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/parser.h>
20 #include <uapi/scsi/fc/fc_fs.h>
21 #include <uapi/scsi/fc/fc_els.h>
22 #include <linux/delay.h>
26 #include <linux/nvme-fc-driver.h>
27 #include <linux/nvme-fc.h>
30 /* *************************** Data Structures/Defines ****************** */
34 * We handle AEN commands ourselves and don't even let the
35 * block layer know about them.
37 #define NVME_FC_NR_AEN_COMMANDS 1
38 #define NVME_FC_AQ_BLKMQ_DEPTH \
39 (NVME_AQ_DEPTH - NVME_FC_NR_AEN_COMMANDS)
40 #define AEN_CMDID_BASE (NVME_FC_AQ_BLKMQ_DEPTH + 1)
42 enum nvme_fc_queue_flags {
43 NVME_FC_Q_CONNECTED = (1 << 0),
46 #define NVMEFC_QUEUE_DELAY 3 /* ms units */
48 struct nvme_fc_queue {
49 struct nvme_fc_ctrl *ctrl;
51 struct blk_mq_hw_ctx *hctx;
54 size_t cmnd_capsule_len;
63 } __aligned(sizeof(u64)); /* alignment for other things alloc'd with */
65 enum nvme_fcop_flags {
66 FCOP_FLAGS_TERMIO = (1 << 0),
67 FCOP_FLAGS_RELEASED = (1 << 1),
68 FCOP_FLAGS_COMPLETE = (1 << 2),
69 FCOP_FLAGS_AEN = (1 << 3),
72 struct nvmefc_ls_req_op {
73 struct nvmefc_ls_req ls_req;
75 struct nvme_fc_rport *rport;
76 struct nvme_fc_queue *queue;
81 struct completion ls_done;
82 struct list_head lsreq_list; /* rport->ls_req_list */
86 enum nvme_fcpop_state {
87 FCPOP_STATE_UNINIT = 0,
89 FCPOP_STATE_ACTIVE = 2,
90 FCPOP_STATE_ABORTED = 3,
91 FCPOP_STATE_COMPLETE = 4,
94 struct nvme_fc_fcp_op {
95 struct nvme_request nreq; /*
98 * the 1st element in the
100 * associated with the
103 struct nvmefc_fcp_req fcp_req;
105 struct nvme_fc_ctrl *ctrl;
106 struct nvme_fc_queue *queue;
114 struct nvme_fc_cmd_iu cmd_iu;
115 struct nvme_fc_ersp_iu rsp_iu;
118 struct nvme_fc_lport {
119 struct nvme_fc_local_port localport;
122 struct list_head port_list; /* nvme_fc_port_list */
123 struct list_head endp_list;
124 struct device *dev; /* physical device for dma */
125 struct nvme_fc_port_template *ops;
127 } __aligned(sizeof(u64)); /* alignment for other things alloc'd with */
129 struct nvme_fc_rport {
130 struct nvme_fc_remote_port remoteport;
132 struct list_head endp_list; /* for lport->endp_list */
133 struct list_head ctrl_list;
134 struct list_head ls_req_list;
135 struct device *dev; /* physical device for dma */
136 struct nvme_fc_lport *lport;
139 } __aligned(sizeof(u64)); /* alignment for other things alloc'd with */
141 enum nvme_fcctrl_flags {
142 FCCTRL_TERMIO = (1 << 0),
145 struct nvme_fc_ctrl {
147 struct nvme_fc_queue *queues;
149 struct nvme_fc_lport *lport;
150 struct nvme_fc_rport *rport;
155 struct list_head ctrl_list; /* rport->ctrl_list */
157 struct blk_mq_tag_set admin_tag_set;
158 struct blk_mq_tag_set tag_set;
160 struct work_struct delete_work;
161 struct delayed_work connect_work;
166 wait_queue_head_t ioabort_wait;
168 struct nvme_fc_fcp_op aen_ops[NVME_FC_NR_AEN_COMMANDS];
170 struct nvme_ctrl ctrl;
173 static inline struct nvme_fc_ctrl *
174 to_fc_ctrl(struct nvme_ctrl *ctrl)
176 return container_of(ctrl, struct nvme_fc_ctrl, ctrl);
179 static inline struct nvme_fc_lport *
180 localport_to_lport(struct nvme_fc_local_port *portptr)
182 return container_of(portptr, struct nvme_fc_lport, localport);
185 static inline struct nvme_fc_rport *
186 remoteport_to_rport(struct nvme_fc_remote_port *portptr)
188 return container_of(portptr, struct nvme_fc_rport, remoteport);
