1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2016 Avago Technologies. All rights reserved.
5 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
6 #include <linux/module.h>
7 #include <linux/slab.h>
8 #include <linux/blk-mq.h>
9 #include <linux/parser.h>
10 #include <linux/random.h>
11 #include <uapi/scsi/fc/fc_fs.h>
12 #include <uapi/scsi/fc/fc_els.h>
15 #include <linux/nvme-fc-driver.h>
16 #include <linux/nvme-fc.h>
17 #include "../host/fc.h"
20 /* *************************** Data Structures/Defines ****************** */
23 #define NVMET_LS_CTX_COUNT 256
25 struct nvmet_fc_tgtport;
26 struct nvmet_fc_tgt_assoc;
28 struct nvmet_fc_ls_iod { /* for an LS RQST RCV */
29 struct nvmefc_ls_rsp *lsrsp;
30 struct nvmefc_tgt_fcp_req *fcpreq; /* only if RS */
32 struct list_head ls_rcv_list; /* tgtport->ls_rcv_list */
34 struct nvmet_fc_tgtport *tgtport;
35 struct nvmet_fc_tgt_assoc *assoc;
38 union nvmefc_ls_requests *rqstbuf;
39 union nvmefc_ls_responses *rspbuf;
43 struct scatterlist sg[2];
45 struct work_struct work;
46 } __aligned(sizeof(unsigned long long));
48 struct nvmet_fc_ls_req_op { /* for an LS RQST XMT */
49 struct nvmefc_ls_req ls_req;
51 struct nvmet_fc_tgtport *tgtport;
55 struct list_head lsreq_list; /* tgtport->ls_req_list */
60 /* desired maximum for a single sequence - if sg list allows it */
61 #define NVMET_FC_MAX_SEQ_LENGTH (256 * 1024)
63 enum nvmet_fcp_datadir {
70 struct nvmet_fc_fcp_iod {
71 struct nvmefc_tgt_fcp_req *fcpreq;
73 struct nvme_fc_cmd_iu cmdiubuf;
74 struct nvme_fc_ersp_iu rspiubuf;
76 struct scatterlist *next_sg;
77 struct scatterlist *data_sg;
80 enum nvmet_fcp_datadir io_dir;
88 struct work_struct defer_work;
90 struct nvmet_fc_tgtport *tgtport;
91 struct nvmet_fc_tgt_queue *queue;
93 struct list_head fcp_list; /* tgtport->fcp_list */
96 struct nvmet_fc_tgtport {
97 struct nvmet_fc_target_port fc_target_port;
99 struct list_head tgt_list; /* nvmet_fc_target_list */
100 struct device *dev; /* dev for dma mapping */
101 struct nvmet_fc_target_template *ops;
103 struct nvmet_fc_ls_iod *iod;
105 struct list_head ls_rcv_list;
106 struct list_head ls_req_list;
107 struct list_head ls_busylist;
108 struct list_head assoc_list;
109 struct list_head host_list;
110 struct ida assoc_cnt;
111 struct nvmet_fc_port_entry *pe;
116 struct nvmet_fc_port_entry {
117 struct nvmet_fc_tgtport *tgtport;
118 struct nvmet_port *port;
121 struct list_head pe_list;
124 struct nvmet_fc_defer_fcp_req {
125 struct list_head req_list;
126 struct nvmefc_tgt_fcp_req *fcp_req;
129 struct nvmet_fc_tgt_queue {
140 struct nvmet_cq nvme_cq;
141 struct nvmet_sq nvme_sq;
142 struct nvmet_fc_tgt_assoc *assoc;
143 struct list_head fod_list;
144 struct list_head pending_cmd_list;
145 struct list_head avail_defer_list;
146 struct workqueue_struct *work_q;
148 struct nvmet_fc_fcp_iod fod[]; /* array of fcp_iods */
149 } __aligned(sizeof(unsigned long long));
151 struct nvmet_fc_hostport {
152 struct nvmet_fc_tgtport *tgtport;
154 struct list_head host_list;
159 struct nvmet_fc_tgt_assoc {
162 atomic_t terminating;
163 struct nvmet_fc_tgtport *tgtport;
164 struct nvmet_fc_hostport *hostport;
165 struct nvmet_fc_ls_iod *rcv_disconn;
166 struct list_head a_list;
167 struct nvmet_fc_tgt_queue *queues[NVMET_NR_QUEUES + 1];
169 struct work_struct del_work;
174 nvmet_fc_iodnum(struct nvmet_fc_ls_iod *iodptr)
176 return (iodptr - iodptr->tgtport->iod);
180 nvmet_fc_fodnum(struct nvmet_fc_fcp_iod *fodptr)
182 return (fodptr - fodptr->queue->fod);
187 * Association and Connection IDs:
189 * Association ID will have random number in upper 6 bytes and zero
192 * Connection IDs will be Association ID with QID or'd in lower 2 bytes
194 * note: Association ID = Connection ID for queue 0
196 #define BYTES_FOR_QID sizeof(u16)
197 #define BYTES_FOR_QID_SHIFT (BYTES_FOR_QID * 8)
198 #define NVMET_FC_QUEUEID_MASK ((u64)((1 << BYTES_FOR_QID_SHIFT) - 1))
201 nvmet_fc_makeconnid(struct nvmet_fc_tgt_assoc *assoc, u16 qid)
203 return (assoc->association_id | qid);
207 nvmet_fc_getassociationid(u64 connectionid)
209 return connectionid & ~NVMET_FC_QUEUEID_MASK;
213 nvmet_fc_getqueueid(u64 connectionid)
215 return (u16)(connectionid & NVMET_FC_QUEUEID_MASK);
218 static inline struct nvmet_fc_tgtport *
219 targetport_to_tgtport(struct nvmet_fc_target_port *targetport)
221 return container_of(targetport, struct nvmet_fc_tgtport,
225 static inline struct nvmet_fc_fcp_iod *
226 nvmet_req_to_fod(struct nvmet_req *nvme_req)
228 return container_of(nvme_req, struct nvmet_fc_fcp_iod, req);
232 /* *************************** Globals **************************** */
235 static DEFINE_SPINLOCK(nvmet_fc_tgtlock);
237 static LIST_HEAD(nvmet_fc_target_list);
238 static DEFINE_IDA(nvmet_fc_tgtport_cnt);
239 static LIST_HEAD(nvmet_fc_portentry_list);
242 static void nvmet_fc_handle_ls_rqst_work(struct work_struct *work);
243 static void nvmet_fc_fcp_rqst_op_defer_work(struct work_struct *work);
244 static void nvmet_fc_tgt_a_put(struct nvmet_fc_tgt_assoc *assoc);
245 static int nvmet_fc_tgt_a_get(struct nvmet_fc_tgt_assoc *assoc);
246 static void nvmet_fc_tgt_q_put(struct nvmet_fc_tgt_queue *queue);
247 static int nvmet_fc_tgt_q_get(struct nvmet_fc_tgt_queue *queue);
248 static void nvmet_fc_tgtport_put(struct nvmet_fc_tgtport *tgtport);
249 static int nvmet_fc_tgtport_get(struct nvmet_fc_tgtport *tgtport);
250 static void nvmet_fc_handle_fcp_rqst(struct nvmet_fc_tgtport *tgtport,
251 struct nvmet_fc_fcp_iod *fod);
252 static void nvmet_fc_delete_target_assoc(struct nvmet_fc_tgt_assoc *assoc);
253 static void nvmet_fc_xmt_ls_rsp(struct nvmet_fc_tgtport *tgtport,
254 struct nvmet_fc_ls_iod *iod);
257 /* *********************** FC-NVME DMA Handling **************************** */
260 * The fcloop device passes in a NULL device pointer. Real LLD's will
261 * pass in a valid device pointer. If NULL is passed to the dma mapping
262 * routines, depending on the platform, it may or may not succeed, and
266 * Wrapper all the dma routines and check the dev pointer.
268 * If simple mappings (return just a dma address, we'll noop them,
269 * returning a dma address of 0.
271 * On more complex mappings (dma_map_sg), a pseudo routine fills
272 * in the scatter list, setting all dma addresses to 0.
275 static inline dma_addr_t
276 fc_dma_map_single(struct device *dev, void *ptr, size_t size,
277 enum dma_data_direction dir)
279 return dev ? dma_map_single(dev, ptr, size, dir) : (dma_addr_t)0L;
283 fc_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
285 return dev ? dma_mapping_error(dev, dma_addr) : 0;
289 fc_dma_unmap_single(struct device *dev, dma_addr_t addr, size_t size,
290 enum dma_data_direction dir)
293 dma_unmap_single(dev, addr, size, dir);
297 fc_dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr, size_t size,
298 enum dma_data_direction dir)
301 dma_sync_single_for_cpu(dev, addr, size, dir);
305 fc_dma_sync_single_for_device(struct device *dev, dma_addr_t addr, size_t size,
306 enum dma_data_direction dir)
309 dma_sync_single_for_device(dev, addr, size, dir);
312 /* pseudo dma_map_sg call */
314 fc_map_sg(struct scatterlist *sg, int nents)
316 struct scatterlist *s;
319 WARN_ON(nents == 0 || sg[0].length == 0);
321 for_each_sg(sg, s, nents, i) {
323 #ifdef CONFIG_NEED_SG_DMA_LENGTH
324 s->dma_length = s->length;
331 fc_dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
332 enum dma_data_direction dir)
334 return dev ? dma_map_sg(dev, sg, nents, dir) : fc_map_sg(sg, nents);
338 fc_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
339 enum dma_data_direction dir)
342 dma_unmap_sg(dev, sg, nents, dir);
346 /* ********************** FC-NVME LS XMT Handling ************************* */
350 __nvmet_fc_finish_ls_req(struct nvmet_fc_ls_req_op *lsop)
352 struct nvmet_fc_tgtport *tgtport = lsop->tgtport;
353 struct nvmefc_ls_req *lsreq = &lsop->ls_req;
356 spin_lock_irqsave(&tgtport->lock, flags);
358 if (!lsop->req_queued) {
359 spin_unlock_irqrestore(&tgtport->lock, flags);
363 list_del(&lsop->lsreq_list);
365 lsop->req_queued = false;
367 spin_unlock_irqrestore(&tgtport->lock, flags);
369 fc_dma_unmap_single(tgtport->dev, lsreq->rqstdma,
370 (lsreq->rqstlen + lsreq->rsplen),
373 nvmet_fc_tgtport_put(tgtport);
377 __nvmet_fc_send_ls_req(struct nvmet_fc_tgtport *tgtport,
378 struct nvmet_fc_ls_req_op *lsop,
379 void (*done)(struct nvmefc_ls_req *req, int status))
381 struct nvmefc_ls_req *lsreq = &lsop->ls_req;
385 if (!tgtport->ops->ls_req)
388 if (!nvmet_fc_tgtport_get(tgtport))
392 lsop->req_queued = false;
393 INIT_LIST_HEAD(&lsop->lsreq_list);
395 lsreq->rqstdma = fc_dma_map_single(tgtport->dev, lsreq->rqstaddr,
396 lsreq->rqstlen + lsreq->rsplen,
398 if (fc_dma_mapping_error(tgtport->dev, lsreq->rqstdma)) {
402 lsreq->rspdma = lsreq->rqstdma + lsreq->rqstlen;
404 spin_lock_irqsave(&tgtport->lock, flags);
406 list_add_tail(&lsop->lsreq_list, &tgtport->ls_req_list);
408 lsop->req_queued = true;
410 spin_unlock_irqrestore(&tgtport->lock, flags);
412 ret = tgtport->ops->ls_req(&tgtport->fc_target_port, lsop->hosthandle,
420 lsop->ls_error = ret;
421 spin_lock_irqsave(&tgtport->lock, flags);
422 lsop->req_queued = false;
423 list_del(&lsop->lsreq_list);
424 spin_unlock_irqrestore(&tgtport->lock, flags);
425 fc_dma_unmap_single(tgtport->dev, lsreq->rqstdma,
426 (lsreq->rqstlen + lsreq->rsplen),
429 nvmet_fc_tgtport_put(tgtport);
435 nvmet_fc_send_ls_req_async(struct nvmet_fc_tgtport *tgtport,
436 struct nvmet_fc_ls_req_op *lsop,
437 void (*done)(struct nvmefc_ls_req *req, int status))
439 /* don't wait for completion */
441 return __nvmet_fc_send_ls_req(tgtport, lsop, done);
445 nvmet_fc_disconnect_assoc_done(struct nvmefc_ls_req *lsreq, int status)
447 struct nvmet_fc_ls_req_op *lsop =
448 container_of(lsreq, struct nvmet_fc_ls_req_op, ls_req);
450 __nvmet_fc_finish_ls_req(lsop);
452 /* fc-nvme target doesn't care about success or failure of cmd */
458 * This routine sends a FC-NVME LS to disconnect (aka terminate)
459 * the FC-NVME Association. Terminating the association also
460 * terminates the FC-NVME connections (per queue, both admin and io
461 * queues) that are part of the association. E.g. things are torn
462 * down, and the related FC-NVME Association ID and Connection IDs
465 * The behavior of the fc-nvme target is such that it's
466 * understanding of the association and connections will implicitly
467 * be torn down. The action is implicit as it may be due to a loss of
468 * connectivity with the fc-nvme host, so the target may never get a
469 * response even if it tried. As such, the action of this routine
470 * is to asynchronously send the LS, ignore any results of the LS, and
471 * continue on with terminating the association. If the fc-nvme host
472 * is present and receives the LS, it too can tear down.
