2 * Copyright (c) 2016 Avago Technologies. All rights reserved.
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of version 2 of the GNU General Public License as
6 * published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful.
9 * ALL EXPRESS OR IMPLIED CONDITIONS, REPRESENTATIONS AND WARRANTIES,
10 * INCLUDING ANY IMPLIED WARRANTY OF MERCHANTABILITY, FITNESS FOR A
11 * PARTICULAR PURPOSE, OR NON-INFRINGEMENT, ARE DISCLAIMED, EXCEPT TO
12 * THE EXTENT THAT SUCH DISCLAIMERS ARE HELD TO BE LEGALLY INVALID.
13 * See the GNU General Public License for more details, a copy of which
14 * can be found in the file COPYING included with this package
17 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
18 #include <linux/module.h>
19 #include <linux/slab.h>
20 #include <linux/blk-mq.h>
21 #include <linux/parser.h>
22 #include <linux/random.h>
23 #include <uapi/scsi/fc/fc_fs.h>
24 #include <uapi/scsi/fc/fc_els.h>
27 #include <linux/nvme-fc-driver.h>
28 #include <linux/nvme-fc.h>
31 /* *************************** Data Structures/Defines ****************** */
34 #define NVMET_LS_CTX_COUNT 4
36 /* for this implementation, assume small single frame rqst/rsp */
37 #define NVME_FC_MAX_LS_BUFFER_SIZE 2048
39 struct nvmet_fc_tgtport
;
40 struct nvmet_fc_tgt_assoc
;
42 struct nvmet_fc_ls_iod
{
43 struct nvmefc_tgt_ls_req
*lsreq
;
44 struct nvmefc_tgt_fcp_req
*fcpreq
; /* only if RS */
46 struct list_head ls_list
; /* tgtport->ls_list */
48 struct nvmet_fc_tgtport
*tgtport
;
49 struct nvmet_fc_tgt_assoc
*assoc
;
56 struct scatterlist sg
[2];
58 struct work_struct work
;
59 } __aligned(sizeof(unsigned long long));
61 #define NVMET_FC_MAX_SEQ_LENGTH (256 * 1024)
62 #define NVMET_FC_MAX_XFR_SGENTS (NVMET_FC_MAX_SEQ_LENGTH / PAGE_SIZE)
64 enum nvmet_fcp_datadir
{
71 struct nvmet_fc_fcp_iod
{
72 struct nvmefc_tgt_fcp_req
*fcpreq
;
74 struct nvme_fc_cmd_iu cmdiubuf
;
75 struct nvme_fc_ersp_iu rspiubuf
;
77 struct scatterlist
*data_sg
;
80 enum nvmet_fcp_datadir io_dir
;
88 struct work_struct work
;
89 struct work_struct done_work
;
91 struct nvmet_fc_tgtport
*tgtport
;
92 struct nvmet_fc_tgt_queue
*queue
;
94 struct list_head fcp_list
; /* tgtport->fcp_list */
97 struct nvmet_fc_tgtport
{
99 struct nvmet_fc_target_port fc_target_port
;
101 struct list_head tgt_list
; /* nvmet_fc_target_list */
102 struct device
*dev
; /* dev for dma mapping */
103 struct nvmet_fc_target_template
*ops
;
105 struct nvmet_fc_ls_iod
*iod
;
107 struct list_head ls_list
;
108 struct list_head ls_busylist
;
109 struct list_head assoc_list
;
110 struct ida assoc_cnt
;
111 struct nvmet_port
*port
;
116 struct nvmet_fc_defer_fcp_req
{
117 struct list_head req_list
;
118 struct nvmefc_tgt_fcp_req
*fcp_req
;
121 struct nvmet_fc_tgt_queue
{
133 struct nvmet_port
*port
;
134 struct nvmet_cq nvme_cq
;
135 struct nvmet_sq nvme_sq
;
136 struct nvmet_fc_tgt_assoc
*assoc
;
137 struct nvmet_fc_fcp_iod
*fod
; /* array of fcp_iods */
138 struct list_head fod_list
;
139 struct list_head pending_cmd_list
;
140 struct list_head avail_defer_list
;
141 struct workqueue_struct
*work_q
;
143 } __aligned(sizeof(unsigned long long));
145 struct nvmet_fc_tgt_assoc
{
148 struct nvmet_fc_tgtport
*tgtport
;
149 struct list_head a_list
;
150 struct nvmet_fc_tgt_queue
*queues
[NVMET_NR_QUEUES
+ 1];
152 struct work_struct del_work
;
157 nvmet_fc_iodnum(struct nvmet_fc_ls_iod
*iodptr
)
159 return (iodptr
- iodptr
->tgtport
->iod
);
163 nvmet_fc_fodnum(struct nvmet_fc_fcp_iod
*fodptr
)
165 return (fodptr
- fodptr
->queue
->fod
);
170 * Association and Connection IDs:
172 * Association ID will have random number in upper 6 bytes and zero
175 * Connection IDs will be Association ID with QID or'd in lower 2 bytes
177 * note: Association ID = Connection ID for queue 0
179 #define BYTES_FOR_QID sizeof(u16)
180 #define BYTES_FOR_QID_SHIFT (BYTES_FOR_QID * 8)
181 #define NVMET_FC_QUEUEID_MASK ((u64)((1 << BYTES_FOR_QID_SHIFT) - 1))
184 nvmet_fc_makeconnid(struct nvmet_fc_tgt_assoc
*assoc
, u16 qid
)
186 return (assoc
->association_id
| qid
);
190 nvmet_fc_getassociationid(u64 connectionid
)
192 return connectionid
& ~NVMET_FC_QUEUEID_MASK
;
196 nvmet_fc_getqueueid(u64 connectionid
)
198 return (u16
)(connectionid
& NVMET_FC_QUEUEID_MASK
);
201 static inline struct nvmet_fc_tgtport
*
202 targetport_to_tgtport(struct nvmet_fc_target_port
*targetport
)
204 return container_of(targetport
, struct nvmet_fc_tgtport
,
208 static inline struct nvmet_fc_fcp_iod
*
209 nvmet_req_to_fod(struct nvmet_req
*nvme_req
)
211 return container_of(nvme_req
, struct nvmet_fc_fcp_iod
, req
);
215 /* *************************** Globals **************************** */
218 static DEFINE_SPINLOCK(nvmet_fc_tgtlock
);
220 static LIST_HEAD(nvmet_fc_target_list
);
221 static DEFINE_IDA(nvmet_fc_tgtport_cnt
);
224 static void nvmet_fc_handle_ls_rqst_work(struct work_struct
*work
);
225 static void nvmet_fc_handle_fcp_rqst_work(struct work_struct
*work
);
226 static void nvmet_fc_fcp_rqst_op_done_work(struct work_struct
*work
);
227 static void nvmet_fc_tgt_a_put(struct nvmet_fc_tgt_assoc
*assoc
);
228 static int nvmet_fc_tgt_a_get(struct nvmet_fc_tgt_assoc
*assoc
);
229 static void nvmet_fc_tgt_q_put(struct nvmet_fc_tgt_queue
*queue
);
230 static int nvmet_fc_tgt_q_get(struct nvmet_fc_tgt_queue
*queue
);
231 static void nvmet_fc_tgtport_put(struct nvmet_fc_tgtport
*tgtport
);
232 static int nvmet_fc_tgtport_get(struct nvmet_fc_tgtport
*tgtport
);
233 static void nvmet_fc_handle_fcp_rqst(struct nvmet_fc_tgtport
*tgtport
,
234 struct nvmet_fc_fcp_iod
*fod
);
235 static void nvmet_fc_delete_target_assoc(struct nvmet_fc_tgt_assoc
*assoc
);
238 /* *********************** FC-NVME DMA Handling **************************** */
241 * The fcloop device passes in a NULL device pointer. Real LLD's will
242 * pass in a valid device pointer. If NULL is passed to the dma mapping
243 * routines, depending on the platform, it may or may not succeed, and
247 * Wrapper all the dma routines and check the dev pointer.
249 * If simple mappings (return just a dma address, we'll noop them,
250 * returning a dma address of 0.
252 * On more complex mappings (dma_map_sg), a pseudo routine fills
253 * in the scatter list, setting all dma addresses to 0.
256 static inline dma_addr_t
257 fc_dma_map_single(struct device
*dev
, void *ptr
, size_t size
,
258 enum dma_data_direction dir
)
260 return dev
? dma_map_single(dev
, ptr
, size
, dir
) : (dma_addr_t
)0L;
264 fc_dma_mapping_error(struct device
*dev
, dma_addr_t dma_addr
)
266 return dev
? dma_mapping_error(dev
, dma_addr
) : 0;
270 fc_dma_unmap_single(struct device
*dev
, dma_addr_t addr
, size_t size
,
271 enum dma_data_direction dir
)
274 dma_unmap_single(dev
, addr
, size
, dir
);
278 fc_dma_sync_single_for_cpu(struct device
*dev
, dma_addr_t addr
, size_t size
,
279 enum dma_data_direction dir
)
282 dma_sync_single_for_cpu(dev
, addr
, size
, dir
);
286 fc_dma_sync_single_for_device(struct device
*dev
, dma_addr_t addr
, size_t size
,
287 enum dma_data_direction dir
)
290 dma_sync_single_for_device(dev
, addr
, size
, dir
);
293 /* pseudo dma_map_sg call */
295 fc_map_sg(struct scatterlist
*sg
, int nents
)
297 struct scatterlist
*s
;
300 WARN_ON(nents
== 0 || sg
[0].length
== 0);
302 for_each_sg(sg
, s
, nents
, i
) {
304 #ifdef CONFIG_NEED_SG_DMA_LENGTH
305 s
->dma_length
= s
->length
;
312 fc_dma_map_sg(struct device
*dev
, struct scatterlist
*sg
, int nents
,
313 enum dma_data_direction dir
)
315 return dev
? dma_map_sg(dev
, sg
, nents
, dir
) : fc_map_sg(sg
, nents
);
319 fc_dma_unmap_sg(struct device
*dev
, struct scatterlist
*sg
, int nents
,
320 enum dma_data_direction dir
)
323 dma_unmap_sg(dev
, sg
, nents
, dir
);
327 /* *********************** FC-NVME Port Management ************************ */
331 nvmet_fc_alloc_ls_iodlist(struct nvmet_fc_tgtport
*tgtport
)
333 struct nvmet_fc_ls_iod
*iod
;
336 iod
= kcalloc(NVMET_LS_CTX_COUNT
, sizeof(struct nvmet_fc_ls_iod
),
343 for (i
= 0; i
< NVMET_LS_CTX_COUNT
; iod
++, i
++) {
344 INIT_WORK(&iod
->work
, nvmet_fc_handle_ls_rqst_work
);
345 iod
->tgtport
= tgtport
;
346 list_add_tail(&iod
->ls_list
, &tgtport
->ls_list
);
348 iod
->rqstbuf
= kcalloc(2, NVME_FC_MAX_LS_BUFFER_SIZE
,
353 iod
->rspbuf
= iod
->rqstbuf
+ NVME_FC_MAX_LS_BUFFER_SIZE
;
355 iod
->rspdma
= fc_dma_map_single(tgtport
->dev
, iod
->rspbuf
,
356 NVME_FC_MAX_LS_BUFFER_SIZE
,
358 if (fc_dma_mapping_error(tgtport
->dev
, iod
->rspdma
))
366 list_del(&iod
->ls_list
);
367 for (iod
--, i
--; i
>= 0; iod
--, i
--) {
368 fc_dma_unmap_single(tgtport
->dev
, iod
->rspdma
,
369 NVME_FC_MAX_LS_BUFFER_SIZE
, DMA_TO_DEVICE
);
371 list_del(&iod
->ls_list
);
380 nvmet_fc_free_ls_iodlist(struct nvmet_fc_tgtport
*tgtport
)
382 struct nvmet_fc_ls_iod
*iod
= tgtport
->iod
;
385 for (i
= 0; i
< NVMET_LS_CTX_COUNT
; iod
++, i
++) {
386 fc_dma_unmap_single(tgtport
