec %lu, cur_nr_sec %u\n", req->nr_sectors,
945 req->current_nr_sectors);
947 /* release the command and kill it */
948 scsi_release_buffers(cmd);
949 scsi_put_command(cmd);
953 static int scsi_prep_fn(struct request_queue *q, struct request *req)
955 struct scsi_device *sdev = q->queuedata;
956 struct scsi_cmnd *cmd;
957 int specials_only = 0;
960 * Just check to see if the device is online. If it isn't, we
961 * refuse to process any commands. The device must be brought
962 * online before trying any recovery commands
964 if (unlikely(!scsi_device_online(sdev))) {
965 printk(KERN_ERR "scsi%d (%d:%d): rejecting I/O to offline device\n",
966 sdev->host->host_no, sdev->id, sdev->lun);
969 if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
970 /* OK, we're not in a running state don't prep
972 if (sdev->sdev_state == SDEV_DEL) {
973 /* Device is fully deleted, no commands
974 * at all allowed down */
975 printk(KERN_ERR "scsi%d (%d:%d): rejecting I/O to dead device\n",
976 sdev->host->host_no, sdev->id, sdev->lun);
979 /* OK, we only allow special commands (i.e. not
980 * user initiated ones */
981 specials_only = sdev->sdev_state;
985 * Find the actual device driver associated with this command.
986 * The SPECIAL requests are things like character device or
987 * ioctls, which did not originate from ll_rw_blk. Note that
988 * the special field is also used to indicate the cmd for
989 * the remainder of a partially fulfilled request that can
990 * come up when there is a medium error. We have to treat
991 * these two cases differently. We differentiate by looking
992 * at request->cmd, as this tells us the real story.
994 if (req->flags & REQ_SPECIAL) {
995 struct scsi_request *sreq = req->special;
997 if (sreq->sr_magic == SCSI_REQ_MAGIC) {
998 cmd = scsi_get_command(sreq->sr_device, GFP_ATOMIC);
1001 scsi_init_cmd_from_req(cmd, sreq);
1004 } else if (req->flags & (REQ_CMD | REQ_BLOCK_PC)) {
1006 if(unlikely(specials_only)) {
1007 if(specials_only == SDEV_QUIESCE)
1008 return BLKPREP_DEFER;
1010 printk(KERN_ERR "scsi%d (%d:%d): rejecting I/O to device being removed\n",
1011 sdev->host->host_no, sdev->id, sdev->lun);
1012 return BLKPREP_KILL;
1017 * Now try and find a command block that we can use.
1019 if (!req->special) {
1020 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1026 /* pull a tag out of the request if we have one */
1027 cmd->tag = req->tag;
1029 blk_dump_rq_flags(req, "SCSI bad req");
1030 return BLKPREP_KILL;
1033 /* note the overloading of req->special. When the tag
1034 * is active it always means cmd. If the tag goes
1035 * back for re-queueing, it may be reset */
1040 * FIXME: drop the lock here because the functions below
1041 * expect to be called without the queue lock held. Also,
1042 * previously, we dequeued the request before dropping the
1043 * lock. We hope REQ_STARTED prevents anything untoward from
1046 if (req->flags & (REQ_CMD | REQ_BLOCK_PC)) {
1047 struct scsi_driver *drv;
1051 * This will do a couple of things:
1052 * 1) Fill in the actual SCSI command.
1053 * 2) Fill in any other upper-level specific fields
1056 * If this returns 0, it means that the request failed
1057 * (reading past end of disk, reading offline device,
1058 * etc). This won't actually talk to the device, but
1059 * some kinds of consistency checking may cause the
1060 * request to be rejected immediately.
1064 * This sets up the scatter-gather table (allocating if
1067 ret = scsi_init_io(cmd);
1068 if (ret) /* BLKPREP_KILL return also releases the command */
1072 * Initialize the actual SCSI command for this request.
1074 drv = *(struct scsi_driver **)req->rq_disk->private_data;
1075 if (unlikely(!drv->init_command(cmd))) {
1076 scsi_release_buffers(cmd);
1077 scsi_put_command(cmd);
1078 return BLKPREP_KILL;
1083 * The request is now prepped, no need to come back here
1085 req->flags |= REQ_DONTPREP;
1089 /* If we defer, the elv_next_request() returns NULL, but the
1090 * queue must be restarted, so we plug here if no returning
1091 * command will automatically do that. */
1092 if (sdev->device_busy == 0)
1094 return BLKPREP_DEFER;
1098 * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1101 * Called with the queue_lock held.
