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 256
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;
90 struct work_struct defer_work;
92 struct nvmet_fc_tgtport *tgtport;
93 struct nvmet_fc_tgt_queue *queue;
95 struct list_head fcp_list; /* tgtport->fcp_list */
98 struct nvmet_fc_tgtport {
100 struct nvmet_fc_target_port fc_target_port;
102 struct list_head tgt_list; /* nvmet_fc_target_list */
103 struct device *dev; /* dev for dma mapping */
104 struct nvmet_fc_target_template *ops;
106 struct nvmet_fc_ls_iod *iod;
108 struct list_head ls_list;
109 struct list_head ls_busylist;
110 struct list_head assoc_list;
111 struct ida assoc_cnt;
112 struct nvmet_port *port;
117 struct nvmet_fc_defer_fcp_req {
118 struct list_head req_list;
119 struct nvmefc_tgt_fcp_req *fcp_req;
122 struct nvmet_fc_tgt_queue {
134 struct nvmet_port *port;
135 struct nvmet_cq nvme_cq;
136 struct nvmet_sq nvme_sq;
137 struct nvmet_fc_tgt_assoc *assoc;
138 struct nvmet_fc_fcp_iod *fod; /* array of fcp_iods */
139 struct list_head fod_list;
140 struct list_head pending_cmd_list;
141 struct list_head avail_defer_list;
142 struct workqueue_struct *work_q;
144 } __aligned(sizeof(unsigned long long));
146 struct nvmet_fc_tgt_assoc {
149 struct nvmet_fc_tgtport *tgtport;
150 struct list_head a_list;
151 struct nvmet_fc_tgt_queue *queues[NVMET_NR_QUEUES + 1];
153 struct work_struct del_work;
158 nvmet_fc_iodnum(struct nvmet_fc_ls_iod *iodptr)
160 return (iodptr - iodptr->tgtport->iod);
164 nvmet_fc_fodnum(struct nvmet_fc_fcp_iod *fodptr)
166 return (fodptr - fodptr->queue->fod);
171 * Association and Connection IDs:
173 * Association ID will have random number in upper 6 bytes and zero
176 * Connection IDs will be Association ID with QID or'd in lower 2 bytes
178 * note: Association ID = Connection ID for queue 0
180 #define BYTES_FOR_QID sizeof(u16)
181 #define BYTES_FOR_QID_SHIFT (BYTES_FOR_QID * 8)
182 #define NVMET_FC_QUEUEID_MASK ((u64)((1 << BYTES_FOR_QID_SHIFT) - 1))
185 nvmet_fc_makeconnid(struct nvmet_fc_tgt_assoc *assoc, u16 qid)
187 return (assoc->association_id | qid);
191 nvmet_fc_getassociationid(u64 connectionid)
193 return connectionid & ~NVMET_FC_QUEUEID_MASK;
197 nvmet_fc_getqueueid(u64 connectionid)
199 return (u16)(connectionid & NVMET_FC_QUEUEID_MASK);
202 static inline struct nvmet_fc_tgtport *
203 targetport_to_tgtport(struct nvmet_fc_target_port *targetport)
205 return container_of(targetport, struct nvmet_fc_tgtport,
209 static inline struct nvmet_fc_fcp_iod *
210 nvmet_req_to_fod(struct nvmet_req *nvme_req)
212 return container_of(nvme_req, struct nvmet_fc_fcp_iod, req);
216 /* *************************** Globals **************************** */
219 static DEFINE_SPINLOCK(nvmet_fc_tgtlock);
221 static LIST_HEAD(nvmet_fc_target_list);
222 static DEFINE_IDA(nvmet_fc_tgtport_cnt);
225 static void nvmet_fc_handle_ls_rqst_work(struct work_struct *work);
226 static void nvmet_fc_handle_fcp_rqst_work(struct work_struct *work);
227 static void nvmet_fc_fcp_rqst_op_done_work(struct work_struct *work);
228 static void nvmet_fc_fcp_rqst_op_defer_work(struct work_struct *work);
229 static void nvmet_fc_tgt_a_put(struct nvmet_fc_tgt_assoc *assoc);
230 static int nvmet_fc_tgt_a_get(struct nvmet_fc_tgt_assoc *assoc);
231 static void nvmet_fc_tgt_q_put(struct nvmet_fc_tgt_queue *queue);
232 static int nvmet_fc_tgt_q_get(struct nvmet_fc_tgt_queue *queue);
233 static void nvmet_fc_tgtport_put(struct nvmet_fc_tgtport *tgtport);
234 static int nvmet_fc_tgtport_get(struct nvmet_fc_tgtport *tgtport);
235 static void nvmet_fc_handle_fcp_rqst(struct nvmet_fc_tgtport *tgtport,
236 struct nvmet_fc_fcp_iod *fod);
237 static void nvmet_fc_delete_target_assoc(struct nvmet_fc_tgt_assoc *assoc);
240 /* *********************** FC-NVME DMA Handling **************************** */
243 * The fcloop device passes in a NULL device pointer. Real LLD's will
244 * pass in a valid device pointer. If NULL is passed to the dma mapping
245 * routines, depending on the platform, it may or may not succeed, and
249 * Wrapper all the dma routines and check the dev pointer.
251 * If simple mappings (return just a dma address, we'll noop them,
252 * returning a dma address of 0.
254 * On more complex mappings (dma_map_sg), a pseudo routine fills
255 * in the scatter list, setting all dma addresses to 0.
258 static inline dma_addr_t
259 fc_dma_map_single(struct device *dev, void *ptr, size_t size,
260 enum dma_data_direction dir)
262 return dev ? dma_map_single(dev, ptr, size, dir) : (dma_addr_t)0L;
266 fc_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
268 return dev ? dma_mapping_error(dev, dma_addr) : 0;
272 fc_dma_unmap_single(struct device *dev, dma_addr_t addr, size_t size,
273 enum dma_data_direction dir)
276 dma_unmap_single(dev, addr, size, dir);
280 fc_dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr, size_t size,
281 enum dma_data_direction dir)
284 dma_sync_single_for_cpu(dev, addr, size, dir);
288 fc_dma_sync_single_for_device(struct device *dev, dma_addr_t addr, size_t size,
289 enum dma_data_direction dir)
292 dma_sync_single_for_device(dev, addr, size, dir);
295 /* pseudo dma_map_sg call */
297 fc_map_sg(struct scatterlist *sg, int nents)
299 struct scatterlist *s;
302 WARN_ON(nents == 0 || sg[0].length == 0);
304 for_each_sg(sg, s, nents, i) {
306 #ifdef CONFIG_NEED_SG_DMA_LENGTH
307 s->dma_length = s->length;
314 fc_dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
315 enum dma_data_direction dir)
317 return dev ? dma_map_sg(dev, sg, nents, dir) : fc_map_sg(sg, nents);
321 fc_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
322 enum dma_data_direction dir)
325 dma_unmap_sg(dev, sg, nents, dir);
329 /* *********************** FC-NVME Port Management ************************ */
333 nvmet_fc_alloc_ls_iodlist(struct nvmet_fc_tgtport *tgtport)
335 struct nvmet_fc_ls_iod *iod;
338 iod = kcalloc(NVMET_LS_CTX_COUNT, sizeof(struct nvmet_fc_ls_iod),
345 for (i = 0; i < NVMET_LS_CTX_COUNT; iod++, i++) {
346 INIT_WORK(&iod->work, nvmet_fc_handle_ls_rqst_work);
347 iod->tgtport = tgtport;
348 list_add_tail(&iod->ls_list, &tgtport->ls_list);
350 iod->rqstbuf = kcalloc(2, NVME_FC_MAX_LS_BUFFER_SIZE,
355 iod->rspbuf = iod->rqstbuf + NVME_FC_MAX_LS_BUFFER_SIZE;
357 iod->rspdma = fc_dma_map_single(tgtport->dev, iod->rspbuf,
358 NVME_FC_MAX_LS_BUFFER_SIZE,
360 if (fc_dma_mapping_error(tgtport->dev, iod->rspdma))
368 list_del(&iod->ls_list);
369 for (iod--, i--; i >= 0; iod--, i--) {
370 fc_dma_unmap_single(tgtport->dev, iod->rspdma,
371 NVME_FC_MAX_LS_BUFFER_SIZE, DMA_TO_DEVICE);
373 list_del(&iod->ls_list);
382 nvmet_fc_free_ls_iodlist(struct nvmet_fc_tgtport *tgtport)
384 struct nvmet_fc_ls_iod *iod = tgtport->iod;
387 for (i = 0; i < NVMET_LS_CTX_COUNT; iod++, i++) {
388 fc_dma_unmap_single(tgtport->dev,
389 iod->rspdma, NVME_FC_MAX_LS_BUFFER_SIZE,
392 list_del(&iod->ls_list);
397 static struct nvmet_fc_ls_iod *
398 nvmet_fc_alloc_ls_iod(struct nvmet_fc_tgtport *tgtport)
400 struct nvmet_fc_ls_iod *iod;
403 spin_lock_irqsave(&tgtport->lock, flags);
404 iod = list_first_entry_or_null(&tgtport->ls_list,
405 struct nvmet_fc_ls_iod, ls_list);
407 list_move_tail(&iod->ls_list, &tgtport->ls_busylist);
408 spin_unlock_irqrestore(&tgtport->lock, flags);
414 nvmet_fc_free_ls_iod(struct nvmet_fc_tgtport *tgtport,
415 struct nvmet_fc_ls_iod *iod)
419 spin_lock_irqsave(&tgtport->lock, flags);
420 list_move(&iod->ls_list, &tgtport->ls_list);
421 spin_unlock_irqrestore(&tgtport->lock, flags);
425 nvmet_fc_prep_fcp_iodlist(struct nvmet_fc_tgtport *tgtport,
426 struct nvmet_fc_tgt_queue *queue)
428 struct nvmet_fc_fcp_iod *fod = queue->fod;
431 for (i = 0; i < queue->sqsize; fod++, i++) {
432 INIT_WORK(&fod->work, nvmet_fc_handle_fcp_rqst_work);
433 INIT_WORK(&fod->done_work, nvmet_fc_fcp_rqst_op_done_work);
434 INIT_WORK(&fod->defer_work, nvmet_fc_fcp_rqst_op_defer_work);
435 fod->tgtport = tgtport;
439 fod->aborted = false;
441 list_add_tail(&fod->fcp_list, &queue->fod_list);
442 spin_lock_init(&fod->flock);
444 fod->rspdma = fc_dma_map_single(tgtport->dev, &fod->rspiubuf,
445 sizeof(fod->rspiubuf), DMA_TO_DEVICE);
446 if (fc_dma_mapping_error(tgtport->dev, fod->rspdma)) {
447 list_del(&fod->fcp_list);
448 for (fod--, i--; i >= 0; fod--, i--) {
449 fc_dma_unmap_single(tgtport->dev, fod->rspdma,
450 sizeof(fod->rspiubuf),
453 list_del(&fod->fcp_list);
462 nvmet_fc_destroy_fcp_iodlist(struct nvmet_fc_tgtport *tgtport,
463 struct nvmet_fc_tgt_queue *queue)
465 struct nvmet_fc_fcp_iod *fod = queue->fod;
468 for (i = 0; i < queue->sqsize; fod++, i++) {
470 fc_dma_unmap_single(tgtport->dev, fod->rspdma,
471 sizeof(fod->rspiubuf), DMA_TO_DEVICE);
475 static struct nvmet_fc_fcp_iod *
476 nvmet_fc_alloc_fcp_iod(struct nvmet_fc_tgt_queue *queue)
478 struct nvmet_fc_fcp_iod *fod;
480 lockdep_assert_held(&queue->qlock);
482 fod = list_first_entry_or_null(&queue->fod_list,
483 struct nvmet_fc_fcp_iod, fcp_list);
485 list_del(&fod->fcp_list);
488 * no queue reference is taken, as it was taken by the
489 * queue lookup just prior to the allocation. The iod
490 * will "inherit" that reference.