->dev
,
387 iod
->rspdma
, NVME_FC_MAX_LS_BUFFER_SIZE
,
390 list_del(&iod
->ls_list
);
395 static struct nvmet_fc_ls_iod
*
396 nvmet_fc_alloc_ls_iod(struct nvmet_fc_tgtport
*tgtport
)
398 struct nvmet_fc_ls_iod
*iod
;
401 spin_lock_irqsave(&tgtport
->lock
, flags
);
402 iod
= list_first_entry_or_null(&tgtport
->ls_list
,
403 struct nvmet_fc_ls_iod
, ls_list
);
405 list_move_tail(&iod
->ls_list
, &tgtport
->ls_busylist
);
406 spin_unlock_irqrestore(&tgtport
->lock
, flags
);
412 nvmet_fc_free_ls_iod(struct nvmet_fc_tgtport
*tgtport
,
413 struct nvmet_fc_ls_iod
*iod
)
417 spin_lock_irqsave(&tgtport
->lock
, flags
);
418 list_move(&iod
->ls_list
, &tgtport
->ls_list
);
419 spin_unlock_irqrestore(&tgtport
->lock
, flags
);
423 nvmet_fc_prep_fcp_iodlist(struct nvmet_fc_tgtport
*tgtport
,
424 struct nvmet_fc_tgt_queue
*queue
)
426 struct nvmet_fc_fcp_iod
*fod
= queue
->fod
;
429 for (i
= 0; i
< queue
->sqsize
; fod
++, i
++) {
430 INIT_WORK(&fod
->work
, nvmet_fc_handle_fcp_rqst_work
);
431 INIT_WORK(&fod
->done_work
, nvmet_fc_fcp_rqst_op_done_work
);
432 fod
->tgtport
= tgtport
;
436 fod
->aborted
= false;
438 list_add_tail(&fod
->fcp_list
, &queue
->fod_list
);
439 spin_lock_init(&fod
->flock
);
441 fod
->rspdma
= fc_dma_map_single(tgtport
->dev
, &fod
->rspiubuf
,
442 sizeof(fod
->rspiubuf
), DMA_TO_DEVICE
);
443 if (fc_dma_mapping_error(tgtport
->dev
, fod
->rspdma
)) {
444 list_del(&fod
->fcp_list
);
445 for (fod
--, i
--; i
>= 0; fod
--, i
--) {
446 fc_dma_unmap_single(tgtport
->dev
, fod
->rspdma
,
447 sizeof(fod
->rspiubuf
),
450 list_del(&fod
->fcp_list
);
459 nvmet_fc_destroy_fcp_iodlist(struct nvmet_fc_tgtport
*tgtport
,
460 struct nvmet_fc_tgt_queue
*queue
)
462 struct nvmet_fc_fcp_iod
*fod
= queue
->fod
;
465 for (i
= 0; i
< queue
->sqsize
; fod
++, i
++) {
467 fc_dma_unmap_single(tgtport
->dev
, fod
->rspdma
,
468 sizeof(fod
->rspiubuf
), DMA_TO_DEVICE
);
472 static struct nvmet_fc_fcp_iod
*
473 nvmet_fc_alloc_fcp_iod(struct nvmet_fc_tgt_queue
*queue
)
475 struct nvmet_fc_fcp_iod
*fod
;
477 lockdep_assert_held(&queue
->qlock
);
479 fod
= list_first_entry_or_null(&queue
->fod_list
,
480 struct nvmet_fc_fcp_iod
, fcp_list
);
482 list_del(&fod
->fcp_list
);
485 * no queue reference is taken, as it was taken by the
486 * queue lookup just prior to the allocation. The iod
487 * will "inherit" that reference.
495 nvmet_fc_queue_fcp_req(struct nvmet_fc_tgtport
*tgtport
,
496 struct nvmet_fc_tgt_queue
*queue
,
497 struct nvmefc_tgt_fcp_req
*fcpreq
)
499 struct nvmet_fc_fcp_iod
*fod
= fcpreq
->nvmet_fc_private
;
502 * put all admin cmds on hw queue id 0. All io commands go to
503 * the respective hw queue based on a modulo basis
505 fcpreq
->hwqid
= queue
->qid
?
506 ((queue
->qid
- 1) % tgtport
->ops
->max_hw_queues
) : 0;
508 if (tgtport
->ops
->target_features
& NVMET_FCTGTFEAT_CMD_IN_ISR
)
509 queue_work_on(queue
->cpu
, queue
->work_q
, &fod
->work
);
511 nvmet_fc_handle_fcp_rqst(tgtport
, fod
);
515 nvmet_fc_free_fcp_iod(struct nvmet_fc_tgt_queue
*queue
,
516 struct nvmet_fc_fcp_iod
*fod
)
518 struct nvmefc_tgt_fcp_req
*fcpreq
= fod
->fcpreq
;
519 struct nvmet_fc_tgtport
*tgtport
= fod
->tgtport
;
520 struct nvmet_fc_defer_fcp_req
*deferfcp
;
523 fc_dma_sync_single_for_cpu(tgtport
->dev
, fod
->rspdma
,
524 sizeof(fod
->rspiubuf
), DMA_TO_DEVICE
);
526 fcpreq
->nvmet_fc_private
= NULL
;
530 fod
->aborted
= false;
531 fod
->writedataactive
= false;
534 tgtport
->ops
->fcp_req_release(&tgtport
->fc_target_port
, fcpreq
);
536 spin_lock_irqsave(&queue
->qlock
, flags
);
537 deferfcp
= list_first_entry_or_null(&queue
->pending_cmd_list
,
538 struct nvmet_fc_defer_fcp_req
, req_list
);
540 list_add_tail(&fod
->fcp_list
, &fod
->queue
->fod_list
);
541 spin_unlock_irqrestore(&queue
->qlock
, flags
);
543 /* Release reference taken at queue lookup and fod allocation */
544 nvmet_fc_tgt_q_put(queue
);
548 /* Re-use the fod for the next pending cmd that was deferred */
549 list_del(&deferfcp
->req_list
);
551 fcpreq
= deferfcp
->fcp_req
;
553 /* deferfcp can be reused for another IO at a later date */
554 list_add_tail(&deferfcp
->req_list
, &queue
->avail_defer_list
);
556 spin_unlock_irqrestore(&queue
->qlock
, flags
);
558 /* Save NVME CMD IO in fod */
559 memcpy(&fod
->cmdiubuf
, fcpreq
->rspaddr
, fcpreq
->rsplen
);
561 /* Setup new fcpreq to be processed */
562 fcpreq
->rspaddr
= NULL
;
564 fcpreq
->nvmet_fc_private
= fod
;
565 fod
->fcpreq
= fcpreq
;
568 /* inform LLDD IO is now being processed */
569 tgtport
->ops
->defer_rcv(&tgtport
->fc_target_port
, fcpreq
);
571 /* Submit deferred IO for processing */
572 nvmet_fc_queue_fcp_req(tgtport
, queue
, fcpreq
);
575 * Leave the queue lookup get reference taken when
576 * fod was originally allocated.
581 nvmet_fc_queue_to_cpu(struct nvmet_fc_tgtport
*tgtport
, int qid
)
585 if (tgtport
->ops
->max_hw_queues
== 1)
586 return WORK_CPU_UNBOUND
;
588 /* Simple cpu selection based on qid modulo active cpu count */
589 idx
= !qid
? 0 : (qid
- 1) % num_active_cpus();
591 /* find the n'th active cpu */
592 for (cpu
= 0, cnt
= 0; ; ) {
593 if (cpu_active(cpu
)) {
598 cpu
= (cpu
+ 1) % num_possible_cpus();
604 static struct nvmet_fc_tgt_queue
*
605 nvmet_fc_alloc_target_queue(struct nvmet_fc_tgt_assoc
*assoc
,
608 struct nvmet_fc_tgt_queue
*queue
;
612 if (qid
> NVMET_NR_QUEUES
)
615 queue
= kzalloc((sizeof(*queue
) +
616 (sizeof(struct nvmet_fc_fcp_iod
) * sqsize
)),
621 if (!nvmet_fc_tgt_a_get(assoc
))
624 queue
->work_q
= alloc_workqueue("ntfc%d.%d.%d", 0, 0,
625 assoc
->tgtport
->fc_target_port
.port_num
,
630 queue
->fod
= (struct nvmet_fc_fcp_iod
*)&queue
[1];
632 queue
->sqsize
= sqsize
;
633 queue
->assoc
= assoc
;
634 queue
->port
= assoc
->tgtport
->port
;
635 queue
->cpu
= nvmet_fc_queue_to_cpu(assoc
->tgtport
, qid
);
636 INIT_LIST_HEAD(&queue
->fod_list
);
637 INIT_LIST_HEAD(&queue
->avail_defer_list
);
638 INIT_LIST_HEAD(&queue
->pending_cmd_list
);
639 atomic_set(&queue
->connected
, 0);
640 atomic_set(&queue
->sqtail
, 0);
641 atomic_set(&queue
->rsn
, 1);
642 atomic_set(&queue
->zrspcnt
, 0);
643 spin_lock_init(&queue
->qlock
);
644 kref_init(&queue
->ref
);
646 nvmet_fc_prep_fcp_iodlist(assoc
->tgtport
, queue
);
648 ret
= nvmet_sq_init(&queue
->nvme_sq
);
650 goto out_fail_iodlist
;
652 WARN_ON(assoc
->queues
[qid
]);
653 spin_lock_irqsave(&assoc
->tgtport
->lock
, flags
);
654 assoc
->queues
[qid
] = queue
;
655 spin_unlock_irqrestore(&assoc
->tgtport
->lock
, flags
);
660 nvmet_fc_destroy_fcp_iodlist(assoc
->tgtport
, queue
);
661 destroy_workqueue(queue
->work_q
);
663 nvmet_fc_tgt_a_put(assoc
);
671 nvmet_fc_tgt_queue_free(struct kref
*ref
)
673 struct nvmet_fc_tgt_queue
*queue
=
674 container_of(ref
, struct nvmet_fc_tgt_queue
, ref
);
677 spin_lock_irqsave(&queue
->assoc
->tgtport
->lock
, flags
);
678 queue
->assoc
->queues
[queue
->qid
] = NULL
;
679 spin_unlock_irqrestore(&queue
->assoc
->tgtport
->lock
, flags
);
681 nvmet_fc_destroy_fcp_iodlist(queue
->assoc
->tgtport
, queue
);
683 nvmet_fc_tgt_a_put(queue
->assoc
);
685 destroy_workqueue(queue
->work_q
);
691 nvmet_fc_tgt_q_put(struct nvmet_fc_tgt_queue
*queue
)
693 kref_put(&queue
->ref
, nvmet_fc_tgt_queue_free
);
697 nvmet_fc_tgt_q_get(struct nvmet_fc_tgt_queue
*queue
)
699 return kref_get_unless_zero(&queue
->ref
);
704 nvmet_fc_delete_target_queue(struct nvmet_fc_tgt_queue
*queue
)
706 struct nvmet_fc_tgtport
*tgtport
= queue
->assoc
->tgtport
;
707 struct nvmet_fc_fcp_iod
*fod
= queue
->fod
;
708 struct nvmet_fc_defer_fcp_req
*deferfcp
, *tempptr
;
710 int i
, writedataactive
;
713 disconnect
= atomic_xchg(&queue
->connected
, 0);
715 spin_lock_irqsave(&queue
->qlock
, flags
);
716 /* about outstanding io's */
717 for (i
= 0; i
< queue
->sqsize
; fod
++, i
++) {
719 spin_lock(&fod
->flock
);
721 writedataactive
= fod
->writedataactive
;
722 spin_unlock(&fod
->flock
);
724 * only call lldd abort routine if waiting for
725 * writedata. other outstanding ops should finish
728 if (writedataactive
) {
729 spin_lock(&fod
->flock
);
731 spin_unlock(&fod
->flock
);
732 tgtport
->ops
->fcp_abort(
733 &tgtport
->fc_target_port
, fod
->fcpreq
);
738 /* Cleanup defer'ed IOs in queue */
739 list_for_each_entry_safe(deferfcp
, tempptr
, &queue
->avail_defer_list
,
741 list_del(&deferfcp
->req_list
);
746 deferfcp
= list_first_entry_or_null(&queue
->pending_cmd_list
,
747 struct nvmet_fc_defer_fcp_req
, req_list
);
751 list_del(&deferfcp
->req_list
);
752 spin_unlock_irqrestore(&queue
->qlock
, flags
);
754 tgtport
->ops
->defer_rcv(&tgtport
->fc_target_port
,
757 tgtport
->ops
->fcp_abort(&tgtport
->fc_target_port
,
760 tgtport
->ops
->fcp_req_release(&tgtport
->fc_target_port
,
765 spin_lock_irqsave(&queue
->qlock
, flags
);
767 spin_unlock_irqrestore(&queue
->qlock
, flags
);
769 flush_workqueue(queue
->work_q
);
772 nvmet_sq_destroy(&queue
->nvme_sq
