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 /* desired maximum for a single sequence - if sg list allows it */
62 #define NVMET_FC_MAX_SEQ_LENGTH (256 * 1024)
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 *next_sg;
78 struct scatterlist *data_sg;
81 enum nvmet_fcp_datadir io_dir;
89 struct work_struct work;
90 struct work_struct done_work;
91 struct work_struct defer_work;
93 struct nvmet_fc_tgtport *tgtport;
94 struct nvmet_fc_tgt_queue *queue;
96 struct list_head fcp_list; /* tgtport->fcp_list */
99 struct nvmet_fc_tgtport {
101 struct nvmet_fc_target_port fc_target_port;
103 struct list_head tgt_list; /* nvmet_fc_target_list */
104 struct device *dev; /* dev for dma mapping */
105 struct nvmet_fc_target_template *ops;
107 struct nvmet_fc_ls_iod *iod;
109 struct list_head ls_list;
110 struct list_head ls_busylist;
111 struct list_head assoc_list;
112 struct ida assoc_cnt;
113 struct nvmet_port *port;
118 struct nvmet_fc_defer_fcp_req {
119 struct list_head req_list;
120 struct nvmefc_tgt_fcp_req *fcp_req;
123 struct nvmet_fc_tgt_queue {
135 struct nvmet_port *port;
136 struct nvmet_cq nvme_cq;
137 struct nvmet_sq nvme_sq;
138 struct nvmet_fc_tgt_assoc *assoc;
139 struct nvmet_fc_fcp_iod *fod; /* array of fcp_iods */
140 struct list_head fod_list;
141 struct list_head pending_cmd_list;
142 struct list_head avail_defer_list;
143 struct workqueue_struct *work_q;
145 } __aligned(sizeof(unsigned long long));
147 struct nvmet_fc_tgt_assoc {
150 struct nvmet_fc_tgtport *tgtport;
151 struct list_head a_list;
152 struct nvmet_fc_tgt_queue *queues[NVMET_NR_QUEUES + 1];
154 struct work_struct del_work;
159 nvmet_fc_iodnum(struct nvmet_fc_ls_iod *iodptr)
161 return (iodptr - iodptr->tgtport->iod);
165 nvmet_fc_fodnum(struct nvmet_fc_fcp_iod *fodptr)
167 return (fodptr - fodptr->queue->fod);
172 * Association and Connection IDs:
174 * Association ID will have random number in upper 6 bytes and zero
177 * Connection IDs will be Association ID with QID or'd in lower 2 bytes
179 * note: Association ID = Connection ID for queue 0
181 #define BYTES_FOR_QID sizeof(u16)
182 #define BYTES_FOR_QID_SHIFT (BYTES_FOR_QID * 8)
183 #define NVMET_FC_QUEUEID_MASK ((u64)((1 << BYTES_FOR_QID_SHIFT) - 1))
186 nvmet_fc_makeconnid(struct nvmet_fc_tgt_assoc *assoc, u16 qid)
188 return (assoc->association_id | qid);
192 nvmet_fc_getassociationid(u64 connectionid)
194 return connectionid & ~NVMET_FC_QUEUEID_MASK;
198 nvmet_fc_getqueueid(u64 connectionid)
200 return (u16)(connectionid & NVMET_FC_QUEUEID_MASK);
203 static inline struct nvmet_fc_tgtport *
204 targetport_to_tgtport(struct nvmet_fc_target_port *targetport)
206 return container_of(targetport, struct nvmet_fc_tgtport,
210 static inline struct nvmet_fc_fcp_iod *
211 nvmet_req_to_fod(struct nvmet_req *nvme_req)
213 return container_of(nvme_req, struct nvmet_fc_fcp_iod, req);
217 /* *************************** Globals **************************** */
220 static DEFINE_SPINLOCK(nvmet_fc_tgtlock);
222 static LIST_HEAD(nvmet_fc_target_list);
223 static DEFINE_IDA(nvmet_fc_tgtport_cnt);
226 static void nvmet_fc_handle_ls_rqst_work(struct work_struct *work);
227 static void nvmet_fc_handle_fcp_rqst_work(struct work_struct *work);
228 static void nvmet_fc_fcp_rqst_op_done_work(struct work_struct *work);
229 static void nvmet_fc_fcp_rqst_op_defer_work(struct work_struct *work);
230 static void nvmet_fc_tgt_a_put(struct nvmet_fc_tgt_assoc *assoc);
231 static int nvmet_fc_tgt_a_get(struct nvmet_fc_tgt_assoc *assoc);
232 static void nvmet_fc_tgt_q_put(struct nvmet_fc_tgt_queue *queue);
233 static int nvmet_fc_tgt_q_get(struct nvmet_fc_tgt_queue *queue);
234 static void nvmet_fc_tgtport_put(struct nvmet_fc_tgtport *tgtport);
235 static int nvmet_fc_tgtport_get(struct nvmet_fc_tgtport *tgtport);
236 static void nvmet_fc_handle_fcp_rqst(struct nvmet_fc_tgtport *tgtport,
237 struct nvmet_fc_fcp_iod *fod);
238 static void nvmet_fc_delete_target_assoc(struct nvmet_fc_tgt_assoc *assoc);
241 /* *********************** FC-NVME DMA Handling **************************** */
244 * The fcloop device passes in a NULL device pointer. Real LLD's will
245 * pass in a valid device pointer. If NULL is passed to the dma mapping
246 * routines, depending on the platform, it may or may not succeed, and
250 * Wrapper all the dma routines and check the dev pointer.
252 * If simple mappings (return just a dma address, we'll noop them,
253 * returning a dma address of 0.
255 * On more complex mappings (dma_map_sg), a pseudo routine fills
256 * in the scatter list, setting all dma addresses to 0.