191 static inline struct nvmefc_ls_req_op *
192 ls_req_to_lsop(struct nvmefc_ls_req *lsreq)
194 return container_of(lsreq, struct nvmefc_ls_req_op, ls_req);
197 static inline struct nvme_fc_fcp_op *
198 fcp_req_to_fcp_op(struct nvmefc_fcp_req *fcpreq)
200 return container_of(fcpreq, struct nvme_fc_fcp_op, fcp_req);
205 /* *************************** Globals **************************** */
208 static DEFINE_SPINLOCK(nvme_fc_lock);
210 static LIST_HEAD(nvme_fc_lport_list);
211 static DEFINE_IDA(nvme_fc_local_port_cnt);
212 static DEFINE_IDA(nvme_fc_ctrl_cnt);
217 * These items are short-term. They will eventually be moved into
218 * a generic FC class. See comments in module init.
220 static struct class *fc_class;
221 static struct device *fc_udev_device;
224 /* *********************** FC-NVME Port Management ************************ */
226 static int __nvme_fc_del_ctrl(struct nvme_fc_ctrl *);
227 static void __nvme_fc_delete_hw_queue(struct nvme_fc_ctrl *,
228 struct nvme_fc_queue *, unsigned int);
231 nvme_fc_free_lport(struct kref *ref)
233 struct nvme_fc_lport *lport =
234 container_of(ref, struct nvme_fc_lport, ref);
237 WARN_ON(lport->localport.port_state != FC_OBJSTATE_DELETED);
238 WARN_ON(!list_empty(&lport->endp_list));
240 /* remove from transport list */
241 spin_lock_irqsave(&nvme_fc_lock, flags);
242 list_del(&lport->port_list);
243 spin_unlock_irqrestore(&nvme_fc_lock, flags);
245 /* let the LLDD know we've finished tearing it down */
246 lport->ops->localport_delete(&lport->localport);
248 ida_simple_remove(&nvme_fc_local_port_cnt, lport->localport.port_num);
249 ida_destroy(&lport->endp_cnt);
251 put_device(lport->dev);
257 nvme_fc_lport_put(struct nvme_fc_lport *lport)
259 kref_put(&lport->ref, nvme_fc_free_lport);
263 nvme_fc_lport_get(struct nvme_fc_lport *lport)
265 return kref_get_unless_zero(&lport->ref);
269 static struct nvme_fc_lport *
270 nvme_fc_attach_to_unreg_lport(struct nvme_fc_port_info *pinfo)
272 struct nvme_fc_lport *lport;
275 spin_lock_irqsave(&nvme_fc_lock, flags);
277 list_for_each_entry(lport, &nvme_fc_lport_list, port_list) {
278 if (lport->localport.node_name != pinfo->node_name ||
279 lport->localport.port_name != pinfo->port_name)
282 if (lport->localport.port_state != FC_OBJSTATE_DELETED) {
283 lport = ERR_PTR(-EEXIST);
287 if (!nvme_fc_lport_get(lport)) {
289 * fails if ref cnt already 0. If so,
290 * act as if lport already deleted
296 /* resume the lport */
298 lport->localport.port_role = pinfo->port_role;
299 lport->localport.port_id = pinfo->port_id;
300 lport->localport.port_state = FC_OBJSTATE_ONLINE;
302 spin_unlock_irqrestore(&nvme_fc_lock, flags);
310 spin_unlock_irqrestore(&nvme_fc_lock, flags);
316 * nvme_fc_register_localport - transport entry point called by an
317 * LLDD to register the existence of a NVME
319 * @pinfo: pointer to information about the port to be registered
320 * @template: LLDD entrypoints and operational parameters for the port
321 * @dev: physical hardware device node port corresponds to. Will be
322 * used for DMA mappings
323 * @lport_p: pointer to a local port pointer. Upon success, the routine
324 * will allocate a nvme_fc_local_port structure and place its
325 * address in the local port pointer. Upon failure, local port
326 * pointer will be set to 0.