1103 static inline int scsi_dev_queue_ready(struct request_queue *q,
1104 struct scsi_device *sdev)
1106 if (sdev->device_busy >= sdev->queue_depth)
1108 if (sdev->device_busy == 0 && sdev->device_blocked) {
1110 * unblock after device_blocked iterates to zero
1112 if (--sdev->device_blocked == 0) {
1114 printk("scsi%d (%d:%d) unblocking device at"
1115 " zero depth\n", sdev->host->host_no,
1116 sdev->id, sdev->lun));
1122 if (sdev->device_blocked)
1129 * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1130 * return 0. We must end up running the queue again whenever 0 is
1131 * returned, else IO can hang.
1133 * Called with host_lock held.
1135 static inline int scsi_host_queue_ready(struct request_queue *q,
1136 struct Scsi_Host *shost,
1137 struct scsi_device *sdev)
1139 if (test_bit(SHOST_RECOVERY, &shost->shost_state))
1141 if (shost->host_busy == 0 && shost->host_blocked) {
1143 * unblock after host_blocked iterates to zero
1145 if (--shost->host_blocked == 0) {
1147 printk("scsi%d unblocking host at zero depth\n",
1154 if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1155 shost->host_blocked || shost->host_self_blocked) {
1156 if (list_empty(&sdev->starved_entry))
1157 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1161 /* We're OK to process the command, so we can't be starved */
1162 if (!list_empty(&sdev->starved_entry))
1163 list_del_init(&sdev->starved_entry);
1169 * Function: scsi_request_fn()
1171 * Purpose: Main strategy routine for SCSI.
1173 * Arguments: q - Pointer to actual queue.
1177 * Lock status: IO request lock assumed to be held when called.
1179 static void scsi_request_fn(struct request_queue *q)
1181 struct scsi_device *sdev = q->queuedata;
1182 struct Scsi_Host *shost = sdev->host;
1183 struct scsi_cmnd *cmd;
1184 struct request *req;
1186 if(!get_device(&sdev->sdev_gendev))
1187 /* We must be tearing the block queue down already */
1191 * To start with, we keep looping until the queue is empty, or until
1192 * the host is no longer able to accept any more requests.
1194 while (!blk_queue_plugged(q)) {
1197 * get next queueable request. We do this early to make sure
1198 * that the request is fully prepared even if we cannot
1201 req = elv_next_request(q);
1202 if (!req || !scsi_dev_queue_ready(q, sdev))
1205 if (unlikely(!scsi_device_online(sdev))) {
1206 printk(KERN_ERR "scsi%d (%d:%d): rejecting I/O to offline device\n",
1207 sdev->host->host_no, sdev->id, sdev->lun);
1208 blkdev_dequeue_request(req);
1209 req->flags |= REQ_QUIET;
1210 while (end_that_request_first(req, 0, req->nr_sectors))
1212 end_that_request_last(req);
1218 * Remove the request from the request list.
1220 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1221 blkdev_dequeue_request(req);
1222 sdev->device_busy++;
1224 spin_unlock(q->queue_lock);
1225 spin_lock(shost->host_lock);
1227 if (!scsi_host_queue_ready(q, shost, sdev))
1229 if (sdev->single_lun) {
1230 if (sdev->sdev_target->starget_sdev_user &&
1231 sdev->sdev_target->starget_sdev_user != sdev)
1233 sdev->sdev_target->starget_sdev_user = sdev;
1238 * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1239 * take the lock again.
1241 spin_unlock_irq(shost->host_lock);
1244 if (unlikely(cmd == NULL)) {
1245 printk(KERN_CRIT "impossible request in %s.\n"
1246 "please mail a stack trace to "
1247 "linux-scsi@vger.kernel.org",
1253 * Finally, initialize any error handling parameters, and set up
1254 * the timers for timeouts.
1256 scsi_init_cmd_errh(cmd);
1259 * Dispatch the command to the low-level driver.
1261 rtn = scsi_dispatch_cmd(cmd);
1262 spin_lock_irq(q->queue_lock);
1264 /* we're refusing the command; because of
1265 * the way locks get dropped, we need to
1266 * check here if plugging is required */
1267 if(sdev->device_busy == 0)
1277 spin_unlock_irq(shost->host_lock);
1280 * lock q, handle tag, requeue req, and decrement device_busy. We
1281 * must return with queue_lock held.
1283 * Decrementing device_busy without checking it is OK, as all such
1284 * cases (host limits or settings) should run the queue at some
1287 spin_lock_irq(q->queue_lock);
1288 blk_requeue_request(q, req);
1289 sdev->device_busy--;
1290 if(sdev->device_busy == 0)
1293 /* must be careful here...if we trigger the ->remove() function
1294 * we cannot be holding the q lock */
1295 spin_unlock_irq(q->queue_lock);
1296 put_device(&sdev->sdev_gendev);
1297 spin_lock_irq(q->queue_lock);
1300 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1302 struct device *host_dev;
1304 if (shost->unchecked_isa_dma)
1305 return BLK_BOUNCE_ISA;
1307 host_dev = scsi_get_device(shost);
1308 if (PCI_DMA_BUS_IS_PHYS && host_dev && host_dev->dma_mask)
1309 return *host_dev->dma_mask;
1312 * Platforms with virtual-DMA translation
1313 * hardware have no practical limit.