475 nvmet_fc_xmt_disconnect_assoc(struct nvmet_fc_tgt_assoc *assoc)
477 struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
478 struct fcnvme_ls_disconnect_assoc_rqst *discon_rqst;
479 struct fcnvme_ls_disconnect_assoc_acc *discon_acc;
480 struct nvmet_fc_ls_req_op *lsop;
481 struct nvmefc_ls_req *lsreq;
485 * If ls_req is NULL or no hosthandle, it's an older lldd and no
486 * message is normal. Otherwise, send unless the hostport has
487 * already been invalidated by the lldd.
489 if (!tgtport->ops->ls_req || !assoc->hostport ||
490 assoc->hostport->invalid)
493 lsop = kzalloc((sizeof(*lsop) +
494 sizeof(*discon_rqst) + sizeof(*discon_acc) +
495 tgtport->ops->lsrqst_priv_sz), GFP_KERNEL);
497 dev_info(tgtport->dev,
498 "{%d:%d} send Disconnect Association failed: ENOMEM\n",
499 tgtport->fc_target_port.port_num, assoc->a_id);
503 discon_rqst = (struct fcnvme_ls_disconnect_assoc_rqst *)&lsop[1];
504 discon_acc = (struct fcnvme_ls_disconnect_assoc_acc *)&discon_rqst[1];
505 lsreq = &lsop->ls_req;
506 if (tgtport->ops->lsrqst_priv_sz)
507 lsreq->private = (void *)&discon_acc[1];
509 lsreq->private = NULL;
511 lsop->tgtport = tgtport;
512 lsop->hosthandle = assoc->hostport->hosthandle;
514 nvmefc_fmt_lsreq_discon_assoc(lsreq, discon_rqst, discon_acc,
515 assoc->association_id);
517 ret = nvmet_fc_send_ls_req_async(tgtport, lsop,
518 nvmet_fc_disconnect_assoc_done);
520 dev_info(tgtport->dev,
521 "{%d:%d} XMT Disconnect Association failed: %d\n",
522 tgtport->fc_target_port.port_num, assoc->a_id, ret);
528 /* *********************** FC-NVME Port Management ************************ */
532 nvmet_fc_alloc_ls_iodlist(struct nvmet_fc_tgtport *tgtport)
534 struct nvmet_fc_ls_iod *iod;
537 iod = kcalloc(NVMET_LS_CTX_COUNT, sizeof(struct nvmet_fc_ls_iod),
544 for (i = 0; i < NVMET_LS_CTX_COUNT; iod++, i++) {
545 INIT_WORK(&iod->work, nvmet_fc_handle_ls_rqst_work);
546 iod->tgtport = tgtport;
547 list_add_tail(&iod->ls_rcv_list, &tgtport->ls_rcv_list);
549 iod->rqstbuf = kzalloc(sizeof(union nvmefc_ls_requests) +
550 sizeof(union nvmefc_ls_responses),
555 iod->rspbuf = (union nvmefc_ls_responses *)&iod->rqstbuf[1];
557 iod->rspdma = fc_dma_map_single(tgtport->dev, iod->rspbuf,
558 sizeof(*iod->rspbuf),
560 if (fc_dma_mapping_error(tgtport->dev, iod->rspdma))
568 list_del(&iod->ls_rcv_list);
569 for (iod--, i--; i >= 0; iod--, i--) {
570 fc_dma_unmap_single(tgtport->dev, iod->rspdma,
571 sizeof(*iod->rspbuf), DMA_TO_DEVICE);
573 list_del(&iod->ls_rcv_list);
582 nvmet_fc_free_ls_iodlist(struct nvmet_fc_tgtport *tgtport)
584 struct nvmet_fc_ls_iod *iod = tgtport->iod;
587 for (i = 0; i < NVMET_LS_CTX_COUNT; iod++, i++) {
588 fc_dma_unmap_single(tgtport->dev,
589 iod->rspdma, sizeof(*iod->rspbuf),
592 list_del(&iod->ls_rcv_list);
597 static struct nvmet_fc_ls_iod *
598 nvmet_fc_alloc_ls_iod(struct nvmet_fc_tgtport *tgtport)
600 struct nvmet_fc_ls_iod *iod;
603 spin_lock_irqsave(&tgtport->lock, flags);
604 iod = list_first_entry_or_null(&tgtport->ls_rcv_list,
605 struct nvmet_fc_ls_iod, ls_rcv_list);
607 list_move_tail(&iod->ls_rcv_list, &tgtport->ls_busylist);
608 spin_unlock_irqrestore(&tgtport->lock, flags);
614 nvmet_fc_free_ls_iod(struct nvmet_fc_tgtport *tgtport,
615 struct nvmet_fc_ls_iod *iod)
619 spin_lock_irqsave(&tgtport->lock, flags);
620 list_move(&iod->ls_rcv_list, &tgtport->ls_rcv_list);
621 spin_unlock_irqrestore(&tgtport->lock, flags);
625 nvmet_fc_prep_fcp_iodlist(struct nvmet_fc_tgtport *tgtport,
626 struct nvmet_fc_tgt_queue *queue)
628 struct nvmet_fc_fcp_iod *fod = queue->fod;
631 for (i = 0; i < queue->sqsize; fod++, i++) {
632 INIT_WORK(&fod->defer_work, nvmet_fc_fcp_rqst_op_defer_work);
633 fod->tgtport = tgtport;
637 fod->aborted = false;
639 list_add_tail(&fod->fcp_list, &queue->fod_list);
640 spin_lock_init(&fod->flock);
642 fod->rspdma = fc_dma_map_single(tgtport->dev, &fod->rspiubuf,
643 sizeof(fod->rspiubuf), DMA_TO_DEVICE);
644 if (fc_dma_mapping_error(tgtport->dev, fod->rspdma)) {
645 list_del(&fod->fcp_list);
646 for (fod--, i--; i >= 0; fod--, i--) {
647 fc_dma_unmap_single(tgtport->dev, fod->rspdma,
648 sizeof(fod->rspiubuf),
651 list_del(&fod->fcp_list);
660 nvmet_fc_destroy_fcp_iodlist(struct nvmet_fc_tgtport *tgtport,
661 struct nvmet_fc_tgt_queue *queue)
663 struct nvmet_fc_fcp_iod *fod = queue->fod;
666 for (i = 0; i < queue->sqsize; fod++, i++) {
668 fc_dma_unmap_single(tgtport->dev, fod->rspdma,
669 sizeof(fod->rspiubuf), DMA_TO_DEVICE);
673 static struct nvmet_fc_fcp_iod *
674 nvmet_fc_alloc_fcp_iod(struct nvmet_fc_tgt_queue *queue)
676 struct nvmet_fc_fcp_iod *fod;
678 lockdep_assert_held(&queue->qlock);
680 fod = list_first_entry_or_null(&queue->fod_list,
681 struct nvmet_fc_fcp_iod, fcp_list);
683 list_del(&fod->fcp_list);
686 * no queue reference is taken, as it was taken by the
687 * queue lookup just prior to the allocation. The iod
688 * will "inherit" that reference.
696 nvmet_fc_queue_fcp_req(struct nvmet_fc_tgtport *tgtport,
697 struct nvmet_fc_tgt_queue *queue,
698 struct nvmefc_tgt_fcp_req *fcpreq)
700 struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
703 * put all admin cmds on hw queue id 0. All io commands go to
704 * the respective hw queue based on a modulo basis
706 fcpreq->hwqid = queue->qid ?
707 ((queue->qid - 1) % tgtport->ops->max_hw_queues) : 0;
709 nvmet_fc_handle_fcp_rqst(tgtport, fod);
713 nvmet_fc_fcp_rqst_op_defer_work(struct work_struct *work)
715 struct nvmet_fc_fcp_iod *fod =
716 container_of(work, struct nvmet_fc_fcp_iod, defer_work);
718 /* Submit deferred IO for processing */
719 nvmet_fc_queue_fcp_req(fod->tgtport, fod->queue, fod->fcpreq);
724 nvmet_fc_free_fcp_iod(struct nvmet_fc_tgt_queue *queue,
725 struct nvmet_fc_fcp_iod *fod)
727 struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
728 struct nvmet_fc_tgtport *tgtport = fod->tgtport;
729 struct nvmet_fc_defer_fcp_req *deferfcp;
732 fc_dma_sync_single_for_cpu(tgtport->dev, fod->rspdma,
733 sizeof(fod->rspiubuf), DMA_TO_DEVICE);
735 fcpreq->nvmet_fc_private = NULL;
739 fod->aborted = false;
740 fod->writedataactive = false;
743 tgtport->ops->fcp_req_release(&tgtport->fc_target_port, fcpreq);
745 /* release the queue lookup reference on the completed IO */
746 nvmet_fc_tgt_q_put(queue);
748 spin_lock_irqsave(&queue->qlock, flags);
749 deferfcp = list_first_entry_or_null(&queue->pending_cmd_list,
750 struct nvmet_fc_defer_fcp_req, req_list);
752 list_add_tail(&fod->fcp_list, &fod->queue->fod_list);
753 spin_unlock_irqrestore(&queue->qlock, flags);
757 /* Re-use the fod for the next pending cmd that was deferred */
758 list_del(&deferfcp->req_list);
760 fcpreq = deferfcp->fcp_req;
762 /* deferfcp can be reused for another IO at a later date */
763 list_add_tail(&deferfcp->req_list, &queue->avail_defer_list);
765 spin_unlock_irqrestore(&queue->qlock, flags);
767 /* Save NVME CMD IO in fod */
768 memcpy(&fod->cmdiubuf, fcpreq->rspaddr, fcpreq->rsplen);
770 /* Setup new fcpreq to be processed */
771 fcpreq->rspaddr = NULL;
773 fcpreq->nvmet_fc_private = fod;
774 fod->fcpreq = fcpreq;
777 /* inform LLDD IO is now being processed */
778 tgtport->ops->defer_rcv(&tgtport->fc_target_port, fcpreq);
781 * Leave the queue lookup get reference taken when
782 * fod was originally allocated.