259 static inline dma_addr_t
260 fc_dma_map_single(struct device *dev, void *ptr, size_t size,
261 enum dma_data_direction dir)
263 return dev ? dma_map_single(dev, ptr, size, dir) : (dma_addr_t)0L;
267 fc_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
269 return dev ? dma_mapping_error(dev, dma_addr) : 0;
273 fc_dma_unmap_single(struct device *dev, dma_addr_t addr, size_t size,
274 enum dma_data_direction dir)
277 dma_unmap_single(dev, addr, size, dir);
281 fc_dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr, size_t size,
282 enum dma_data_direction dir)
285 dma_sync_single_for_cpu(dev, addr, size, dir);
289 fc_dma_sync_single_for_device(struct device *dev, dma_addr_t addr, size_t size,
290 enum dma_data_direction dir)
293 dma_sync_single_for_device(dev, addr, size, dir);
296 /* pseudo dma_map_sg call */
298 fc_map_sg(struct scatterlist *sg, int nents)
300 struct scatterlist *s;
303 WARN_ON(nents == 0 || sg[0].length == 0);
305 for_each_sg(sg, s, nents, i) {
307 #ifdef CONFIG_NEED_SG_DMA_LENGTH
308 s->dma_length = s->length;
315 fc_dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
316 enum dma_data_direction dir)
318 return dev ? dma_map_sg(dev, sg, nents, dir) : fc_map_sg(sg, nents);
322 fc_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
323 enum dma_data_direction dir)
326 dma_unmap_sg(dev, sg, nents, dir);
330 /* *********************** FC-NVME Port Management ************************ */
334 nvmet_fc_alloc_ls_iodlist(struct nvmet_fc_tgtport *tgtport)
336 struct nvmet_fc_ls_iod *iod;
339 iod = kcalloc(NVMET_LS_CTX_COUNT, sizeof(struct nvmet_fc_ls_iod),
346 for (i = 0; i < NVMET_LS_CTX_COUNT; iod++, i++) {
347 INIT_WORK(&iod->work, nvmet_fc_handle_ls_rqst_work);
348 iod->tgtport = tgtport;
349 list_add_tail(&iod->ls_list, &tgtport->ls_list);
351 iod->rqstbuf = kcalloc(2, NVME_FC_MAX_LS_BUFFER_SIZE,
356 iod->rspbuf = iod->rqstbuf + NVME_FC_MAX_LS_BUFFER_SIZE;
358 iod->rspdma = fc_dma_map_single(tgtport->dev, iod->rspbuf,
359 NVME_FC_MAX_LS_BUFFER_SIZE,
361 if (fc_dma_mapping_error(tgtport->dev, iod->rspdma))
369 list_del(&iod->ls_list);
370 for (iod--, i--; i >= 0; iod--, i--) {
371 fc_dma_unmap_single(tgtport->dev, iod->rspdma,
372 NVME_FC_MAX_LS_BUFFER_SIZE, DMA_TO_DEVICE);
374 list_del(&iod->ls_list);
383 nvmet_fc_free_ls_iodlist(struct nvmet_fc_tgtport *tgtport)
385 struct nvmet_fc_ls_iod *iod = tgtport->iod;
388 for (i = 0; i < NVMET_LS_CTX_COUNT; iod++, i++) {
389 fc_dma_unmap_single(tgtport->dev,
390 iod->rspdma, NVME_FC_MAX_LS_BUFFER_SIZE,
393 list_del(&iod->ls_list);
398 static struct nvmet_fc_ls_iod *
399 nvmet_fc_alloc_ls_iod(struct nvmet_fc_tgtport *tgtport)
401 struct nvmet_fc_ls_iod *iod;
404 spin_lock_irqsave(&tgtport->lock, flags);
405 iod = list_first_entry_or_null(&tgtport->ls_list,
406 struct nvmet_fc_ls_iod, ls_list);
408 list_move_tail(&iod->ls_list, &tgtport->ls_busylist);
409 spin_unlock_irqrestore(&tgtport->lock, flags);
415 nvmet_fc_free_ls_iod(struct nvmet_fc_tgtport *tgtport,
416 struct nvmet_fc_ls_iod *iod)
420 spin_lock_irqsave(&tgtport->lock, flags);
421 list_move(&iod->ls_list, &tgtport->ls_list);
422 spin_unlock_irqrestore(&tgtport->lock, flags);
426 nvmet_fc_prep_fcp_iodlist(struct nvmet_fc_tgtport *tgtport,
427 struct nvmet_fc_tgt_queue *queue)
429 struct nvmet_fc_fcp_iod *fod = queue->fod;
432 for (i = 0; i < queue->sqsize; fod++, i++) {
433 INIT_WORK(&fod->work, nvmet_fc_handle_fcp_rqst_work);
434 INIT_WORK(&fod->done_work, nvmet_fc_fcp_rqst_op_done_work);
435 INIT_WORK(&fod->defer_work, nvmet_fc_fcp_rqst_op_defer_work);
436 fod->tgtport = tgtport;
440 fod->aborted = false;
442 list_add_tail(&fod->fcp_list, &queue->fod_list);
443 spin_lock_init(&fod->flock);
445 fod->rspdma = fc_dma_map_single(tgtport->dev, &fod->rspiubuf,
446 sizeof(fod->rspiubuf), DMA_TO_DEVICE);
447 if (fc_dma_mapping_error(tgtport->dev, fod->rspdma)) {
448 list_del(&fod->fcp_list);
449 for (fod--, i--; i >= 0; fod--, i--) {
450 fc_dma_unmap_single(tgtport->dev, fod->rspdma,
451 sizeof(fod->rspiubuf),
454 list_del(&fod->fcp_list);
463 nvmet_fc_destroy_fcp_iodlist(struct nvmet_fc_tgtport *tgtport,
464 struct nvmet_fc_tgt_queue *queue)
466 struct nvmet_fc_fcp_iod *fod = queue->fod;
469 for (i = 0; i < queue->sqsize; fod++, i++) {
471 fc_dma_unmap_single(tgtport->dev, fod->rspdma,
472 sizeof(fod->rspiubuf), DMA_TO_DEVICE);
476 static struct nvmet_fc_fcp_iod *
477 nvmet_fc_alloc_fcp_iod(struct nvmet_fc_tgt_queue *queue)
479 struct nvmet_fc_fcp_iod *fod;
481 lockdep_assert_held(&queue->qlock);
483 fod = list_first_entry_or_null(&queue->fod_list,
484 struct nvmet_fc_fcp_iod, fcp_list);
486 list_del(&fod->fcp_list);
489 * no queue reference is taken, as it was taken by the
490 * queue lookup just prior to the allocation. The iod
491 * will "inherit" that reference.
499 nvmet_fc_queue_fcp_req(struct nvmet_fc_tgtport *tgtport,
500 struct nvmet_fc_tgt_queue *queue,
501 struct nvmefc_tgt_fcp_req *fcpreq)
503 struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
506 * put all admin cmds on hw queue id 0. All io commands go to
507 * the respective hw queue based on a modulo basis
509 fcpreq->hwqid = queue->qid ?
510 ((queue->qid - 1) % tgtport->ops->max_hw_queues) : 0;
512 if (tgtport->ops->target_features & NVMET_FCTGTFEAT_CMD_IN_ISR)
513 queue_work_on(queue->cpu, queue->work_q, &fod->work);
515 nvmet_fc_handle_fcp_rqst(tgtport, fod);
519 nvmet_fc_fcp_rqst_op_defer_work(struct work_struct *work)
521 struct nvmet_fc_fcp_iod *fod =
522 container_of(work, struct nvmet_fc_fcp_iod, defer_work);
524 /* Submit deferred IO for processing */
525 nvmet_fc_queue_fcp_req(fod->tgtport, fod->queue, fod->fcpreq);
530 nvmet_fc_free_fcp_iod(struct nvmet_fc_tgt_queue *queue,
531 struct nvmet_fc_fcp_iod *fod)
533 struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
534 struct nvmet_fc_tgtport *tgtport = fod->tgtport;
535 struct nvmet_fc_defer_fcp_req *deferfcp;
538 fc_dma_sync_single_for_cpu(tgtport->dev, fod->rspdma,
539 sizeof(fod->rspiubuf), DMA_TO_DEVICE);
541 fcpreq->nvmet_fc_private = NULL;
545 fod->aborted = false;
546 fod->writedataactive = false;
549 tgtport->ops->fcp_req_release(&tgtport->fc_target_port, fcpreq);
551 /* release the queue lookup reference on the completed IO */
552 nvmet_fc_tgt_q_put(queue);
554 spin_lock_irqsave(&queue->qlock, flags);
555 deferfcp = list_first_entry_or_null(&queue->pending_cmd_list,
556 struct nvmet_fc_defer_fcp_req, req_list);
558 list_add_tail(&fod->fcp_list, &fod->queue->fod_list);
559 spin_unlock_irqrestore(&queue->qlock, flags);
563 /* Re-use the fod for the next pending cmd that was deferred */
564 list_del(&deferfcp->req_list);
566 fcpreq = deferfcp->fcp_req;
568 /* deferfcp can be reused for another IO at a later date */
569 list_add_tail(&deferfcp->req_list, &queue->avail_defer_list);
571 spin_unlock_irqrestore(&queue->qlock, flags);
573 /* Save NVME CMD IO in fod */
574 memcpy(&fod->cmdiubuf, fcpreq->rspaddr, fcpreq->rsplen);
576 /* Setup new fcpreq to be processed */
577 fcpreq->rspaddr = NULL;
579 fcpreq->nvmet_fc_private = fod;
580 fod->fcpreq = fcpreq;
583 /* inform LLDD IO is now being processed */
584 tgtport->ops->defer_rcv(&tgtport->fc_target_port, fcpreq);
587 * Leave the queue lookup get reference taken when
588 * fod was originally allocated.