329 * a completion status. Must be 0 upon success; a negative errno
330 * (ex: -ENXIO) upon failure.
333 nvme_fc_register_localport(struct nvme_fc_port_info *pinfo,
334 struct nvme_fc_port_template *template,
336 struct nvme_fc_local_port **portptr)
338 struct nvme_fc_lport *newrec;
342 if (!template->localport_delete || !template->remoteport_delete ||
343 !template->ls_req || !template->fcp_io ||
344 !template->ls_abort || !template->fcp_abort ||
345 !template->max_hw_queues || !template->max_sgl_segments ||
346 !template->max_dif_sgl_segments || !template->dma_boundary) {
348 goto out_reghost_failed;
352 * look to see if there is already a localport that had been
353 * deregistered and in the process of waiting for all the
354 * references to fully be removed. If the references haven't
355 * expired, we can simply re-enable the localport. Remoteports
356 * and controller reconnections should resume naturally.
358 newrec = nvme_fc_attach_to_unreg_lport(pinfo);
360 /* found an lport, but something about its state is bad */
361 if (IS_ERR(newrec)) {
362 ret = PTR_ERR(newrec);
363 goto out_reghost_failed;
365 /* found existing lport, which was resumed */
367 *portptr = &newrec->localport;
371 /* nothing found - allocate a new localport struct */
373 newrec = kmalloc((sizeof(*newrec) + template->local_priv_sz),
377 goto out_reghost_failed;
380 idx = ida_simple_get(&nvme_fc_local_port_cnt, 0, 0, GFP_KERNEL);
386 if (!get_device(dev) && dev) {
391 INIT_LIST_HEAD(&newrec->port_list);
392 INIT_LIST_HEAD(&newrec->endp_list);
393 kref_init(&newrec->ref);
394 newrec->ops = template;
396 ida_init(&newrec->endp_cnt);
397 newrec->localport.private = &newrec[1];
398 newrec->localport.node_name = pinfo->node_name;
399 newrec->localport.port_name = pinfo->port_name;
400 newrec->localport.port_role = pinfo->port_role;
401 newrec->localport.port_id = pinfo->port_id;
402 newrec->localport.port_state = FC_OBJSTATE_ONLINE;
403 newrec->localport.port_num = idx;
405 spin_lock_irqsave(&nvme_fc_lock, flags);
406 list_add_tail(&newrec->port_list, &nvme_fc_lport_list);
407 spin_unlock_irqrestore(&nvme_fc_lock, flags);
410 dma_set_seg_boundary(dev, template->dma_boundary);
412 *portptr = &newrec->localport;
416 ida_simple_remove(&nvme_fc_local_port_cnt, idx);
424 EXPORT_SYMBOL_GPL(nvme_fc_register_localport);
427 * nvme_fc_unregister_localport - transport entry point called by an
428 * LLDD to deregister/remove a previously
429 * registered a NVME host FC port.
430 * @localport: pointer to the (registered) local port that is to be
434 * a completion status. Must be 0 upon success; a negative errno
435 * (ex: -ENXIO) upon failure.