785 queue_work(queue->work_q, &fod->defer_work);
788 static struct nvmet_fc_tgt_queue *
789 nvmet_fc_alloc_target_queue(struct nvmet_fc_tgt_assoc *assoc,
792 struct nvmet_fc_tgt_queue *queue;
796 if (qid > NVMET_NR_QUEUES)
799 queue = kzalloc(struct_size(queue, fod, sqsize), GFP_KERNEL);
803 if (!nvmet_fc_tgt_a_get(assoc))
806 queue->work_q = alloc_workqueue("ntfc%d.%d.%d", 0, 0,
807 assoc->tgtport->fc_target_port.port_num,
813 queue->sqsize = sqsize;
814 queue->assoc = assoc;
815 INIT_LIST_HEAD(&queue->fod_list);
816 INIT_LIST_HEAD(&queue->avail_defer_list);
817 INIT_LIST_HEAD(&queue->pending_cmd_list);
818 atomic_set(&queue->connected, 0);
819 atomic_set(&queue->sqtail, 0);
820 atomic_set(&queue->rsn, 1);
821 atomic_set(&queue->zrspcnt, 0);
822 spin_lock_init(&queue->qlock);
823 kref_init(&queue->ref);
825 nvmet_fc_prep_fcp_iodlist(assoc->tgtport, queue);
827 ret = nvmet_sq_init(&queue->nvme_sq);
829 goto out_fail_iodlist;
831 WARN_ON(assoc->queues[qid]);
832 spin_lock_irqsave(&assoc->tgtport->lock, flags);
833 assoc->queues[qid] = queue;
834 spin_unlock_irqrestore(&assoc->tgtport->lock, flags);
839 nvmet_fc_destroy_fcp_iodlist(assoc->tgtport, queue);
840 destroy_workqueue(queue->work_q);
842 nvmet_fc_tgt_a_put(assoc);
850 nvmet_fc_tgt_queue_free(struct kref *ref)
852 struct nvmet_fc_tgt_queue *queue =
853 container_of(ref, struct nvmet_fc_tgt_queue, ref);
856 spin_lock_irqsave(&queue->assoc->tgtport->lock, flags);
857 queue->assoc->queues[queue->qid] = NULL;
858 spin_unlock_irqrestore(&queue->assoc->tgtport->lock, flags);
860 nvmet_fc_destroy_fcp_iodlist(queue->assoc->tgtport, queue);
862 nvmet_fc_tgt_a_put(queue->assoc);
864 destroy_workqueue(queue->work_q);
870 nvmet_fc_tgt_q_put(struct nvmet_fc_tgt_queue *queue)
872 kref_put(&queue->ref, nvmet_fc_tgt_queue_free);
876 nvmet_fc_tgt_q_get(struct nvmet_fc_tgt_queue *queue)
878 return kref_get_unless_zero(&queue->ref);
883 nvmet_fc_delete_target_queue(struct nvmet_fc_tgt_queue *queue)
885 struct nvmet_fc_tgtport *tgtport = queue->assoc->tgtport;
886 struct nvmet_fc_fcp_iod *fod = queue->fod;
887 struct nvmet_fc_defer_fcp_req *deferfcp, *tempptr;
892 disconnect = atomic_xchg(&queue->connected, 0);
894 /* if not connected, nothing to do */
898 spin_lock_irqsave(&queue->qlock, flags);
899 /* abort outstanding io's */
900 for (i = 0; i < queue->sqsize; fod++, i++) {
902 spin_lock(&fod->flock);
905 * only call lldd abort routine if waiting for
906 * writedata. other outstanding ops should finish
909 if (fod->writedataactive) {
911 spin_unlock(&fod->flock);
912 tgtport->ops->fcp_abort(
913 &tgtport->fc_target_port, fod->fcpreq);
915 spin_unlock(&fod->flock);
919 /* Cleanup defer'ed IOs in queue */
920 list_for_each_entry_safe(deferfcp, tempptr, &queue->avail_defer_list,
922 list_del(&deferfcp->req_list);
927 deferfcp = list_first_entry_or_null(&queue->pending_cmd_list,
928 struct nvmet_fc_defer_fcp_req, req_list);
932 list_del(&deferfcp->req_list);
933 spin_unlock_irqrestore(&queue->qlock, flags);
935 tgtport->ops->defer_rcv(&tgtport->fc_target_port,
938 tgtport->ops->fcp_abort(&tgtport->fc_target_port,
941 tgtport->ops->fcp_req_release(&tgtport->fc_target_port,
944 /* release the queue lookup reference */
945 nvmet_fc_tgt_q_put(queue);
949 spin_lock_irqsave(&queue->qlock, flags);
951 spin_unlock_irqrestore(&queue->qlock, flags);
953 flush_workqueue(queue->work_q);
955 nvmet_sq_destroy(&queue->nvme_sq);
957 nvmet_fc_tgt_q_put(queue);
960 static struct nvmet_fc_tgt_queue *
961 nvmet_fc_find_target_queue(struct nvmet_fc_tgtport *tgtport,
964 struct nvmet_fc_tgt_assoc *assoc;
965 struct nvmet_fc_tgt_queue *queue;
966 u64 association_id = nvmet_fc_getassociationid(connection_id);
967 u16 qid = nvmet_fc_getqueueid(connection_id);
970 if (qid > NVMET_NR_QUEUES)
973 spin_lock_irqsave(&tgtport->lock, flags);
974 list_for_each_entry(assoc, &tgtport->assoc_list, a_list) {
975 if (association_id == assoc->association_id) {
976 queue = assoc->queues[qid];
978 (!atomic_read(&queue->connected) ||
979 !nvmet_fc_tgt_q_get(queue)))
981 spin_unlock_irqrestore(&tgtport->lock, flags);
985 spin_unlock_irqrestore(&tgtport->lock, flags);
990 nvmet_fc_hostport_free(struct kref *ref)
992 struct nvmet_fc_hostport *hostport =
993 container_of(ref, struct nvmet_fc_hostport, ref);
994 struct nvmet_fc_tgtport *tgtport = hostport->tgtport;
997 spin_lock_irqsave(&tgtport->lock, flags);
998 list_del(&hostport->host_list);
999 spin_unlock_irqrestore(&tgtport->lock, flags);
1000 if (tgtport->ops->host_release && hostport->invalid)
1001 tgtport->ops->host_release(hostport->hosthandle);
1003 nvmet_fc_tgtport_put(tgtport);
1007 nvmet_fc_hostport_put(struct nvmet_fc_hostport *hostport)
1009 kref_put(&hostport->ref, nvmet_fc_hostport_free);
1013 nvmet_fc_hostport_get(struct nvmet_fc_hostport *hostport)
1015 return kref_get_unless_zero(&hostport->ref);
1019 nvmet_fc_free_hostport(struct nvmet_fc_hostport *hostport)
1021 /* if LLDD not implemented, leave as NULL */
1022 if (!hostport || !hostport->hosthandle)
1025 nvmet_fc_hostport_put(hostport);
1028 static struct nvmet_fc_hostport *
1029 nvmet_fc_alloc_hostport(struct nvmet_fc_tgtport *tgtport, void *hosthandle)
1031 struct nvmet_fc_hostport *newhost, *host, *match = NULL;
1032 unsigned long flags;
1034 /* if LLDD not implemented, leave as NULL */
1038 /* take reference for what will be the newly allocated hostport */
1039 if (!nvmet_fc_tgtport_get(tgtport))
1040 return ERR_PTR(-EINVAL);
1042 newhost = kzalloc(sizeof(*newhost), GFP_KERNEL);
1044 spin_lock_irqsave(&tgtport->lock, flags);
1045 list_for_each_entry(host, &tgtport->host_list, host_list) {
1046 if (host->hosthandle == hosthandle && !host->invalid) {
1047 if (nvmet_fc_hostport_get(host)) {
1053 spin_unlock_irqrestore(&tgtport->lock, flags);
1054 /* no allocation - release reference */
1055 nvmet_fc_tgtport_put(tgtport);
1056 return (match) ? match : ERR_PTR(-ENOMEM);
1059 newhost->tgtport = tgtport;
1060 newhost->hosthandle = hosthandle;
1061 INIT_LIST_HEAD(&newhost->host_list);
1062 kref_init(&newhost->ref);
1064 spin_lock_irqsave(&tgtport->lock, flags);
1065 list_for_each_entry(host, &tgtport->host_list, host_list) {
1066 if (host->hosthandle == hosthandle && !host->invalid) {
1067 if (nvmet_fc_hostport_get(host)) {
1076 /* releasing allocation - release reference */
1077 nvmet_fc_tgtport_put(tgtport);
1079 list_add_tail(&newhost->host_list, &tgtport->host_list);
1080 spin_unlock_irqrestore(&tgtport->lock, flags);
1082 return (match) ? match : newhost;
1086 nvmet_fc_delete_assoc(struct work_struct *work)
1088 struct nvmet_fc_tgt_assoc *assoc =
1089 container_of(work, struct nvmet_fc_tgt_assoc, del_work);
1091 nvmet_fc_delete_target_assoc(assoc);
1092 nvmet_fc_tgt_a_put(assoc);
1095 static struct nvmet_fc_tgt_assoc *
1096 nvmet_fc_alloc_target_assoc(struct nvmet_fc_tgtport *tgtport, void *hosthandle)
1098 struct nvmet_fc_tgt_assoc *assoc, *tmpassoc;
1099 unsigned long flags;
1102 bool needrandom = true;
1104 assoc = kzalloc(sizeof(*assoc), GFP_KERNEL);
1108 idx = ida_simple_get(&tgtport->assoc_cnt, 0, 0, GFP_KERNEL);
1110 goto out_free_assoc;
1112 if (!nvmet_fc_tgtport_get(tgtport))
1115 assoc->hostport = nvmet_fc_alloc_hostport(tgtport, hosthandle);
1116 if (IS_ERR(assoc->hostport))
1119 assoc->tgtport = tgtport;
1121 INIT_LIST_HEAD(&assoc->a_list);
1122 kref_init(&assoc->ref);
1123 INIT_WORK(&assoc->del_work, nvmet_fc_delete_assoc);
1124 atomic_set(&assoc->terminating, 0);
1126 while (needrandom) {
1127 get_random_bytes(&ran, sizeof(ran) - BYTES_FOR_QID);
1128 ran = ran << BYTES_FOR_QID_SHIFT;
1130 spin_lock_irqsave(&tgtport->lock, flags);
1132 list_for_each_entry(tmpassoc, &tgtport->assoc_list, a_list) {
1133 if (ran == tmpassoc->association_id) {
1139 assoc->association_id = ran;
1140 list_add_tail(&assoc->a_list, &tgtport->assoc_list);
1142 spin_unlock_irqrestore(&tgtport->lock, flags);
1148 nvmet_fc_tgtport_put(tgtport);
1150 ida_simple_remove(&tgtport->assoc_cnt, idx);
1157 nvmet_fc_target_assoc_free(struct kref *ref)
1159 struct nvmet_fc_tgt_assoc *assoc =
1160 container_of(ref, struct nvmet_fc_tgt_assoc, ref);
1161 struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
1162 struct nvmet_fc_ls_iod *oldls;
1163 unsigned long flags;
1165 /* Send Disconnect now that all i/o has completed */
1166 nvmet_fc_xmt_disconnect_assoc(assoc);
1168 nvmet_fc_free_hostport(assoc->hostport);
1169 spin_lock_irqsave(&tgtport->lock, flags);
1170 list_del(&assoc->a_list);
1171 oldls = assoc->rcv_disconn;
1172 spin_unlock_irqrestore(&tgtport->lock, flags);
1173 /* if pending Rcv Disconnect Association LS, send rsp now */
1175 nvmet_fc_xmt_ls_rsp(tgtport, oldls);
1176 ida_simple_remove(&tgtport->assoc_cnt, assoc->a_id);
1177 dev_info(tgtport->dev,
1178 "{%d:%d} Association freed\n",
1179 tgtport->fc_target_port.port_num, assoc->a_id);
1181 nvmet_fc_tgtport_put(tgtport);
1185 nvmet_fc_tgt_a_put(struct nvmet_fc_tgt_assoc *assoc)
1187 kref_put(&assoc->ref, nvmet_fc_target_assoc_free);
1191 nvmet_fc_tgt_a_get(struct nvmet_fc_tgt_assoc *assoc)
1193 return kref_get_unless_zero(&assoc->ref);
1197 nvmet_fc_delete_target_assoc(struct nvmet_fc_tgt_assoc *assoc)
1199 struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
1200 struct nvmet_fc_tgt_queue *queue;
1201 unsigned long flags;
1204 terminating = atomic_xchg(&assoc->terminating, 1);
1206 /* if already terminating, do nothing */
1210 spin_lock_irqsave(&tgtport->lock, flags);
1211 for (i = NVMET_NR_QUEUES; i >= 0; i--) {
1212 queue = assoc->queues[i];
1214 if (!nvmet_fc_tgt_q_get(queue))
1216 spin_unlock_irqrestore(&tgtport->lock, flags);
1217 nvmet_fc_delete_target_queue(queue);
1218 nvmet_fc_tgt_q_put(queue);
1219 spin_lock_irqsave(&tgtport->lock, flags);
1222 spin_unlock_irqrestore(&tgtport->lock, flags);
1224 dev_info(tgtport->dev,
1225 "{%d:%d} Association deleted\n",
1226 tgtport->fc_target_port.port_num, assoc->a_id);
1228 nvmet_fc_tgt_a_put(assoc);
1231 static struct nvmet_fc_tgt_assoc *
1232 nvmet_fc_find_target_assoc(struct nvmet_fc_tgtport *tgtport,
1235 struct nvmet_fc_tgt_assoc *assoc;
1236 struct nvmet_fc_tgt_assoc *ret = NULL;
1237 unsigned long flags;
1239 spin_lock_irqsave(&tgtport->lock, flags);
1240 list_for_each_entry(assoc, &tgtport->assoc_list, a_list) {
1241 if (association_id == assoc->association_id) {
1243 if (!nvmet_fc_tgt_a_get(assoc))
1248 spin_unlock_irqrestore(&tgtport->lock, flags);
1254 nvmet_fc_portentry_bind(struct nvmet_fc_tgtport *tgtport,
1255 struct nvmet_fc_port_entry *pe,
1256 struct nvmet_port *port)
1258 lockdep_assert_held(&nvmet_fc_tgtlock);
1260 pe->tgtport = tgtport;
1266 pe->node_name = tgtport->fc_target_port.node_name;
1267 pe->port_name = tgtport->fc_target_port.port_name;
1268 INIT_LIST_HEAD(&pe->pe_list);
1270 list_add_tail(&pe->pe_list, &nvmet_fc_portentry_list);
1274 nvmet_fc_portentry_unbind(struct nvmet_fc_port_entry *pe)
1276 unsigned long flags;
1278 spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1280 pe->tgtport->pe = NULL;
1281 list_del(&pe->pe_list);
1282 spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1286 * called when a targetport deregisters. Breaks the relationship
1287 * with the nvmet port, but leaves the port_entry in place so that
1288 * re-registration can resume operation.