591 queue_work(queue->work_q, &fod->defer_work);
595 nvmet_fc_queue_to_cpu(struct nvmet_fc_tgtport *tgtport, int qid)
599 if (tgtport->ops->max_hw_queues == 1)
600 return WORK_CPU_UNBOUND;
602 /* Simple cpu selection based on qid modulo active cpu count */
603 idx = !qid ? 0 : (qid - 1) % num_active_cpus();
605 /* find the n'th active cpu */
606 for (cpu = 0, cnt = 0; ; ) {
607 if (cpu_active(cpu)) {
612 cpu = (cpu + 1) % num_possible_cpus();
618 static struct nvmet_fc_tgt_queue *
619 nvmet_fc_alloc_target_queue(struct nvmet_fc_tgt_assoc *assoc,
622 struct nvmet_fc_tgt_queue *queue;
626 if (qid > NVMET_NR_QUEUES)
629 queue = kzalloc((sizeof(*queue) +
630 (sizeof(struct nvmet_fc_fcp_iod) * sqsize)),
635 if (!nvmet_fc_tgt_a_get(assoc))
638 queue->work_q = alloc_workqueue("ntfc%d.%d.%d", 0, 0,
639 assoc->tgtport->fc_target_port.port_num,
644 queue->fod = (struct nvmet_fc_fcp_iod *)&queue[1];
646 queue->sqsize = sqsize;
647 queue->assoc = assoc;
648 queue->port = assoc->tgtport->port;
649 queue->cpu = nvmet_fc_queue_to_cpu(assoc->tgtport, qid);
650 INIT_LIST_HEAD(&queue->fod_list);
651 INIT_LIST_HEAD(&queue->avail_defer_list);
652 INIT_LIST_HEAD(&queue->pending_cmd_list);
653 atomic_set(&queue->connected, 0);
654 atomic_set(&queue->sqtail, 0);
655 atomic_set(&queue->rsn, 1);
656 atomic_set(&queue->zrspcnt, 0);
657 spin_lock_init(&queue->qlock);
658 kref_init(&queue->ref);
660 nvmet_fc_prep_fcp_iodlist(assoc->tgtport, queue);
662 ret = nvmet_sq_init(&queue->nvme_sq);
664 goto out_fail_iodlist;
666 WARN_ON(assoc->queues[qid]);
667 spin_lock_irqsave(&assoc->tgtport->lock, flags);
668 assoc->queues[qid] = queue;
669 spin_unlock_irqrestore(&assoc->tgtport->lock, flags);
674 nvmet_fc_destroy_fcp_iodlist(assoc->tgtport, queue);
675 destroy_workqueue(queue->work_q);
677 nvmet_fc_tgt_a_put(assoc);
685 nvmet_fc_tgt_queue_free(struct kref *ref)
687 struct nvmet_fc_tgt_queue *queue =
688 container_of(ref, struct nvmet_fc_tgt_queue, ref);
691 spin_lock_irqsave(&queue->assoc->tgtport->lock, flags);
692 queue->assoc->queues[queue->qid] = NULL;
693 spin_unlock_irqrestore(&queue->assoc->tgtport->lock, flags);
695 nvmet_fc_destroy_fcp_iodlist(queue->assoc->tgtport, queue);
697 nvmet_fc_tgt_a_put(queue->assoc);
699 destroy_workqueue(queue->work_q);
705 nvmet_fc_tgt_q_put(struct nvmet_fc_tgt_queue *queue)
707 kref_put(&queue->ref, nvmet_fc_tgt_queue_free);
711 nvmet_fc_tgt_q_get(struct nvmet_fc_tgt_queue *queue)
713 return kref_get_unless_zero(&queue->ref);
718 nvmet_fc_delete_target_queue(struct nvmet_fc_tgt_queue *queue)
720 struct nvmet_fc_tgtport *tgtport = queue->assoc->tgtport;
721 struct nvmet_fc_fcp_iod *fod = queue->fod;
722 struct nvmet_fc_defer_fcp_req *deferfcp, *tempptr;
724 int i, writedataactive;
727 disconnect = atomic_xchg(&queue->connected, 0);
729 spin_lock_irqsave(&queue->qlock, flags);
730 /* about outstanding io's */
731 for (i = 0; i < queue->sqsize; fod++, i++) {
733 spin_lock(&fod->flock);
735 writedataactive = fod->writedataactive;
736 spin_unlock(&fod->flock);
738 * only call lldd abort routine if waiting for
739 * writedata. other outstanding ops should finish
742 if (writedataactive) {
743 spin_lock(&fod->flock);
745 spin_unlock(&fod->flock);
746 tgtport->ops->fcp_abort(
747 &tgtport->fc_target_port, fod->fcpreq);
752 /* Cleanup defer'ed IOs in queue */
753 list_for_each_entry_safe(deferfcp, tempptr, &queue->avail_defer_list,
755 list_del(&deferfcp->req_list);
760 deferfcp = list_first_entry_or_null(&queue->pending_cmd_list,
761 struct nvmet_fc_defer_fcp_req, req_list);
765 list_del(&deferfcp->req_list);
766 spin_unlock_irqrestore(&queue->qlock, flags);
768 tgtport->ops->defer_rcv(&tgtport->fc_target_port,
771 tgtport->ops->fcp_abort(&tgtport->fc_target_port,
774 tgtport->ops->fcp_req_release(&tgtport->fc_target_port,
777 /* release the queue lookup reference */
778 nvmet_fc_tgt_q_put(queue);
782 spin_lock_irqsave(&queue->qlock, flags);
784 spin_unlock_irqrestore(&queue->qlock, flags);
786 flush_workqueue(queue->work_q);
789 nvmet_sq_destroy(&queue->nvme_sq);
791 nvmet_fc_tgt_q_put(queue);
794 static struct nvmet_fc_tgt_queue *
795 nvmet_fc_find_target_queue(struct nvmet_fc_tgtport *tgtport,
798 struct nvmet_fc_tgt_assoc *assoc;
799 struct nvmet_fc_tgt_queue *queue;
800 u64 association_id = nvmet_fc_getassociationid(connection_id);
801 u16 qid = nvmet_fc_getqueueid(connection_id);
804 if (qid > NVMET_NR_QUEUES)
807 spin_lock_irqsave(&tgtport->lock, flags);
808 list_for_each_entry(assoc, &tgtport->assoc_list, a_list) {
809 if (association_id == assoc->association_id) {
810 queue = assoc->queues[qid];
812 (!atomic_read(&queue->connected) ||
813 !nvmet_fc_tgt_q_get(queue)))
815 spin_unlock_irqrestore(&tgtport->lock, flags);
819 spin_unlock_irqrestore(&tgtport->lock, flags);
824 nvmet_fc_delete_assoc(struct work_struct *work)
826 struct nvmet_fc_tgt_assoc *assoc =
827 container_of(work, struct nvmet_fc_tgt_assoc, del_work);
829 nvmet_fc_delete_target_assoc(assoc);