);
774 nvmet_fc_tgt_q_put(queue
);
777 static struct nvmet_fc_tgt_queue
*
778 nvmet_fc_find_target_queue(struct nvmet_fc_tgtport
*tgtport
,
781 struct nvmet_fc_tgt_assoc
*assoc
;
782 struct nvmet_fc_tgt_queue
*queue
;
783 u64 association_id
= nvmet_fc_getassociationid(connection_id
);
784 u16 qid
= nvmet_fc_getqueueid(connection_id
);
787 if (qid
> NVMET_NR_QUEUES
)
790 spin_lock_irqsave(&tgtport
->lock
, flags
);
791 list_for_each_entry(assoc
, &tgtport
->assoc_list
, a_list
) {
792 if (association_id
== assoc
->association_id
) {
793 queue
= assoc
->queues
[qid
];
795 (!atomic_read(&queue
->connected
) ||
796 !nvmet_fc_tgt_q_get(queue
)))
798 spin_unlock_irqrestore(&tgtport
->lock
, flags
);
802 spin_unlock_irqrestore(&tgtport
->lock
, flags
);
807 nvmet_fc_delete_assoc(struct work_struct
*work
)
809 struct nvmet_fc_tgt_assoc
*assoc
=
810 container_of(work
, struct nvmet_fc_tgt_assoc
, del_work
);
812 nvmet_fc_delete_target_assoc(assoc
);
813 nvmet_fc_tgt_a_put(assoc
);
816 static struct nvmet_fc_tgt_assoc
*
817 nvmet_fc_alloc_target_assoc(struct nvmet_fc_tgtport
*tgtport
)
819 struct nvmet_fc_tgt_assoc
*assoc
, *tmpassoc
;
823 bool needrandom
= true;
825 assoc
= kzalloc(sizeof(*assoc
), GFP_KERNEL
);
829 idx
= ida_simple_get(&tgtport
->assoc_cnt
, 0, 0, GFP_KERNEL
);
833 if (!nvmet_fc_tgtport_get(tgtport
))
836 assoc
->tgtport
= tgtport
;
838 INIT_LIST_HEAD(&assoc
->a_list
);
839 kref_init(&assoc
->ref
);
840 INIT_WORK(&assoc
->del_work
, nvmet_fc_delete_assoc
);
843 get_random_bytes(&ran
, sizeof(ran
) - BYTES_FOR_QID
);
844 ran
= ran
<< BYTES_FOR_QID_SHIFT
;
846 spin_lock_irqsave(&tgtport
->lock
, flags
);
848 list_for_each_entry(tmpassoc
, &tgtport
->assoc_list
, a_list
)
849 if (ran
== tmpassoc
->association_id
) {
854 assoc
->association_id
= ran
;
855 list_add_tail(&assoc
->a_list
, &tgtport
->assoc_list
);
857 spin_unlock_irqrestore(&tgtport
->lock
, flags
);
863 ida_simple_remove(&tgtport
->assoc_cnt
, idx
);
870 nvmet_fc_target_assoc_free(struct kref
*ref
)
872 struct nvmet_fc_tgt_assoc
*assoc
=
873 container_of(ref
, struct nvmet_fc_tgt_assoc
, ref
);
874 struct nvmet_fc_tgtport
*tgtport
= assoc
->tgtport
;
877 spin_lock_irqsave(&tgtport
->lock
, flags
);
878 list_del(&assoc
->a_list
);
879 spin_unlock_irqrestore(&tgtport
->lock
, flags
);
880 ida_simple_remove(&tgtport
->assoc_cnt
, assoc
->a_id
);
882 nvmet_fc_tgtport_put(tgtport
);
886 nvmet_fc_tgt_a_put(struct nvmet_fc_tgt_assoc
*assoc
)
888 kref_put(&assoc
->ref
, nvmet_fc_target_assoc_free
);
892 nvmet_fc_tgt_a_get(struct nvmet_fc_tgt_assoc
*assoc
)
894 return kref_get_unless_zero(&assoc
->ref
);
898 nvmet_fc_delete_target_assoc(struct nvmet_fc_tgt_assoc
*assoc
)
900 struct nvmet_fc_tgtport
*tgtport
= assoc
->tgtport
;
901 struct nvmet_fc_tgt_queue
*queue
;
905 spin_lock_irqsave(&tgtport
->lock
, flags
);
906 for (i
= NVMET_NR_QUEUES
; i
>= 0; i
--) {
907 queue
= assoc
->queues
[i
];
909 if (!nvmet_fc_tgt_q_get(queue
))
911 spin_unlock_irqrestore(&tgtport
->lock
, flags
);
912 nvmet_fc_delete_target_queue(queue
);
913 nvmet_fc_tgt_q_put(queue
);
914 spin_lock_irqsave(&tgtport
->lock
, flags
);
917 spin_unlock_irqrestore(&tgtport
->lock
, flags
);
919 nvmet_fc_tgt_a_put(assoc
);
922 static struct nvmet_fc_tgt_assoc
*
923 nvmet_fc_find_target_assoc(struct nvmet_fc_tgtport
*tgtport
,
926 struct nvmet_fc_tgt_assoc
*assoc
;
927 struct nvmet_fc_tgt_assoc
*ret
= NULL
;
930 spin_lock_irqsave(&tgtport
->lock
, flags
);
931 list_for_each_entry(assoc
, &tgtport
->assoc_list
, a_list
) {
932 if (association_id
== assoc
->association_id
) {
934 nvmet_fc_tgt_a_get(assoc
);
938 spin_unlock_irqrestore(&tgtport
->lock
, flags
);
945 * nvme_fc_register_targetport - transport entry point called by an
946 * LLDD to register the existence of a local
947 * NVME subystem FC port.
948 * @pinfo: pointer to information about the port to be registered
949 * @template: LLDD entrypoints and operational parameters for the port
950 * @dev: physical hardware device node port corresponds to. Will be
951 * used for DMA mappings
952 * @portptr: pointer to a local port pointer. Upon success, the routine
953 * will allocate a nvme_fc_local_port structure and place its
954 * address in the local port pointer. Upon failure, local port
955 * pointer will be set to NULL.
958 * a completion status. Must be 0 upon success; a negative errno
959 * (ex: -ENXIO) upon failure.
962 nvmet_fc_register_targetport(struct nvmet_fc_port_info
*pinfo
,
963 struct nvmet_fc_target_template
*template,
965 struct nvmet_fc_target_port
**portptr
)
967 struct nvmet_fc_tgtport
*newrec
;
971 if (!template->xmt_ls_rsp
|| !template->fcp_op
||
972 !template->fcp_abort
||
973 !template->fcp_req_release
|| !template->targetport_delete
||
974 !template->max_hw_queues
|| !template->max_sgl_segments
||
975 !template->max_dif_sgl_segments
|| !template->dma_boundary
) {
977 goto out_regtgt_failed
;
980 newrec
= kzalloc((sizeof(*newrec
) + template->target_priv_sz
),
984 goto out_regtgt_failed
;
987 idx
= ida_simple_get(&nvmet_fc_tgtport_cnt
, 0, 0, GFP_KERNEL
);
993 if (!get_device(dev
) && dev
) {
998 newrec
->fc_target_port
.node_name
= pinfo
->node_name
;
999 newrec
->fc_target_port
.port_name
= pinfo
->port_name
;
1000 newrec
->fc_target_port
.private = &newrec
[1];
1001 newrec
->fc_target_port
.port_id
= pinfo
->port_id
;
1002 newrec
->fc_target_port
.port_num
= idx
;
1003 INIT_LIST_HEAD(&newrec
->tgt_list
);
1005 newrec
->ops
= template;
1006 spin_lock_init(&newrec
->lock
);
1007 INIT_LIST_HEAD(&newrec
->ls_list
);
1008 INIT_LIST_HEAD(&newrec
->ls_busylist
);
1009 INIT_LIST_HEAD(&newrec
->assoc_list
);
1010 kref_init(&newrec
->ref
);
1011 ida_init(&newrec
->assoc_cnt
);
1012 newrec
->max_sg_cnt
= min_t(u32
, NVMET_FC_MAX_XFR_SGENTS
,
1013 template->max_sgl_segments
);
1015 ret
= nvmet_fc_alloc_ls_iodlist(newrec
);
1018 goto out_free_newrec
;
1021 spin_lock_irqsave(&nvmet_fc_tgtlock
, flags
);
1022 list_add_tail(&newrec
->tgt_list
, &nvmet_fc_target_list
);
1023 spin_unlock_irqrestore(&nvmet_fc_tgtlock
, flags
);
1025 *portptr
= &newrec
->fc_target_port
;
1031 ida_simple_remove(&nvmet_fc_tgtport_cnt
, idx
);
1038 EXPORT_SYMBOL_GPL(nvmet_fc_register_targetport
);
1042 nvmet_fc_free_tgtport(struct kref
*ref
)
1044 struct nvmet_fc_tgtport
*tgtport
=
1045 container_of(ref
, struct nvmet_fc_tgtport
, ref
);
1046 struct device
*dev
= tgtport
->dev
;
1047 unsigned long flags
;
1049 spin_lock_irqsave(&nvmet_fc_tgtlock
, flags
);
1050 list_del(&tgtport
->tgt_list
);
1051 spin_unlock_irqrestore(&nvmet_fc_tgtlock
, flags
);
1053 nvmet_fc_free_ls_iodlist(tgtport
);
1055 /* let the LLDD know we've finished tearing it down */
1056 tgtport
->ops
->targetport_delete(&tgtport
->fc_target_port
);
1058 ida_simple_remove(&nvmet_fc_tgtport_cnt
,
1059 tgtport
->fc_target_port
.port_num
);
1061 ida_destroy(&tgtport
->assoc_cnt
);
1069 nvmet_fc_tgtport_put(struct nvmet_fc_tgtport
*tgtport
)
1071 kref_put(&tgtport
->ref
, nvmet_fc_free_tgtport
);
1075 nvmet_fc_tgtport_get(struct nvmet_fc_tgtport
*tgtport
)
1077 return kref_get_unless_zero(&tgtport
->ref
);
1081 __nvmet_fc_free_assocs(struct nvmet_fc_tgtport
*tgtport
)
1083 struct nvmet_fc_tgt_assoc
*assoc
, *next
;
1084 unsigned long flags
;
1086 spin_lock_irqsave(&tgtport
->lock
, flags
);
1087 list_for_each_entry_safe(assoc
, next
,
1088 &tgtport
->assoc_list
, a_list
) {
1089 if (!nvmet_fc_tgt_a_get(assoc
))
1091 spin_unlock_irqrestore(&tgtport
->lock
, flags
);
1092 nvmet_fc_delete_target_assoc(assoc
);
1093 nvmet_fc_tgt_a_put(assoc
);
1094 spin_lock_irqsave(&tgtport
->lock
, flags
);
1096 spin_unlock_irqrestore(&tgtport
->lock
, flags
);
1100 * nvmet layer has called to terminate an association
1103 nvmet_fc_delete_ctrl(struct nvmet_ctrl
*ctrl
)
1105 struct nvmet_fc_tgtport
*tgtport
, *next
;
1106 struct nvmet_fc_tgt_assoc
*assoc
;
1107 struct nvmet_fc_tgt_queue
*queue
;
1108 unsigned long flags
;
1109 bool found_ctrl
= false;
1111 /* this is a bit ugly, but don't want to make locks layered */
1112 spin_lock_irqsave(&nvmet_fc_tgtlock
, flags
);
1113 list_for_each_entry_safe(tgtport
, next
, &nvmet_fc_target_list
,
1115 if (!nvmet_fc_tgtport_get(tgtport
))
1117 spin_unlock_irqrestore(&nvmet_fc_tgtlock
, flags
);
1119 spin_lock_irqsave(&tgtport
->lock
, flags
);
1120 list_for_each_entry(assoc
, &tgtport
->assoc_list
, a_list
) {
1121 queue
= assoc
->queues
[0];
1122 if (queue
&& queue
->nvme_sq
.ctrl
== ctrl
) {
1123 if (nvmet_fc_tgt_a_get(assoc
))
1128 spin_unlock_irqrestore(&tgtport
->lock
, flags
);
1130 nvmet_fc_tgtport_put(tgtport
);
1133 schedule_work(&assoc
->del_work
);
1137 spin_lock_irqsave(&nvmet_fc_tgtlock
, flags
);
1139 spin_unlock_irqrestore(&nvmet_fc_tgtlock
, flags
);
1143 * nvme_fc_unregister_targetport - transport entry point called by an
1144 * LLDD to deregister/remove a previously
1145 * registered a local NVME subsystem FC port.