1291 nvmet_fc_portentry_unbind_tgt(struct nvmet_fc_tgtport *tgtport)
1293 struct nvmet_fc_port_entry *pe;
1294 unsigned long flags;
1296 spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1301 spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1305 * called when a new targetport is registered. Looks in the
1306 * existing nvmet port_entries to see if the nvmet layer is
1307 * configured for the targetport's wwn's. (the targetport existed,
1308 * nvmet configured, the lldd unregistered the tgtport, and is now
1309 * reregistering the same targetport). If so, set the nvmet port
1310 * port entry on the targetport.
1313 nvmet_fc_portentry_rebind_tgt(struct nvmet_fc_tgtport *tgtport)
1315 struct nvmet_fc_port_entry *pe;
1316 unsigned long flags;
1318 spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1319 list_for_each_entry(pe, &nvmet_fc_portentry_list, pe_list) {
1320 if (tgtport->fc_target_port.node_name == pe->node_name &&
1321 tgtport->fc_target_port.port_name == pe->port_name) {
1322 WARN_ON(pe->tgtport);
1324 pe->tgtport = tgtport;
1328 spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1332 * nvme_fc_register_targetport - transport entry point called by an
1333 * LLDD to register the existence of a local
1334 * NVME subystem FC port.
1335 * @pinfo: pointer to information about the port to be registered
1336 * @template: LLDD entrypoints and operational parameters for the port
1337 * @dev: physical hardware device node port corresponds to. Will be
1338 * used for DMA mappings
1339 * @portptr: pointer to a local port pointer. Upon success, the routine
1340 * will allocate a nvme_fc_local_port structure and place its
1341 * address in the local port pointer. Upon failure, local port
1342 * pointer will be set to NULL.
1345 * a completion status. Must be 0 upon success; a negative errno
1346 * (ex: -ENXIO) upon failure.
1349 nvmet_fc_register_targetport(struct nvmet_fc_port_info *pinfo,
1350 struct nvmet_fc_target_template *template,
1352 struct nvmet_fc_target_port **portptr)
1354 struct nvmet_fc_tgtport *newrec;
1355 unsigned long flags;
1358 if (!template->xmt_ls_rsp || !template->fcp_op ||
1359 !template->fcp_abort ||
1360 !template->fcp_req_release || !template->targetport_delete ||
1361 !template->max_hw_queues || !template->max_sgl_segments ||
1362 !template->max_dif_sgl_segments || !template->dma_boundary) {
1364 goto out_regtgt_failed;
1367 newrec = kzalloc((sizeof(*newrec) + template->target_priv_sz),
1371 goto out_regtgt_failed;
1374 idx = ida_simple_get(&nvmet_fc_tgtport_cnt, 0, 0, GFP_KERNEL);
1377 goto out_fail_kfree;
1380 if (!get_device(dev) && dev) {
1385 newrec->fc_target_port.node_name = pinfo->node_name;
1386 newrec->fc_target_port.port_name = pinfo->port_name;
1387 if (template->target_priv_sz)
1388 newrec->fc_target_port.private = &newrec[1];
1390 newrec->fc_target_port.private = NULL;
1391 newrec->fc_target_port.port_id = pinfo->port_id;
1392 newrec->fc_target_port.port_num = idx;
1393 INIT_LIST_HEAD(&newrec->tgt_list);
1395 newrec->ops = template;
1396 spin_lock_init(&newrec->lock);
1397 INIT_LIST_HEAD(&newrec->ls_rcv_list);
1398 INIT_LIST_HEAD(&newrec->ls_req_list);
1399 INIT_LIST_HEAD(&newrec->ls_busylist);
1400 INIT_LIST_HEAD(&newrec->assoc_list);
1401 INIT_LIST_HEAD(&newrec->host_list);
1402 kref_init(&newrec->ref);
1403 ida_init(&newrec->assoc_cnt);
1404 newrec->max_sg_cnt = template->max_sgl_segments;
1406 ret = nvmet_fc_alloc_ls_iodlist(newrec);
1409 goto out_free_newrec;
1412 nvmet_fc_portentry_rebind_tgt(newrec);
1414 spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1415 list_add_tail(&newrec->tgt_list, &nvmet_fc_target_list);
1416 spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1418 *portptr = &newrec->fc_target_port;
1424 ida_simple_remove(&nvmet_fc_tgtport_cnt, idx);
1431 EXPORT_SYMBOL_GPL(nvmet_fc_register_targetport);
1435 nvmet_fc_free_tgtport(struct kref *ref)
1437 struct nvmet_fc_tgtport *tgtport =
1438 container_of(ref, struct nvmet_fc_tgtport, ref);
1439 struct device *dev = tgtport->dev;
1440 unsigned long flags;
1442 spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1443 list_del(&tgtport->tgt_list);
1444 spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1446 nvmet_fc_free_ls_iodlist(tgtport);
1448 /* let the LLDD know we've finished tearing it down */
1449 tgtport->ops->targetport_delete(&tgtport->fc_target_port);
1451 ida_simple_remove(&nvmet_fc_tgtport_cnt,
1452 tgtport->fc_target_port.port_num);
1454 ida_destroy(&tgtport->assoc_cnt);
1462 nvmet_fc_tgtport_put(struct nvmet_fc_tgtport *tgtport)
1464 kref_put(&tgtport->ref, nvmet_fc_free_tgtport);
1468 nvmet_fc_tgtport_get(struct nvmet_fc_tgtport *tgtport)
1470 return kref_get_unless_zero(&tgtport->ref);
1474 __nvmet_fc_free_assocs(struct nvmet_fc_tgtport *tgtport)
1476 struct nvmet_fc_tgt_assoc *assoc, *next;
1477 unsigned long flags;
1479 spin_lock_irqsave(&tgtport->lock, flags);
1480 list_for_each_entry_safe(assoc, next,
1481 &tgtport->assoc_list, a_list) {
1482 if (!nvmet_fc_tgt_a_get(assoc))
1484 if (!schedule_work(&assoc->del_work))
1485 /* already deleting - release local reference */
1486 nvmet_fc_tgt_a_put(assoc);
1488 spin_unlock_irqrestore(&tgtport->lock, flags);
1492 * nvmet_fc_invalidate_host - transport entry point called by an LLDD
1493 * to remove references to a hosthandle for LS's.
1495 * The nvmet-fc layer ensures that any references to the hosthandle
1496 * on the targetport are forgotten (set to NULL). The LLDD will
1497 * typically call this when a login with a remote host port has been
1498 * lost, thus LS's for the remote host port are no longer possible.
1500 * If an LS request is outstanding to the targetport/hosthandle (or
1501 * issued concurrently with the call to invalidate the host), the
1502 * LLDD is responsible for terminating/aborting the LS and completing
1503 * the LS request. It is recommended that these terminations/aborts
1504 * occur after calling to invalidate the host handle to avoid additional
1505 * retries by the nvmet-fc transport. The nvmet-fc transport may
1506 * continue to reference host handle while it cleans up outstanding
1507 * NVME associations. The nvmet-fc transport will call the
1508 * ops->host_release() callback to notify the LLDD that all references
1509 * are complete and the related host handle can be recovered.
1510 * Note: if there are no references, the callback may be called before
1511 * the invalidate host call returns.
1513 * @target_port: pointer to the (registered) target port that a prior
1514 * LS was received on and which supplied the transport the
1516 * @hosthandle: the handle (pointer) that represents the host port
1517 * that no longer has connectivity and that LS's should
1518 * no longer be directed to.
1521 nvmet_fc_invalidate_host(struct nvmet_fc_target_port *target_port,
1524 struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
1525 struct nvmet_fc_tgt_assoc *assoc, *next;
1526 unsigned long flags;
1527 bool noassoc = true;
1529 spin_lock_irqsave(&tgtport->lock, flags);
1530 list_for_each_entry_safe(assoc, next,
1531 &tgtport->assoc_list, a_list) {
1532 if (!assoc->hostport ||
1533 assoc->hostport->hosthandle != hosthandle)
1535 if (!nvmet_fc_tgt_a_get(assoc))
1537 assoc->hostport->invalid = 1;
1539 if (!schedule_work(&assoc->del_work))
1540 /* already deleting - release local reference */
1541 nvmet_fc_tgt_a_put(assoc);
1543 spin_unlock_irqrestore(&tgtport->lock, flags);
1545 /* if there's nothing to wait for - call the callback */
1546 if (noassoc && tgtport->ops->host_release)
1547 tgtport->ops->host_release(hosthandle);
1549 EXPORT_SYMBOL_GPL(nvmet_fc_invalidate_host);
1552 * nvmet layer has called to terminate an association
1555 nvmet_fc_delete_ctrl(struct nvmet_ctrl *ctrl)
1557 struct nvmet_fc_tgtport *tgtport, *next;
1558 struct nvmet_fc_tgt_assoc *assoc;
1559 struct nvmet_fc_tgt_queue *queue;
1560 unsigned long flags;
1561 bool found_ctrl = false;
1563 /* this is a bit ugly, but don't want to make locks layered */
1564 spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1565 list_for_each_entry_safe(tgtport, next, &nvmet_fc_target_list,
1567 if (!nvmet_fc_tgtport_get(tgtport))
1569 spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1571 spin_lock_irqsave(&tgtport->lock, flags);
1572 list_for_each_entry(assoc, &tgtport->assoc_list, a_list) {
1573 queue = assoc->queues[0];
1574 if (queue && queue->nvme_sq.ctrl == ctrl) {
1575 if (nvmet_fc_tgt_a_get(assoc))
1580 spin_unlock_irqrestore(&tgtport->lock, flags);
1582 nvmet_fc_tgtport_put(tgtport);
1585 if (!schedule_work(&assoc->del_work))
1586 /* already deleting - release local reference */
1587 nvmet_fc_tgt_a_put(assoc);
1591 spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1593 spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1597 * nvme_fc_unregister_targetport - transport entry point called by an
1598 * LLDD to deregister/remove a previously
1599 * registered a local NVME subsystem FC port.