830 nvmet_fc_tgt_a_put(assoc);
833 static struct nvmet_fc_tgt_assoc *
834 nvmet_fc_alloc_target_assoc(struct nvmet_fc_tgtport *tgtport)
836 struct nvmet_fc_tgt_assoc *assoc, *tmpassoc;
840 bool needrandom = true;
842 assoc = kzalloc(sizeof(*assoc), GFP_KERNEL);
846 idx = ida_simple_get(&tgtport->assoc_cnt, 0, 0, GFP_KERNEL);
850 if (!nvmet_fc_tgtport_get(tgtport))
853 assoc->tgtport = tgtport;
855 INIT_LIST_HEAD(&assoc->a_list);
856 kref_init(&assoc->ref);
857 INIT_WORK(&assoc->del_work, nvmet_fc_delete_assoc);
860 get_random_bytes(&ran, sizeof(ran) - BYTES_FOR_QID);
861 ran = ran << BYTES_FOR_QID_SHIFT;
863 spin_lock_irqsave(&tgtport->lock, flags);
865 list_for_each_entry(tmpassoc, &tgtport->assoc_list, a_list)
866 if (ran == tmpassoc->association_id) {
871 assoc->association_id = ran;
872 list_add_tail(&assoc->a_list, &tgtport->assoc_list);
874 spin_unlock_irqrestore(&tgtport->lock, flags);
880 ida_simple_remove(&tgtport->assoc_cnt, idx);
887 nvmet_fc_target_assoc_free(struct kref *ref)
889 struct nvmet_fc_tgt_assoc *assoc =
890 container_of(ref, struct nvmet_fc_tgt_assoc, ref);
891 struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
894 spin_lock_irqsave(&tgtport->lock, flags);
895 list_del(&assoc->a_list);
896 spin_unlock_irqrestore(&tgtport->lock, flags);
897 ida_simple_remove(&tgtport->assoc_cnt, assoc->a_id);
899 nvmet_fc_tgtport_put(tgtport);
903 nvmet_fc_tgt_a_put(struct nvmet_fc_tgt_assoc *assoc)
905 kref_put(&assoc->ref, nvmet_fc_target_assoc_free);
909 nvmet_fc_tgt_a_get(struct nvmet_fc_tgt_assoc *assoc)
911 return kref_get_unless_zero(&assoc->ref);
915 nvmet_fc_delete_target_assoc(struct nvmet_fc_tgt_assoc *assoc)
917 struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
918 struct nvmet_fc_tgt_queue *queue;
922 spin_lock_irqsave(&tgtport->lock, flags);
923 for (i = NVMET_NR_QUEUES; i >= 0; i--) {
924 queue = assoc->queues[i];
926 if (!nvmet_fc_tgt_q_get(queue))
928 spin_unlock_irqrestore(&tgtport->lock, flags);
929 nvmet_fc_delete_target_queue(queue);
930 nvmet_fc_tgt_q_put(queue);
931 spin_lock_irqsave(&tgtport->lock, flags);
934 spin_unlock_irqrestore(&tgtport->lock, flags);
936 nvmet_fc_tgt_a_put(assoc);
939 static struct nvmet_fc_tgt_assoc *
940 nvmet_fc_find_target_assoc(struct nvmet_fc_tgtport *tgtport,
943 struct nvmet_fc_tgt_assoc *assoc;
944 struct nvmet_fc_tgt_assoc *ret = NULL;
947 spin_lock_irqsave(&tgtport->lock, flags);
948 list_for_each_entry(assoc, &tgtport->assoc_list, a_list) {
949 if (association_id == assoc->association_id) {
951 nvmet_fc_tgt_a_get(assoc);
955 spin_unlock_irqrestore(&tgtport->lock, flags);
962 * nvme_fc_register_targetport - transport entry point called by an
963 * LLDD to register the existence of a local
964 * NVME subystem FC port.
965 * @pinfo: pointer to information about the port to be registered
966 * @template: LLDD entrypoints and operational parameters for the port
967 * @dev: physical hardware device node port corresponds to. Will be
968 * used for DMA mappings
969 * @portptr: pointer to a local port pointer. Upon success, the routine
970 * will allocate a nvme_fc_local_port structure and place its
971 * address in the local port pointer. Upon failure, local port
972 * pointer will be set to NULL.
975 * a completion status. Must be 0 upon success; a negative errno
976 * (ex: -ENXIO) upon failure.
979 nvmet_fc_register_targetport(struct nvmet_fc_port_info *pinfo,
980 struct nvmet_fc_target_template *template,
982 struct nvmet_fc_target_port **portptr)
984 struct nvmet_fc_tgtport *newrec;
988 if (!template->xmt_ls_rsp || !template->fcp_op ||
989 !template->fcp_abort ||
990 !template->fcp_req_release || !template->targetport_delete ||
991 !template->max_hw_queues || !template->max_sgl_segments ||
992 !template->max_dif_sgl_segments || !template->dma_boundary) {
994 goto out_regtgt_failed;
997 newrec = kzalloc((sizeof(*newrec) + template->target_priv_sz),
1001 goto out_regtgt_failed;
1004 idx = ida_simple_get(&nvmet_fc_tgtport_cnt, 0, 0, GFP_KERNEL);
1007 goto out_fail_kfree;
1010 if (!get_device(dev) && dev) {
1015 newrec->fc_target_port.node_name = pinfo->node_name;
1016 newrec->fc_target_port.port_name = pinfo->port_name;
1017 newrec->fc_target_port.private = &newrec[1];
1018 newrec->fc_target_port.port_id = pinfo->port_id;
1019 newrec->fc_target_port.port_num = idx;
1020 INIT_LIST_HEAD(&newrec->tgt_list);
1022 newrec->ops = template;
1023 spin_lock_init(&newrec->lock);
1024 INIT_LIST_HEAD(&newrec->ls_list);
1025 INIT_LIST_HEAD(&newrec->ls_busylist);
1026 INIT_LIST_HEAD(&newrec->assoc_list);
1027 kref_init(&newrec->ref);
1028 ida_init(&newrec->assoc_cnt);
1029 newrec->max_sg_cnt = 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 */
1725 fod->next_sg = fod->data_sg;
1730 return NVME_SC_INTERNAL;
1734 nvmet_fc_free_tgt_pgs(struct nvmet_fc_fcp_iod *fod)
1736 if (!fod->data_sg || !fod->data_sg_cnt)
1739 fc_dma_unmap_sg(fod->tgtport->dev, fod->data_sg, fod->data_sg_cnt,
1740 ((fod->io_dir == NVMET_FCP_WRITE) ?