1146 * @tgtport: pointer to the (registered) target port that is to be
1150 * a completion status. Must be 0 upon success; a negative errno
1151 * (ex: -ENXIO) upon failure.
1154 nvmet_fc_unregister_targetport(struct nvmet_fc_target_port
*target_port
)
1156 struct nvmet_fc_tgtport
*tgtport
= targetport_to_tgtport(target_port
);
1158 /* terminate any outstanding associations */
1159 __nvmet_fc_free_assocs(tgtport
);
1161 nvmet_fc_tgtport_put(tgtport
);
1165 EXPORT_SYMBOL_GPL(nvmet_fc_unregister_targetport
);
1168 /* *********************** FC-NVME LS Handling **************************** */
1172 nvmet_fc_format_rsp_hdr(void *buf
, u8 ls_cmd
, __be32 desc_len
, u8 rqst_ls_cmd
)
1174 struct fcnvme_ls_acc_hdr
*acc
= buf
;
1176 acc
->w0
.ls_cmd
= ls_cmd
;
1177 acc
->desc_list_len
= desc_len
;
1178 acc
->rqst
.desc_tag
= cpu_to_be32(FCNVME_LSDESC_RQST
);
1179 acc
->rqst
.desc_len
=
1180 fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_rqst
));
1181 acc
->rqst
.w0
.ls_cmd
= rqst_ls_cmd
;
1185 nvmet_fc_format_rjt(void *buf
, u16 buflen
, u8 ls_cmd
,
1186 u8 reason
, u8 explanation
, u8 vendor
)
1188 struct fcnvme_ls_rjt
*rjt
= buf
;
1190 nvmet_fc_format_rsp_hdr(buf
, FCNVME_LSDESC_RQST
,
1191 fcnvme_lsdesc_len(sizeof(struct fcnvme_ls_rjt
)),
1193 rjt
->rjt
.desc_tag
= cpu_to_be32(FCNVME_LSDESC_RJT
);
1194 rjt
->rjt
.desc_len
= fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_rjt
));
1195 rjt
->rjt
.reason_code
= reason
;
1196 rjt
->rjt
.reason_explanation
= explanation
;
1197 rjt
->rjt
.vendor
= vendor
;
1199 return sizeof(struct fcnvme_ls_rjt
);
1202 /* Validation Error indexes into the string table below */
1205 VERR_CR_ASSOC_LEN
= 1,
1206 VERR_CR_ASSOC_RQST_LEN
= 2,
1207 VERR_CR_ASSOC_CMD
= 3,
1208 VERR_CR_ASSOC_CMD_LEN
= 4,
1209 VERR_ERSP_RATIO
= 5,
1210 VERR_ASSOC_ALLOC_FAIL
= 6,
1211 VERR_QUEUE_ALLOC_FAIL
= 7,
1212 VERR_CR_CONN_LEN
= 8,
1213 VERR_CR_CONN_RQST_LEN
= 9,
1215 VERR_ASSOC_ID_LEN
= 11,
1218 VERR_CONN_ID_LEN
= 14,
1220 VERR_CR_CONN_CMD
= 16,
1221 VERR_CR_CONN_CMD_LEN
= 17,
1222 VERR_DISCONN_LEN
= 18,
1223 VERR_DISCONN_RQST_LEN
= 19,
1224 VERR_DISCONN_CMD
= 20,
1225 VERR_DISCONN_CMD_LEN
= 21,
1226 VERR_DISCONN_SCOPE
= 22,
1228 VERR_RS_RQST_LEN
= 24,
1230 VERR_RS_CMD_LEN
= 26,
1235 static char *validation_errors
[] = {
1237 "Bad CR_ASSOC Length",
1238 "Bad CR_ASSOC Rqst Length",
1240 "Bad CR_ASSOC Cmd Length",
1242 "Association Allocation Failed",
1243 "Queue Allocation Failed",
1244 "Bad CR_CONN Length",
1245 "Bad CR_CONN Rqst Length",
1246 "Not Association ID",
1247 "Bad Association ID Length",
1249 "Not Connection ID",
1250 "Bad Connection ID Length",
1253 "Bad CR_CONN Cmd Length",
1254 "Bad DISCONN Length",
1255 "Bad DISCONN Rqst Length",
1257 "Bad DISCONN Cmd Length",
1258 "Bad Disconnect Scope",
1260 "Bad RS Rqst Length",
1262 "Bad RS Cmd Length",
1264 "Bad RS Relative Offset",
1268 nvmet_fc_ls_create_association(struct nvmet_fc_tgtport
*tgtport
,
1269 struct nvmet_fc_ls_iod
*iod
)
1271 struct fcnvme_ls_cr_assoc_rqst
*rqst
=
1272 (struct fcnvme_ls_cr_assoc_rqst
*)iod
->rqstbuf
;
1273 struct fcnvme_ls_cr_assoc_acc
*acc
=
1274 (struct fcnvme_ls_cr_assoc_acc
*)iod
->rspbuf
;
1275 struct nvmet_fc_tgt_queue
*queue
;
1278 memset(acc
, 0, sizeof(*acc
));
1281 * FC-NVME spec changes. There are initiators sending different
1282 * lengths as padding sizes for Create Association Cmd descriptor
1284 * Accept anything of "minimum" length. Assume format per 1.15
1285 * spec (with HOSTID reduced to 16 bytes), ignore how long the
1286 * trailing pad length is.
1288 if (iod
->rqstdatalen
< FCNVME_LSDESC_CRA_RQST_MINLEN
)
1289 ret
= VERR_CR_ASSOC_LEN
;
1290 else if (be32_to_cpu(rqst
->desc_list_len
) <
1291 FCNVME_LSDESC_CRA_RQST_MIN_LISTLEN
)
1292 ret
= VERR_CR_ASSOC_RQST_LEN
;
1293 else if (rqst
->assoc_cmd
.desc_tag
!=
1294 cpu_to_be32(FCNVME_LSDESC_CREATE_ASSOC_CMD
))
1295 ret
= VERR_CR_ASSOC_CMD
;
1296 else if (be32_to_cpu(rqst
->assoc_cmd
.desc_len
) <
1297 FCNVME_LSDESC_CRA_CMD_DESC_MIN_DESCLEN
)
1298 ret
= VERR_CR_ASSOC_CMD_LEN
;
1299 else if (!rqst
->assoc_cmd
.ersp_ratio
||
1300 (be16_to_cpu(rqst
->assoc_cmd
.ersp_ratio
) >=
1301 be16_to_cpu(rqst
->assoc_cmd
.sqsize
)))
1302 ret
= VERR_ERSP_RATIO
;
1305 /* new association w/ admin queue */
1306 iod
->assoc
= nvmet_fc_alloc_target_assoc(tgtport
);
1308 ret
= VERR_ASSOC_ALLOC_FAIL
;
1310 queue
= nvmet_fc_alloc_target_queue(iod
->assoc
, 0,
1311 be16_to_cpu(rqst
->assoc_cmd
.sqsize
));
1313 ret
= VERR_QUEUE_ALLOC_FAIL
;
1318 dev_err(tgtport
->dev
,
1319 "Create Association LS failed: %s\n",
1320 validation_errors
[ret
]);
1321 iod
->lsreq
->rsplen
= nvmet_fc_format_rjt(acc
,
1322 NVME_FC_MAX_LS_BUFFER_SIZE
, rqst
->w0
.ls_cmd
,
1323 FCNVME_RJT_RC_LOGIC
,
1324 FCNVME_RJT_EXP_NONE
, 0);
1328 queue
->ersp_ratio
= be16_to_cpu(rqst
->assoc_cmd
.ersp_ratio
);
1329 atomic_set(&queue
->connected
, 1);
1330 queue
->sqhd
= 0; /* best place to init value */
1332 /* format a response */
1334 iod
->lsreq
->rsplen
= sizeof(*acc
);
1336 nvmet_fc_format_rsp_hdr(acc
, FCNVME_LS_ACC
,
1338 sizeof(struct fcnvme_ls_cr_assoc_acc
)),
1339 FCNVME_LS_CREATE_ASSOCIATION
);
1340 acc
->associd
.desc_tag
= cpu_to_be32(FCNVME_LSDESC_ASSOC_ID
);
1341 acc
->associd
.desc_len
=
1343 sizeof(struct fcnvme_lsdesc_assoc_id
));
1344 acc
->associd
.association_id
=
1345 cpu_to_be64(nvmet_fc_makeconnid(iod
->assoc
, 0));
1346 acc
->connectid
.desc_tag
= cpu_to_be32(FCNVME_LSDESC_CONN_ID
);
1347 acc
->connectid
.desc_len
=
1349 sizeof(struct fcnvme_lsdesc_conn_id
));
1350 acc
->connectid
.connection_id
= acc
->associd
.association_id
;
1354 nvmet_fc_ls_create_connection(struct nvmet_fc_tgtport
*tgtport
,
1355 struct nvmet_fc_ls_iod
*iod
)
1357 struct fcnvme_ls_cr_conn_rqst
*rqst
=
1358 (struct fcnvme_ls_cr_conn_rqst
*)iod
->rqstbuf
;
1359 struct fcnvme_ls_cr_conn_acc
*acc
=
1360 (struct fcnvme_ls_cr_conn_acc
*)iod
->rspbuf
;
1361 struct nvmet_fc_tgt_queue
*queue
;
1364 memset(acc
, 0, sizeof(*acc
));
1366 if (iod
->rqstdatalen
< sizeof(struct fcnvme_ls_cr_conn_rqst
))
1367 ret
= VERR_CR_CONN_LEN
;
1368 else if (rqst
->desc_list_len
!=
1370 sizeof(struct fcnvme_ls_cr_conn_rqst
)))
1371 ret
= VERR_CR_CONN_RQST_LEN
;
1372 else if (rqst
->associd
.desc_tag
!= cpu_to_be32(FCNVME_LSDESC_ASSOC_ID
))
1373 ret
= VERR_ASSOC_ID
;
1374 else if (rqst
->associd
.desc_len
!=
1376 sizeof(struct fcnvme_lsdesc_assoc_id
)))
1377 ret
= VERR_ASSOC_ID_LEN
;
1378 else if (rqst
->connect_cmd
.desc_tag
!=
1379 cpu_to_be32(FCNVME_LSDESC_CREATE_CONN_CMD
))
1380 ret
= VERR_CR_CONN_CMD
;
1381 else if (rqst
->connect_cmd
.desc_len
!=
1383 sizeof(struct fcnvme_lsdesc_cr_conn_cmd
)))
1384 ret
= VERR_CR_CONN_CMD_LEN
;
1385 else if (!rqst
->connect_cmd
.ersp_ratio
||
1386 (be16_to_cpu(rqst
->connect_cmd
.ersp_ratio
) >=
1387 be16_to_cpu(rqst
->connect_cmd
.sqsize
)))
1388 ret
= VERR_ERSP_RATIO
;
1392 iod
->assoc
= nvmet_fc_find_target_assoc(tgtport
,
1393 be64_to_cpu(rqst
->associd
.association_id
));
1395 ret
= VERR_NO_ASSOC
;
1397 queue
= nvmet_fc_alloc_target_queue(iod
->assoc
,
1398 be16_to_cpu(rqst
->connect_cmd
.qid
),
1399 be16_to_cpu(rqst
->connect_cmd
.sqsize
));
1401 ret
= VERR_QUEUE_ALLOC_FAIL
;
1403 /* release get taken in nvmet_fc_find_target_assoc */
1404 nvmet_fc_tgt_a_put(iod
->assoc
);
1409 dev_err(tgtport
->dev
,
1410 "Create Connection LS failed: %s\n",
1411 validation_errors
[ret
]);
1412 iod
->lsreq
->rsplen
= nvmet_fc_format_rjt(acc
,
1413 NVME_FC_MAX_LS_BUFFER_SIZE
, rqst
->w0
.ls_cmd
,
1414 (ret
== VERR_NO_ASSOC
) ?