1600 * @target_port: pointer to the (registered) target port that is to be
1604 * a completion status. Must be 0 upon success; a negative errno
1605 * (ex: -ENXIO) upon failure.
1608 nvmet_fc_unregister_targetport(struct nvmet_fc_target_port *target_port)
1610 struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
1612 nvmet_fc_portentry_unbind_tgt(tgtport);
1614 /* terminate any outstanding associations */
1615 __nvmet_fc_free_assocs(tgtport);
1618 * should terminate LS's as well. However, LS's will be generated
1619 * at the tail end of association termination, so they likely don't
1620 * exist yet. And even if they did, it's worthwhile to just let
1621 * them finish and targetport ref counting will clean things up.
1624 nvmet_fc_tgtport_put(tgtport);
1628 EXPORT_SYMBOL_GPL(nvmet_fc_unregister_targetport);
1631 /* ********************** FC-NVME LS RCV Handling ************************* */
1635 nvmet_fc_ls_create_association(struct nvmet_fc_tgtport *tgtport,
1636 struct nvmet_fc_ls_iod *iod)
1638 struct fcnvme_ls_cr_assoc_rqst *rqst = &iod->rqstbuf->rq_cr_assoc;
1639 struct fcnvme_ls_cr_assoc_acc *acc = &iod->rspbuf->rsp_cr_assoc;
1640 struct nvmet_fc_tgt_queue *queue;
1643 memset(acc, 0, sizeof(*acc));
1646 * FC-NVME spec changes. There are initiators sending different
1647 * lengths as padding sizes for Create Association Cmd descriptor
1649 * Accept anything of "minimum" length. Assume format per 1.15
1650 * spec (with HOSTID reduced to 16 bytes), ignore how long the
1651 * trailing pad length is.
1653 if (iod->rqstdatalen < FCNVME_LSDESC_CRA_RQST_MINLEN)
1654 ret = VERR_CR_ASSOC_LEN;
1655 else if (be32_to_cpu(rqst->desc_list_len) <
1656 FCNVME_LSDESC_CRA_RQST_MIN_LISTLEN)
1657 ret = VERR_CR_ASSOC_RQST_LEN;
1658 else if (rqst->assoc_cmd.desc_tag !=
1659 cpu_to_be32(FCNVME_LSDESC_CREATE_ASSOC_CMD))
1660 ret = VERR_CR_ASSOC_CMD;
1661 else if (be32_to_cpu(rqst->assoc_cmd.desc_len) <
1662 FCNVME_LSDESC_CRA_CMD_DESC_MIN_DESCLEN)
1663 ret = VERR_CR_ASSOC_CMD_LEN;
1664 else if (!rqst->assoc_cmd.ersp_ratio ||
1665 (be16_to_cpu(rqst->assoc_cmd.ersp_ratio) >=
1666 be16_to_cpu(rqst->assoc_cmd.sqsize)))
1667 ret = VERR_ERSP_RATIO;
1670 /* new association w/ admin queue */
1671 iod->assoc = nvmet_fc_alloc_target_assoc(
1672 tgtport, iod->hosthandle);
1674 ret = VERR_ASSOC_ALLOC_FAIL;
1676 queue = nvmet_fc_alloc_target_queue(iod->assoc, 0,
1677 be16_to_cpu(rqst->assoc_cmd.sqsize));
1679 ret = VERR_QUEUE_ALLOC_FAIL;
1684 dev_err(tgtport->dev,
1685 "Create Association LS failed: %s\n",
1686 validation_errors[ret]);
1687 iod->lsrsp->rsplen = nvme_fc_format_rjt(acc,
1688 sizeof(*acc), rqst->w0.ls_cmd,
1689 FCNVME_RJT_RC_LOGIC,
1690 FCNVME_RJT_EXP_NONE, 0);
1694 queue->ersp_ratio = be16_to_cpu(rqst->assoc_cmd.ersp_ratio);
1695 atomic_set(&queue->connected, 1);
1696 queue->sqhd = 0; /* best place to init value */
1698 dev_info(tgtport->dev,
1699 "{%d:%d} Association created\n",
1700 tgtport->fc_target_port.port_num, iod->assoc->a_id);
1702 /* format a response */
1704 iod->lsrsp->rsplen = sizeof(*acc);
1706 nvme_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1708 sizeof(struct fcnvme_ls_cr_assoc_acc)),
1709 FCNVME_LS_CREATE_ASSOCIATION);
1710 acc->associd.desc_tag = cpu_to_be32(FCNVME_LSDESC_ASSOC_ID);
1711 acc->associd.desc_len =
1713 sizeof(struct fcnvme_lsdesc_assoc_id));
1714 acc->associd.association_id =
1715 cpu_to_be64(nvmet_fc_makeconnid(iod->assoc, 0));
1716 acc->connectid.desc_tag = cpu_to_be32(FCNVME_LSDESC_CONN_ID);
1717 acc->connectid.desc_len =
1719 sizeof(struct fcnvme_lsdesc_conn_id));
1720 acc->connectid.connection_id = acc->associd.association_id;
1724 nvmet_fc_ls_create_connection(struct nvmet_fc_tgtport *tgtport,
1725 struct nvmet_fc_ls_iod *iod)
1727 struct fcnvme_ls_cr_conn_rqst *rqst = &iod->rqstbuf->rq_cr_conn;
1728 struct fcnvme_ls_cr_conn_acc *acc = &iod->rspbuf->rsp_cr_conn;
1729 struct nvmet_fc_tgt_queue *queue;
1732 memset(acc, 0, sizeof(*acc));
1734 if (iod->rqstdatalen < sizeof(struct fcnvme_ls_cr_conn_rqst))
1735 ret = VERR_CR_CONN_LEN;
1736 else if (rqst->desc_list_len !=
1738 sizeof(struct fcnvme_ls_cr_conn_rqst)))
1739 ret = VERR_CR_CONN_RQST_LEN;
1740 else if (rqst->associd.desc_tag != cpu_to_be32(FCNVME_LSDESC_ASSOC_ID))
1741 ret = VERR_ASSOC_ID;
1742 else if (rqst->associd.desc_len !=
1744 sizeof(struct fcnvme_lsdesc_assoc_id)))
1745 ret = VERR_ASSOC_ID_LEN;
1746 else if (rqst->connect_cmd.desc_tag !=
1747 cpu_to_be32(FCNVME_LSDESC_CREATE_CONN_CMD))
1748 ret = VERR_CR_CONN_CMD;
1749 else if (rqst->connect_cmd.desc_len !=
1751 sizeof(struct fcnvme_lsdesc_cr_conn_cmd)))
1752 ret = VERR_CR_CONN_CMD_LEN;
1753 else if (!rqst->connect_cmd.ersp_ratio ||
1754 (be16_to_cpu(rqst->connect_cmd.ersp_ratio) >=
1755 be16_to_cpu(rqst->connect_cmd.sqsize)))
1756 ret = VERR_ERSP_RATIO;
1760 iod->assoc = nvmet_fc_find_target_assoc(tgtport,
1761 be64_to_cpu(rqst->associd.association_id));
1763 ret = VERR_NO_ASSOC;
1765 queue = nvmet_fc_alloc_target_queue(iod->assoc,
1766 be16_to_cpu(rqst->connect_cmd.qid),
1767 be16_to_cpu(rqst->connect_cmd.sqsize));
1769 ret = VERR_QUEUE_ALLOC_FAIL;
1771 /* release get taken in nvmet_fc_find_target_assoc */
1772 nvmet_fc_tgt_a_put(iod->assoc);
1777 dev_err(tgtport->dev,
1778 "Create Connection LS failed: %s\n",
1779 validation_errors[ret]);
1780 iod->lsrsp->rsplen = nvme_fc_format_rjt(acc,
1781 sizeof(*acc), rqst->w0.ls_cmd,
1782 (ret == VERR_NO_ASSOC) ?
1783 FCNVME_RJT_RC_INV_ASSOC :
1784 FCNVME_RJT_RC_LOGIC,
1785 FCNVME_RJT_EXP_NONE, 0);
1789 queue->ersp_ratio = be16_to_cpu(rqst->connect_cmd.ersp_ratio);
1790 atomic_set(&queue->connected, 1);
1791 queue->sqhd = 0; /* best place to init value */
1793 /* format a response */
1795 iod->lsrsp->rsplen = sizeof(*acc);
1797 nvme_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1798 fcnvme_lsdesc_len(sizeof(struct fcnvme_ls_cr_conn_acc)),
1799 FCNVME_LS_CREATE_CONNECTION);
1800 acc->connectid.desc_tag = cpu_to_be32(FCNVME_LSDESC_CONN_ID);
1801 acc->connectid.desc_len =
1803 sizeof(struct fcnvme_lsdesc_conn_id));
1804 acc->connectid.connection_id =
1805 cpu_to_be64(nvmet_fc_makeconnid(iod->assoc,
1806 be16_to_cpu(rqst->connect_cmd.qid)));
1810 * Returns true if the LS response is to be transmit
1811 * Returns false if the LS response is to be delayed
1814 nvmet_fc_ls_disconnect(struct nvmet_fc_tgtport *tgtport,
1815 struct nvmet_fc_ls_iod *iod)
1817 struct fcnvme_ls_disconnect_assoc_rqst *rqst =
1818 &iod->rqstbuf->rq_dis_assoc;
1819 struct fcnvme_ls_disconnect_assoc_acc *acc =
1820 &iod->rspbuf->rsp_dis_assoc;
1821 struct nvmet_fc_tgt_assoc *assoc = NULL;
1822 struct nvmet_fc_ls_iod *oldls = NULL;
1823 unsigned long flags;
1826 memset(acc, 0, sizeof(*acc));
1828 ret = nvmefc_vldt_lsreq_discon_assoc(iod->rqstdatalen, rqst);
1830 /* match an active association - takes an assoc ref if !NULL */
1831 assoc = nvmet_fc_find_target_assoc(tgtport,
1832 be64_to_cpu(rqst->associd.association_id));
1835 ret = VERR_NO_ASSOC;
1838 if (ret || !assoc) {
1839 dev_err(tgtport->dev,
1840 "Disconnect LS failed: %s\n",
1841 validation_errors[ret]);
1842 iod->lsrsp->rsplen = nvme_fc_format_rjt(acc,
1843 sizeof(*acc), rqst->w0.ls_cmd,
1844 (ret == VERR_NO_ASSOC) ?
1845 FCNVME_RJT_RC_INV_ASSOC :
1846 FCNVME_RJT_RC_LOGIC,
1847 FCNVME_RJT_EXP_NONE, 0);
1851 /* format a response */
1853 iod->lsrsp->rsplen = sizeof(*acc);
1855 nvme_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1857 sizeof(struct fcnvme_ls_disconnect_assoc_acc)),
1858 FCNVME_LS_DISCONNECT_ASSOC);
1860 /* release get taken in nvmet_fc_find_target_assoc */
1861 nvmet_fc_tgt_a_put(assoc);
1864 * The rules for LS response says the response cannot
1865 * go back until ABTS's have been sent for all outstanding
1866 * I/O and a Disconnect Association LS has been sent.