1741 DMA_FROM_DEVICE : DMA_TO_DEVICE));
1742 sgl_free(fod->data_sg);
1743 fod->data_sg = NULL;
1744 fod->data_sg_cnt = 0;
1749 queue_90percent_full(struct nvmet_fc_tgt_queue *q, u32 sqhd)
1753 /* egad, this is ugly. And sqtail is just a best guess */
1754 sqtail = atomic_read(&q->sqtail) % q->sqsize;
1756 used = (sqtail < sqhd) ? (sqtail + q->sqsize - sqhd) : (sqtail - sqhd);
1757 return ((used * 10) >= (((u32)(q->sqsize - 1) * 9)));
1762 * May be a NVMET_FCOP_RSP or NVMET_FCOP_READDATA_RSP op
1765 nvmet_fc_prep_fcp_rsp(struct nvmet_fc_tgtport *tgtport,
1766 struct nvmet_fc_fcp_iod *fod)
1768 struct nvme_fc_ersp_iu *ersp = &fod->rspiubuf;
1769 struct nvme_common_command *sqe = &fod->cmdiubuf.sqe.common;
1770 struct nvme_completion *cqe = &ersp->cqe;
1771 u32 *cqewd = (u32 *)cqe;
1772 bool send_ersp = false;
1773 u32 rsn, rspcnt, xfr_length;
1775 if (fod->fcpreq->op == NVMET_FCOP_READDATA_RSP)
1776 xfr_length = fod->req.transfer_len;
1778 xfr_length = fod->offset;
1781 * check to see if we can send a 0's rsp.
1782 * Note: to send a 0's response, the NVME-FC host transport will
1783 * recreate the CQE. The host transport knows: sq id, SQHD (last
1784 * seen in an ersp), and command_id. Thus it will create a
1785 * zero-filled CQE with those known fields filled in. Transport
1786 * must send an ersp for any condition where the cqe won't match
1789 * Here are the FC-NVME mandated cases where we must send an ersp:
1790 * every N responses, where N=ersp_ratio
1791 * force fabric commands to send ersp's (not in FC-NVME but good
1793 * normal cmds: any time status is non-zero, or status is zero
1794 * but words 0 or 1 are non-zero.
1795 * the SQ is 90% or more full
1796 * the cmd is a fused command
1797 * transferred data length not equal to cmd iu length
1799 rspcnt = atomic_inc_return(&fod->queue->zrspcnt);
1800 if (!(rspcnt % fod->queue->ersp_ratio) ||
1801 sqe->opcode == nvme_fabrics_command ||
1802 xfr_length != fod->req.transfer_len ||
1803 (le16_to_cpu(cqe->status) & 0xFFFE) || cqewd[0] || cqewd[1] ||
1804 (sqe->flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND)) ||
1805 queue_90percent_full(fod->queue, le16_to_cpu(cqe->sq_head)))
1808 /* re-set the fields */
1809 fod->fcpreq->rspaddr = ersp;
1810 fod->fcpreq->rspdma = fod->rspdma;
1813 memset(ersp, 0, NVME_FC_SIZEOF_ZEROS_RSP);
1814 fod->fcpreq->rsplen = NVME_FC_SIZEOF_ZEROS_RSP;
1816 ersp->iu_len = cpu_to_be16(sizeof(*ersp)/sizeof(u32));
1817 rsn = atomic_inc_return(&fod->queue->rsn);
1818 ersp->rsn = cpu_to_be32(rsn);
1819 ersp->xfrd_len = cpu_to_be32(xfr_length);
1820 fod->fcpreq->rsplen = sizeof(*ersp);
1823 fc_dma_sync_single_for_device(tgtport->dev, fod->rspdma,
1824 sizeof(fod->rspiubuf), DMA_TO_DEVICE);
1827 static void nvmet_fc_xmt_fcp_op_done(struct nvmefc_tgt_fcp_req *fcpreq);
1830 nvmet_fc_abort_op(struct nvmet_fc_tgtport *tgtport,
1831 struct nvmet_fc_fcp_iod *fod)
1833 struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
1835 /* data no longer needed */
1836 nvmet_fc_free_tgt_pgs(fod);
1839 * if an ABTS was received or we issued the fcp_abort early
1840 * don't call abort routine again.
1842 /* no need to take lock - lock was taken earlier to get here */
1844 tgtport->ops->fcp_abort(&tgtport->fc_target_port, fcpreq);
1846 nvmet_fc_free_fcp_iod(fod->queue, fod);
1850 nvmet_fc_xmt_fcp_rsp(struct nvmet_fc_tgtport *tgtport,
1851 struct nvmet_fc_fcp_iod *fod)
1855 fod->fcpreq->op = NVMET_FCOP_RSP;
1856 fod->fcpreq->timeout = 0;
1858 nvmet_fc_prep_fcp_rsp(tgtport, fod);
1860 ret = tgtport->ops->fcp_op(&tgtport->fc_target_port, fod->fcpreq);
1862 nvmet_fc_abort_op(tgtport, fod);
1866 nvmet_fc_transfer_fcp_data(struct nvmet_fc_tgtport *tgtport,
1867 struct nvmet_fc_fcp_iod *fod, u8 op)
1869 struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
1870 struct scatterlist *sg = fod->next_sg;
1871 unsigned long flags;
1872 u32 remaininglen = fod->req.transfer_len - fod->offset;
1877 fcpreq->offset = fod->offset;
1878 fcpreq->timeout = NVME_FC_TGTOP_TIMEOUT_SEC;
1881 * for next sequence:
1882 * break at a sg element boundary
1883 * attempt to keep sequence length capped at
1884 * NVMET_FC_MAX_SEQ_LENGTH but allow sequence to
1885 * be longer if a single sg element is larger
1886 * than that amount. This is done to avoid creating
1887 * a new sg list to use for the tgtport api.
1891 while (tlen < remaininglen &&
1892 fcpreq->sg_cnt < tgtport->max_sg_cnt &&
1893 tlen + sg_dma_len(sg) < NVMET_FC_MAX_SEQ_LENGTH) {
1895 tlen += sg_dma_len(sg);
1898 if (tlen < remaininglen && fcpreq->sg_cnt == 0) {
1900 tlen += min_t(u32, sg_dma_len(sg), remaininglen);
1903 if (tlen < remaininglen)
1906 fod->next_sg = NULL;
1908 fcpreq->transfer_length = tlen;
1909 fcpreq->transferred_length = 0;
1910 fcpreq->fcp_error = 0;
1914 * If the last READDATA request: check if LLDD supports
1915 * combined xfr with response.