498 nvmet_fc_queue_fcp_req(struct nvmet_fc_tgtport *tgtport,
499 struct nvmet_fc_tgt_queue *queue,
500 struct nvmefc_tgt_fcp_req *fcpreq)
502 struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
505 * put all admin cmds on hw queue id 0. All io commands go to
506 * the respective hw queue based on a modulo basis
508 fcpreq->hwqid = queue->qid ?
509 ((queue->qid - 1) % tgtport->ops->max_hw_queues) : 0;
511 if (tgtport->ops->target_features & NVMET_FCTGTFEAT_CMD_IN_ISR)
512 queue_work_on(queue->cpu, queue->work_q, &fod->work);
514 nvmet_fc_handle_fcp_rqst(tgtport, fod);
518 nvmet_fc_fcp_rqst_op_defer_work(struct work_struct *work)
520 struct nvmet_fc_fcp_iod *fod =
521 container_of(work, struct nvmet_fc_fcp_iod, defer_work);
523 /* Submit deferred IO for processing */
524 nvmet_fc_queue_fcp_req(fod->tgtport, fod->queue, fod->fcpreq);
529 nvmet_fc_free_fcp_iod(struct nvmet_fc_tgt_queue *queue,
530 struct nvmet_fc_fcp_iod *fod)
532 struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
533 struct nvmet_fc_tgtport *tgtport = fod->tgtport;
534 struct nvmet_fc_defer_fcp_req *deferfcp;
537 fc_dma_sync_single_for_cpu(tgtport->dev, fod->rspdma,
538 sizeof(fod->rspiubuf), DMA_TO_DEVICE);
540 fcpreq->nvmet_fc_private = NULL;
544 fod->aborted = false;
545 fod->writedataactive = false;
548 tgtport->ops->fcp_req_release(&tgtport->fc_target_port, fcpreq);
550 /* release the queue lookup reference on the completed IO */
551 nvmet_fc_tgt_q_put(queue);
553 spin_lock_irqsave(&queue->qlock, flags);
554 deferfcp = list_first_entry_or_null(&queue->pending_cmd_list,
555 struct nvmet_fc_defer_fcp_req, req_list);
557 list_add_tail(&fod->fcp_list, &fod->queue->fod_list);
558 spin_unlock_irqrestore(&queue->qlock, flags);
562 /* Re-use the fod for the next pending cmd that was deferred */
563 list_del(&deferfcp->req_list);
565 fcpreq = deferfcp->fcp_req;
567 /* deferfcp can be reused for another IO at a later date */
568 list_add_tail(&deferfcp->req_list, &queue->avail_defer_list);
570 spin_unlock_irqrestore(&queue->qlock, flags);
572 /* Save NVME CMD IO in fod */
573 memcpy(&fod->cmdiubuf, fcpreq->rspaddr, fcpreq->rsplen);
575 /* Setup new fcpreq to be processed */
576 fcpreq->rspaddr = NULL;
578 fcpreq->nvmet_fc_private = fod;
579 fod->fcpreq = fcpreq;
582 /* inform LLDD IO is now being processed */
583 tgtport->ops->defer_rcv(&tgtport->fc_target_port, fcpreq);
586 * Leave the queue lookup get reference taken when
587 * fod was originally allocated.
590 queue_work(queue->work_q, &fod->defer_work);
594 nvmet_fc_queue_to_cpu(struct nvmet_fc_tgtport *tgtport, int qid)
598 if (tgtport->ops->max_hw_queues == 1)
599 return WORK_CPU_UNBOUND;
601 /* Simple cpu selection based on qid modulo active cpu count */
602 idx = !qid ? 0 : (qid - 1) % num_active_cpus();
604 /* find the n'th active cpu */
605 for (cpu = 0, cnt = 0; ; ) {
606 if (cpu_active(cpu)) {
611 cpu = (cpu + 1) % num_possible_cpus();
617 static struct nvmet_fc_tgt_queue *
618 nvmet_fc_alloc_target_queue(struct nvmet_fc_tgt_assoc *assoc,
621 struct nvmet_fc_tgt_queue *queue;
625 if (qid > NVMET_NR_QUEUES)
628 queue = kzalloc((sizeof(*queue) +
629 (sizeof(struct nvmet_fc_fcp_iod) * sqsize)),
634 if (!nvmet_fc_tgt_a_get(assoc))
637 queue->work_q = alloc_workqueue("ntfc%d.%d.%d", 0, 0,
638 assoc->tgtport->fc_target_port.port_num,
643 queue->fod = (struct nvmet_fc_fcp_iod *)&queue[1];
645 queue->sqsize = sqsize;
646 queue->assoc = assoc;
647 queue->port = assoc->tgtport->port;
648 queue->cpu = nvmet_fc_queue_to_cpu(assoc->tgtport, qid);
649 INIT_LIST_HEAD(&queue->fod_list);
650 INIT_LIST_HEAD(&queue->avail_defer_list);
651 INIT_LIST_HEAD(&queue->pending_cmd_list);
652 atomic_set(&queue->connected, 0);
653 atomic_set(&queue->sqtail, 0);
654 atomic_set(&queue->rsn, 1);
655 atomic_set(&queue->zrspcnt, 0);
656 spin_lock_init(&queue->qlock);
657 kref_init(&queue->ref);
659 nvmet_fc_prep_fcp_iodlist(assoc->tgtport, queue);
661 ret = nvmet_sq_init(&queue->nvme_sq);
663 goto out_fail_iodlist;
665 WARN_ON(assoc->queues[qid]);
666 spin_lock_irqsave(&assoc->tgtport->lock, flags);
667 assoc->queues[qid] = queue;
668 spin_unlock_irqrestore(&assoc->tgtport->lock, flags);
673 nvmet_fc_destroy_fcp_iodlist(assoc->tgtport, queue);
674 destroy_workqueue(queue->work_q);
676 nvmet_fc_tgt_a_put(assoc);
684 nvmet_fc_tgt_queue_free(struct kref *ref)
686 struct nvmet_fc_tgt_queue *queue =
687 container_of(ref, struct nvmet_fc_tgt_queue, ref);
690 spin_lock_irqsave(&queue->assoc->tgtport->lock, flags);
691 queue->assoc->queues[queue->qid] = NULL;
692 spin_unlock_irqrestore(&queue->assoc->tgtport->lock, flags);
694 nvmet_fc_destroy_fcp_iodlist(queue->assoc->tgtport, queue);
696 nvmet_fc_tgt_a_put(queue->assoc);
698 destroy_workqueue(queue->work_q);
704 nvmet_fc_tgt_q_put(struct nvmet_fc_tgt_queue *queue)
706 kref_put(&queue->ref, nvmet_fc_tgt_queue_free);
710 nvmet_fc_tgt_q_get(struct nvmet_fc_tgt_queue *queue)
712 return kref_get_unless_zero(&queue->ref);
717 nvmet_fc_delete_target_queue(struct nvmet_fc_tgt_queue *queue)
719 struct nvmet_fc_tgtport *tgtport = queue->assoc->tgtport;
720 struct nvmet_fc_fcp_iod *fod = queue->fod;
721 struct nvmet_fc_defer_fcp_req *deferfcp, *tempptr;
723 int i, writedataactive;
726 disconnect = atomic_xchg(&queue->connected, 0);
728 spin_lock_irqsave(&queue->qlock, flags);
729 /* about outstanding io's */
730 for (i = 0; i < queue->sqsize; fod++, i++) {
732 spin_lock(&fod->flock);
734 writedataactive = fod->writedataactive;
735 spin_unlock(&fod->flock);
737 * only call lldd abort routine if waiting for
738 * writedata. other outstanding ops should finish
741 if (writedataactive) {
742 spin_lock(&fod->flock);
744 spin_unlock(&fod->flock);
745 tgtport->ops->fcp_abort(
746 &tgtport->fc_target_port, fod->fcpreq);
751 /* Cleanup defer'ed IOs in queue */
752 list_for_each_entry_safe(deferfcp, tempptr, &queue->avail_defer_list,
754 list_del(&deferfcp->req_list);
759 deferfcp = list_first_entry_or_null(&queue->pending_cmd_list,
760 struct nvmet_fc_defer_fcp_req, req_list);
764 list_del(&deferfcp->req_list);
765 spin_unlock_irqrestore(&queue->qlock, flags);
767 tgtport->ops->defer_rcv(&tgtport->fc_target_port,
770 tgtport->ops->fcp_abort(&tgtport->fc_target_port,
773 tgtport->ops->fcp_req_release(&tgtport->fc_target_port,
776 /* release the queue lookup reference */
777 nvmet_fc_tgt_q_put(queue);
781 spin_lock_irqsave(&queue->qlock, flags);
783 spin_unlock_irqrestore(&queue->qlock, flags);
785 flush_workqueue(queue->work_q);
788 nvmet_sq_destroy(&queue->nvme_sq);
790 nvmet_fc_tgt_q_put(queue);
793 static struct nvmet_fc_tgt_queue *
794 nvmet_fc_find_target_queue(struct nvmet_fc_tgtport *tgtport,
797 struct nvmet_fc_tgt_assoc *assoc;
798 struct nvmet_fc_tgt_queue *queue;
799 u64 association_id = nvmet_fc_getassociationid(connection_id);
800 u16 qid = nvmet_fc_getqueueid(connection_id);
803 if (qid > NVMET_NR_QUEUES)
806 spin_lock_irqsave(&tgtport->lock, flags);
807 list_for_each_entry(assoc, &tgtport->assoc_list, a_list) {
808 if (association_id == assoc->association_id) {
809 queue = assoc->queues[qid];
811 (!atomic_read(&queue->connected) ||
812 !nvmet_fc_tgt_q_get(queue)))
814 spin_unlock_irqrestore(&tgtport->lock, flags);
818 spin_unlock_irqrestore(&tgtport->lock, flags);
823 nvmet_fc_delete_assoc(struct work_struct *work)
825 struct nvmet_fc_tgt_assoc *assoc =
826 container_of(work, struct nvmet_fc_tgt_assoc, del_work);
828 nvmet_fc_delete_target_assoc(assoc);
829 nvmet_fc_tgt_a_put(assoc);
832 static struct nvmet_fc_tgt_assoc *
833 nvmet_fc_alloc_target_assoc(struct nvmet_fc_tgtport *tgtport)