438 nvme_fc_unregister_localport(struct nvme_fc_local_port *portptr)
440 struct nvme_fc_lport *lport = localport_to_lport(portptr);
446 spin_lock_irqsave(&nvme_fc_lock, flags);
448 if (portptr->port_state != FC_OBJSTATE_ONLINE) {
449 spin_unlock_irqrestore(&nvme_fc_lock, flags);
452 portptr->port_state = FC_OBJSTATE_DELETED;
454 spin_unlock_irqrestore(&nvme_fc_lock, flags);
456 nvme_fc_lport_put(lport);
460 EXPORT_SYMBOL_GPL(nvme_fc_unregister_localport);
463 * TRADDR strings, per FC-NVME are fixed format:
464 * "nn-0x<16hexdigits>:pn-0x<16hexdigits>" - 43 characters
465 * udev event will only differ by prefix of what field is
467 * "NVMEFC_HOST_TRADDR=" or "NVMEFC_TRADDR=" - 19 max characters
468 * 19 + 43 + null_fudge = 64 characters
470 #define FCNVME_TRADDR_LENGTH 64
473 nvme_fc_signal_discovery_scan(struct nvme_fc_lport *lport,
474 struct nvme_fc_rport *rport)
476 char hostaddr[FCNVME_TRADDR_LENGTH]; /* NVMEFC_HOST_TRADDR=...*/
477 char tgtaddr[FCNVME_TRADDR_LENGTH]; /* NVMEFC_TRADDR=...*/
478 char *envp[4] = { "FC_EVENT=nvmediscovery", hostaddr, tgtaddr, NULL };
480 if (!(rport->remoteport.port_role & FC_PORT_ROLE_NVME_DISCOVERY))
483 snprintf(hostaddr, sizeof(hostaddr),
484 "NVMEFC_HOST_TRADDR=nn-0x%016llx:pn-0x%016llx",
485 lport->localport.node_name, lport->localport.port_name);
486 snprintf(tgtaddr, sizeof(tgtaddr),
487 "NVMEFC_TRADDR=nn-0x%016llx:pn-0x%016llx",
488 rport->remoteport.node_name, rport->remoteport.port_name);
489 kobject_uevent_env(&fc_udev_device->kobj, KOBJ_CHANGE, envp);
493 nvme_fc_free_rport(struct kref *ref)
495 struct nvme_fc_rport *rport =
496 container_of(ref, struct nvme_fc_rport, ref);
497 struct nvme_fc_lport *lport =
498 localport_to_lport(rport->remoteport.localport);
501 WARN_ON(rport->remoteport.port_state != FC_OBJSTATE_DELETED);
502 WARN_ON(!list_empty(&rport->ctrl_list));
504 /* remove from lport list */
505 spin_lock_irqsave(&nvme_fc_lock, flags);
506 list_del(&rport->endp_list);
507 spin_unlock_irqrestore(&nvme_fc_lock, flags);
509 /* let the LLDD know we've finished tearing it down */
510 lport->ops->remoteport_delete(&rport->remoteport);
512 ida_simple_remove(&lport->endp_cnt, rport->remoteport.port_num);
516 nvme_fc_lport_put(lport);
520 nvme_fc_rport_put(struct nvme_fc_rport *rport)
522 kref_put(&rport->ref, nvme_fc_free_rport);
526 nvme_fc_rport_get(struct nvme_fc_rport *rport)
528 return kref_get_unless_zero(&rport->ref);
532 * nvme_fc_register_remoteport - transport entry point called by an
533 * LLDD to register the existence of a NVME
534 * subsystem FC port on its fabric.
535 * @localport: pointer to the (registered) local port that the remote
536 * subsystem port is connected to.
537 * @pinfo: pointer to information about the port to be registered
538 * @rport_p: pointer to a remote port pointer. Upon success, the routine
539 * will allocate a nvme_fc_remote_port structure and place its
540 * address in the remote port pointer. Upon failure, remote port
541 * pointer will be set to 0.
544 * a completion status. Must be 0 upon success; a negative errno
545 * (ex: -ENXIO) upon failure.