1867 * So... save off the Disconnect LS to send the response
1868 * later. If there was a prior LS already saved, replace
1869 * it with the newer one and send a can't perform reject
1872 spin_lock_irqsave(&tgtport->lock, flags);
1873 oldls = assoc->rcv_disconn;
1874 assoc->rcv_disconn = iod;
1875 spin_unlock_irqrestore(&tgtport->lock, flags);
1877 nvmet_fc_delete_target_assoc(assoc);
1880 dev_info(tgtport->dev,
1881 "{%d:%d} Multiple Disconnect Association LS's "
1883 tgtport->fc_target_port.port_num, assoc->a_id);
1884 /* overwrite good response with bogus failure */
1885 oldls->lsrsp->rsplen = nvme_fc_format_rjt(oldls->rspbuf,
1886 sizeof(*iod->rspbuf),
1887 /* ok to use rqst, LS is same */
1890 FCNVME_RJT_EXP_NONE, 0);
1891 nvmet_fc_xmt_ls_rsp(tgtport, oldls);
1898 /* *********************** NVME Ctrl Routines **************************** */
1901 static void nvmet_fc_fcp_nvme_cmd_done(struct nvmet_req *nvme_req);
1903 static const struct nvmet_fabrics_ops nvmet_fc_tgt_fcp_ops;
1906 nvmet_fc_xmt_ls_rsp_done(struct nvmefc_ls_rsp *lsrsp)
1908 struct nvmet_fc_ls_iod *iod = lsrsp->nvme_fc_private;
1909 struct nvmet_fc_tgtport *tgtport = iod->tgtport;
1911 fc_dma_sync_single_for_cpu(tgtport->dev, iod->rspdma,
1912 sizeof(*iod->rspbuf), DMA_TO_DEVICE);
1913 nvmet_fc_free_ls_iod(tgtport, iod);
1914 nvmet_fc_tgtport_put(tgtport);
1918 nvmet_fc_xmt_ls_rsp(struct nvmet_fc_tgtport *tgtport,
1919 struct nvmet_fc_ls_iod *iod)
1923 fc_dma_sync_single_for_device(tgtport->dev, iod->rspdma,
1924 sizeof(*iod->rspbuf), DMA_TO_DEVICE);
1926 ret = tgtport->ops->xmt_ls_rsp(&tgtport->fc_target_port, iod->lsrsp);
1928 nvmet_fc_xmt_ls_rsp_done(iod->lsrsp);
1932 * Actual processing routine for received FC-NVME LS Requests from the LLD
1935 nvmet_fc_handle_ls_rqst(struct nvmet_fc_tgtport *tgtport,
1936 struct nvmet_fc_ls_iod *iod)
1938 struct fcnvme_ls_rqst_w0 *w0 = &iod->rqstbuf->rq_cr_assoc.w0;
1939 bool sendrsp = true;
1941 iod->lsrsp->nvme_fc_private = iod;
1942 iod->lsrsp->rspbuf = iod->rspbuf;
1943 iod->lsrsp->rspdma = iod->rspdma;
1944 iod->lsrsp->done = nvmet_fc_xmt_ls_rsp_done;
1945 /* Be preventative. handlers will later set to valid length */
1946 iod->lsrsp->rsplen = 0;
1952 * parse request input, execute the request, and format the
1955 switch (w0->ls_cmd) {
1956 case FCNVME_LS_CREATE_ASSOCIATION:
1957 /* Creates Association and initial Admin Queue/Connection */
1958 nvmet_fc_ls_create_association(tgtport, iod);
1960 case FCNVME_LS_CREATE_CONNECTION:
1961 /* Creates an IO Queue/Connection */
1962 nvmet_fc_ls_create_connection(tgtport, iod);
1964 case FCNVME_LS_DISCONNECT_ASSOC:
1965 /* Terminate a Queue/Connection or the Association */
1966 sendrsp = nvmet_fc_ls_disconnect(tgtport, iod);
1969 iod->lsrsp->rsplen = nvme_fc_format_rjt(iod->rspbuf,
1970 sizeof(*iod->rspbuf), w0->ls_cmd,
1971 FCNVME_RJT_RC_INVAL, FCNVME_RJT_EXP_NONE, 0);
1975 nvmet_fc_xmt_ls_rsp(tgtport, iod);
1979 * Actual processing routine for received FC-NVME LS Requests from the LLD
1982 nvmet_fc_handle_ls_rqst_work(struct work_struct *work)
1984 struct nvmet_fc_ls_iod *iod =
1985 container_of(work, struct nvmet_fc_ls_iod, work);
1986 struct nvmet_fc_tgtport *tgtport = iod->tgtport;
1988 nvmet_fc_handle_ls_rqst(tgtport, iod);
1993 * nvmet_fc_rcv_ls_req - transport entry point called by an LLDD
1994 * upon the reception of a NVME LS request.
1996 * The nvmet-fc layer will copy payload to an internal structure for
1997 * processing. As such, upon completion of the routine, the LLDD may
1998 * immediately free/reuse the LS request buffer passed in the call.
2000 * If this routine returns error, the LLDD should abort the exchange.
2002 * @target_port: pointer to the (registered) target port the LS was
2004 * @lsrsp: pointer to a lsrsp structure to be used to reference
2005 * the exchange corresponding to the LS.
2006 * @lsreqbuf: pointer to the buffer containing the LS Request
2007 * @lsreqbuf_len: length, in bytes, of the received LS request
2010 nvmet_fc_rcv_ls_req(struct nvmet_fc_target_port *target_port,
2012 struct nvmefc_ls_rsp *lsrsp,
2013 void *lsreqbuf, u32 lsreqbuf_len)
2015 struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
2016 struct nvmet_fc_ls_iod *iod;
2017 struct fcnvme_ls_rqst_w0 *w0 = (struct fcnvme_ls_rqst_w0 *)lsreqbuf;
2019 if (lsreqbuf_len > sizeof(union nvmefc_ls_requests)) {
2020 dev_info(tgtport->dev,
2021 "RCV %s LS failed: payload too large (%d)\n",
2022 (w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
2023 nvmefc_ls_names[w0->ls_cmd] : "",
2028 if (!nvmet_fc_tgtport_get(tgtport)) {
2029 dev_info(tgtport->dev,
2030 "RCV %s LS failed: target deleting\n",
2031 (w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
2032 nvmefc_ls_names[w0->ls_cmd] : "");
2036 iod = nvmet_fc_alloc_ls_iod(tgtport);
2038 dev_info(tgtport->dev,
2039 "RCV %s LS failed: context allocation failed\n",
2040 (w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
2041 nvmefc_ls_names[w0->ls_cmd] : "");
2042 nvmet_fc_tgtport_put(tgtport);
2048 memcpy(iod->rqstbuf, lsreqbuf, lsreqbuf_len);
2049 iod->rqstdatalen = lsreqbuf_len;
2050 iod->hosthandle = hosthandle;
2052 schedule_work(&iod->work);
2056 EXPORT_SYMBOL_GPL(nvmet_fc_rcv_ls_req);
2060 * **********************
2061 * Start of FCP handling
2062 * **********************
2066 nvmet_fc_alloc_tgt_pgs(struct nvmet_fc_fcp_iod *fod)
2068 struct scatterlist *sg;
2071 sg = sgl_alloc(fod->req.transfer_len, GFP_KERNEL, &nent);
2076 fod->data_sg_cnt = nent;
2077 fod->data_sg_cnt = fc_dma_map_sg(fod->tgtport->dev, sg, nent,
2078 ((fod->io_dir == NVMET_FCP_WRITE) ?
2079 DMA_FROM_DEVICE : DMA_TO_DEVICE));
2080 /* note: write from initiator perspective */
2081 fod->next_sg = fod->data_sg;
2086 return NVME_SC_INTERNAL;
2090 nvmet_fc_free_tgt_pgs(struct nvmet_fc_fcp_iod *fod)
2092 if (!fod->data_sg || !fod->data_sg_cnt)
2095 fc_dma_unmap_sg(fod->tgtport->dev, fod->data_sg, fod->data_sg_cnt,
2096 ((fod->io_dir == NVMET_FCP_WRITE) ?
2097 DMA_FROM_DEVICE : DMA_TO_DEVICE));
2098 sgl_free(fod->data_sg);
2099 fod->data_sg = NULL;
2100 fod->data_sg_cnt = 0;
2105 queue_90percent_full(struct nvmet_fc_tgt_queue *q, u32 sqhd)
2109 /* egad, this is ugly. And sqtail is just a best guess */
2110 sqtail = atomic_read(&q->sqtail) % q->sqsize;
2112 used = (sqtail < sqhd) ? (sqtail + q->sqsize - sqhd) : (sqtail - sqhd);
2113 return ((used * 10) >= (((u32)(q->sqsize - 1) * 9)));
2118 * May be a NVMET_FCOP_RSP or NVMET_FCOP_READDATA_RSP op
2121 nvmet_fc_prep_fcp_rsp(struct nvmet_fc_tgtport *tgtport,
2122 struct nvmet_fc_fcp_iod *fod)
2124 struct nvme_fc_ersp_iu *ersp = &fod->rspiubuf;
2125 struct nvme_common_command *sqe = &fod->cmdiubuf.sqe.common;
2126 struct nvme_completion *cqe = &ersp->cqe;
2127 u32 *cqewd = (u32 *)cqe;
2128 bool send_ersp = false;
2129 u32 rsn, rspcnt, xfr_length;
2131 if (fod->fcpreq->op == NVMET_FCOP_READDATA_RSP)
2132 xfr_length = fod->req.transfer_len;
2134 xfr_length = fod->offset;
2137 * check to see if we can send a 0's rsp.
2138 * Note: to send a 0's response, the NVME-FC host transport will
2139 * recreate the CQE. The host transport knows: sq id, SQHD (last
2140 * seen in an ersp), and command_id. Thus it will create a
2141 * zero-filled CQE with those known fields filled in. Transport
2142 * must send an ersp for any condition where the cqe won't match
2145 * Here are the FC-NVME mandated cases where we must send an ersp:
2146 * every N responses, where N=ersp_ratio
2147 * force fabric commands to send ersp's (not in FC-NVME but good
2149 * normal cmds: any time status is non-zero, or status is zero
2150 * but words 0 or 1 are non-zero.
2151 * the SQ is 90% or more full
2152 * the cmd is a fused command
2153 * transferred data length not equal to cmd iu length
2155 rspcnt = atomic_inc_return(&fod->queue->zrspcnt);
2156 if (!(rspcnt % fod->queue->ersp_ratio) ||
2157 nvme_is_fabrics((struct nvme_command *) sqe) ||
2158 xfr_length != fod->req.transfer_len ||
2159 (le16_to_cpu(cqe->status) & 0xFFFE) || cqewd[0] || cqewd[1] ||
2160 (sqe->flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND)) ||
2161 queue_90percent_full(fod->queue, le16_to_cpu(cqe->sq_head)))
2164 /* re-set the fields */
2165 fod->fcpreq->rspaddr = ersp;
2166 fod->fcpreq->rspdma = fod->rspdma;
2169 memset(ersp, 0, NVME_FC_SIZEOF_ZEROS_RSP);
2170 fod->fcpreq->rsplen = NVME_FC_SIZEOF_ZEROS_RSP;
2172 ersp->iu_len = cpu_to_be16(sizeof(*ersp)/sizeof(u32));
2173 rsn = atomic_inc_return(&fod->queue->rsn);
2174 ersp->rsn = cpu_to_be32(rsn);
2175 ersp->xfrd_len = cpu_to_be32(xfr_length);
2176 fod->fcpreq->rsplen = sizeof(*ersp);
2179 fc_dma_sync_single_for_device(tgtport->dev, fod->rspdma,
2180 sizeof(fod->rspiubuf), DMA_TO_DEVICE);
2183 static void nvmet_fc_xmt_fcp_op_done(struct nvmefc_tgt_fcp_req *fcpreq);
2186 nvmet_fc_abort_op(struct nvmet_fc_tgtport *tgtport,
2187 struct nvmet_fc_fcp_iod *fod)
2189 struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2191 /* data no longer needed */
2192 nvmet_fc_free_tgt_pgs(fod);
2195 * if an ABTS was received or we issued the fcp_abort early
2196 * don't call abort routine again.