1315 return BLK_BOUNCE_ANY;
1318 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1320 struct Scsi_Host *shost = sdev->host;
1321 struct request_queue *q;
1323 q = blk_init_queue(scsi_request_fn, &sdev->sdev_lock);
1327 blk_queue_prep_rq(q, scsi_prep_fn);
1329 blk_queue_max_hw_segments(q, shost->sg_tablesize);
1330 blk_queue_max_phys_segments(q, SCSI_MAX_PHYS_SEGMENTS);
1331 blk_queue_max_sectors(q, shost->max_sectors);
1332 blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1333 blk_queue_segment_boundary(q, shost->dma_boundary);
1335 if (!shost->use_clustering)
1336 clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
1340 void scsi_free_queue(struct request_queue *q)
1342 blk_cleanup_queue(q);
1346 * Function: scsi_block_requests()
1348 * Purpose: Utility function used by low-level drivers to prevent further
1349 * commands from being queued to the device.
1351 * Arguments: shost - Host in question
1355 * Lock status: No locks are assumed held.
1357 * Notes: There is no timer nor any other means by which the requests
1358 * get unblocked other than the low-level driver calling
1359 * scsi_unblock_requests().
1361 void scsi_block_requests(struct Scsi_Host *shost)
1363 shost->host_self_blocked = 1;
1367 * Function: scsi_unblock_requests()
1369 * Purpose: Utility function used by low-level drivers to allow further
1370 * commands from being queued to the device.
1372 * Arguments: shost - Host in question
1376 * Lock status: No locks are assumed held.
1378 * Notes: There is no timer nor any other means by which the requests
1379 * get unblocked other than the low-level driver calling
1380 * scsi_unblock_requests().
1382 * This is done as an API function so that changes to the
1383 * internals of the scsi mid-layer won't require wholesale
1384 * changes to drivers that use this feature.
1386 void scsi_unblock_requests(struct Scsi_Host *shost)
1388 shost->host_self_blocked = 0;
1389 scsi_run_host_queues(shost);
1392 int __init scsi_init_queue(void)
1396 for (i = 0; i < SG_MEMPOOL_NR; i++) {
1397 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1398 int size = sgp->size * sizeof(struct scatterlist);
1400 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1401 SLAB_HWCACHE_ALIGN, NULL, NULL);
1403 printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1407 sgp->pool = mempool_create(SG_MEMPOOL_SIZE,
1408 mempool_alloc_slab, mempool_free_slab,
1411 printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1419 void scsi_exit_queue(void)
1423 for (i = 0; i < SG_MEMPOOL_NR; i++) {
1424 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1425 mempool_destroy(sgp->pool);
1426 kmem_cache_destroy(sgp->slab);
1430 * __scsi_mode_sense - issue a mode sense, falling back from 10 to
1431 * six bytes if necessary.
1432 * @sreq: SCSI request to fill in with the MODE_SENSE
1433 * @dbd: set if mode sense will allow block descriptors to be returned
1434 * @modepage: mode page being requested
1435 * @buffer: request buffer (may not be smaller than eight bytes)
1436 * @len: length of request buffer.
1437 * @timeout: command timeout
1438 * @retries: number of retries before failing
1439 * @data: returns a structure abstracting the mode header data
1441 * Returns zero if unsuccessful, or the header offset (either 4
1442 * or 8 depending on whether a six or ten byte command was
1443 * issued) if successful.