835 struct nvmet_fc_tgt_assoc *assoc, *tmpassoc;
839 bool needrandom = true;
841 assoc = kzalloc(sizeof(*assoc), GFP_KERNEL);
845 idx = ida_simple_get(&tgtport->assoc_cnt, 0, 0, GFP_KERNEL);
849 if (!nvmet_fc_tgtport_get(tgtport))
852 assoc->tgtport = tgtport;
854 INIT_LIST_HEAD(&assoc->a_list);
855 kref_init(&assoc->ref);
856 INIT_WORK(&assoc->del_work, nvmet_fc_delete_assoc);
859 get_random_bytes(&ran, sizeof(ran) - BYTES_FOR_QID);
860 ran = ran << BYTES_FOR_QID_SHIFT;
862 spin_lock_irqsave(&tgtport->lock, flags);
864 list_for_each_entry(tmpassoc, &tgtport->assoc_list, a_list)
865 if (ran == tmpassoc->association_id) {
870 assoc->association_id = ran;
871 list_add_tail(&assoc->a_list, &tgtport->assoc_list);
873 spin_unlock_irqrestore(&tgtport->lock, flags);
879 ida_simple_remove(&tgtport->assoc_cnt, idx);
886 nvmet_fc_target_assoc_free(struct kref *ref)
888 struct nvmet_fc_tgt_assoc *assoc =
889 container_of(ref, struct nvmet_fc_tgt_assoc, ref);
890 struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
893 spin_lock_irqsave(&tgtport->lock, flags);
894 list_del(&assoc->a_list);
895 spin_unlock_irqrestore(&tgtport->lock, flags);
896 ida_simple_remove(&tgtport->assoc_cnt, assoc->a_id);
898 nvmet_fc_tgtport_put(tgtport);
902 nvmet_fc_tgt_a_put(struct nvmet_fc_tgt_assoc *assoc)
904 kref_put(&assoc->ref, nvmet_fc_target_assoc_free);
908 nvmet_fc_tgt_a_get(struct nvmet_fc_tgt_assoc *assoc)
910 return kref_get_unless_zero(&assoc->ref);
914 nvmet_fc_delete_target_assoc(struct nvmet_fc_tgt_assoc *assoc)
916 struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
917 struct nvmet_fc_tgt_queue *queue;
921 spin_lock_irqsave(&tgtport->lock, flags);
922 for (i = NVMET_NR_QUEUES; i >= 0; i--) {
923 queue = assoc->queues[i];
925 if (!nvmet_fc_tgt_q_get(queue))
927 spin_unlock_irqrestore(&tgtport->lock, flags);
928 nvmet_fc_delete_target_queue(queue);
929 nvmet_fc_tgt_q_put(queue);
930 spin_lock_irqsave(&tgtport->lock, flags);
933 spin_unlock_irqrestore(&tgtport->lock, flags);
935 nvmet_fc_tgt_a_put(assoc);
938 static struct nvmet_fc_tgt_assoc *
939 nvmet_fc_find_target_assoc(struct nvmet_fc_tgtport *tgtport,
942 struct nvmet_fc_tgt_assoc *assoc;
943 struct nvmet_fc_tgt_assoc *ret = NULL;
946 spin_lock_irqsave(&tgtport->lock, flags);
947 list_for_each_entry(assoc, &tgtport->assoc_list, a_list) {
948 if (association_id == assoc->association_id) {
950 nvmet_fc_tgt_a_get(assoc);
954 spin_unlock_irqrestore(&tgtport->lock, flags);
961 * nvme_fc_register_targetport - transport entry point called by an
962 * LLDD to register the existence of a local
963 * NVME subystem FC port.
964 * @pinfo: pointer to information about the port to be registered
965 * @template: LLDD entrypoints and operational parameters for the port
966 * @dev: physical hardware device node port corresponds to. Will be
967 * used for DMA mappings
968 * @portptr: pointer to a local port pointer. Upon success, the routine
969 * will allocate a nvme_fc_local_port structure and place its
970 * address in the local port pointer. Upon failure, local port
971 * pointer will be set to NULL.
974 * a completion status. Must be 0 upon success; a negative errno
975 * (ex: -ENXIO) upon failure.
978 nvmet_fc_register_targetport(struct nvmet_fc_port_info *pinfo,
979 struct nvmet_fc_target_template *template,
981 struct nvmet_fc_target_port **portptr)
983 struct nvmet_fc_tgtport *newrec;
987 if (!template->xmt_ls_rsp || !template->fcp_op ||
988 !template->fcp_abort ||
989 !template->fcp_req_release || !template->targetport_delete ||
990 !template->max_hw_queues || !template->max_sgl_segments ||
991 !template->max_dif_sgl_segments || !template->dma_boundary) {
993 goto out_regtgt_failed;
996 newrec = kzalloc((sizeof(*newrec) + template->target_priv_sz),
1000 goto out_regtgt_failed;
1003 idx = ida_simple_get(&nvmet_fc_tgtport_cnt, 0, 0, GFP_KERNEL);
1006 goto out_fail_kfree;
1009 if (!get_device(dev) && dev) {
1014 newrec->fc_target_port.node_name = pinfo->node_name;
1015 newrec->fc_target_port.port_name = pinfo->port_name;
1016 newrec->fc_target_port.private = &newrec[1];
1017 newrec->fc_target_port.port_id = pinfo->port_id;
1018 newrec->fc_target_port.port_num = idx;
1019 INIT_LIST_HEAD(&newrec->tgt_list);
1021 newrec->ops = template;
1022 spin_lock_init(&newrec->lock);
1023 INIT_LIST_HEAD(&newrec->ls_list);
1024 INIT_LIST_HEAD(&newrec->ls_busylist);
1025 INIT_LIST_HEAD(&newrec->assoc_list);
1026 kref_init(&newrec->ref);
1027 ida_init(&newrec->assoc_cnt);
1028 newrec->max_sg_cnt = min_t(u32, NVMET_FC_MAX_XFR_SGENTS,
1029 template->max_sgl_segments);
1031 ret = nvmet_fc_alloc_ls_iodlist(newrec);
1034 goto out_free_newrec;
1037 spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1038 list_add_tail(&newrec->tgt_list, &nvmet_fc_target_list);
1039 spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1041 *portptr = &newrec->fc_target_port;
1047 ida_simple_remove(&nvmet_fc_tgtport_cnt, idx);
1054 EXPORT_SYMBOL_GPL(nvmet_fc_register_targetport);
1058 nvmet_fc_free_tgtport(struct kref *ref)
1060 struct nvmet_fc_tgtport *tgtport =
1061 container_of(ref, struct nvmet_fc_tgtport, ref);
1062 struct device *dev = tgtport->dev;
1063 unsigned long flags;
1065 spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1066 list_del(&tgtport->tgt_list);
1067 spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1069 nvmet_fc_free_ls_iodlist(tgtport);
1071 /* let the LLDD know we've finished tearing it down */
1072 tgtport->ops->targetport_delete(&tgtport->fc_target_port);
1074 ida_simple_remove(&nvmet_fc_tgtport_cnt,
1075 tgtport->fc_target_port.port_num);
1077 ida_destroy(&tgtport->assoc_cnt);
1085 nvmet_fc_tgtport_put(struct nvmet_fc_tgtport *tgtport)
1087 kref_put(&tgtport->ref, nvmet_fc_free_tgtport);
1091 nvmet_fc_tgtport_get(struct nvmet_fc_tgtport *tgtport)
1093 return kref_get_unless_zero(&tgtport->ref);
1097 __nvmet_fc_free_assocs(struct nvmet_fc_tgtport *tgtport)
1099 struct nvmet_fc_tgt_assoc *assoc, *next;
1100 unsigned long flags;
1102 spin_lock_irqsave(&tgtport->lock, flags);
1103 list_for_each_entry_safe(assoc, next,
1104 &tgtport->assoc_list, a_list) {
1105 if (!nvmet_fc_tgt_a_get(assoc))
1107 spin_unlock_irqrestore(&tgtport->lock, flags);
1108 nvmet_fc_delete_target_assoc(assoc);
1109 nvmet_fc_tgt_a_put(assoc);
1110 spin_lock_irqsave(&tgtport->lock, flags);
1112 spin_unlock_irqrestore(&tgtport->lock, flags);
1116 * nvmet layer has called to terminate an association
1119 nvmet_fc_delete_ctrl(struct nvmet_ctrl *ctrl)
1121 struct nvmet_fc_tgtport *tgtport, *next;
1122 struct nvmet_fc_tgt_assoc *assoc;
1123 struct nvmet_fc_tgt_queue *queue;
1124 unsigned long flags;
1125 bool found_ctrl = false;
1127 /* this is a bit ugly, but don't want to make locks layered */
1128 spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1129 list_for_each_entry_safe(tgtport, next, &nvmet_fc_target_list,
1131 if (!nvmet_fc_tgtport_get(tgtport))
1133 spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1135 spin_lock_irqsave(&tgtport->lock, flags);
1136 list_for_each_entry(assoc, &tgtport->assoc_list, a_list) {
1137 queue = assoc->queues[0];
1138 if (queue && queue->nvme_sq.ctrl == ctrl) {
1139 if (nvmet_fc_tgt_a_get(assoc))
1144 spin_unlock_irqrestore(&tgtport->lock, flags);
1146 nvmet_fc_tgtport_put(tgtport);
1149 schedule_work(&assoc->del_work);
1153 spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1155 spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1159 * nvme_fc_unregister_targetport - transport entry point called by an
1160 * LLDD to deregister/remove a previously
1161 * registered a local NVME subsystem FC port.