1415 FCNVME_RJT_RC_INV_ASSOC
:
1416 FCNVME_RJT_RC_LOGIC
,
1417 FCNVME_RJT_EXP_NONE
, 0);
1421 queue
->ersp_ratio
= be16_to_cpu(rqst
->connect_cmd
.ersp_ratio
);
1422 atomic_set(&queue
->connected
, 1);
1423 queue
->sqhd
= 0; /* best place to init value */
1425 /* format a response */
1427 iod
->lsreq
->rsplen
= sizeof(*acc
);
1429 nvmet_fc_format_rsp_hdr(acc
, FCNVME_LS_ACC
,
1430 fcnvme_lsdesc_len(sizeof(struct fcnvme_ls_cr_conn_acc
)),
1431 FCNVME_LS_CREATE_CONNECTION
);
1432 acc
->connectid
.desc_tag
= cpu_to_be32(FCNVME_LSDESC_CONN_ID
);
1433 acc
->connectid
.desc_len
=
1435 sizeof(struct fcnvme_lsdesc_conn_id
));
1436 acc
->connectid
.connection_id
=
1437 cpu_to_be64(nvmet_fc_makeconnid(iod
->assoc
,
1438 be16_to_cpu(rqst
->connect_cmd
.qid
)));
1442 nvmet_fc_ls_disconnect(struct nvmet_fc_tgtport
*tgtport
,
1443 struct nvmet_fc_ls_iod
*iod
)
1445 struct fcnvme_ls_disconnect_rqst
*rqst
=
1446 (struct fcnvme_ls_disconnect_rqst
*)iod
->rqstbuf
;
1447 struct fcnvme_ls_disconnect_acc
*acc
=
1448 (struct fcnvme_ls_disconnect_acc
*)iod
->rspbuf
;
1449 struct nvmet_fc_tgt_queue
*queue
= NULL
;
1450 struct nvmet_fc_tgt_assoc
*assoc
;
1452 bool del_assoc
= false;
1454 memset(acc
, 0, sizeof(*acc
));
1456 if (iod
->rqstdatalen
< sizeof(struct fcnvme_ls_disconnect_rqst
))
1457 ret
= VERR_DISCONN_LEN
;
1458 else if (rqst
->desc_list_len
!=
1460 sizeof(struct fcnvme_ls_disconnect_rqst
)))
1461 ret
= VERR_DISCONN_RQST_LEN
;
1462 else if (rqst
->associd
.desc_tag
!= cpu_to_be32(FCNVME_LSDESC_ASSOC_ID
))
1463 ret
= VERR_ASSOC_ID
;
1464 else if (rqst
->associd
.desc_len
!=
1466 sizeof(struct fcnvme_lsdesc_assoc_id
)))
1467 ret
= VERR_ASSOC_ID_LEN
;
1468 else if (rqst
->discon_cmd
.desc_tag
!=
1469 cpu_to_be32(FCNVME_LSDESC_DISCONN_CMD
))
1470 ret
= VERR_DISCONN_CMD
;
1471 else if (rqst
->discon_cmd
.desc_len
!=
1473 sizeof(struct fcnvme_lsdesc_disconn_cmd
)))
1474 ret
= VERR_DISCONN_CMD_LEN
;
1475 else if ((rqst
->discon_cmd
.scope
!= FCNVME_DISCONN_ASSOCIATION
) &&
1476 (rqst
->discon_cmd
.scope
!= FCNVME_DISCONN_CONNECTION
))
1477 ret
= VERR_DISCONN_SCOPE
;
1479 /* match an active association */
1480 assoc
= nvmet_fc_find_target_assoc(tgtport
,
1481 be64_to_cpu(rqst
->associd
.association_id
));
1484 if (rqst
->discon_cmd
.scope
==
1485 FCNVME_DISCONN_CONNECTION
) {
1486 queue
= nvmet_fc_find_target_queue(tgtport
,
1488 rqst
->discon_cmd
.id
));
1490 nvmet_fc_tgt_a_put(assoc
);
1495 ret
= VERR_NO_ASSOC
;
1499 dev_err(tgtport
->dev
,
1500 "Disconnect LS failed: %s\n",
1501 validation_errors
[ret
]);
1502 iod
->lsreq
->rsplen
= nvmet_fc_format_rjt(acc
,
1503 NVME_FC_MAX_LS_BUFFER_SIZE
, rqst
->w0
.ls_cmd
,
1504 (ret
== VERR_NO_ASSOC
) ?
1505 FCNVME_RJT_RC_INV_ASSOC
:
1506 (ret
== VERR_NO_CONN
) ?
1507 FCNVME_RJT_RC_INV_CONN
:
1508 FCNVME_RJT_RC_LOGIC
,
1509 FCNVME_RJT_EXP_NONE
, 0);
1513 /* format a response */
1515 iod
->lsreq
->rsplen
= sizeof(*acc
);
1517 nvmet_fc_format_rsp_hdr(acc
, FCNVME_LS_ACC
,
1519 sizeof(struct fcnvme_ls_disconnect_acc
)),
1520 FCNVME_LS_DISCONNECT
);
1523 /* are we to delete a Connection ID (queue) */
1525 int qid
= queue
->qid
;
1527 nvmet_fc_delete_target_queue(queue
);
1529 /* release the get taken by find_target_queue */
1530 nvmet_fc_tgt_q_put(queue
);
1532 /* tear association down if io queue terminated */
1537 /* release get taken in nvmet_fc_find_target_assoc */
1538 nvmet_fc_tgt_a_put(iod
->assoc
);
1541 nvmet_fc_delete_target_assoc(iod
->assoc
);
1545 /* *********************** NVME Ctrl Routines **************************** */
1548 static void nvmet_fc_fcp_nvme_cmd_done(struct nvmet_req
*nvme_req
);
1550 static struct nvmet_fabrics_ops nvmet_fc_tgt_fcp_ops
;
1553 nvmet_fc_xmt_ls_rsp_done(struct nvmefc_tgt_ls_req
*lsreq
)
1555 struct nvmet_fc_ls_iod
*iod
= lsreq
->nvmet_fc_private
;
1556 struct nvmet_fc_tgtport
*tgtport
= iod
->tgtport
;
1558 fc_dma_sync_single_for_cpu(tgtport
->dev
, iod
->rspdma
,
1559 NVME_FC_MAX_LS_BUFFER_SIZE
, DMA_TO_DEVICE
);
1560 nvmet_fc_free_ls_iod(tgtport
, iod
);
1561 nvmet_fc_tgtport_put(tgtport
);
1565 nvmet_fc_xmt_ls_rsp(struct nvmet_fc_tgtport
*tgtport
,
1566 struct nvmet_fc_ls_iod
*iod
)
1570 fc_dma_sync_single_for_device(tgtport
->dev
, iod
->rspdma
,
1571 NVME_FC_MAX_LS_BUFFER_SIZE
, DMA_TO_DEVICE
);
1573 ret
= tgtport
->ops
->xmt_ls_rsp(&tgtport
->fc_target_port
, iod
->lsreq
);
1575 nvmet_fc_xmt_ls_rsp_done(iod
->lsreq
);
1579 * Actual processing routine for received FC-NVME LS Requests from the LLD
1582 nvmet_fc_handle_ls_rqst(struct nvmet_fc_tgtport
*tgtport
,
1583 struct nvmet_fc_ls_iod
*iod
)
1585 struct fcnvme_ls_rqst_w0
*w0
=
1586 (struct fcnvme_ls_rqst_w0
*)iod
->rqstbuf
;
1588 iod
->lsreq
->nvmet_fc_private
= iod
;
1589 iod
->lsreq
->rspbuf
= iod
->rspbuf
;
1590 iod
->lsreq
->rspdma
= iod
->rspdma
;
1591 iod
->lsreq
->done
= nvmet_fc_xmt_ls_rsp_done
;
1592 /* Be preventative. handlers will later set to valid length */
1593 iod
->lsreq
->rsplen
= 0;
1599 * parse request input, execute the request, and format the
1602 switch (w0
->ls_cmd
) {
1603 case FCNVME_LS_CREATE_ASSOCIATION
:
1604 /* Creates Association and initial Admin Queue/Connection */
1605 nvmet_fc_ls_create_association(tgtport
, iod
);
1607 case FCNVME_LS_CREATE_CONNECTION
:
1608 /* Creates an IO Queue/Connection */
1609 nvmet_fc_ls_create_connection(tgtport
, iod
);
1611 case FCNVME_LS_DISCONNECT
:
1612 /* Terminate a Queue/Connection or the Association */
1613 nvmet_fc_ls_disconnect(tgtport
, iod
);
1616 iod
->lsreq
->rsplen
= nvmet_fc_format_rjt(iod
->rspbuf
,
1617 NVME_FC_MAX_LS_BUFFER_SIZE
, w0
->ls_cmd
,
1618 FCNVME_RJT_RC_INVAL
, FCNVME_RJT_EXP_NONE
, 0);
1621 nvmet_fc_xmt_ls_rsp(tgtport
, iod
);
1625 * Actual processing routine for received FC-NVME LS Requests from the LLD
1628 nvmet_fc_handle_ls_rqst_work(struct work_struct
*work
)
1630 struct nvmet_fc_ls_iod
*iod
=
1631 container_of(work
, struct nvmet_fc_ls_iod
, work
);
1632 struct nvmet_fc_tgtport
*tgtport
= iod
->tgtport
;
1634 nvmet_fc_handle_ls_rqst(tgtport
, iod
);
1639 * nvmet_fc_rcv_ls_req - transport entry point called by an LLDD
1640 * upon the reception of a NVME LS request.
1642 * The nvmet-fc layer will copy payload to an internal structure for
1643 * processing. As such, upon completion of the routine, the LLDD may
1644 * immediately free/reuse the LS request buffer passed in the call.
1646 * If this routine returns error, the LLDD should abort the exchange.