1917 if ((op == NVMET_FCOP_READDATA) &&
1918 ((fod->offset + fcpreq->transfer_length) == fod->req.transfer_len) &&
1919 (tgtport->ops->target_features & NVMET_FCTGTFEAT_READDATA_RSP)) {
1920 fcpreq->op = NVMET_FCOP_READDATA_RSP;
1921 nvmet_fc_prep_fcp_rsp(tgtport, fod);
1924 ret = tgtport->ops->fcp_op(&tgtport->fc_target_port, fod->fcpreq);
1927 * should be ok to set w/o lock as its in the thread of
1928 * execution (not an async timer routine) and doesn't
1929 * contend with any clearing action
1933 if (op == NVMET_FCOP_WRITEDATA) {
1934 spin_lock_irqsave(&fod->flock, flags);
1935 fod->writedataactive = false;
1936 spin_unlock_irqrestore(&fod->flock, flags);
1937 nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
1938 } else /* NVMET_FCOP_READDATA or NVMET_FCOP_READDATA_RSP */ {
1939 fcpreq->fcp_error = ret;
1940 fcpreq->transferred_length = 0;
1941 nvmet_fc_xmt_fcp_op_done(fod->fcpreq);
1947 __nvmet_fc_fod_op_abort(struct nvmet_fc_fcp_iod *fod, bool abort)
1949 struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
1950 struct nvmet_fc_tgtport *tgtport = fod->tgtport;
1952 /* if in the middle of an io and we need to tear down */
1954 if (fcpreq->op == NVMET_FCOP_WRITEDATA) {
1955 nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
1959 nvmet_fc_abort_op(tgtport, fod);
1967 * actual done handler for FCP operations when completed by the lldd
1970 nvmet_fc_fod_op_done(struct nvmet_fc_fcp_iod *fod)
1972 struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
1973 struct nvmet_fc_tgtport *tgtport = fod->tgtport;
1974 unsigned long flags;
1977 spin_lock_irqsave(&fod->flock, flags);
1979 fod->writedataactive = false;
1980 spin_unlock_irqrestore(&fod->flock, flags);
1982 switch (fcpreq->op) {
1984 case NVMET_FCOP_WRITEDATA:
1985 if (__nvmet_fc_fod_op_abort(fod, abort))
1987 if (fcpreq->fcp_error ||
1988 fcpreq->transferred_length != fcpreq->transfer_length) {
1989 spin_lock_irqsave(&fod->flock, flags);
1991 spin_unlock_irqrestore(&fod->flock, flags);
1993 nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
1997 fod->offset += fcpreq->transferred_length;
1998 if (fod->offset != fod->req.transfer_len) {
1999 spin_lock_irqsave(&fod->flock, flags);
2000 fod->writedataactive = true;
2001 spin_unlock_irqrestore(&fod->flock, flags);
2003 /* transfer the next chunk */
2004 nvmet_fc_transfer_fcp_data(tgtport, fod,
2005 NVMET_FCOP_WRITEDATA);
2009 /* data transfer complete, resume with nvmet layer */
2010 nvmet_req_execute(&fod->req);
2013 case NVMET_FCOP_READDATA:
2014 case NVMET_FCOP_READDATA_RSP:
2015 if (__nvmet_fc_fod_op_abort(fod, abort))
2017 if (fcpreq->fcp_error ||
2018 fcpreq->transferred_length != fcpreq->transfer_length) {
2019 nvmet_fc_abort_op(tgtport, fod);
2025 if (fcpreq->op == NVMET_FCOP_READDATA_RSP) {
2026 /* data no longer needed */
2027 nvmet_fc_free_tgt_pgs(fod);
2028 nvmet_fc_free_fcp_iod(fod->queue, fod);
2032 fod->offset += fcpreq->transferred_length;
2033 if (fod->offset != fod->req.transfer_len) {
2034 /* transfer the next chunk */
2035 nvmet_fc_transfer_fcp_data(tgtport, fod,
2036 NVMET_FCOP_READDATA);
2040 /* data transfer complete, send response */
2042 /* data no longer needed */
2043 nvmet_fc_free_tgt_pgs(fod);
2045 nvmet_fc_xmt_fcp_rsp(tgtport, fod);
2049 case NVMET_FCOP_RSP:
2050 if (__nvmet_fc_fod_op_abort(fod, abort))
2052 nvmet_fc_free_fcp_iod(fod->queue, fod);
2061 nvmet_fc_fcp_rqst_op_done_work(struct work_struct *work)
2063 struct nvmet_fc_fcp_iod *fod =
2064 container_of(work, struct nvmet_fc_fcp_iod, done_work);
2066 nvmet_fc_fod_op_done(fod);
2070 nvmet_fc_xmt_fcp_op_done(struct nvmefc_tgt_fcp_req *fcpreq)
2072 struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
2073 struct nvmet_fc_tgt_queue *queue = fod->queue;
2075 if (fod->tgtport->ops->target_features & NVMET_FCTGTFEAT_OPDONE_IN_ISR)
2076 /* context switch so completion is not in ISR context */
2077 queue_work_on(queue->cpu, queue->work_q, &fod->done_work);
2079 nvmet_fc_fod_op_done(fod);
2083 * actual completion handler after execution by the nvmet layer
2086 __nvmet_fc_fcp_nvme_cmd_done(struct nvmet_fc_tgtport *tgtport,
2087 struct nvmet_fc_fcp_iod *fod, int status)
2089 struct nvme_common_command *sqe = &fod->cmdiubuf.sqe.common;
2090 struct nvme_completion *cqe = &fod->rspiubuf.cqe;
2091 unsigned long flags;
2094 spin_lock_irqsave(&fod->flock, flags);
2096 spin_unlock_irqrestore(&fod->flock, flags);
2098 /* if we have a CQE, snoop the last sq_head value */
2100 fod->queue->sqhd = cqe->sq_head;
2103 nvmet_fc_abort_op(tgtport, fod);
2107 /* if an error handling the cmd post initial parsing */
2109 /* fudge up a failed CQE status for our transport error */
2110 memset(cqe, 0, sizeof(*cqe));
2111 cqe->sq_head = fod->queue->sqhd; /* echo last cqe sqhd */
2112 cqe->sq_id = cpu_to_le16(fod->queue->qid);
2113 cqe->command_id = sqe->command_id;
2114 cqe->status = cpu_to_le16(status);
2118 * try to push the data even if the SQE status is non-zero.
2119 * There may be a status where data still was intended to
2122 if ((fod->io_dir == NVMET_FCP_READ) && (fod->data_sg_cnt)) {
2123 /* push the data over before sending rsp */
2124 nvmet_fc_transfer_fcp_data(tgtport, fod,
2125 NVMET_FCOP_READDATA);
2129 /* writes & no data - fall thru */
2132 /* data no longer needed */
2133 nvmet_fc_free_tgt_pgs(fod);
2135 nvmet_fc_xmt_fcp_rsp(tgtport, fod);
2140 nvmet_fc_fcp_nvme_cmd_done(struct nvmet_req *nvme_req)
2142 struct nvmet_fc_fcp_iod *fod = nvmet_req_to_fod(nvme_req);
2143 struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2145 __nvmet_fc_fcp_nvme_cmd_done(tgtport, fod, 0);
2150 * Actual processing routine for received FC-NVME LS Requests from the LLD
2153 nvmet_fc_handle_fcp_rqst(struct nvmet_fc_tgtport *tgtport,
2154 struct nvmet_fc_fcp_iod *fod)
2156 struct nvme_fc_cmd_iu *cmdiu = &fod->cmdiubuf;
2157 u32 xfrlen = be32_to_cpu(cmdiu->data_len);
2161 * Fused commands are currently not supported in the linux
2164 * As such, the implementation of the FC transport does not
2165 * look at the fused commands and order delivery to the upper
2166 * layer until we have both based on csn.