1162 * @tgtport: pointer to the (registered) target port that is to be
1166 * a completion status. Must be 0 upon success; a negative errno
1167 * (ex: -ENXIO) upon failure.
1170 nvmet_fc_unregister_targetport(struct nvmet_fc_target_port *target_port)
1172 struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
1174 /* terminate any outstanding associations */
1175 __nvmet_fc_free_assocs(tgtport);
1177 nvmet_fc_tgtport_put(tgtport);
1181 EXPORT_SYMBOL_GPL(nvmet_fc_unregister_targetport);
1184 /* *********************** FC-NVME LS Handling **************************** */
1188 nvmet_fc_format_rsp_hdr(void *buf, u8 ls_cmd, __be32 desc_len, u8 rqst_ls_cmd)
1190 struct fcnvme_ls_acc_hdr *acc = buf;
1192 acc->w0.ls_cmd = ls_cmd;
1193 acc->desc_list_len = desc_len;
1194 acc->rqst.desc_tag = cpu_to_be32(FCNVME_LSDESC_RQST);
1195 acc->rqst.desc_len =
1196 fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_rqst));
1197 acc->rqst.w0.ls_cmd = rqst_ls_cmd;
1201 nvmet_fc_format_rjt(void *buf, u16 buflen, u8 ls_cmd,
1202 u8 reason, u8 explanation, u8 vendor)
1204 struct fcnvme_ls_rjt *rjt = buf;
1206 nvmet_fc_format_rsp_hdr(buf, FCNVME_LSDESC_RQST,
1207 fcnvme_lsdesc_len(sizeof(struct fcnvme_ls_rjt)),
1209 rjt->rjt.desc_tag = cpu_to_be32(FCNVME_LSDESC_RJT);
1210 rjt->rjt.desc_len = fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_rjt));
1211 rjt->rjt.reason_code = reason;
1212 rjt->rjt.reason_explanation = explanation;
1213 rjt->rjt.vendor = vendor;
1215 return sizeof(struct fcnvme_ls_rjt);
1218 /* Validation Error indexes into the string table below */
1221 VERR_CR_ASSOC_LEN = 1,
1222 VERR_CR_ASSOC_RQST_LEN = 2,
1223 VERR_CR_ASSOC_CMD = 3,
1224 VERR_CR_ASSOC_CMD_LEN = 4,
1225 VERR_ERSP_RATIO = 5,
1226 VERR_ASSOC_ALLOC_FAIL = 6,
1227 VERR_QUEUE_ALLOC_FAIL = 7,
1228 VERR_CR_CONN_LEN = 8,
1229 VERR_CR_CONN_RQST_LEN = 9,
1231 VERR_ASSOC_ID_LEN = 11,
1234 VERR_CONN_ID_LEN = 14,
1236 VERR_CR_CONN_CMD = 16,
1237 VERR_CR_CONN_CMD_LEN = 17,
1238 VERR_DISCONN_LEN = 18,
1239 VERR_DISCONN_RQST_LEN = 19,
1240 VERR_DISCONN_CMD = 20,
1241 VERR_DISCONN_CMD_LEN = 21,
1242 VERR_DISCONN_SCOPE = 22,
1244 VERR_RS_RQST_LEN = 24,
1246 VERR_RS_CMD_LEN = 26,
1251 static char *validation_errors[] = {
1253 "Bad CR_ASSOC Length",
1254 "Bad CR_ASSOC Rqst Length",
1256 "Bad CR_ASSOC Cmd Length",
1258 "Association Allocation Failed",
1259 "Queue Allocation Failed",
1260 "Bad CR_CONN Length",
1261 "Bad CR_CONN Rqst Length",
1262 "Not Association ID",
1263 "Bad Association ID Length",
1265 "Not Connection ID",
1266 "Bad Connection ID Length",
1269 "Bad CR_CONN Cmd Length",
1270 "Bad DISCONN Length",
1271 "Bad DISCONN Rqst Length",
1273 "Bad DISCONN Cmd Length",
1274 "Bad Disconnect Scope",
1276 "Bad RS Rqst Length",
1278 "Bad RS Cmd Length",
1280 "Bad RS Relative Offset",
1284 nvmet_fc_ls_create_association(struct nvmet_fc_tgtport *tgtport,
1285 struct nvmet_fc_ls_iod *iod)
1287 struct fcnvme_ls_cr_assoc_rqst *rqst =
1288 (struct fcnvme_ls_cr_assoc_rqst *)iod->rqstbuf;
1289 struct fcnvme_ls_cr_assoc_acc *acc =
1290 (struct fcnvme_ls_cr_assoc_acc *)iod->rspbuf;
1291 struct nvmet_fc_tgt_queue *queue;
1294 memset(acc, 0, sizeof(*acc));
1297 * FC-NVME spec changes. There are initiators sending different
1298 * lengths as padding sizes for Create Association Cmd descriptor
1300 * Accept anything of "minimum" length. Assume format per 1.15
1301 * spec (with HOSTID reduced to 16 bytes), ignore how long the
1302 * trailing pad length is.
1304 if (iod->rqstdatalen < FCNVME_LSDESC_CRA_RQST_MINLEN)
1305 ret = VERR_CR_ASSOC_LEN;
1306 else if (be32_to_cpu(rqst->desc_list_len) <
1307 FCNVME_LSDESC_CRA_RQST_MIN_LISTLEN)
1308 ret = VERR_CR_ASSOC_RQST_LEN;
1309 else if (rqst->assoc_cmd.desc_tag !=
1310 cpu_to_be32(FCNVME_LSDESC_CREATE_ASSOC_CMD))
1311 ret = VERR_CR_ASSOC_CMD;
1312 else if (be32_to_cpu(rqst->assoc_cmd.desc_len) <
1313 FCNVME_LSDESC_CRA_CMD_DESC_MIN_DESCLEN)
1314 ret = VERR_CR_ASSOC_CMD_LEN;
1315 else if (!rqst->assoc_cmd.ersp_ratio ||
1316 (be16_to_cpu(rqst->assoc_cmd.ersp_ratio) >=
1317 be16_to_cpu(rqst->assoc_cmd.sqsize)))
1318 ret = VERR_ERSP_RATIO;
1321 /* new association w/ admin queue */
1322 iod->assoc = nvmet_fc_alloc_target_assoc(tgtport);
1324 ret = VERR_ASSOC_ALLOC_FAIL;
1326 queue = nvmet_fc_alloc_target_queue(iod->assoc, 0,
1327 be16_to_cpu(rqst->assoc_cmd.sqsize));
1329 ret = VERR_QUEUE_ALLOC_FAIL;
1334 dev_err(tgtport->dev,
1335 "Create Association LS failed: %s\n",
1336 validation_errors[ret]);
1337 iod->lsreq->rsplen = nvmet_fc_format_rjt(acc,
1338 NVME_FC_MAX_LS_BUFFER_SIZE, rqst->w0.ls_cmd,
1339 FCNVME_RJT_RC_LOGIC,
1340 FCNVME_RJT_EXP_NONE, 0);
1344 queue->ersp_ratio = be16_to_cpu(rqst->assoc_cmd.ersp_ratio);
1345 atomic_set(&queue->connected, 1);
1346 queue->sqhd = 0; /* best place to init value */
1348 /* format a response */
1350 iod->lsreq->rsplen = sizeof(*acc);
1352 nvmet_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1354 sizeof(struct fcnvme_ls_cr_assoc_acc)),
1355 FCNVME_LS_CREATE_ASSOCIATION);
1356 acc->associd.desc_tag = cpu_to_be32(FCNVME_LSDESC_ASSOC_ID);
1357 acc->associd.desc_len =
1359 sizeof(struct fcnvme_lsdesc_assoc_id));
1360 acc->associd.association_id =
1361 cpu_to_be64(nvmet_fc_makeconnid(iod->assoc, 0));
1362 acc->connectid.desc_tag = cpu_to_be32(FCNVME_LSDESC_CONN_ID);
1363 acc->connectid.desc_len =
1365 sizeof(struct fcnvme_lsdesc_conn_id));
1366 acc->connectid.connection_id = acc->associd.association_id;
1370 nvmet_fc_ls_create_connection(struct nvmet_fc_tgtport *tgtport,
1371 struct nvmet_fc_ls_iod *iod)
1373 struct fcnvme_ls_cr_conn_rqst *rqst =
1374 (struct fcnvme_ls_cr_conn_rqst *)iod->rqstbuf;
1375 struct fcnvme_ls_cr_conn_acc *acc =
1376 (struct fcnvme_ls_cr_conn_acc *)iod->rspbuf;
1377 struct nvmet_fc_tgt_queue *queue;
1380 memset(acc, 0, sizeof(*acc));
1382 if (iod->rqstdatalen < sizeof(struct fcnvme_ls_cr_conn_rqst))
1383 ret = VERR_CR_CONN_LEN;
1384 else if (rqst->desc_list_len !=
1386 sizeof(struct fcnvme_ls_cr_conn_rqst)))
1387 ret = VERR_CR_CONN_RQST_LEN;
1388 else if (rqst->associd.desc_tag != cpu_to_be32(FCNVME_LSDESC_ASSOC_ID))
1389 ret = VERR_ASSOC_ID;
1390 else if (rqst->associd.desc_len !=
1392 sizeof(struct fcnvme_lsdesc_assoc_id)))
1393 ret = VERR_ASSOC_ID_LEN;
1394 else if (rqst->connect_cmd.desc_tag !=
1395 cpu_to_be32(FCNVME_LSDESC_CREATE_CONN_CMD))
1396 ret = VERR_CR_CONN_CMD;
1397 else if (rqst->connect_cmd.desc_len !=
1399 sizeof(struct fcnvme_lsdesc_cr_conn_cmd)))
1400 ret = VERR_CR_CONN_CMD_LEN;
1401 else if (!rqst->connect_cmd.ersp_ratio ||
1402 (be16_to_cpu(rqst->connect_cmd.ersp_ratio) >=
1403 be16_to_cpu(rqst->connect_cmd.sqsize)))
1404 ret = VERR_ERSP_RATIO;
1408 iod->assoc = nvmet_fc_find_target_assoc(tgtport,
1409 be64_to_cpu(rqst->associd.association_id));
1411 ret = VERR_NO_ASSOC;
1413 queue = nvmet_fc_alloc_target_queue(iod->assoc,
1414 be16_to_cpu(rqst->connect_cmd.qid),
1415 be16_to_cpu(rqst->connect_cmd.sqsize));
1417 ret = VERR_QUEUE_ALLOC_FAIL;
1419 /* release get taken in nvmet_fc_find_target_assoc */
1420 nvmet_fc_tgt_a_put(iod->assoc);
1425 dev_err(tgtport->dev,
1426 "Create Connection LS failed: %s\n",
1427 validation_errors[ret]);
1428 iod->lsreq->rsplen = nvmet_fc_format_rjt(acc,
1429 NVME_FC_MAX_LS_BUFFER_SIZE, rqst->w0.ls_cmd,
1430 (ret == VERR_NO_ASSOC) ?