1446 __scsi_mode_sense(struct scsi_request *sreq, int dbd, int modepage,
1447 unsigned char *buffer, int len, int timeout, int retries,
1448 struct scsi_mode_data *data) {
1449 unsigned char cmd[12];
1453 memset(data, 0, sizeof(*data));
1454 memset(&cmd[0], 0, 12);
1455 cmd[1] = dbd & 0x18; /* allows DBD and LLBA bits */
1459 use_10_for_ms = sreq->sr_device->use_10_for_ms;
1461 if (use_10_for_ms) {
1465 cmd[0] = MODE_SENSE_10;
1472 cmd[0] = MODE_SENSE;
1477 sreq->sr_cmd_len = 0;
1478 sreq->sr_sense_buffer[0] = 0;
1479 sreq->sr_sense_buffer[2] = 0;
1480 sreq->sr_data_direction = DMA_FROM_DEVICE;
1482 memset(buffer, 0, len);
1484 scsi_wait_req(sreq, cmd, buffer, len, timeout, retries);
1486 /* This code looks awful: what it's doing is making sure an
1487 * ILLEGAL REQUEST sense return identifies the actual command
1488 * byte as the problem. MODE_SENSE commands can return
1489 * ILLEGAL REQUEST if the code page isn't supported */
1490 if (use_10_for_ms && ! scsi_status_is_good(sreq->sr_result) &&
1491 (driver_byte(sreq->sr_result) & DRIVER_SENSE) &&
1492 sreq->sr_sense_buffer[2] == ILLEGAL_REQUEST &&
1493 (sreq->sr_sense_buffer[4] & 0x40) == 0x40 &&
1494 sreq->sr_sense_buffer[5] == 0 &&
1495 sreq->sr_sense_buffer[6] == 0 ) {
1496 sreq->sr_device->use_10_for_ms = 0;
1500 if(scsi_status_is_good(sreq->sr_result)) {
1501 data->header_length = header_length;
1503 data->length = buffer[0]*256 + buffer[1] + 2;
1504 data->medium_type = buffer[2];
1505 data->device_specific = buffer[3];
1506 data->longlba = buffer[4] & 0x01;
1507 data->block_descriptor_length = buffer[6]*256
1510 data->length = buffer[0] + 1;
1511 data->medium_type = buffer[1];
1512 data->device_specific = buffer[2];
1513 data->block_descriptor_length = buffer[3];
1517 return sreq->sr_result;
1521 * scsi_mode_sense - issue a mode sense, falling back from 10 to
1522 * six bytes if necessary.
1523 * @sdev: scsi device to send command to.
1524 * @dbd: set if mode sense will disable block descriptors in the return
1525 * @modepage: mode page being requested
1526 * @buffer: request buffer (may not be smaller than eight bytes)
1527 * @len: length of request buffer.
1528 * @timeout: command timeout
1529 * @retries: number of retries before failing
1531 * Returns zero if unsuccessful, or the header offset (either 4
1532 * or 8 depending on whether a six or ten byte command was
1533 * issued) if successful.
1536 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1537 unsigned char *buffer, int len, int timeout, int retries,
1538 struct scsi_mode_data *data)
1540 struct scsi_request *sreq = scsi_allocate_request(sdev, GFP_KERNEL);
1546 ret = __scsi_mode_sense(sreq, dbd, modepage, buffer, len,
1547 timeout, retries, data);
1549 scsi_release_request(sreq);
1555 * scsi_device_set_state - Take the given device through the device
1557 * @sdev: scsi device to change the state of.
1558 * @state: state to change to.
1560 * Returns zero if unsuccessful or an error if the requested
1561 * transition is illegal.
1564 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
1566 enum scsi_device_state oldstate = sdev->sdev_state;
1568 if (state == oldstate)
1573 /* There are no legal states that come back to
1574 * created. This is the manually initialised start
1631 sdev->sdev_state = state;
1635 dev_printk(KERN_ERR, &sdev->sdev_gendev,
1636 "Illegal state transition %s->%s\n",
1637 scsi_device_state_name(oldstate),
1638 scsi_device_state_name(state));
1642 EXPORT_SYMBOL(scsi_device_set_state);
1645 * scsi_device_quiesce - Block user issued commands.
1646 * @sdev: scsi device to quiesce.
1648 * This works by trying to transition to the SDEV_QUIESCE state
1649 * (which must be a legal transition). When the device is in this
1650 * state, only special requests will be accepted, all others will
1651 * be deferred. Since special requests may also be requeued requests,
1652 * a successful return doesn't guarantee the device will be
1653 * totally quiescent.
1655 * Must be called with user context, may sleep.
1657 * Returns zero if unsuccessful or an error if not.
1660 scsi_device_quiesce(struct scsi_device *sdev)
1662 int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
1666 scsi_run_queue(sdev->request_queue);
1667 while (sdev->device_busy) {
1668 schedule_timeout(HZ/5);
1669 scsi_run_queue(sdev->request_queue);
1673 EXPORT_SYMBOL(scsi_device_quiesce);
1676 * scsi_device_resume - Restart user issued commands to a quiesced device.
1677 * @sdev: scsi device to resume.
1679 * Moves the device from quiesced back to running and restarts the
1682 * Must be called with user context, may sleep.
1685 scsi_device_resume(struct scsi_device *sdev)
1687 if(scsi_device_set_state(sdev, SDEV_RUNNING))
1689 scsi_run_queue(sdev->request_queue);
1691 EXPORT_SYMBOL(scsi_device_resume);