1648 * @tgtport: pointer to the (registered) target port the LS was
1650 * @lsreq: pointer to a lsreq request structure to be used to reference
1651 * the exchange corresponding to the LS.
1652 * @lsreqbuf: pointer to the buffer containing the LS Request
1653 * @lsreqbuf_len: length, in bytes, of the received LS request
1656 nvmet_fc_rcv_ls_req(struct nvmet_fc_target_port
*target_port
,
1657 struct nvmefc_tgt_ls_req
*lsreq
,
1658 void *lsreqbuf
, u32 lsreqbuf_len
)
1660 struct nvmet_fc_tgtport
*tgtport
= targetport_to_tgtport(target_port
);
1661 struct nvmet_fc_ls_iod
*iod
;
1663 if (lsreqbuf_len
> NVME_FC_MAX_LS_BUFFER_SIZE
)
1666 if (!nvmet_fc_tgtport_get(tgtport
))
1669 iod
= nvmet_fc_alloc_ls_iod(tgtport
);
1671 nvmet_fc_tgtport_put(tgtport
);
1677 memcpy(iod
->rqstbuf
, lsreqbuf
, lsreqbuf_len
);
1678 iod
->rqstdatalen
= lsreqbuf_len
;
1680 schedule_work(&iod
->work
);
1684 EXPORT_SYMBOL_GPL(nvmet_fc_rcv_ls_req
);
1688 * **********************
1689 * Start of FCP handling
1690 * **********************
1694 nvmet_fc_alloc_tgt_pgs(struct nvmet_fc_fcp_iod
*fod
)
1696 struct scatterlist
*sg
;
1699 u32 page_len
, length
;
1702 length
= fod
->req
.transfer_len
;
1703 nent
= DIV_ROUND_UP(length
, PAGE_SIZE
);
1704 sg
= kmalloc_array(nent
, sizeof(struct scatterlist
), GFP_KERNEL
);
1708 sg_init_table(sg
, nent
);
1711 page_len
= min_t(u32
, length
, PAGE_SIZE
);
1713 page
= alloc_page(GFP_KERNEL
);
1715 goto out_free_pages
;
1717 sg_set_page(&sg
[i
], page
, page_len
, 0);
1723 fod
->data_sg_cnt
= nent
;
1724 fod
->data_sg_cnt
= fc_dma_map_sg(fod
->tgtport
->dev
, sg
, nent
,
1725 ((fod
->io_dir
== NVMET_FCP_WRITE
) ?
1726 DMA_FROM_DEVICE
: DMA_TO_DEVICE
));
1727 /* note: write from initiator perspective */
1734 __free_page(sg_page(&sg
[i
]));
1737 fod
->data_sg
= NULL
;
1738 fod
->data_sg_cnt
= 0;
1740 return NVME_SC_INTERNAL
;
1744 nvmet_fc_free_tgt_pgs(struct nvmet_fc_fcp_iod
*fod
)
1746 struct scatterlist
*sg
;
1749 if (!fod
->data_sg
|| !fod
->data_sg_cnt
)
1752 fc_dma_unmap_sg(fod
->tgtport
->dev
, fod
->data_sg
, fod
->data_sg_cnt
,
1753 ((fod
->io_dir
== NVMET_FCP_WRITE
) ?
1754 DMA_FROM_DEVICE
: DMA_TO_DEVICE
));
1755 for_each_sg(fod
->data_sg
, sg
, fod
->data_sg_cnt
, count
)
1756 __free_page(sg_page(sg
));
1757 kfree(fod
->data_sg
);
1758 fod
->data_sg
= NULL
;
1759 fod
->data_sg_cnt
= 0;
1764 queue_90percent_full(struct nvmet_fc_tgt_queue
*q
, u32 sqhd
)
1768 /* egad, this is ugly. And sqtail is just a best guess */
1769 sqtail
= atomic_read(&q
->sqtail
) % q
->sqsize
;
1771 used
= (sqtail
< sqhd
) ? (sqtail
+ q
->sqsize
- sqhd
) : (sqtail
- sqhd
);
1772 return ((used
* 10) >= (((u32
)(q
->sqsize
- 1) * 9)));
1777 * May be a NVMET_FCOP_RSP or NVMET_FCOP_READDATA_RSP op
1780 nvmet_fc_prep_fcp_rsp(struct nvmet_fc_tgtport
*tgtport
,
1781 struct nvmet_fc_fcp_iod
*fod
)
1783 struct nvme_fc_ersp_iu
*ersp
= &fod
->rspiubuf
;
1784 struct nvme_common_command
*sqe
= &fod
->cmdiubuf
.sqe
.common
;
1785 struct nvme_completion
*cqe
= &ersp
->cqe
;
1786 u32
*cqewd
= (u32
*)cqe
;
1787 bool send_ersp
= false;
1788 u32 rsn
, rspcnt
, xfr_length
;
1790 if (fod
->fcpreq
->op
== NVMET_FCOP_READDATA_RSP
)
1791 xfr_length
= fod
->req
.transfer_len
;
1793 xfr_length
= fod
->offset
;
1796 * check to see if we can send a 0's rsp.
1797 * Note: to send a 0's response, the NVME-FC host transport will
1798 * recreate the CQE. The host transport knows: sq id, SQHD (last
1799 * seen in an ersp), and command_id. Thus it will create a
1800 * zero-filled CQE with those known fields filled in. Transport
1801 * must send an ersp for any condition where the cqe won't match
1804 * Here are the FC-NVME mandated cases where we must send an ersp:
1805 * every N responses, where N=ersp_ratio
1806 * force fabric commands to send ersp's (not in FC-NVME but good
1808 * normal cmds: any time status is non-zero, or status is zero
1809 * but words 0 or 1 are non-zero.
1810 * the SQ is 90% or more full
1811 * the cmd is a fused command
1812 * transferred data length not equal to cmd iu length
1814 rspcnt
= atomic_inc_return(&fod
->queue
->zrspcnt
);
1815 if (!(rspcnt
% fod
->queue
->ersp_ratio
) ||
1816 sqe
->opcode
== nvme_fabrics_command
||
1817 xfr_length
!= fod
->req
.transfer_len
||
1818 (le16_to_cpu(cqe
->status
) & 0xFFFE) || cqewd
[0] || cqewd
[1] ||
1819 (sqe
->flags
& (NVME_CMD_FUSE_FIRST
| NVME_CMD_FUSE_SECOND
)) ||
1820 queue_90percent_full(fod
->queue
, le16_to_cpu(cqe
->sq_head
)))
1823 /* re-set the fields */
1824 fod
->fcpreq
->rspaddr
= ersp
;
1825 fod
->fcpreq
->rspdma
= fod
->rspdma
;
1828 memset(ersp
, 0, NVME_FC_SIZEOF_ZEROS_RSP
);
1829 fod
->fcpreq
->rsplen
= NVME_FC_SIZEOF_ZEROS_RSP
;
1831 ersp
->iu_len
= cpu_to_be16(sizeof(*ersp
)/sizeof(u32
));
1832 rsn
= atomic_inc_return(&fod
->queue
->rsn
);
1833 ersp
->rsn
= cpu_to_be32(rsn
);
1834 ersp
->xfrd_len
= cpu_to_be32(xfr_length
);
1835 fod
->fcpreq
->rsplen
= sizeof(*ersp
);
1838 fc_dma_sync_single_for_device(tgtport
->dev
, fod
->rspdma
,
1839 sizeof(fod
->rspiubuf
), DMA_TO_DEVICE
);
1842 static void nvmet_fc_xmt_fcp_op_done(struct nvmefc_tgt_fcp_req
*fcpreq
);
1845 nvmet_fc_abort_op(struct nvmet_fc_tgtport
*tgtport
,
1846 struct nvmet_fc_fcp_iod
*fod
)
1848 struct nvmefc_tgt_fcp_req
*fcpreq
= fod
->fcpreq
;
1850 /* data no longer needed */
1851 nvmet_fc_free_tgt_pgs(fod
);
1854 * if an ABTS was received or we issued the fcp_abort early
1855 * don't call abort routine again.
1857 /* no need to take lock - lock was taken earlier to get here */
1859 tgtport
->ops
->fcp_abort(&tgtport
->fc_target_port
, fcpreq
);
1861 nvmet_fc_free_fcp_iod(fod
->queue
, fod
);
1865 nvmet_fc_xmt_fcp_rsp(struct nvmet_fc_tgtport
*tgtport
,
1866 struct nvmet_fc_fcp_iod
*fod
)
1870 fod
->fcpreq
->op
= NVMET_FCOP_RSP
;
1871 fod
->fcpreq
->timeout
= 0;
1873 nvmet_fc_prep_fcp_rsp(tgtport
, fod
);
1875 ret
= tgtport
->ops
->fcp_op(&tgtport
->fc_target_port
, fod
->fcpreq
);
1877 nvmet_fc_abort_op(tgtport
, fod
);
1881 nvmet_fc_transfer_fcp_data(struct nvmet_fc_tgtport
*tgtport
,
1882 struct nvmet_fc_fcp_iod
*fod
, u8 op
)
1884 struct nvmefc_tgt_fcp_req
*fcpreq
= fod
->fcpreq
;
1885 unsigned long flags
;
1890 fcpreq
->offset
= fod
->offset
;
1891 fcpreq
->timeout
= NVME_FC_TGTOP_TIMEOUT_SEC
;
1893 tlen
= min_t(u32
, tgtport
->max_sg_cnt
* PAGE_SIZE
,
1894 (fod
->req
.transfer_len
- fod
->offset
));
1895 fcpreq
->transfer_length
= tlen
;
1896 fcpreq
->transferred_length
= 0;
1897 fcpreq
->fcp_error
= 0;
1900 fcpreq
->sg
= &fod
->data_sg
[fod
->offset
/ PAGE_SIZE
];
1901 fcpreq
->sg_cnt
= DIV_ROUND_UP(tlen
, PAGE_SIZE
);
1904 * If the last READDATA request: check if LLDD supports
1905 * combined xfr with response.