1431 FCNVME_RJT_RC_INV_ASSOC :
1432 FCNVME_RJT_RC_LOGIC,
1433 FCNVME_RJT_EXP_NONE, 0);
1437 queue->ersp_ratio = be16_to_cpu(rqst->connect_cmd.ersp_ratio);
1438 atomic_set(&queue->connected, 1);
1439 queue->sqhd = 0; /* best place to init value */
1441 /* format a response */
1443 iod->lsreq->rsplen = sizeof(*acc);
1445 nvmet_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1446 fcnvme_lsdesc_len(sizeof(struct fcnvme_ls_cr_conn_acc)),
1447 FCNVME_LS_CREATE_CONNECTION);
1448 acc->connectid.desc_tag = cpu_to_be32(FCNVME_LSDESC_CONN_ID);
1449 acc->connectid.desc_len =
1451 sizeof(struct fcnvme_lsdesc_conn_id));
1452 acc->connectid.connection_id =
1453 cpu_to_be64(nvmet_fc_makeconnid(iod->assoc,
1454 be16_to_cpu(rqst->connect_cmd.qid)));
1458 nvmet_fc_ls_disconnect(struct nvmet_fc_tgtport *tgtport,
1459 struct nvmet_fc_ls_iod *iod)
1461 struct fcnvme_ls_disconnect_rqst *rqst =
1462 (struct fcnvme_ls_disconnect_rqst *)iod->rqstbuf;
1463 struct fcnvme_ls_disconnect_acc *acc =
1464 (struct fcnvme_ls_disconnect_acc *)iod->rspbuf;
1465 struct nvmet_fc_tgt_queue *queue = NULL;
1466 struct nvmet_fc_tgt_assoc *assoc;
1468 bool del_assoc = false;
1470 memset(acc, 0, sizeof(*acc));
1472 if (iod->rqstdatalen < sizeof(struct fcnvme_ls_disconnect_rqst))
1473 ret = VERR_DISCONN_LEN;
1474 else if (rqst->desc_list_len !=
1476 sizeof(struct fcnvme_ls_disconnect_rqst)))
1477 ret = VERR_DISCONN_RQST_LEN;
1478 else if (rqst->associd.desc_tag != cpu_to_be32(FCNVME_LSDESC_ASSOC_ID))
1479 ret = VERR_ASSOC_ID;
1480 else if (rqst->associd.desc_len !=
1482 sizeof(struct fcnvme_lsdesc_assoc_id)))
1483 ret = VERR_ASSOC_ID_LEN;
1484 else if (rqst->discon_cmd.desc_tag !=
1485 cpu_to_be32(FCNVME_LSDESC_DISCONN_CMD))
1486 ret = VERR_DISCONN_CMD;
1487 else if (rqst->discon_cmd.desc_len !=
1489 sizeof(struct fcnvme_lsdesc_disconn_cmd)))
1490 ret = VERR_DISCONN_CMD_LEN;
1491 else if ((rqst->discon_cmd.scope != FCNVME_DISCONN_ASSOCIATION) &&
1492 (rqst->discon_cmd.scope != FCNVME_DISCONN_CONNECTION))
1493 ret = VERR_DISCONN_SCOPE;
1495 /* match an active association */
1496 assoc = nvmet_fc_find_target_assoc(tgtport,
1497 be64_to_cpu(rqst->associd.association_id));
1500 if (rqst->discon_cmd.scope ==
1501 FCNVME_DISCONN_CONNECTION) {
1502 queue = nvmet_fc_find_target_queue(tgtport,
1504 rqst->discon_cmd.id));
1506 nvmet_fc_tgt_a_put(assoc);
1511 ret = VERR_NO_ASSOC;
1515 dev_err(tgtport->dev,
1516 "Disconnect LS failed: %s\n",
1517 validation_errors[ret]);
1518 iod->lsreq->rsplen = nvmet_fc_format_rjt(acc,
1519 NVME_FC_MAX_LS_BUFFER_SIZE, rqst->w0.ls_cmd,
1520 (ret == VERR_NO_ASSOC) ?
1521 FCNVME_RJT_RC_INV_ASSOC :
1522 (ret == VERR_NO_CONN) ?
1523 FCNVME_RJT_RC_INV_CONN :
1524 FCNVME_RJT_RC_LOGIC,
1525 FCNVME_RJT_EXP_NONE, 0);
1529 /* format a response */
1531 iod->lsreq->rsplen = sizeof(*acc);
1533 nvmet_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1535 sizeof(struct fcnvme_ls_disconnect_acc)),
1536 FCNVME_LS_DISCONNECT);
1539 /* are we to delete a Connection ID (queue) */
1541 int qid = queue->qid;
1543 nvmet_fc_delete_target_queue(queue);
1545 /* release the get taken by find_target_queue */
1546 nvmet_fc_tgt_q_put(queue);
1548 /* tear association down if io queue terminated */
1553 /* release get taken in nvmet_fc_find_target_assoc */
1554 nvmet_fc_tgt_a_put(iod->assoc);
1557 nvmet_fc_delete_target_assoc(iod->assoc);
1561 /* *********************** NVME Ctrl Routines **************************** */
1564 static void nvmet_fc_fcp_nvme_cmd_done(struct nvmet_req *nvme_req);
1566 static const struct nvmet_fabrics_ops nvmet_fc_tgt_fcp_ops;
1569 nvmet_fc_xmt_ls_rsp_done(struct nvmefc_tgt_ls_req *lsreq)
1571 struct nvmet_fc_ls_iod *iod = lsreq->nvmet_fc_private;
1572 struct nvmet_fc_tgtport *tgtport = iod->tgtport;
1574 fc_dma_sync_single_for_cpu(tgtport->dev, iod->rspdma,
1575 NVME_FC_MAX_LS_BUFFER_SIZE, DMA_TO_DEVICE);
1576 nvmet_fc_free_ls_iod(tgtport, iod);
1577 nvmet_fc_tgtport_put(tgtport);
1581 nvmet_fc_xmt_ls_rsp(struct nvmet_fc_tgtport *tgtport,
1582 struct nvmet_fc_ls_iod *iod)
1586 fc_dma_sync_single_for_device(tgtport->dev, iod->rspdma,
1587 NVME_FC_MAX_LS_BUFFER_SIZE, DMA_TO_DEVICE);
1589 ret = tgtport->ops->xmt_ls_rsp(&tgtport->fc_target_port, iod->lsreq);
1591 nvmet_fc_xmt_ls_rsp_done(iod->lsreq);
1595 * Actual processing routine for received FC-NVME LS Requests from the LLD
1598 nvmet_fc_handle_ls_rqst(struct nvmet_fc_tgtport *tgtport,
1599 struct nvmet_fc_ls_iod *iod)
1601 struct fcnvme_ls_rqst_w0 *w0 =
1602 (struct fcnvme_ls_rqst_w0 *)iod->rqstbuf;
1604 iod->lsreq->nvmet_fc_private = iod;
1605 iod->lsreq->rspbuf = iod->rspbuf;
1606 iod->lsreq->rspdma = iod->rspdma;
1607 iod->lsreq->done = nvmet_fc_xmt_ls_rsp_done;
1608 /* Be preventative. handlers will later set to valid length */
1609 iod->lsreq->rsplen = 0;
1615 * parse request input, execute the request, and format the
1618 switch (w0->ls_cmd) {
1619 case FCNVME_LS_CREATE_ASSOCIATION:
1620 /* Creates Association and initial Admin Queue/Connection */
1621 nvmet_fc_ls_create_association(tgtport, iod);
1623 case FCNVME_LS_CREATE_CONNECTION:
1624 /* Creates an IO Queue/Connection */
1625 nvmet_fc_ls_create_connection(tgtport, iod);
1627 case FCNVME_LS_DISCONNECT:
1628 /* Terminate a Queue/Connection or the Association */
1629 nvmet_fc_ls_disconnect(tgtport, iod);
1632 iod->lsreq->rsplen = nvmet_fc_format_rjt(iod->rspbuf,
1633 NVME_FC_MAX_LS_BUFFER_SIZE, w0->ls_cmd,
1634 FCNVME_RJT_RC_INVAL, FCNVME_RJT_EXP_NONE, 0);
1637 nvmet_fc_xmt_ls_rsp(tgtport, iod);
1641 * Actual processing routine for received FC-NVME LS Requests from the LLD
1644 nvmet_fc_handle_ls_rqst_work(struct work_struct *work)
1646 struct nvmet_fc_ls_iod *iod =
1647 container_of(work, struct nvmet_fc_ls_iod, work);
1648 struct nvmet_fc_tgtport *tgtport = iod->tgtport;
1650 nvmet_fc_handle_ls_rqst(tgtport, iod);
1655 * nvmet_fc_rcv_ls_req - transport entry point called by an LLDD
1656 * upon the reception of a NVME LS request.