2169 fod->fcpreq->done = nvmet_fc_xmt_fcp_op_done;
2171 if (cmdiu->flags & FCNVME_CMD_FLAGS_WRITE) {
2172 fod->io_dir = NVMET_FCP_WRITE;
2173 if (!nvme_is_write(&cmdiu->sqe))
2174 goto transport_error;
2175 } else if (cmdiu->flags & FCNVME_CMD_FLAGS_READ) {
2176 fod->io_dir = NVMET_FCP_READ;
2177 if (nvme_is_write(&cmdiu->sqe))
2178 goto transport_error;
2180 fod->io_dir = NVMET_FCP_NODATA;
2182 goto transport_error;
2185 fod->req.cmd = &fod->cmdiubuf.sqe;
2186 fod->req.rsp = &fod->rspiubuf.cqe;
2187 fod->req.port = fod->queue->port;
2189 /* clear any response payload */
2190 memset(&fod->rspiubuf, 0, sizeof(fod->rspiubuf));
2192 fod->data_sg = NULL;
2193 fod->data_sg_cnt = 0;
2195 ret = nvmet_req_init(&fod->req,
2196 &fod->queue->nvme_cq,
2197 &fod->queue->nvme_sq,
2198 &nvmet_fc_tgt_fcp_ops);
2200 /* bad SQE content or invalid ctrl state */
2201 /* nvmet layer has already called op done to send rsp. */
2205 fod->req.transfer_len = xfrlen;
2207 /* keep a running counter of tail position */
2208 atomic_inc(&fod->queue->sqtail);
2210 if (fod->req.transfer_len) {
2211 ret = nvmet_fc_alloc_tgt_pgs(fod);
2213 nvmet_req_complete(&fod->req, ret);
2217 fod->req.sg = fod->data_sg;
2218 fod->req.sg_cnt = fod->data_sg_cnt;
2221 if (fod->io_dir == NVMET_FCP_WRITE) {
2222 /* pull the data over before invoking nvmet layer */
2223 nvmet_fc_transfer_fcp_data(tgtport, fod, NVMET_FCOP_WRITEDATA);
2230 * can invoke the nvmet_layer now. If read data, cmd completion will
2233 nvmet_req_execute(&fod->req);
2237 nvmet_fc_abort_op(tgtport, fod);
2241 * Actual processing routine for received FC-NVME LS Requests from the LLD
2244 nvmet_fc_handle_fcp_rqst_work(struct work_struct *work)
2246 struct nvmet_fc_fcp_iod *fod =
2247 container_of(work, struct nvmet_fc_fcp_iod, work);
2248 struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2250 nvmet_fc_handle_fcp_rqst(tgtport, fod);
2254 * nvmet_fc_rcv_fcp_req - transport entry point called by an LLDD
2255 * upon the reception of a NVME FCP CMD IU.
2257 * Pass a FC-NVME FCP CMD IU received from the FC link to the nvmet-fc
2258 * layer for processing.
2260 * The nvmet_fc layer allocates a local job structure (struct
2261 * nvmet_fc_fcp_iod) from the queue for the io and copies the
2262 * CMD IU buffer to the job structure. As such, on a successful
2263 * completion (returns 0), the LLDD may immediately free/reuse
2264 * the CMD IU buffer passed in the call.
2266 * However, in some circumstances, due to the packetized nature of FC
2267 * and the api of the FC LLDD which may issue a hw command to send the
2268 * response, but the LLDD may not get the hw completion for that command
2269 * and upcall the nvmet_fc layer before a new command may be
2270 * asynchronously received - its possible for a command to be received
2271 * before the LLDD and nvmet_fc have recycled the job structure. It gives
2272 * the appearance of more commands received than fits in the sq.
2273 * To alleviate this scenario, a temporary queue is maintained in the
2274 * transport for pending LLDD requests waiting for a queue job structure.
2275 * In these "overrun" cases, a temporary queue element is allocated
2276 * the LLDD request and CMD iu buffer information remembered, and the
2277 * routine returns a -EOVERFLOW status. Subsequently, when a queue job
2278 * structure is freed, it is immediately reallocated for anything on the
2279 * pending request list. The LLDDs defer_rcv() callback is called,
2280 * informing the LLDD that it may reuse the CMD IU buffer, and the io
2281 * is then started normally with the transport.
2283 * The LLDD, when receiving an -EOVERFLOW completion status, is to treat
2284 * the completion as successful but must not reuse the CMD IU buffer
2285 * until the LLDD's defer_rcv() callback has been called for the
2286 * corresponding struct nvmefc_tgt_fcp_req pointer.
2288 * If there is any other condition in which an error occurs, the
2289 * transport will return a non-zero status indicating the error.
2290 * In all cases other than -EOVERFLOW, the transport has not accepted the
2291 * request and the LLDD should abort the exchange.
2293 * @target_port: pointer to the (registered) target port the FCP CMD IU
2295 * @fcpreq: pointer to a fcpreq request structure to be used to reference
2296 * the exchange corresponding to the FCP Exchange.
2297 * @cmdiubuf: pointer to the buffer containing the FCP CMD IU
2298 * @cmdiubuf_len: length, in bytes, of the received FCP CMD IU
2301 nvmet_fc_rcv_fcp_req(struct nvmet_fc_target_port *target_port,
2302 struct nvmefc_tgt_fcp_req *fcpreq,
2303 void *cmdiubuf, u32 cmdiubuf_len)
2305 struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
2306 struct nvme_fc_cmd_iu *cmdiu = cmdiubuf;
2307 struct nvmet_fc_tgt_queue *queue;
2308 struct nvmet_fc_fcp_iod *fod;
2309 struct nvmet_fc_defer_fcp_req *deferfcp;
2310 unsigned long flags;
2312 /* validate iu, so the connection id can be used to find the queue */
2313 if ((cmdiubuf_len != sizeof(*cmdiu)) ||
2314 (cmdiu->scsi_id != NVME_CMD_SCSI_ID) ||
2315 (cmdiu->fc_id != NVME_CMD_FC_ID) ||
2316 (be16_to_cpu(cmdiu->iu_len) != (sizeof(*cmdiu)/4)))
2319 queue = nvmet_fc_find_target_queue(tgtport,
2320 be64_to_cpu(cmdiu->connection_id));
2325 * note: reference taken by find_target_queue
2326 * After successful fod allocation, the fod will inherit the
2327 * ownership of that reference and will remove the reference
2328 * when the fod is freed.
2331 spin_lock_irqsave(&queue->qlock, flags);
2333 fod = nvmet_fc_alloc_fcp_iod(queue);
2335 spin_unlock_irqrestore(&queue->qlock, flags);
2337 fcpreq->nvmet_fc_private = fod;
2338 fod->fcpreq = fcpreq;
2340 memcpy(&fod->cmdiubuf, cmdiubuf, cmdiubuf_len);
2342 nvmet_fc_queue_fcp_req(tgtport, queue, fcpreq);
2347 if (!tgtport->ops->defer_rcv) {
2348 spin_unlock_irqrestore(&queue->qlock, flags);
2349 /* release the queue lookup reference */
2350 nvmet_fc_tgt_q_put(queue);
2354 deferfcp = list_first_entry_or_null(&queue->avail_defer_list,
2355 struct nvmet_fc_defer_fcp_req, req_list);
2357 /* Just re-use one that was previously allocated */
2358 list_del(&deferfcp->req_list);
2360 spin_unlock_irqrestore(&queue->qlock, flags);
2362 /* Now we need to dynamically allocate one */
2363 deferfcp = kmalloc(sizeof(*deferfcp), GFP_KERNEL);
2365 /* release the queue lookup reference */
2366 nvmet_fc_tgt_q_put(queue);
2369 spin_lock_irqsave(&queue->qlock, flags);
2372 /* For now, use rspaddr / rsplen to save payload information */
2373 fcpreq->rspaddr = cmdiubuf;
2374 fcpreq->rsplen = cmdiubuf_len;
2375 deferfcp->fcp_req = fcpreq;
2377 /* defer processing till a fod becomes available */
2378 list_add_tail(&deferfcp->req_list, &queue->pending_cmd_list);
2380 /* NOTE: the queue lookup reference is still valid */
2382 spin_unlock_irqrestore(&queue->qlock, flags);
2386 EXPORT_SYMBOL_GPL(nvmet_fc_rcv_fcp_req);
2389 * nvmet_fc_rcv_fcp_abort - transport entry point called by an LLDD
2390 * upon the reception of an ABTS for a FCP command
2392 * Notify the transport that an ABTS has been received for a FCP command
2393 * that had been given to the transport via nvmet_fc_rcv_fcp_req(). The
2394 * LLDD believes the command is still being worked on
2395 * (template_ops->fcp_req_release() has not been called).