2198 /* no need to take lock - lock was taken earlier to get here */
2200 tgtport->ops->fcp_abort(&tgtport->fc_target_port, fcpreq);
2202 nvmet_fc_free_fcp_iod(fod->queue, fod);
2206 nvmet_fc_xmt_fcp_rsp(struct nvmet_fc_tgtport *tgtport,
2207 struct nvmet_fc_fcp_iod *fod)
2211 fod->fcpreq->op = NVMET_FCOP_RSP;
2212 fod->fcpreq->timeout = 0;
2214 nvmet_fc_prep_fcp_rsp(tgtport, fod);
2216 ret = tgtport->ops->fcp_op(&tgtport->fc_target_port, fod->fcpreq);
2218 nvmet_fc_abort_op(tgtport, fod);
2222 nvmet_fc_transfer_fcp_data(struct nvmet_fc_tgtport *tgtport,
2223 struct nvmet_fc_fcp_iod *fod, u8 op)
2225 struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2226 struct scatterlist *sg = fod->next_sg;
2227 unsigned long flags;
2228 u32 remaininglen = fod->req.transfer_len - fod->offset;
2233 fcpreq->offset = fod->offset;
2234 fcpreq->timeout = NVME_FC_TGTOP_TIMEOUT_SEC;
2237 * for next sequence:
2238 * break at a sg element boundary
2239 * attempt to keep sequence length capped at
2240 * NVMET_FC_MAX_SEQ_LENGTH but allow sequence to
2241 * be longer if a single sg element is larger
2242 * than that amount. This is done to avoid creating
2243 * a new sg list to use for the tgtport api.
2247 while (tlen < remaininglen &&
2248 fcpreq->sg_cnt < tgtport->max_sg_cnt &&
2249 tlen + sg_dma_len(sg) < NVMET_FC_MAX_SEQ_LENGTH) {
2251 tlen += sg_dma_len(sg);
2254 if (tlen < remaininglen && fcpreq->sg_cnt == 0) {
2256 tlen += min_t(u32, sg_dma_len(sg), remaininglen);
2259 if (tlen < remaininglen)
2262 fod->next_sg = NULL;
2264 fcpreq->transfer_length = tlen;
2265 fcpreq->transferred_length = 0;
2266 fcpreq->fcp_error = 0;
2270 * If the last READDATA request: check if LLDD supports
2271 * combined xfr with response.
2273 if ((op == NVMET_FCOP_READDATA) &&
2274 ((fod->offset + fcpreq->transfer_length) == fod->req.transfer_len) &&
2275 (tgtport->ops->target_features & NVMET_FCTGTFEAT_READDATA_RSP)) {
2276 fcpreq->op = NVMET_FCOP_READDATA_RSP;
2277 nvmet_fc_prep_fcp_rsp(tgtport, fod);
2280 ret = tgtport->ops->fcp_op(&tgtport->fc_target_port, fod->fcpreq);
2283 * should be ok to set w/o lock as its in the thread of
2284 * execution (not an async timer routine) and doesn't
2285 * contend with any clearing action
2289 if (op == NVMET_FCOP_WRITEDATA) {
2290 spin_lock_irqsave(&fod->flock, flags);
2291 fod->writedataactive = false;
2292 spin_unlock_irqrestore(&fod->flock, flags);
2293 nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
2294 } else /* NVMET_FCOP_READDATA or NVMET_FCOP_READDATA_RSP */ {
2295 fcpreq->fcp_error = ret;
2296 fcpreq->transferred_length = 0;
2297 nvmet_fc_xmt_fcp_op_done(fod->fcpreq);
2303 __nvmet_fc_fod_op_abort(struct nvmet_fc_fcp_iod *fod, bool abort)
2305 struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2306 struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2308 /* if in the middle of an io and we need to tear down */
2310 if (fcpreq->op == NVMET_FCOP_WRITEDATA) {
2311 nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
2315 nvmet_fc_abort_op(tgtport, fod);
2323 * actual done handler for FCP operations when completed by the lldd
2326 nvmet_fc_fod_op_done(struct nvmet_fc_fcp_iod *fod)
2328 struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2329 struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2330 unsigned long flags;
2333 spin_lock_irqsave(&fod->flock, flags);
2335 fod->writedataactive = false;
2336 spin_unlock_irqrestore(&fod->flock, flags);
2338 switch (fcpreq->op) {
2340 case NVMET_FCOP_WRITEDATA:
2341 if (__nvmet_fc_fod_op_abort(fod, abort))
2343 if (fcpreq->fcp_error ||
2344 fcpreq->transferred_length != fcpreq->transfer_length) {
2345 spin_lock_irqsave(&fod->flock, flags);
2347 spin_unlock_irqrestore(&fod->flock, flags);
2349 nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
2353 fod->offset += fcpreq->transferred_length;
2354 if (fod->offset != fod->req.transfer_len) {
2355 spin_lock_irqsave(&fod->flock, flags);
2356 fod->writedataactive = true;
2357 spin_unlock_irqrestore(&fod->flock, flags);
2359 /* transfer the next chunk */
2360 nvmet_fc_transfer_fcp_data(tgtport, fod,
2361 NVMET_FCOP_WRITEDATA);
2365 /* data transfer complete, resume with nvmet layer */
2366 fod->req.execute(&fod->req);
2369 case NVMET_FCOP_READDATA:
2370 case NVMET_FCOP_READDATA_RSP:
2371 if (__nvmet_fc_fod_op_abort(fod, abort))
2373 if (fcpreq->fcp_error ||
2374 fcpreq->transferred_length != fcpreq->transfer_length) {
2375 nvmet_fc_abort_op(tgtport, fod);
2381 if (fcpreq->op == NVMET_FCOP_READDATA_RSP) {
2382 /* data no longer needed */
2383 nvmet_fc_free_tgt_pgs(fod);
2384 nvmet_fc_free_fcp_iod(fod->queue, fod);
2388 fod->offset += fcpreq->transferred_length;
2389 if (fod->offset != fod->req.transfer_len) {
2390 /* transfer the next chunk */
2391 nvmet_fc_transfer_fcp_data(tgtport, fod,
2392 NVMET_FCOP_READDATA);
2396 /* data transfer complete, send response */
2398 /* data no longer needed */
2399 nvmet_fc_free_tgt_pgs(fod);
2401 nvmet_fc_xmt_fcp_rsp(tgtport, fod);
2405 case NVMET_FCOP_RSP:
2406 if (__nvmet_fc_fod_op_abort(fod, abort))
2408 nvmet_fc_free_fcp_iod(fod->queue, fod);
2417 nvmet_fc_xmt_fcp_op_done(struct nvmefc_tgt_fcp_req *fcpreq)
2419 struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
2421 nvmet_fc_fod_op_done(fod);
2425 * actual completion handler after execution by the nvmet layer
2428 __nvmet_fc_fcp_nvme_cmd_done(struct nvmet_fc_tgtport *tgtport,
2429 struct nvmet_fc_fcp_iod *fod, int status)
2431 struct nvme_common_command *sqe = &fod->cmdiubuf.sqe.common;
2432 struct nvme_completion *cqe = &fod->rspiubuf.cqe;
2433 unsigned long flags;
2436 spin_lock_irqsave(&fod->flock, flags);
2438 spin_unlock_irqrestore(&fod->flock, flags);
2440 /* if we have a CQE, snoop the last sq_head value */
2442 fod->queue->sqhd = cqe->sq_head;
2445 nvmet_fc_abort_op(tgtport, fod);
2449 /* if an error handling the cmd post initial parsing */
2451 /* fudge up a failed CQE status for our transport error */
2452 memset(cqe, 0, sizeof(*cqe));
2453 cqe->sq_head = fod->queue->sqhd; /* echo last cqe sqhd */
2454 cqe->sq_id = cpu_to_le16(fod->queue->qid);
2455 cqe->command_id = sqe->command_id;
2456 cqe->status = cpu_to_le16(status);
2460 * try to push the data even if the SQE status is non-zero.
2461 * There may be a status where data still was intended to
2464 if ((fod->io_dir == NVMET_FCP_READ) && (fod->data_sg_cnt)) {
2465 /* push the data over before sending rsp */
2466 nvmet_fc_transfer_fcp_data(tgtport, fod,
2467 NVMET_FCOP_READDATA);
2471 /* writes & no data - fall thru */
2474 /* data no longer needed */
2475 nvmet_fc_free_tgt_pgs(fod);
2477 nvmet_fc_xmt_fcp_rsp(tgtport, fod);
2482 nvmet_fc_fcp_nvme_cmd_done(struct nvmet_req *nvme_req)
2484 struct nvmet_fc_fcp_iod *fod = nvmet_req_to_fod(nvme_req);
2485 struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2487 __nvmet_fc_fcp_nvme_cmd_done(tgtport, fod, 0);
2492 * Actual processing routine for received FC-NVME I/O Requests from the LLD
2495 nvmet_fc_handle_fcp_rqst(struct nvmet_fc_tgtport *tgtport,
2496 struct nvmet_fc_fcp_iod *fod)
2498 struct nvme_fc_cmd_iu *cmdiu = &fod->cmdiubuf;
2499 u32 xfrlen = be32_to_cpu(cmdiu->data_len);
2503 * if there is no nvmet mapping to the targetport there
2504 * shouldn't be requests. just terminate them.
2507 goto transport_error;
2510 * Fused commands are currently not supported in the linux
2513 * As such, the implementation of the FC transport does not
2514 * look at the fused commands and order delivery to the upper
2515 * layer until we have both based on csn.
2518 fod->fcpreq->done = nvmet_fc_xmt_fcp_op_done;
2520 if (cmdiu->flags & FCNVME_CMD_FLAGS_WRITE) {
2521 fod->io_dir = NVMET_FCP_WRITE;
2522 if (!nvme_is_write(&cmdiu->sqe))
2523 goto transport_error;
2524 } else if (cmdiu->flags & FCNVME_CMD_FLAGS_READ) {
2525 fod->io_dir = NVMET_FCP_READ;
2526 if (nvme_is_write(&cmdiu->sqe))
2527 goto transport_error;
2529 fod->io_dir = NVMET_FCP_NODATA;
2531 goto transport_error;
2534 fod->req.cmd = &fod->cmdiubuf.sqe;
2535 fod->req.cqe = &fod->rspiubuf.cqe;
2536 fod->req.port = tgtport->pe->port;
2538 /* clear any response payload */
2539 memset(&fod->rspiubuf, 0, sizeof(fod->rspiubuf));
2541 fod->data_sg = NULL;
2542 fod->data_sg_cnt = 0;
2544 ret = nvmet_req_init(&fod->req,
2545 &fod->queue->nvme_cq,
2546 &fod->queue->nvme_sq,
2547 &nvmet_fc_tgt_fcp_ops);
2549 /* bad SQE content or invalid ctrl state */
2550 /* nvmet layer has already called op done to send rsp. */
2554 fod->req.transfer_len = xfrlen;
2556 /* keep a running counter of tail position */
2557 atomic_inc(&fod->queue->sqtail);
2559 if (fod->req.transfer_len) {
2560 ret = nvmet_fc_alloc_tgt_pgs(fod);
2562 nvmet_req_complete(&fod->req, ret);
2566 fod->req.sg = fod->data_sg;
2567 fod->req.sg_cnt = fod->data_sg_cnt;
2570 if (fod->io_dir == NVMET_FCP_WRITE) {
2571 /* pull the data over before invoking nvmet layer */
2572 nvmet_fc_transfer_fcp_data(tgtport, fod, NVMET_FCOP_WRITEDATA);
2579 * can invoke the nvmet_layer now. If read data, cmd completion will
2582 fod->req.execute(&fod->req);
2586 nvmet_fc_abort_op(tgtport, fod);
2590 * nvmet_fc_rcv_fcp_req - transport entry point called by an LLDD
2591 * upon the reception of a NVME FCP CMD IU.
2593 * Pass a FC-NVME FCP CMD IU received from the FC link to the nvmet-fc
2594 * layer for processing.
2596 * The nvmet_fc layer allocates a local job structure (struct
2597 * nvmet_fc_fcp_iod) from the queue for the io and copies the
2598 * CMD IU buffer to the job structure. As such, on a successful
2599 * completion (returns 0), the LLDD may immediately free/reuse
2600 * the CMD IU buffer passed in the call.
2602 * However, in some circumstances, due to the packetized nature of FC
2603 * and the api of the FC LLDD which may issue a hw command to send the
2604 * response, but the LLDD may not get the hw completion for that command
2605 * and upcall the nvmet_fc layer before a new command may be
2606 * asynchronously received - its possible for a command to be received
2607 * before the LLDD and nvmet_fc have recycled the job structure. It gives
2608 * the appearance of more commands received than fits in the sq.