1658 * The nvmet-fc layer will copy payload to an internal structure for
1659 * processing. As such, upon completion of the routine, the LLDD may
1660 * immediately free/reuse the LS request buffer passed in the call.
1662 * If this routine returns error, the LLDD should abort the exchange.
1664 * @tgtport: pointer to the (registered) target port the LS was
1666 * @lsreq: pointer to a lsreq request structure to be used to reference
1667 * the exchange corresponding to the LS.
1668 * @lsreqbuf: pointer to the buffer containing the LS Request
1669 * @lsreqbuf_len: length, in bytes, of the received LS request
1672 nvmet_fc_rcv_ls_req(struct nvmet_fc_target_port *target_port,
1673 struct nvmefc_tgt_ls_req *lsreq,
1674 void *lsreqbuf, u32 lsreqbuf_len)
1676 struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
1677 struct nvmet_fc_ls_iod *iod;
1679 if (lsreqbuf_len > NVME_FC_MAX_LS_BUFFER_SIZE)
1682 if (!nvmet_fc_tgtport_get(tgtport))
1685 iod = nvmet_fc_alloc_ls_iod(tgtport);
1687 nvmet_fc_tgtport_put(tgtport);
1693 memcpy(iod->rqstbuf, lsreqbuf, lsreqbuf_len);
1694 iod->rqstdatalen = lsreqbuf_len;
1696 schedule_work(&iod->work);
1700 EXPORT_SYMBOL_GPL(nvmet_fc_rcv_ls_req);
1704 * **********************
1705 * Start of FCP handling
1706 * **********************
1710 nvmet_fc_alloc_tgt_pgs(struct nvmet_fc_fcp_iod *fod)
1712 struct scatterlist *sg;
1715 sg = sgl_alloc(fod->req.transfer_len, GFP_KERNEL, &nent);
1720 fod->data_sg_cnt = nent;
1721 fod->data_sg_cnt = fc_dma_map_sg(fod->tgtport->dev, sg, nent,
1722 ((fod->io_dir == NVMET_FCP_WRITE) ?
1723 DMA_FROM_DEVICE : DMA_TO_DEVICE));
1724 /* note: write from initiator perspective */
1729 return NVME_SC_INTERNAL;
1733 nvmet_fc_free_tgt_pgs(struct nvmet_fc_fcp_iod *fod)
1735 if (!fod->data_sg || !fod->data_sg_cnt)
1738 fc_dma_unmap_sg(fod->tgtport->dev, fod->data_sg, fod->data_sg_cnt,
1739 ((fod->io_dir == NVMET_FCP_WRITE) ?
1740 DMA_FROM_DEVICE : DMA_TO_DEVICE));
1741 sgl_free(fod->data_sg);
1742 fod->data_sg = NULL;
1743 fod->data_sg_cnt = 0;
1748 queue_90percent_full(struct nvmet_fc_tgt_queue *q, u32 sqhd)
1752 /* egad, this is ugly. And sqtail is just a best guess */
1753 sqtail = atomic_read(&q->sqtail) % q->sqsize;
1755 used = (sqtail < sqhd) ? (sqtail + q->sqsize - sqhd) : (sqtail - sqhd);
1756 return ((used * 10) >= (((u32)(q->sqsize - 1) * 9)));
1761 * May be a NVMET_FCOP_RSP or NVMET_FCOP_READDATA_RSP op
1764 nvmet_fc_prep_fcp_rsp(struct nvmet_fc_tgtport *tgtport,
1765 struct nvmet_fc_fcp_iod *fod)
1767 struct nvme_fc_ersp_iu *ersp = &fod->rspiubuf;
1768 struct nvme_common_command *sqe = &fod->cmdiubuf.sqe.common;
1769 struct nvme_completion *cqe = &ersp->cqe;
1770 u32 *cqewd = (u32 *)cqe;
1771 bool send_ersp = false;
1772 u32 rsn, rspcnt, xfr_length;
1774 if (fod->fcpreq->op == NVMET_FCOP_READDATA_RSP)
1775 xfr_length = fod->req.transfer_len;
1777 xfr_length = fod->offset;
1780 * check to see if we can send a 0's rsp.
1781 * Note: to send a 0's response, the NVME-FC host transport will
1782 * recreate the CQE. The host transport knows: sq id, SQHD (last
1783 * seen in an ersp), and command_id. Thus it will create a
1784 * zero-filled CQE with those known fields filled in. Transport
1785 * must send an ersp for any condition where the cqe won't match
1788 * Here are the FC-NVME mandated cases where we must send an ersp:
1789 * every N responses, where N=ersp_ratio
1790 * force fabric commands to send ersp's (not in FC-NVME but good
1792 * normal cmds: any time status is non-zero, or status is zero
1793 * but words 0 or 1 are non-zero.
1794 * the SQ is 90% or more full
1795 * the cmd is a fused command
1796 * transferred data length not equal to cmd iu length
1798 rspcnt = atomic_inc_return(&fod->queue->zrspcnt);
1799 if (!(rspcnt % fod->queue->ersp_ratio) ||
1800 sqe->opcode == nvme_fabrics_command ||
1801 xfr_length != fod->req.transfer_len ||
1802 (le16_to_cpu(cqe->status) & 0xFFFE) || cqewd[0] || cqewd[1] ||
1803 (sqe->flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND)) ||
1804 queue_90percent_full(fod->queue, le16_to_cpu(cqe->sq_head)))
1807 /* re-set the fields */
1808 fod->fcpreq->rspaddr = ersp;
1809 fod->fcpreq->rspdma = fod->rspdma;
1812 memset(ersp, 0, NVME_FC_SIZEOF_ZEROS_RSP);
1813 fod->fcpreq->rsplen = NVME_FC_SIZEOF_ZEROS_RSP;
1815 ersp->iu_len = cpu_to_be16(sizeof(*ersp)/sizeof(u32));
1816 rsn = atomic_inc_return(&fod->queue->rsn);
1817 ersp->rsn = cpu_to_be32(rsn);
1818 ersp->xfrd_len = cpu_to_be32(xfr_length);
1819 fod->fcpreq->rsplen = sizeof(*ersp);
1822 fc_dma_sync_single_for_device(tgtport->dev, fod->rspdma,
1823 sizeof(fod->rspiubuf), DMA_TO_DEVICE);
1826 static void nvmet_fc_xmt_fcp_op_done(struct nvmefc_tgt_fcp_req *fcpreq);
1829 nvmet_fc_abort_op(struct nvmet_fc_tgtport *tgtport,
1830 struct nvmet_fc_fcp_iod *fod)
1832 struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
1834 /* data no longer needed */
1835 nvmet_fc_free_tgt_pgs(fod);
1838 * if an ABTS was received or we issued the fcp_abort early
1839 * don't call abort routine again.
1841 /* no need to take lock - lock was taken earlier to get here */
1843 tgtport->ops->fcp_abort(&tgtport->fc_target_port, fcpreq);
1845 nvmet_fc_free_fcp_iod(fod->queue, fod);
1849 nvmet_fc_xmt_fcp_rsp(struct nvmet_fc_tgtport *tgtport,
1850 struct nvmet_fc_fcp_iod *fod)
1854 fod->fcpreq->op = NVMET_FCOP_RSP;
1855 fod->fcpreq->timeout = 0;
1857 nvmet_fc_prep_fcp_rsp(tgtport, fod);
1859 ret = tgtport->ops->fcp_op(&tgtport->fc_target_port, fod->fcpreq);
1861 nvmet_fc_abort_op(tgtport, fod);
1865 nvmet_fc_transfer_fcp_data(struct nvmet_fc_tgtport *tgtport,
1866 struct nvmet_fc_fcp_iod *fod, u8 op)
1868 struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
1869 unsigned long flags;
1874 fcpreq->offset = fod->offset;
1875 fcpreq->timeout = NVME_FC_TGTOP_TIMEOUT_SEC;
1877 tlen = min_t(u32, tgtport->max_sg_cnt * PAGE_SIZE,
1878 (fod->req.transfer_len - fod->offset));
1879 fcpreq->transfer_length = tlen;
1880 fcpreq->transferred_length = 0;
1881 fcpreq->fcp_error = 0;
1884 fcpreq->sg = &fod->data_sg[fod->offset / PAGE_SIZE];
1885 fcpreq->sg_cnt = DIV_ROUND_UP(tlen, PAGE_SIZE);
1888 * If the last READDATA request: check if LLDD supports
1889 * combined xfr with response.