2397 * The transport will wait for any outstanding work (an op to the LLDD,
2398 * which the lldd should complete with error due to the ABTS; or the
2399 * completion from the nvmet layer of the nvme command), then will
2400 * stop processing and call the nvmet_fc_rcv_fcp_req() callback to
2401 * return the i/o context to the LLDD. The LLDD may send the BA_ACC
2402 * to the ABTS either after return from this function (assuming any
2403 * outstanding op work has been terminated) or upon the callback being
2406 * @target_port: pointer to the (registered) target port the FCP CMD IU
2408 * @fcpreq: pointer to the fcpreq request structure that corresponds
2409 * to the exchange that received the ABTS.
2412 nvmet_fc_rcv_fcp_abort(struct nvmet_fc_target_port *target_port,
2413 struct nvmefc_tgt_fcp_req *fcpreq)
2415 struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
2416 struct nvmet_fc_tgt_queue *queue;
2417 unsigned long flags;
2419 if (!fod || fod->fcpreq != fcpreq)
2420 /* job appears to have already completed, ignore abort */
2425 spin_lock_irqsave(&queue->qlock, flags);
2428 * mark as abort. The abort handler, invoked upon completion
2429 * of any work, will detect the aborted status and do the
2432 spin_lock(&fod->flock);
2434 fod->aborted = true;
2435 spin_unlock(&fod->flock);
2437 spin_unlock_irqrestore(&queue->qlock, flags);
2439 EXPORT_SYMBOL_GPL(nvmet_fc_rcv_fcp_abort);
2442 struct nvmet_fc_traddr {
2448 __nvme_fc_parse_u64(substring_t *sstr, u64 *val)
2452 if (match_u64(sstr, &token64))
2460 * This routine validates and extracts the WWN's from the TRADDR string.
2461 * As kernel parsers need the 0x to determine number base, universally
2462 * build string to parse with 0x prefix before parsing name strings.
2465 nvme_fc_parse_traddr(struct nvmet_fc_traddr *traddr, char *buf, size_t blen)
2467 char name[2 + NVME_FC_TRADDR_HEXNAMELEN + 1];
2468 substring_t wwn = { name, &name[sizeof(name)-1] };
2469 int nnoffset, pnoffset;
2471 /* validate it string one of the 2 allowed formats */
2472 if (strnlen(buf, blen) == NVME_FC_TRADDR_MAXLENGTH &&
2473 !strncmp(buf, "nn-0x", NVME_FC_TRADDR_OXNNLEN) &&
2474 !strncmp(&buf[NVME_FC_TRADDR_MAX_PN_OFFSET],
2475 "pn-0x", NVME_FC_TRADDR_OXNNLEN)) {
2476 nnoffset = NVME_FC_TRADDR_OXNNLEN;
2477 pnoffset = NVME_FC_TRADDR_MAX_PN_OFFSET +
2478 NVME_FC_TRADDR_OXNNLEN;
2479 } else if ((strnlen(buf, blen) == NVME_FC_TRADDR_MINLENGTH &&
2480 !strncmp(buf, "nn-", NVME_FC_TRADDR_NNLEN) &&
2481 !strncmp(&buf[NVME_FC_TRADDR_MIN_PN_OFFSET],
2482 "pn-", NVME_FC_TRADDR_NNLEN))) {
2483 nnoffset = NVME_FC_TRADDR_NNLEN;
2484 pnoffset = NVME_FC_TRADDR_MIN_PN_OFFSET + NVME_FC_TRADDR_NNLEN;
2490 name[2 + NVME_FC_TRADDR_HEXNAMELEN] = 0;
2492 memcpy(&name[2], &buf[nnoffset], NVME_FC_TRADDR_HEXNAMELEN);
2493 if (__nvme_fc_parse_u64(&wwn, &traddr->nn))
2496 memcpy(&name[2], &buf[pnoffset], NVME_FC_TRADDR_HEXNAMELEN);
2497 if (__nvme_fc_parse_u64(&wwn, &traddr->pn))
2503 pr_warn("%s: bad traddr string\n", __func__);
2508 nvmet_fc_add_port(struct nvmet_port *port)
2510 struct nvmet_fc_tgtport *tgtport;
2511 struct nvmet_fc_traddr traddr = { 0L, 0L };
2512 unsigned long flags;
2515 /* validate the address info */
2516 if ((port->disc_addr.trtype != NVMF_TRTYPE_FC) ||
2517 (port->disc_addr.adrfam != NVMF_ADDR_FAMILY_FC))
2520 /* map the traddr address info to a target port */
2522 ret = nvme_fc_parse_traddr(&traddr, port->disc_addr.traddr,
2523 sizeof(port->disc_addr.traddr));
2528 spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
2529 list_for_each_entry(tgtport, &nvmet_fc_target_list, tgt_list) {
2530 if ((tgtport->fc_target_port.node_name == traddr.nn) &&
2531 (tgtport->fc_target_port.port_name == traddr.pn)) {
2532 tgtport->port = port;
2537 spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
2542 nvmet_fc_remove_port(struct nvmet_port *port)
2547 static const struct nvmet_fabrics_ops nvmet_fc_tgt_fcp_ops = {
2548 .owner = THIS_MODULE,
2549 .type = NVMF_TRTYPE_FC,
2551 .add_port = nvmet_fc_add_port,
2552 .remove_port = nvmet_fc_remove_port,
2553 .queue_response = nvmet_fc_fcp_nvme_cmd_done,
2554 .delete_ctrl = nvmet_fc_delete_ctrl,
2557 static int __init nvmet_fc_init_module(void)
2559 return nvmet_register_transport(&nvmet_fc_tgt_fcp_ops);
2562 static void __exit nvmet_fc_exit_module(void)
2564 /* sanity check - all lports should be removed */
2565 if (!list_empty(&nvmet_fc_target_list))
2566 pr_warn("%s: targetport list not empty\n", __func__);
2568 nvmet_unregister_transport(&nvmet_fc_tgt_fcp_ops);
2570 ida_destroy(&nvmet_fc_tgtport_cnt);
2573 module_init(nvmet_fc_init_module);
2574 module_exit(nvmet_fc_exit_module);
2576 MODULE_LICENSE("GPL v2");