2609 * To alleviate this scenario, a temporary queue is maintained in the
2610 * transport for pending LLDD requests waiting for a queue job structure.
2611 * In these "overrun" cases, a temporary queue element is allocated
2612 * the LLDD request and CMD iu buffer information remembered, and the
2613 * routine returns a -EOVERFLOW status. Subsequently, when a queue job
2614 * structure is freed, it is immediately reallocated for anything on the
2615 * pending request list. The LLDDs defer_rcv() callback is called,
2616 * informing the LLDD that it may reuse the CMD IU buffer, and the io
2617 * is then started normally with the transport.
2619 * The LLDD, when receiving an -EOVERFLOW completion status, is to treat
2620 * the completion as successful but must not reuse the CMD IU buffer
2621 * until the LLDD's defer_rcv() callback has been called for the
2622 * corresponding struct nvmefc_tgt_fcp_req pointer.
2624 * If there is any other condition in which an error occurs, the
2625 * transport will return a non-zero status indicating the error.
2626 * In all cases other than -EOVERFLOW, the transport has not accepted the
2627 * request and the LLDD should abort the exchange.
2629 * @target_port: pointer to the (registered) target port the FCP CMD IU
2631 * @fcpreq: pointer to a fcpreq request structure to be used to reference
2632 * the exchange corresponding to the FCP Exchange.
2633 * @cmdiubuf: pointer to the buffer containing the FCP CMD IU
2634 * @cmdiubuf_len: length, in bytes, of the received FCP CMD IU
2637 nvmet_fc_rcv_fcp_req(struct nvmet_fc_target_port *target_port,
2638 struct nvmefc_tgt_fcp_req *fcpreq,
2639 void *cmdiubuf, u32 cmdiubuf_len)
2641 struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
2642 struct nvme_fc_cmd_iu *cmdiu = cmdiubuf;
2643 struct nvmet_fc_tgt_queue *queue;
2644 struct nvmet_fc_fcp_iod *fod;
2645 struct nvmet_fc_defer_fcp_req *deferfcp;
2646 unsigned long flags;
2648 /* validate iu, so the connection id can be used to find the queue */
2649 if ((cmdiubuf_len != sizeof(*cmdiu)) ||
2650 (cmdiu->format_id != NVME_CMD_FORMAT_ID) ||
2651 (cmdiu->fc_id != NVME_CMD_FC_ID) ||
2652 (be16_to_cpu(cmdiu->iu_len) != (sizeof(*cmdiu)/4)))
2655 queue = nvmet_fc_find_target_queue(tgtport,
2656 be64_to_cpu(cmdiu->connection_id));
2661 * note: reference taken by find_target_queue
2662 * After successful fod allocation, the fod will inherit the
2663 * ownership of that reference and will remove the reference
2664 * when the fod is freed.
2667 spin_lock_irqsave(&queue->qlock, flags);
2669 fod = nvmet_fc_alloc_fcp_iod(queue);
2671 spin_unlock_irqrestore(&queue->qlock, flags);
2673 fcpreq->nvmet_fc_private = fod;
2674 fod->fcpreq = fcpreq;
2676 memcpy(&fod->cmdiubuf, cmdiubuf, cmdiubuf_len);
2678 nvmet_fc_queue_fcp_req(tgtport, queue, fcpreq);
2683 if (!tgtport->ops->defer_rcv) {
2684 spin_unlock_irqrestore(&queue->qlock, flags);
2685 /* release the queue lookup reference */
2686 nvmet_fc_tgt_q_put(queue);
2690 deferfcp = list_first_entry_or_null(&queue->avail_defer_list,
2691 struct nvmet_fc_defer_fcp_req, req_list);
2693 /* Just re-use one that was previously allocated */
2694 list_del(&deferfcp->req_list);
2696 spin_unlock_irqrestore(&queue->qlock, flags);
2698 /* Now we need to dynamically allocate one */
2699 deferfcp = kmalloc(sizeof(*deferfcp), GFP_KERNEL);
2701 /* release the queue lookup reference */
2702 nvmet_fc_tgt_q_put(queue);
2705 spin_lock_irqsave(&queue->qlock, flags);
2708 /* For now, use rspaddr / rsplen to save payload information */
2709 fcpreq->rspaddr = cmdiubuf;
2710 fcpreq->rsplen = cmdiubuf_len;
2711 deferfcp->fcp_req = fcpreq;
2713 /* defer processing till a fod becomes available */
2714 list_add_tail(&deferfcp->req_list, &queue->pending_cmd_list);
2716 /* NOTE: the queue lookup reference is still valid */
2718 spin_unlock_irqrestore(&queue->qlock, flags);
2722 EXPORT_SYMBOL_GPL(nvmet_fc_rcv_fcp_req);
2725 * nvmet_fc_rcv_fcp_abort - transport entry point called by an LLDD
2726 * upon the reception of an ABTS for a FCP command
2728 * Notify the transport that an ABTS has been received for a FCP command
2729 * that had been given to the transport via nvmet_fc_rcv_fcp_req(). The
2730 * LLDD believes the command is still being worked on
2731 * (template_ops->fcp_req_release() has not been called).
2733 * The transport will wait for any outstanding work (an op to the LLDD,
2734 * which the lldd should complete with error due to the ABTS; or the
2735 * completion from the nvmet layer of the nvme command), then will
2736 * stop processing and call the nvmet_fc_rcv_fcp_req() callback to
2737 * return the i/o context to the LLDD. The LLDD may send the BA_ACC
2738 * to the ABTS either after return from this function (assuming any
2739 * outstanding op work has been terminated) or upon the callback being
2742 * @target_port: pointer to the (registered) target port the FCP CMD IU
2744 * @fcpreq: pointer to the fcpreq request structure that corresponds
2745 * to the exchange that received the ABTS.
2748 nvmet_fc_rcv_fcp_abort(struct nvmet_fc_target_port *target_port,
2749 struct nvmefc_tgt_fcp_req *fcpreq)
2751 struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
2752 struct nvmet_fc_tgt_queue *queue;
2753 unsigned long flags;
2755 if (!fod || fod->fcpreq != fcpreq)
2756 /* job appears to have already completed, ignore abort */
2761 spin_lock_irqsave(&queue->qlock, flags);
2764 * mark as abort. The abort handler, invoked upon completion
2765 * of any work, will detect the aborted status and do the
2768 spin_lock(&fod->flock);
2770 fod->aborted = true;
2771 spin_unlock(&fod->flock);
2773 spin_unlock_irqrestore(&queue->qlock, flags);
2775 EXPORT_SYMBOL_GPL(nvmet_fc_rcv_fcp_abort);
2778 struct nvmet_fc_traddr {
2784 __nvme_fc_parse_u64(substring_t *sstr, u64 *val)
2788 if (match_u64(sstr, &token64))
2796 * This routine validates and extracts the WWN's from the TRADDR string.
2797 * As kernel parsers need the 0x to determine number base, universally
2798 * build string to parse with 0x prefix before parsing name strings.
2801 nvme_fc_parse_traddr(struct nvmet_fc_traddr *traddr, char *buf, size_t blen)
2803 char name[2 + NVME_FC_TRADDR_HEXNAMELEN + 1];
2804 substring_t wwn = { name, &name[sizeof(name)-1] };
2805 int nnoffset, pnoffset;
2807 /* validate if string is one of the 2 allowed formats */
2808 if (strnlen(buf, blen) == NVME_FC_TRADDR_MAXLENGTH &&
2809 !strncmp(buf, "nn-0x", NVME_FC_TRADDR_OXNNLEN) &&
2810 !strncmp(&buf[NVME_FC_TRADDR_MAX_PN_OFFSET],
2811 "pn-0x", NVME_FC_TRADDR_OXNNLEN)) {
2812 nnoffset = NVME_FC_TRADDR_OXNNLEN;
2813 pnoffset = NVME_FC_TRADDR_MAX_PN_OFFSET +
2814 NVME_FC_TRADDR_OXNNLEN;
2815 } else if ((strnlen(buf, blen) == NVME_FC_TRADDR_MINLENGTH &&
2816 !strncmp(buf, "nn-", NVME_FC_TRADDR_NNLEN) &&
2817 !strncmp(&buf[NVME_FC_TRADDR_MIN_PN_OFFSET],
2818 "pn-", NVME_FC_TRADDR_NNLEN))) {
2819 nnoffset = NVME_FC_TRADDR_NNLEN;
2820 pnoffset = NVME_FC_TRADDR_MIN_PN_OFFSET + NVME_FC_TRADDR_NNLEN;
2826 name[2 + NVME_FC_TRADDR_HEXNAMELEN] = 0;
2828 memcpy(&name[2], &buf[nnoffset], NVME_FC_TRADDR_HEXNAMELEN);
2829 if (__nvme_fc_parse_u64(&wwn, &traddr->nn))
2832 memcpy(&name[2], &buf[pnoffset], NVME_FC_TRADDR_HEXNAMELEN);
2833 if (__nvme_fc_parse_u64(&wwn, &traddr->pn))
2839 pr_warn("%s: bad traddr string\n", __func__);
2844 nvmet_fc_add_port(struct nvmet_port *port)
2846 struct nvmet_fc_tgtport *tgtport;
2847 struct nvmet_fc_port_entry *pe;
2848 struct nvmet_fc_traddr traddr = { 0L, 0L };
2849 unsigned long flags;
2852 /* validate the address info */
2853 if ((port->disc_addr.trtype != NVMF_TRTYPE_FC) ||
2854 (port->disc_addr.adrfam != NVMF_ADDR_FAMILY_FC))
2857 /* map the traddr address info to a target port */
2859 ret = nvme_fc_parse_traddr(&traddr, port->disc_addr.traddr,
2860 sizeof(port->disc_addr.traddr));
2864 pe = kzalloc(sizeof(*pe), GFP_KERNEL);
2869 spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
2870 list_for_each_entry(tgtport, &nvmet_fc_target_list, tgt_list) {
2871 if ((tgtport->fc_target_port.node_name == traddr.nn) &&
2872 (tgtport->fc_target_port.port_name == traddr.pn)) {
2873 /* a FC port can only be 1 nvmet port id */
2875 nvmet_fc_portentry_bind(tgtport, pe, port);
2882 spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
2891 nvmet_fc_remove_port(struct nvmet_port *port)
2893 struct nvmet_fc_port_entry *pe = port->priv;
2895 nvmet_fc_portentry_unbind(pe);
2901 nvmet_fc_discovery_chg(struct nvmet_port *port)
2903 struct nvmet_fc_port_entry *pe = port->priv;
2904 struct nvmet_fc_tgtport *tgtport = pe->tgtport;
2906 if (tgtport && tgtport->ops->discovery_event)
2907 tgtport->ops->discovery_event(&tgtport->fc_target_port);
2910 static const struct nvmet_fabrics_ops nvmet_fc_tgt_fcp_ops = {
2911 .owner = THIS_MODULE,
2912 .type = NVMF_TRTYPE_FC,
2914 .add_port = nvmet_fc_add_port,
2915 .remove_port = nvmet_fc_remove_port,
2916 .queue_response = nvmet_fc_fcp_nvme_cmd_done,
2917 .delete_ctrl = nvmet_fc_delete_ctrl,
2918 .discovery_chg = nvmet_fc_discovery_chg,
2921 static int __init nvmet_fc_init_module(void)
2923 return nvmet_register_transport(&nvmet_fc_tgt_fcp_ops);
2926 static void __exit nvmet_fc_exit_module(void)
2928 /* sanity check - all lports should be removed */
2929 if (!list_empty(&nvmet_fc_target_list))
2930 pr_warn("%s: targetport list not empty\n", __func__);
2932 nvmet_unregister_transport(&nvmet_fc_tgt_fcp_ops);
2934 ida_destroy(&nvmet_fc_tgtport_cnt);
2937 module_init(nvmet_fc_init_module);
2938 module_exit(nvmet_fc_exit_module);
2940 MODULE_LICENSE("GPL v2");