1907 if ((op
== NVMET_FCOP_READDATA
) &&
1908 ((fod
->offset
+ fcpreq
->transfer_length
) == fod
->req
.transfer_len
) &&
1909 (tgtport
->ops
->target_features
& NVMET_FCTGTFEAT_READDATA_RSP
)) {
1910 fcpreq
->op
= NVMET_FCOP_READDATA_RSP
;
1911 nvmet_fc_prep_fcp_rsp(tgtport
, fod
);
1914 ret
= tgtport
->ops
->fcp_op(&tgtport
->fc_target_port
, fod
->fcpreq
);
1917 * should be ok to set w/o lock as its in the thread of
1918 * execution (not an async timer routine) and doesn't
1919 * contend with any clearing action
1923 if (op
== NVMET_FCOP_WRITEDATA
) {
1924 spin_lock_irqsave(&fod
->flock
, flags
);
1925 fod
->writedataactive
= false;
1926 spin_unlock_irqrestore(&fod
->flock
, flags
);
1927 nvmet_req_complete(&fod
->req
, NVME_SC_INTERNAL
);
1928 } else /* NVMET_FCOP_READDATA or NVMET_FCOP_READDATA_RSP */ {
1929 fcpreq
->fcp_error
= ret
;
1930 fcpreq
->transferred_length
= 0;
1931 nvmet_fc_xmt_fcp_op_done(fod
->fcpreq
);
1937 __nvmet_fc_fod_op_abort(struct nvmet_fc_fcp_iod
*fod
, bool abort
)
1939 struct nvmefc_tgt_fcp_req
*fcpreq
= fod
->fcpreq
;
1940 struct nvmet_fc_tgtport
*tgtport
= fod
->tgtport
;
1942 /* if in the middle of an io and we need to tear down */
1944 if (fcpreq
->op
== NVMET_FCOP_WRITEDATA
) {
1945 nvmet_req_complete(&fod
->req
, NVME_SC_INTERNAL
);
1949 nvmet_fc_abort_op(tgtport
, fod
);
1957 * actual done handler for FCP operations when completed by the lldd
1960 nvmet_fc_fod_op_done(struct nvmet_fc_fcp_iod
*fod
)
1962 struct nvmefc_tgt_fcp_req
*fcpreq
= fod
->fcpreq
;
1963 struct nvmet_fc_tgtport
*tgtport
= fod
->tgtport
;
1964 unsigned long flags
;
1967 spin_lock_irqsave(&fod
->flock
, flags
);
1969 fod
->writedataactive
= false;
1970 spin_unlock_irqrestore(&fod
->flock
, flags
);
1972 switch (fcpreq
->op
) {
1974 case NVMET_FCOP_WRITEDATA
:
1975 if (__nvmet_fc_fod_op_abort(fod
, abort
))
1977 if (fcpreq
->fcp_error
||
1978 fcpreq
->transferred_length
!= fcpreq
->transfer_length
) {
1979 spin_lock(&fod
->flock
);
1981 spin_unlock(&fod
->flock
);
1983 nvmet_req_complete(&fod
->req
, NVME_SC_INTERNAL
);
1987 fod
->offset
+= fcpreq
->transferred_length
;
1988 if (fod
->offset
!= fod
->req
.transfer_len
) {
1989 spin_lock_irqsave(&fod
->flock
, flags
);
1990 fod
->writedataactive
= true;
1991 spin_unlock_irqrestore(&fod
->flock
, flags
);
1993 /* transfer the next chunk */
1994 nvmet_fc_transfer_fcp_data(tgtport
, fod
,
1995 NVMET_FCOP_WRITEDATA
);
1999 /* data transfer complete, resume with nvmet layer */
2000 nvmet_req_execute(&fod
->req
);
2003 case NVMET_FCOP_READDATA
:
2004 case NVMET_FCOP_READDATA_RSP
:
2005 if (__nvmet_fc_fod_op_abort(fod
, abort
))
2007 if (fcpreq
->fcp_error
||
2008 fcpreq
->transferred_length
!= fcpreq
->transfer_length
) {
2009 nvmet_fc_abort_op(tgtport
, fod
);
2015 if (fcpreq
->op
== NVMET_FCOP_READDATA_RSP
) {
2016 /* data no longer needed */
2017 nvmet_fc_free_tgt_pgs(fod
);
2018 nvmet_fc_free_fcp_iod(fod
->queue
, fod
);
2022 fod
->offset
+= fcpreq
->transferred_length
;
2023 if (fod
->offset
!= fod
->req
.transfer_len
) {
2024 /* transfer the next chunk */
2025 nvmet_fc_transfer_fcp_data(tgtport
, fod
,
2026 NVMET_FCOP_READDATA
);
2030 /* data transfer complete, send response */
2032 /* data no longer needed */
2033 nvmet_fc_free_tgt_pgs(fod
);
2035 nvmet_fc_xmt_fcp_rsp(tgtport
, fod
);
2039 case NVMET_FCOP_RSP
:
2040 if (__nvmet_fc_fod_op_abort(fod
, abort
))
2042 nvmet_fc_free_fcp_iod(fod
->queue
, fod
);
2051 nvmet_fc_fcp_rqst_op_done_work(struct work_struct
*work
)
2053 struct nvmet_fc_fcp_iod
*fod
=
2054 container_of(work
, struct nvmet_fc_fcp_iod
, done_work
);
2056 nvmet_fc_fod_op_done(fod
);
2060 nvmet_fc_xmt_fcp_op_done(struct nvmefc_tgt_fcp_req
*fcpreq
)
2062 struct nvmet_fc_fcp_iod
*fod
= fcpreq
->nvmet_fc_private
;
2063 struct nvmet_fc_tgt_queue
*queue
= fod
->queue
;
2065 if (fod
->tgtport
->ops
->target_features
& NVMET_FCTGTFEAT_OPDONE_IN_ISR
)
2066 /* context switch so completion is not in ISR context */
2067 queue_work_on(queue
->cpu
, queue
->work_q
, &fod
->done_work
);
2069 nvmet_fc_fod_op_done(fod
);
2073 * actual completion handler after execution by the nvmet layer
2076 __nvmet_fc_fcp_nvme_cmd_done(struct nvmet_fc_tgtport
*tgtport
,
2077 struct nvmet_fc_fcp_iod
*fod
, int status
)
2079 struct nvme_common_command
*sqe
= &fod
->cmdiubuf
.sqe
.common
;
2080 struct nvme_completion
*cqe
= &fod
->rspiubuf
.cqe
;
2081 unsigned long flags
;
2084 spin_lock_irqsave(&fod
->flock
, flags
);
2086 spin_unlock_irqrestore(&fod
->flock
, flags
);
2088 /* if we have a CQE, snoop the last sq_head value */
2090 fod
->queue
->sqhd
= cqe
->sq_head
;
2093 nvmet_fc_abort_op(tgtport
, fod
);
2097 /* if an error handling the cmd post initial parsing */
2099 /* fudge up a failed CQE status for our transport error */
2100 memset(cqe
, 0, sizeof(*cqe
));
2101 cqe
->sq_head
= fod
->queue
->sqhd
; /* echo last cqe sqhd */
2102 cqe
->sq_id
= cpu_to_le16(fod
->queue
->qid
);
2103 cqe
->command_id
= sqe
->command_id
;
2104 cqe
->status
= cpu_to_le16(status
);
2108 * try to push the data even if the SQE status is non-zero.
2109 * There may be a status where data still was intended to
2112 if ((fod
->io_dir
== NVMET_FCP_READ
) && (fod
->data_sg_cnt
)) {
2113 /* push the data over before sending rsp */
2114 nvmet_fc_transfer_fcp_data(tgtport
, fod
,
2115 NVMET_FCOP_READDATA
);
2119 /* writes & no data - fall thru */
2122 /* data no longer needed */
2123 nvmet_fc_free_tgt_pgs(fod
);
2125 nvmet_fc_xmt_fcp_rsp(tgtport
, fod
);
2130 nvmet_fc_fcp_nvme_cmd_done(struct nvmet_req
*nvme_req
)
2132 struct nvmet_fc_fcp_iod
*fod
= nvmet_req_to_fod(nvme_req
);
2133 struct nvmet_fc_tgtport
*tgtport
= fod
->tgtport
;
2135 __nvmet_fc_fcp_nvme_cmd_done(tgtport
, fod
, 0);
2140 * Actual processing routine for received FC-NVME LS Requests from the LLD
2143 nvmet_fc_handle_fcp_rqst(struct nvmet_fc_tgtport
*tgtport
,
2144 struct nvmet_fc_fcp_iod
*fod
)
2146 struct nvme_fc_cmd_iu
*cmdiu
= &fod
->cmdiubuf
;
2147 u32 xfrlen
= be32_to_cpu(cmdiu
->data_len
);
2151 * Fused commands are currently not supported in the linux
2154 * As such, the implementation of the FC transport does not
2155 * look at the fused commands and order delivery to the upper
2156 * layer until we have both based on csn.
2159 fod
->fcpreq
->done
= nvmet_fc_xmt_fcp_op_done
;
2161 if (cmdiu
->flags
& FCNVME_CMD_FLAGS_WRITE
) {
2162 fod
->io_dir
= NVMET_FCP_WRITE
;
2163 if (!nvme_is_write(&cmdiu
->sqe
))
2164 goto transport_error
;
2165 } else if (cmdiu
->flags
& FCNVME_CMD_FLAGS_READ
) {
2166 fod
->io_dir
= NVMET_FCP_READ
;
2167 if (nvme_is_write(&cmdiu
->sqe
))
2168 goto transport_error
;
2170 fod
->io_dir
= NVMET_FCP_NODATA
;
2172 goto transport_error
;
2175 fod
->req
.cmd
= &fod
->cmdiubuf
.sqe
;
2176 fod
->req
.rsp
= &fod
->rspiubuf
.cqe
;
2177 fod
->req
.port
= fod
->queue
->port
;
2179 /* clear any response payload */
2180 memset(&fod
->rspiubuf
, 0, sizeof(fod
->rspiubuf
));
2182 fod
->data_sg
= NULL
;
2183 fod
->data_sg_cnt
= 0;
2185 ret
= nvmet_req_init(&fod
->req
,
2186 &fod
->queue
->nvme_cq
,
2187 &fod
->queue
->nvme_sq
,
2188 &nvmet_fc_tgt_fcp_ops
);
2190 /* bad SQE content or invalid ctrl state */
2191 /* nvmet layer has already called op done to send rsp. */
2195 fod
->req
.transfer_len
= xfrlen
;
2197 /* keep a running counter of tail position */
2198 atomic_inc(&fod
->queue
->sqtail
);
2200 if (fod
->req
.transfer_len
) {
2201 ret
= nvmet_fc_alloc_tgt_pgs(fod
);
2203 nvmet_req_complete(&fod
->req
, ret
);
2207 fod
->req
.sg
= fod
->data_sg
;
2208 fod
->req
.sg_cnt
= fod
->data_sg_cnt
;
2211 if (fod
->io_dir
== NVMET_FCP_WRITE
) {
2212 /* pull the data over before invoking nvmet layer */
2213 nvmet_fc_transfer_fcp_data(tgtport
, fod
, NVMET_FCOP_WRITEDATA
);
2220 * can invoke the nvmet_layer now. If read data, cmd completion will
2223 nvmet_req_execute(&fod
->req
);
2227 nvmet_fc_abort_op(tgtport
, fod
);
2231 * Actual processing routine for received FC-NVME LS Requests from the LLD
2234 nvmet_fc_handle_fcp_rqst_work(struct work_struct
*work
)
2236 struct nvmet_fc_fcp_iod
*fod
=
2237 container_of(work
, struct nvmet_fc_fcp_iod
, work
);
2238 struct nvmet_fc_tgtport
*tgtport
= fod
->tgtport
;
2240 nvmet_fc_handle_fcp_rqst(tgtport
, fod
);
2244 * nvmet_fc_rcv_fcp_req - transport entry point called by an LLDD
2245 * upon the reception of a NVME FCP CMD IU.
2247 * Pass a FC-NVME FCP CMD IU received from the FC link to the nvmet-fc
2248 * layer for processing.
2250 * The nvmet_fc layer allocates a local job structure (struct
2251 * nvmet_fc_fcp_iod) from the queue for the io and copies the
2252 * CMD IU buffer to the job structure. As such, on a successful
2253 * completion (returns 0), the LLDD may immediately free/reuse
2254 * the CMD IU buffer passed in the call.
2256 * However, in some circumstances, due to the packetized nature of FC
2257 * and the api of the FC LLDD which may issue a hw command to send the
2258 * response, but the LLDD may not get the hw completion for that command
2259 * and upcall the nvmet_fc layer before a new command may be
2260 * asynchronously received - its possible for a command to be received
2261 * before the LLDD and nvmet_fc have recycled the job structure. It gives
2262 * the appearance of more commands received than fits in the sq.
2263 * To alleviate this scenario, a temporary queue is maintained in the
2264 * transport for pending LLDD requests waiting for a queue job structure.
2265 * In these "overrun" cases, a temporary queue element is allocated
2266 * the LLDD request and CMD iu buffer information remembered, and the
2267 * routine returns a -EOVERFLOW status. Subsequently, when a queue job
2268 * structure is freed, it is immediately reallocated for anything on the
2269 * pending request list. The LLDDs defer_rcv() callback is called,
2270 * informing the LLDD that it may reuse the CMD IU buffer, and the io
2271 * is then started normally with the transport.
2273 * The LLDD, when receiving an -EOVERFLOW completion status, is to treat
2274 * the completion as successful but must not reuse the CMD IU buffer
2275 * until the LLDD's defer_rcv() callback has been called for the
2276 * corresponding struct nvmefc_tgt_fcp_req pointer.