1891 if ((op == NVMET_FCOP_READDATA) &&
1892 ((fod->offset + fcpreq->transfer_length) == fod->req.transfer_len) &&
1893 (tgtport->ops->target_features & NVMET_FCTGTFEAT_READDATA_RSP)) {
1894 fcpreq->op = NVMET_FCOP_READDATA_RSP;
1895 nvmet_fc_prep_fcp_rsp(tgtport, fod);
1898 ret = tgtport->ops->fcp_op(&tgtport->fc_target_port, fod->fcpreq);
1901 * should be ok to set w/o lock as its in the thread of
1902 * execution (not an async timer routine) and doesn't
1903 * contend with any clearing action
1907 if (op == NVMET_FCOP_WRITEDATA) {
1908 spin_lock_irqsave(&fod->flock, flags);
1909 fod->writedataactive = false;
1910 spin_unlock_irqrestore(&fod->flock, flags);
1911 nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
1912 } else /* NVMET_FCOP_READDATA or NVMET_FCOP_READDATA_RSP */ {
1913 fcpreq->fcp_error = ret;
1914 fcpreq->transferred_length = 0;
1915 nvmet_fc_xmt_fcp_op_done(fod->fcpreq);
1921 __nvmet_fc_fod_op_abort(struct nvmet_fc_fcp_iod *fod, bool abort)
1923 struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
1924 struct nvmet_fc_tgtport *tgtport = fod->tgtport;
1926 /* if in the middle of an io and we need to tear down */
1928 if (fcpreq->op == NVMET_FCOP_WRITEDATA) {
1929 nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
1933 nvmet_fc_abort_op(tgtport, fod);
1941 * actual done handler for FCP operations when completed by the lldd
1944 nvmet_fc_fod_op_done(struct nvmet_fc_fcp_iod *fod)
1946 struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
1947 struct nvmet_fc_tgtport *tgtport = fod->tgtport;
1948 unsigned long flags;
1951 spin_lock_irqsave(&fod->flock, flags);
1953 fod->writedataactive = false;
1954 spin_unlock_irqrestore(&fod->flock, flags);
1956 switch (fcpreq->op) {
1958 case NVMET_FCOP_WRITEDATA:
1959 if (__nvmet_fc_fod_op_abort(fod, abort))
1961 if (fcpreq->fcp_error ||
1962 fcpreq->transferred_length != fcpreq->transfer_length) {
1963 spin_lock(&fod->flock);
1965 spin_unlock(&fod->flock);
1967 nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
1971 fod->offset += fcpreq->transferred_length;
1972 if (fod->offset != fod->req.transfer_len) {
1973 spin_lock_irqsave(&fod->flock, flags);
1974 fod->writedataactive = true;
1975 spin_unlock_irqrestore(&fod->flock, flags);
1977 /* transfer the next chunk */
1978 nvmet_fc_transfer_fcp_data(tgtport, fod,
1979 NVMET_FCOP_WRITEDATA);
1983 /* data transfer complete, resume with nvmet layer */
1984 nvmet_req_execute(&fod->req);
1987 case NVMET_FCOP_READDATA:
1988 case NVMET_FCOP_READDATA_RSP:
1989 if (__nvmet_fc_fod_op_abort(fod, abort))
1991 if (fcpreq->fcp_error ||
1992 fcpreq->transferred_length != fcpreq->transfer_length) {
1993 nvmet_fc_abort_op(tgtport, fod);
1999 if (fcpreq->op == NVMET_FCOP_READDATA_RSP) {
2000 /* data no longer needed */
2001 nvmet_fc_free_tgt_pgs(fod);
2002 nvmet_fc_free_fcp_iod(fod->queue, fod);
2006 fod->offset += fcpreq->transferred_length;
2007 if (fod->offset != fod->req.transfer_len) {
2008 /* transfer the next chunk */
2009 nvmet_fc_transfer_fcp_data(tgtport, fod,
2010 NVMET_FCOP_READDATA);
2014 /* data transfer complete, send response */
2016 /* data no longer needed */
2017 nvmet_fc_free_tgt_pgs(fod);
2019 nvmet_fc_xmt_fcp_rsp(tgtport, fod);
2023 case NVMET_FCOP_RSP:
2024 if (__nvmet_fc_fod_op_abort(fod, abort))
2026 nvmet_fc_free_fcp_iod(fod->queue, fod);
2035 nvmet_fc_fcp_rqst_op_done_work(struct work_struct *work)
2037 struct nvmet_fc_fcp_iod *fod =
2038 container_of(work, struct nvmet_fc_fcp_iod, done_work);
2040 nvmet_fc_fod_op_done(fod);
2044 nvmet_fc_xmt_fcp_op_done(struct nvmefc_tgt_fcp_req *fcpreq)
2046 struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
2047 struct nvmet_fc_tgt_queue *queue = fod->queue;
2049 if (fod->tgtport->ops->target_features & NVMET_FCTGTFEAT_OPDONE_IN_ISR)
2050 /* context switch so completion is not in ISR context */
2051 queue_work_on(queue->cpu, queue->work_q, &fod->done_work);
2053 nvmet_fc_fod_op_done(fod);
2057 * actual completion handler after execution by the nvmet layer
2060 __nvmet_fc_fcp_nvme_cmd_done(struct nvmet_fc_tgtport *tgtport,
2061 struct nvmet_fc_fcp_iod *fod, int status)
2063 struct nvme_common_command *sqe = &fod->cmdiubuf.sqe.common;
2064 struct nvme_completion *cqe = &fod->rspiubuf.cqe;
2065 unsigned long flags;
2068 spin_lock_irqsave(&fod->flock, flags);
2070 spin_unlock_irqrestore(&fod->flock, flags);
2072 /* if we have a CQE, snoop the last sq_head value */
2074 fod->queue->sqhd = cqe->sq_head;
2077 nvmet_fc_abort_op(tgtport, fod);
2081 /* if an error handling the cmd post initial parsing */
2083 /* fudge up a failed CQE status for our transport error */
2084 memset(cqe, 0, sizeof(*cqe));
2085 cqe->sq_head = fod->queue->sqhd; /* echo last cqe sqhd */
2086 cqe->sq_id = cpu_to_le16(fod->queue->qid);
2087 cqe->command_id = sqe->command_id;
2088 cqe->status = cpu_to_le16(status);
2092 * try to push the data even if the SQE status is non-zero.
2093 * There may be a status where data still was intended to
2096 if ((fod->io_dir == NVMET_FCP_READ) && (fod->data_sg_cnt)) {
2097 /* push the data over before sending rsp */
2098 nvmet_fc_transfer_fcp_data(tgtport, fod,
2099 NVMET_FCOP_READDATA);
2103 /* writes & no data - fall thru */
2106 /* data no longer needed */
2107 nvmet_fc_free_tgt_pgs(fod);
2109 nvmet_fc_xmt_fcp_rsp(tgtport, fod);
2114 nvmet_fc_fcp_nvme_cmd_done(struct nvmet_req *nvme_req)
2116 struct nvmet_fc_fcp_iod *fod = nvmet_req_to_fod(nvme_req);
2117 struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2119 __nvmet_fc_fcp_nvme_cmd_done(tgtport, fod, 0);
2124 * Actual processing routine for received FC-NVME LS Requests from the LLD
2127 nvmet_fc_handle_fcp_rqst(struct nvmet_fc_tgtport *tgtport,
2128 struct nvmet_fc_fcp_iod *fod)
2130 struct nvme_fc_cmd_iu *cmdiu = &fod->cmdiubuf;
2131 u32 xfrlen = be32_to_cpu(cmdiu->data_len);
2135 * Fused commands are currently not supported in the linux
2138 * As such, the implementation of the FC transport does not
2139 * look at the fused commands and order delivery to the upper
2140 * layer until we have both based on csn.
2143 fod->fcpreq->done = nvmet_fc_xmt_fcp_op_done;
2145 if (cmdiu->flags & FCNVME_CMD_FLAGS_WRITE) {
2146 fod->io_dir = NVMET_FCP_WRITE;
2147 if (!nvme_is_write(&cmdiu->sqe))
2148 goto transport_error;
2149 } else if (cmdiu->flags & FCNVME_CMD_FLAGS_READ) {
2150 fod->io_dir = NVMET_FCP_READ;
2151 if (nvme_is_write(&cmdiu->sqe))
2152 goto transport_error;
2154 fod->io_dir = NVMET_FCP_NODATA;
2156 goto transport_error;
2159 fod->req.cmd = &fod->cmdiubuf.sqe;
2160 fod->req.rsp = &fod->rspiubuf.cqe;
2161 fod->req.port = fod->queue->port;
2163 /* clear any response payload */
2164 memset(&fod->rspiubuf, 0, sizeof(fod->rspiubuf));
2166 fod->data_sg = NULL;
2167 fod->data_sg_cnt = 0;
2169 ret = nvmet_req_init(&fod->req,
2170 &fod->queue->nvme_cq,
2171 &fod->queue->nvme_sq,
2172 &nvmet_fc_tgt_fcp_ops);
2174 /* bad SQE content or invalid ctrl state */
2175 /* nvmet layer has already called op done to send rsp. */
2179 fod->req.transfer_len = xfrlen;
2181 /* keep a running counter of tail position */
2182 atomic_inc(&fod->queue->sqtail);
2184 if (fod->req.transfer_len) {
2185 ret = nvmet_fc_alloc_tgt_pgs(fod);
2187 nvmet_req_complete(&fod->req, ret);
2191 fod->req.sg = fod->data_sg;
2192 fod->req.sg_cnt = fod->data_sg_cnt;
2195 if (fod->io_dir == NVMET_FCP_WRITE) {
2196 /* pull the data over before invoking nvmet layer */
2197 nvmet_fc_transfer_fcp_data(tgtport, fod, NVMET_FCOP_WRITEDATA);
2204 * can invoke the nvmet_layer now. If read data, cmd completion will
2207 nvmet_req_execute(&fod->req);
2211 nvmet_fc_abort_op(tgtport, fod);
2215 * Actual processing routine for received FC-NVME LS Requests from the LLD
2218 nvmet_fc_handle_fcp_rqst_work(struct work_struct *work)
2220 struct nvmet_fc_fcp_iod *fod =
2221 container_of(work, struct nvmet_fc_fcp_iod, work);
2222 struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2224 nvmet_fc_handle_fcp_rqst(tgtport, fod);
2228 * nvmet_fc_rcv_fcp_req - transport entry point called by an LLDD
2229 * upon the reception of a NVME FCP CMD IU.
2231 * Pass a FC-NVME FCP CMD IU received from the FC link to the nvmet-fc
2232 * layer for processing.
2234 * The nvmet_fc layer allocates a local job structure (struct
2235 * nvmet_fc_fcp_iod) from the queue for the io and copies the
2236 * CMD IU buffer to the job structure. As such, on a successful
2237 * completion (returns 0), the LLDD may immediately free/reuse
2238 * the CMD IU buffer passed in the call.
2240 * However, in some circumstances, due to the packetized nature of FC
2241 * and the api of the FC LLDD which may issue a hw command to send the
2242 * response, but the LLDD may not get the hw completion for that command
2243 * and upcall the nvmet_fc layer before a new command may be
2244 * asynchronously received - its possible for a command to be received
2245 * before the LLDD and nvmet_fc have recycled the job structure. It gives
2246 * the appearance of more commands received than fits in the sq.
2247 * To alleviate this scenario, a temporary queue is maintained in the
2248 * transport for pending LLDD requests waiting for a queue job structure.
2249 * In these "overrun" cases, a temporary queue element is allocated
2250 * the LLDD request and CMD iu buffer information remembered, and the
2251 * routine returns a -EOVERFLOW status. Subsequently, when a queue job
2252 * structure is freed, it is immediately reallocated for anything on the
2253 * pending request list. The LLDDs defer_rcv() callback is called,
2254 * informing the LLDD that it may reuse the CMD IU buffer, and the io
2255 * is then started normally with the transport.
2257 * The LLDD, when receiving an -EOVERFLOW completion status, is to treat
2258 * the completion as successful but must not reuse the CMD IU buffer
2259 * until the LLDD's defer_rcv() callback has been called for the
2260 * corresponding struct nvmefc_tgt_fcp_req pointer.