548 nvme_fc_register_remoteport(struct nvme_fc_local_port *localport,
549 struct nvme_fc_port_info *pinfo,
550 struct nvme_fc_remote_port **portptr)
552 struct nvme_fc_lport *lport = localport_to_lport(localport);
553 struct nvme_fc_rport *newrec;
557 newrec = kmalloc((sizeof(*newrec) + lport->ops->remote_priv_sz),
561 goto out_reghost_failed;
564 if (!nvme_fc_lport_get(lport)) {
566 goto out_kfree_rport;
569 idx = ida_simple_get(&lport->endp_cnt, 0, 0, GFP_KERNEL);
575 INIT_LIST_HEAD(&newrec->endp_list);
576 INIT_LIST_HEAD(&newrec->ctrl_list);
577 INIT_LIST_HEAD(&newrec->ls_req_list);
578 kref_init(&newrec->ref);
579 spin_lock_init(&newrec->lock);
580 newrec->remoteport.localport = &lport->localport;
581 newrec->dev = lport->dev;
582 newrec->lport = lport;
583 newrec->remoteport.private = &newrec[1];
584 newrec->remoteport.port_role = pinfo->port_role;
585 newrec->remoteport.node_name = pinfo->node_name;
586 newrec->remoteport.port_name = pinfo->port_name;
587 newrec->remoteport.port_id = pinfo->port_id;
588 newrec->remoteport.port_state = FC_OBJSTATE_ONLINE;
589 newrec->remoteport.port_num = idx;
591 spin_lock_irqsave(&nvme_fc_lock, flags);
592 list_add_tail(&newrec->endp_list, &lport->endp_list);
593 spin_unlock_irqrestore(&nvme_fc_lock, flags);
595 nvme_fc_signal_discovery_scan(lport, newrec);
597 *portptr = &newrec->remoteport;
601 nvme_fc_lport_put(lport);
608 EXPORT_SYMBOL_GPL(nvme_fc_register_remoteport);
611 nvme_fc_abort_lsops(struct nvme_fc_rport *rport)
613 struct nvmefc_ls_req_op *lsop;
617 spin_lock_irqsave(&rport->lock, flags);
619 list_for_each_entry(lsop, &rport->ls_req_list, lsreq_list) {
620 if (!(lsop->flags & FCOP_FLAGS_TERMIO)) {
621 lsop->flags |= FCOP_FLAGS_TERMIO;
622 spin_unlock_irqrestore(&rport->lock, flags);
623 rport->lport->ops->ls_abort(&rport->lport->localport,
629 spin_unlock_irqrestore(&rport->lock, flags);
635 * nvme_fc_unregister_remoteport - transport entry point called by an
636 * LLDD to deregister/remove a previously
637 * registered a NVME subsystem FC port.
638 * @remoteport: pointer to the (registered) remote port that is to be
642 * a completion status. Must be 0 upon success; a negative errno
643 * (ex: -ENXIO) upon failure.
646 nvme_fc_unregister_remoteport(struct nvme_fc_remote_port *portptr)
648 struct nvme_fc_rport *rport = remoteport_to_rport(portptr);
649 struct nvme_fc_ctrl *ctrl;
655 spin_lock_irqsave(&rport->lock, flags);
657 if (portptr->port_state != FC_OBJSTATE_ONLINE) {
658 spin_unlock_irqrestore(&rport->lock, flags);
661 portptr->port_state = FC_OBJSTATE_DELETED;
663 /* tear down all associations to the remote port */
664 list_for_each_entry(ctrl, &rport->ctrl_list, ctrl_list)
665 __nvme_fc_del_ctrl(ctrl);
667 spin_unlock_irqrestore(&rport->lock, flags);
669 nvme_fc_abort_lsops(rport);
671 nvme_fc_rport_put(rport);
674 EXPORT_SYMBOL_GPL(nvme_fc_unregister_remoteport);
677 * nvme_fc_rescan_remoteport - transport entry point called by an
678 * LLDD to request a nvme device rescan.
679 * @remoteport: pointer to the (registered) remote port that is to be
685 nvme_fc_rescan_remoteport(struct nvme_fc_remote_port *remoteport)
687 struct nvme_fc_rport *rport = remoteport_to_rport(remoteport);
689 nvme_fc_signal_discovery_scan(rport->lport, rport);
691 EXPORT_SYMBOL_GPL(nvme_fc_rescan_remoteport);
694 /* *********************** FC-NVME DMA Handling **************************** */
697 * The fcloop device passes in a NULL device pointer. Real LLD's will
698 * pass in a valid device pointer. If NULL is passed to the dma mapping
699 * routines, depending on the platform, it may or may not succeed, and
703 * Wrapper all the dma routines and check the dev pointer.
705 * If simple mappings (return just a dma address, we'll noop them,
706 * returning a dma address of 0.
708 * On more complex mappings (dma_map_sg), a pseudo routine fills
709 * in the scatter list, setting all dma addresses to 0.