2278 * If there is any other condition in which an error occurs, the
2279 * transport will return a non-zero status indicating the error.
2280 * In all cases other than -EOVERFLOW, the transport has not accepted the
2281 * request and the LLDD should abort the exchange.
2283 * @target_port: pointer to the (registered) target port the FCP CMD IU
2285 * @fcpreq: pointer to a fcpreq request structure to be used to reference
2286 * the exchange corresponding to the FCP Exchange.
2287 * @cmdiubuf: pointer to the buffer containing the FCP CMD IU
2288 * @cmdiubuf_len: length, in bytes, of the received FCP CMD IU
2291 nvmet_fc_rcv_fcp_req(struct nvmet_fc_target_port
*target_port
,
2292 struct nvmefc_tgt_fcp_req
*fcpreq
,
2293 void *cmdiubuf
, u32 cmdiubuf_len
)
2295 struct nvmet_fc_tgtport
*tgtport
= targetport_to_tgtport(target_port
);
2296 struct nvme_fc_cmd_iu
*cmdiu
= cmdiubuf
;
2297 struct nvmet_fc_tgt_queue
*queue
;
2298 struct nvmet_fc_fcp_iod
*fod
;
2299 struct nvmet_fc_defer_fcp_req
*deferfcp
;
2300 unsigned long flags
;
2302 /* validate iu, so the connection id can be used to find the queue */
2303 if ((cmdiubuf_len
!= sizeof(*cmdiu
)) ||
2304 (cmdiu
->scsi_id
!= NVME_CMD_SCSI_ID
) ||
2305 (cmdiu
->fc_id
!= NVME_CMD_FC_ID
) ||
2306 (be16_to_cpu(cmdiu
->iu_len
) != (sizeof(*cmdiu
)/4)))
2309 queue
= nvmet_fc_find_target_queue(tgtport
,
2310 be64_to_cpu(cmdiu
->connection_id
));
2315 * note: reference taken by find_target_queue
2316 * After successful fod allocation, the fod will inherit the
2317 * ownership of that reference and will remove the reference
2318 * when the fod is freed.
2321 spin_lock_irqsave(&queue
->qlock
, flags
);
2323 fod
= nvmet_fc_alloc_fcp_iod(queue
);
2325 spin_unlock_irqrestore(&queue
->qlock
, flags
);
2327 fcpreq
->nvmet_fc_private
= fod
;
2328 fod
->fcpreq
= fcpreq
;
2330 memcpy(&fod
->cmdiubuf
, cmdiubuf
, cmdiubuf_len
);
2332 nvmet_fc_queue_fcp_req(tgtport
, queue
, fcpreq
);
2337 if (!tgtport
->ops
->defer_rcv
) {
2338 spin_unlock_irqrestore(&queue
->qlock
, flags
);
2339 /* release the queue lookup reference */
2340 nvmet_fc_tgt_q_put(queue
);
2344 deferfcp
= list_first_entry_or_null(&queue
->avail_defer_list
,
2345 struct nvmet_fc_defer_fcp_req
, req_list
);
2347 /* Just re-use one that was previously allocated */
2348 list_del(&deferfcp
->req_list
);
2350 spin_unlock_irqrestore(&queue
->qlock
, flags
);
2352 /* Now we need to dynamically allocate one */
2353 deferfcp
= kmalloc(sizeof(*deferfcp
), GFP_KERNEL
);
2355 /* release the queue lookup reference */
2356 nvmet_fc_tgt_q_put(queue
);
2359 spin_lock_irqsave(&queue
->qlock
, flags
);
2362 /* For now, use rspaddr / rsplen to save payload information */
2363 fcpreq
->rspaddr
= cmdiubuf
;
2364 fcpreq
->rsplen
= cmdiubuf_len
;
2365 deferfcp
->fcp_req
= fcpreq
;
2367 /* defer processing till a fod becomes available */
2368 list_add_tail(&deferfcp
->req_list
, &queue
->pending_cmd_list
);
2370 /* NOTE: the queue lookup reference is still valid */
2372 spin_unlock_irqrestore(&queue
->qlock
, flags
);
2376 EXPORT_SYMBOL_GPL(nvmet_fc_rcv_fcp_req
);
2379 * nvmet_fc_rcv_fcp_abort - transport entry point called by an LLDD
2380 * upon the reception of an ABTS for a FCP command
2382 * Notify the transport that an ABTS has been received for a FCP command
2383 * that had been given to the transport via nvmet_fc_rcv_fcp_req(). The
2384 * LLDD believes the command is still being worked on
2385 * (template_ops->fcp_req_release() has not been called).
2387 * The transport will wait for any outstanding work (an op to the LLDD,
2388 * which the lldd should complete with error due to the ABTS; or the
2389 * completion from the nvmet layer of the nvme command), then will
2390 * stop processing and call the nvmet_fc_rcv_fcp_req() callback to
2391 * return the i/o context to the LLDD. The LLDD may send the BA_ACC
2392 * to the ABTS either after return from this function (assuming any
2393 * outstanding op work has been terminated) or upon the callback being
2396 * @target_port: pointer to the (registered) target port the FCP CMD IU
2398 * @fcpreq: pointer to the fcpreq request structure that corresponds
2399 * to the exchange that received the ABTS.
2402 nvmet_fc_rcv_fcp_abort(struct nvmet_fc_target_port
*target_port
,
2403 struct nvmefc_tgt_fcp_req
*fcpreq
)
2405 struct nvmet_fc_fcp_iod
*fod
= fcpreq
->nvmet_fc_private
;
2406 struct nvmet_fc_tgt_queue
*queue
;
2407 unsigned long flags
;
2409 if (!fod
|| fod
->fcpreq
!= fcpreq
)
2410 /* job appears to have already completed, ignore abort */
2415 spin_lock_irqsave(&queue
->qlock
, flags
);
2418 * mark as abort. The abort handler, invoked upon completion
2419 * of any work, will detect the aborted status and do the
2422 spin_lock(&fod
->flock
);
2424 fod
->aborted
= true;
2425 spin_unlock(&fod
->flock
);
2427 spin_unlock_irqrestore(&queue
->qlock
, flags
);
2429 EXPORT_SYMBOL_GPL(nvmet_fc_rcv_fcp_abort
);
2432 struct nvmet_fc_traddr
{
2438 __nvme_fc_parse_u64(substring_t
*sstr
, u64
*val
)
2442 if (match_u64(sstr
, &token64
))
2450 * This routine validates and extracts the WWN's from the TRADDR string.
2451 * As kernel parsers need the 0x to determine number base, universally
2452 * build string to parse with 0x prefix before parsing name strings.
2455 nvme_fc_parse_traddr(struct nvmet_fc_traddr
*traddr
, char *buf
, size_t blen
)
2457 char name
[2 + NVME_FC_TRADDR_HEXNAMELEN
+ 1];
2458 substring_t wwn
= { name
, &name
[sizeof(name
)-1] };
2459 int nnoffset
, pnoffset
;
2461 /* validate it string one of the 2 allowed formats */
2462 if (strnlen(buf
, blen
) == NVME_FC_TRADDR_MAXLENGTH
&&
2463 !strncmp(buf
, "nn-0x", NVME_FC_TRADDR_OXNNLEN
) &&
2464 !strncmp(&buf
[NVME_FC_TRADDR_MAX_PN_OFFSET
],
2465 "pn-0x", NVME_FC_TRADDR_OXNNLEN
)) {
2466 nnoffset
= NVME_FC_TRADDR_OXNNLEN
;
2467 pnoffset
= NVME_FC_TRADDR_MAX_PN_OFFSET
+
2468 NVME_FC_TRADDR_OXNNLEN
;
2469 } else if ((strnlen(buf
, blen
) == NVME_FC_TRADDR_MINLENGTH
&&
2470 !strncmp(buf
, "nn-", NVME_FC_TRADDR_NNLEN
) &&
2471 !strncmp(&buf
[NVME_FC_TRADDR_MIN_PN_OFFSET
],
2472 "pn-", NVME_FC_TRADDR_NNLEN
))) {
2473 nnoffset
= NVME_FC_TRADDR_NNLEN
;
2474 pnoffset
= NVME_FC_TRADDR_MIN_PN_OFFSET
+ NVME_FC_TRADDR_NNLEN
;
2480 name
[2 + NVME_FC_TRADDR_HEXNAMELEN
] = 0;
2482 memcpy(&name
[2], &buf
[nnoffset
], NVME_FC_TRADDR_HEXNAMELEN
);
2483 if (__nvme_fc_parse_u64(&wwn
, &traddr
->nn
))
2486 memcpy(&name
[2], &buf
[pnoffset
], NVME_FC_TRADDR_HEXNAMELEN
);
2487 if (__nvme_fc_parse_u64(&wwn
, &traddr
->pn
))
2493 pr_warn("%s: bad traddr string\n", __func__
);
2498 nvmet_fc_add_port(struct nvmet_port
*port
)
2500 struct nvmet_fc_tgtport
*tgtport
;
2501 struct nvmet_fc_traddr traddr
= { 0L, 0L };
2502 unsigned long flags
;
2505 /* validate the address info */
2506 if ((port
->disc_addr
.trtype
!= NVMF_TRTYPE_FC
) ||
2507 (port
->disc_addr
.adrfam
!= NVMF_ADDR_FAMILY_FC
))
2510 /* map the traddr address info to a target port */
2512 ret
= nvme_fc_parse_traddr(&traddr
, port
->disc_addr
.traddr
,
2513 sizeof(port
->disc_addr
.traddr
));
2518 spin_lock_irqsave(&nvmet_fc_tgtlock
, flags
);
2519 list_for_each_entry(tgtport
, &nvmet_fc_target_list
, tgt_list
) {
2520 if ((tgtport
->fc_target_port
.node_name
== traddr
.nn
) &&
2521 (tgtport
->fc_target_port
.port_name
== traddr
.pn
)) {
2522 /* a FC port can only be 1 nvmet port id */
2523 if (!tgtport
->port
) {
2524 tgtport
->port
= port
;
2525 port
->priv
= tgtport
;
2526 nvmet_fc_tgtport_get(tgtport
);
2533 spin_unlock_irqrestore(&nvmet_fc_tgtlock
, flags
);
2538 nvmet_fc_remove_port(struct nvmet_port
*port
)
2540 struct nvmet_fc_tgtport
*tgtport
= port
->priv
;
2541 unsigned long flags
;
2542 bool matched
= false;
2544 spin_lock_irqsave(&nvmet_fc_tgtlock
, flags
);
2545 if (tgtport
->port
== port
) {
2547 tgtport
->port
= NULL
;
2549 spin_unlock_irqrestore(&nvmet_fc_tgtlock
, flags
);
2552 nvmet_fc_tgtport_put(tgtport
);
2555 static struct nvmet_fabrics_ops nvmet_fc_tgt_fcp_ops
= {
2556 .owner
= THIS_MODULE
,
2557 .type
= NVMF_TRTYPE_FC
,
2559 .add_port
= nvmet_fc_add_port
,
2560 .remove_port
= nvmet_fc_remove_port
,
2561 .queue_response
= nvmet_fc_fcp_nvme_cmd_done
,
2562 .delete_ctrl
= nvmet_fc_delete_ctrl
,
2565 static int __init
nvmet_fc_init_module(void)
2567 return nvmet_register_transport(&nvmet_fc_tgt_fcp_ops
);
2570 static void __exit
nvmet_fc_exit_module(void)
2572 /* sanity check - all lports should be removed */
2573 if (!list_empty(&nvmet_fc_target_list
))
2574 pr_warn("%s: targetport list not empty\n", __func__
);
2576 nvmet_unregister_transport(&nvmet_fc_tgt_fcp_ops
);
2578 ida_destroy(&nvmet_fc_tgtport_cnt
);
2581 module_init(nvmet_fc_init_module
);
2582 module_exit(nvmet_fc_exit_module
);
2584 MODULE_LICENSE("GPL v2");