2262 * If there is any other condition in which an error occurs, the
2263 * transport will return a non-zero status indicating the error.
2264 * In all cases other than -EOVERFLOW, the transport has not accepted the
2265 * request and the LLDD should abort the exchange.
2267 * @target_port: pointer to the (registered) target port the FCP CMD IU
2269 * @fcpreq: pointer to a fcpreq request structure to be used to reference
2270 * the exchange corresponding to the FCP Exchange.
2271 * @cmdiubuf: pointer to the buffer containing the FCP CMD IU
2272 * @cmdiubuf_len: length, in bytes, of the received FCP CMD IU
2275 nvmet_fc_rcv_fcp_req(struct nvmet_fc_target_port *target_port,
2276 struct nvmefc_tgt_fcp_req *fcpreq,
2277 void *cmdiubuf, u32 cmdiubuf_len)
2279 struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
2280 struct nvme_fc_cmd_iu *cmdiu = cmdiubuf;
2281 struct nvmet_fc_tgt_queue *queue;
2282 struct nvmet_fc_fcp_iod *fod;
2283 struct nvmet_fc_defer_fcp_req *deferfcp;
2284 unsigned long flags;
2286 /* validate iu, so the connection id can be used to find the queue */
2287 if ((cmdiubuf_len != sizeof(*cmdiu)) ||
2288 (cmdiu->scsi_id != NVME_CMD_SCSI_ID) ||
2289 (cmdiu->fc_id != NVME_CMD_FC_ID) ||
2290 (be16_to_cpu(cmdiu->iu_len) != (sizeof(*cmdiu)/4)))
2293 queue = nvmet_fc_find_target_queue(tgtport,
2294 be64_to_cpu(cmdiu->connection_id));
2299 * note: reference taken by find_target_queue
2300 * After successful fod allocation, the fod will inherit the
2301 * ownership of that reference and will remove the reference
2302 * when the fod is freed.
2305 spin_lock_irqsave(&queue->qlock, flags);
2307 fod = nvmet_fc_alloc_fcp_iod(queue);
2309 spin_unlock_irqrestore(&queue->qlock, flags);
2311 fcpreq->nvmet_fc_private = fod;
2312 fod->fcpreq = fcpreq;
2314 memcpy(&fod->cmdiubuf, cmdiubuf, cmdiubuf_len);
2316 nvmet_fc_queue_fcp_req(tgtport, queue, fcpreq);
2321 if (!tgtport->ops->defer_rcv) {
2322 spin_unlock_irqrestore(&queue->qlock, flags);
2323 /* release the queue lookup reference */
2324 nvmet_fc_tgt_q_put(queue);
2328 deferfcp = list_first_entry_or_null(&queue->avail_defer_list,
2329 struct nvmet_fc_defer_fcp_req, req_list);
2331 /* Just re-use one that was previously allocated */
2332 list_del(&deferfcp->req_list);
2334 spin_unlock_irqrestore(&queue->qlock, flags);
2336 /* Now we need to dynamically allocate one */
2337 deferfcp = kmalloc(sizeof(*deferfcp), GFP_KERNEL);
2339 /* release the queue lookup reference */
2340 nvmet_fc_tgt_q_put(queue);
2343 spin_lock_irqsave(&queue->qlock, flags);
2346 /* For now, use rspaddr / rsplen to save payload information */
2347 fcpreq->rspaddr = cmdiubuf;
2348 fcpreq->rsplen = cmdiubuf_len;
2349 deferfcp->fcp_req = fcpreq;
2351 /* defer processing till a fod becomes available */
2352 list_add_tail(&deferfcp->req_list, &queue->pending_cmd_list);
2354 /* NOTE: the queue lookup reference is still valid */
2356 spin_unlock_irqrestore(&queue->qlock, flags);
2360 EXPORT_SYMBOL_GPL(nvmet_fc_rcv_fcp_req);
2363 * nvmet_fc_rcv_fcp_abort - transport entry point called by an LLDD
2364 * upon the reception of an ABTS for a FCP command
2366 * Notify the transport that an ABTS has been received for a FCP command
2367 * that had been given to the transport via nvmet_fc_rcv_fcp_req(). The
2368 * LLDD believes the command is still being worked on
2369 * (template_ops->fcp_req_release() has not been called).
2371 * The transport will wait for any outstanding work (an op to the LLDD,
2372 * which the lldd should complete with error due to the ABTS; or the
2373 * completion from the nvmet layer of the nvme command), then will
2374 * stop processing and call the nvmet_fc_rcv_fcp_req() callback to
2375 * return the i/o context to the LLDD. The LLDD may send the BA_ACC
2376 * to the ABTS either after return from this function (assuming any
2377 * outstanding op work has been terminated) or upon the callback being
2380 * @target_port: pointer to the (registered) target port the FCP CMD IU
2382 * @fcpreq: pointer to the fcpreq request structure that corresponds
2383 * to the exchange that received the ABTS.
2386 nvmet_fc_rcv_fcp_abort(struct nvmet_fc_target_port *target_port,
2387 struct nvmefc_tgt_fcp_req *fcpreq)
2389 struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
2390 struct nvmet_fc_tgt_queue *queue;
2391 unsigned long flags;
2393 if (!fod || fod->fcpreq != fcpreq)
2394 /* job appears to have already completed, ignore abort */
2399 spin_lock_irqsave(&queue->qlock, flags);
2402 * mark as abort. The abort handler, invoked upon completion
2403 * of any work, will detect the aborted status and do the
2406 spin_lock(&fod->flock);
2408 fod->aborted = true;
2409 spin_unlock(&fod->flock);
2411 spin_unlock_irqrestore(&queue->qlock, flags);
2413 EXPORT_SYMBOL_GPL(nvmet_fc_rcv_fcp_abort);
2416 struct nvmet_fc_traddr {
2422 __nvme_fc_parse_u64(substring_t *sstr, u64 *val)
2426 if (match_u64(sstr, &token64))
2434 * This routine validates and extracts the WWN's from the TRADDR string.
2435 * As kernel parsers need the 0x to determine number base, universally
2436 * build string to parse with 0x prefix before parsing name strings.
2439 nvme_fc_parse_traddr(struct nvmet_fc_traddr *traddr, char *buf, size_t blen)
2441 char name[2 + NVME_FC_TRADDR_HEXNAMELEN + 1];
2442 substring_t wwn = { name, &name[sizeof(name)-1] };
2443 int nnoffset, pnoffset;
2445 /* validate it string one of the 2 allowed formats */
2446 if (strnlen(buf, blen) == NVME_FC_TRADDR_MAXLENGTH &&
2447 !strncmp(buf, "nn-0x", NVME_FC_TRADDR_OXNNLEN) &&
2448 !strncmp(&buf[NVME_FC_TRADDR_MAX_PN_OFFSET],
2449 "pn-0x", NVME_FC_TRADDR_OXNNLEN)) {
2450 nnoffset = NVME_FC_TRADDR_OXNNLEN;
2451 pnoffset = NVME_FC_TRADDR_MAX_PN_OFFSET +
2452 NVME_FC_TRADDR_OXNNLEN;
2453 } else if ((strnlen(buf, blen) == NVME_FC_TRADDR_MINLENGTH &&
2454 !strncmp(buf, "nn-", NVME_FC_TRADDR_NNLEN) &&
2455 !strncmp(&buf[NVME_FC_TRADDR_MIN_PN_OFFSET],
2456 "pn-", NVME_FC_TRADDR_NNLEN))) {
2457 nnoffset = NVME_FC_TRADDR_NNLEN;
2458 pnoffset = NVME_FC_TRADDR_MIN_PN_OFFSET + NVME_FC_TRADDR_NNLEN;
2464 name[2 + NVME_FC_TRADDR_HEXNAMELEN] = 0;
2466 memcpy(&name[2], &buf[nnoffset], NVME_FC_TRADDR_HEXNAMELEN);
2467 if (__nvme_fc_parse_u64(&wwn, &traddr->nn))
2470 memcpy(&name[2], &buf[pnoffset], NVME_FC_TRADDR_HEXNAMELEN);
2471 if (__nvme_fc_parse_u64(&wwn, &traddr->pn))
2477 pr_warn("%s: bad traddr string\n", __func__);
2482 nvmet_fc_add_port(struct nvmet_port *port)
2484 struct nvmet_fc_tgtport *tgtport;
2485 struct nvmet_fc_traddr traddr = { 0L, 0L };
2486 unsigned long flags;
2489 /* validate the address info */
2490 if ((port->disc_addr.trtype != NVMF_TRTYPE_FC) ||
2491 (port->disc_addr.adrfam != NVMF_ADDR_FAMILY_FC))
2494 /* map the traddr address info to a target port */
2496 ret = nvme_fc_parse_traddr(&traddr, port->disc_addr.traddr,
2497 sizeof(port->disc_addr.traddr));
2502 spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
2503 list_for_each_entry(tgtport, &nvmet_fc_target_list, tgt_list) {
2504 if ((tgtport->fc_target_port.node_name == traddr.nn) &&
2505 (tgtport->fc_target_port.port_name == traddr.pn)) {
2506 tgtport->port = port;
2511 spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
2516 nvmet_fc_remove_port(struct nvmet_port *port)
2521 static const struct nvmet_fabrics_ops nvmet_fc_tgt_fcp_ops = {
2522 .owner = THIS_MODULE,
2523 .type = NVMF_TRTYPE_FC,
2525 .add_port = nvmet_fc_add_port,
2526 .remove_port = nvmet_fc_remove_port,
2527 .queue_response = nvmet_fc_fcp_nvme_cmd_done,
2528 .delete_ctrl = nvmet_fc_delete_ctrl,
2531 static int __init nvmet_fc_init_module(void)
2533 return nvmet_register_transport(&nvmet_fc_tgt_fcp_ops);
2536 static void __exit nvmet_fc_exit_module(void)
2538 /* sanity check - all lports should be removed */
2539 if (!list_empty(&nvmet_fc_target_list))
2540 pr_warn("%s: targetport list not empty\n", __func__);
2542 nvmet_unregister_transport(&nvmet_fc_tgt_fcp_ops);
2544 ida_destroy(&nvmet_fc_tgtport_cnt);
2547 module_init(nvmet_fc_init_module);
2548 module_exit(nvmet_fc_exit_module);
2550 MODULE_LICENSE("GPL v2");