712 static inline dma_addr_t
713 fc_dma_map_single(struct device *dev, void *ptr, size_t size,
714 enum dma_data_direction dir)
716 return dev ? dma_map_single(dev, ptr, size, dir) : (dma_addr_t)0L;
720 fc_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
722 return dev ? dma_mapping_error(dev, dma_addr) : 0;
726 fc_dma_unmap_single(struct device *dev, dma_addr_t addr, size_t size,
727 enum dma_data_direction dir)
730 dma_unmap_single(dev, addr, size, dir);
734 fc_dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr, size_t size,
735 enum dma_data_direction dir)
738 dma_sync_single_for_cpu(dev, addr, size, dir);
742 fc_dma_sync_single_for_device(struct device *dev, dma_addr_t addr, size_t size,
743 enum dma_data_direction dir)
746 dma_sync_single_for_device(dev, addr, size, dir);
749 /* pseudo dma_map_sg call */
751 fc_map_sg(struct scatterlist *sg, int nents)
753 struct scatterlist *s;
756 WARN_ON(nents == 0 || sg[0].length == 0);
758 for_each_sg(sg, s, nents, i) {
760 #ifdef CONFIG_NEED_SG_DMA_LENGTH
761 s->dma_length = s->length;
768 fc_dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
769 enum dma_data_direction dir)
771 return dev ? dma_map_sg(dev, sg, nents, dir) : fc_map_sg(sg, nents);
775 fc_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
776 enum dma_data_direction dir)
779 dma_unmap_sg(dev, sg, nents, dir);
783 /* *********************** FC-NVME LS Handling **************************** */
785 static void nvme_fc_ctrl_put(struct nvme_fc_ctrl *);
786 static int nvme_fc_ctrl_get(struct nvme_fc_ctrl *);
790 __nvme_fc_finish_ls_req(struct nvmefc_ls_req_op *lsop)
792 struct nvme_fc_rport *rport = lsop->rport;
793 struct nvmefc_ls_req *lsreq = &lsop->ls_req;
796 spin_lock_irqsave(&rport->lock, flags);
798 if (!lsop->req_queued) {
799 spin_unlock_irqrestore(&rport->lock, flags);
803 list_del(&lsop->lsreq_list);
805 lsop->req_queued = false;
807 spin_unlock_irqrestore(&rport->lock, flags);
809 fc_dma_unmap_single(rport->dev, lsreq->rqstdma,
810 (lsreq->rqstlen + lsreq->rsplen),
813 nvme_fc_rport_put(rport);
817 __nvme_fc_send_ls_req(struct nvme_fc_rport *rport,
818 struct nvmefc_ls_req_op *lsop,
819 void (*done)(struct nvmefc_ls_req *req, int status))
821 struct nvmefc_ls_req *lsreq = &lsop->ls_req;
825 if (rport->remoteport.port_state != FC_OBJSTATE_ONLINE)
826 return -ECONNREFUSED;
828 if (!nvme_fc_rport_get(rport))
833 lsop->req_queued = false;
834 INIT_LIST_HEAD(&lsop->lsreq_list);
835 init_completion(&lsop->ls_done);
837 lsreq->rqstdma = fc_dma_map_single(rport->dev, lsreq->rqstaddr,
838 lsreq->rqstlen + lsreq->rsplen,
840 if (fc_dma_mapping_error(rport->dev, lsreq->rqstdma)) {
844 lsreq->rspdma = lsreq->rqstdma + lsreq->rqstlen;
846 spin_lock_irqsave(&rport->lock, flags);
848 list_add_tail(&lsop->lsreq_list, &rport->ls_req_list);
850 lsop->req_queued = true;
852 spin_unlock_irqrestore(&rport->lock, flags);
854 ret = rport->lport->ops->ls_req(&rport->lport->localport,
855 &rport->remoteport, lsreq);
862 lsop->ls_error = ret;
863 spin_lock_irqsave(&rport->lock, flags);
864 lsop->req_queued = false;
865 list_del(&lsop->lsreq_list);
866 spin_unlock_irqrestore(&rport->lock, flags);
867 fc_dma_unmap_single(rport->dev, lsreq->rqstdma,
868 (lsreq->rqstlen + lsreq->rsplen),
871 nvme_fc_rport_put(rport);
877 nvme_fc_send_ls_req_done(struct nvmefc_ls_req *lsreq, int status)
879 struct nvmefc_ls_req_op *lsop = ls_req_to_lsop(lsreq);
881 lsop->ls_error = status;
882 complete(&lsop->ls_done);
886 nvme_fc_send_ls_req(struct nvme_fc_rport *rport, struct nvmefc_ls_req_op *lsop)
888 struct nvmefc_ls_req *lsreq = &lsop->ls_req;
889 struct fcnvme_ls_rjt *rjt = lsreq->rspaddr;
892 ret = __nvme_fc_send_ls_req(rport, lsop, nvme_fc_send_ls_req_done);
896 * No timeout/not interruptible as we need the struct
897 * to exist until the lldd calls us back. Thus mandate
898 * wait until driver calls back. lldd responsible for
901 wait_for_completion(&lsop->ls_done);
903 __nvme_fc_finish_ls_req(lsop);
905 ret = lsop->ls_error;
911 /* ACC or RJT payload ? */
912 if (rjt->w0.ls_cmd == FCNVME_LS_RJT)
919 nvme_fc_send_ls_req_async(struct nvme_fc_rport *rport,
920 struct nvmefc_ls_req_op *lsop,
921 void (*done)(struct nvmefc_ls_req *req, int status))
923 /* don't wait for completion */
925 return __nvme_fc_send_ls_req(rport, lsop, done);
928 /* Validation Error indexes into the string table below */
932 VERR_LSDESC_RQST = 2,
933 VERR_LSDESC_RQST_LEN = 3,
935 VERR_ASSOC_ID_LEN = 5,
937 VERR_CONN_ID_LEN = 7,
939 VERR_CR_ASSOC_ACC_LEN = 9,
941 VERR_CR_CONN_ACC_LEN = 11,
943 VERR_DISCONN_ACC_LEN = 13,
946 static char *validation_errors[] = {
950 "Bad LSDESC_RQST Length",
951 "Not Association ID",
952 "Bad Association ID Length",
954 "Bad Connection ID Length",
956 "Bad CR_ASSOC ACC Length",
958 "Bad CR_CONN ACC Length",
959 "Not Disconnect Rqst",
960 "Bad Disconnect ACC Length",
964 nvme_fc_connect_admin_queue(struct nvme_fc_ctrl *ctrl,
965 struct nvme_fc_queue *queue, u16 qsize, u16 ersp_ratio)
967 struct nvmefc_ls_req_op *lsop;
968 struct nvmefc_ls_req *lsreq;
969 struct fcnvme_ls_cr_assoc_rqst *assoc_rqst;
970 struct fcnvme_ls_cr_assoc_acc *assoc_acc;
973 lsop = kzalloc((sizeof(*lsop) +
974 ctrl->lport->ops->lsrqst_priv_sz +
975 sizeof(*assoc_rqst) + sizeof(*assoc_acc)), GFP_KERNEL);
980 lsreq = &lsop->ls_req;
982 lsreq->private = (void *)&lsop[1];
983 assoc_rqst = (struct fcnvme_ls_cr_assoc_rqst *)
984 (lsreq->private + ctrl->lport->ops->lsrqst_priv_sz);
985 assoc_acc = (struct fcnvme_ls_cr_assoc_acc *)&assoc_rqst[1];
987 assoc_rqst->w0.ls_cmd = FCNVME_LS_CREATE_ASSOCIATION;
988 assoc_rqst->desc_list_len =
989 cpu_to_be32(sizeof(struct fcnvme_lsdesc_cr_assoc_cmd));
991 assoc_rqst->assoc_cmd.desc_tag =
992 cpu_to_be32(FCNVME_LSDESC_CREATE_ASSOC_CMD);
993 assoc_rqst->assoc_cmd.desc_len =
995 sizeof(struct fcnvme_lsdesc_cr_assoc_cmd));
997 assoc_rqst->assoc_cmd.ersp_ratio = cpu_to_be16(ersp_ratio);
998 assoc_rqst->assoc_cmd.sqsize = cpu_to_be16(qsize);
999 /* Linux supports only Dynamic controllers */
1000 assoc_rqst->assoc_cmd.cntlid = cpu_to_be16(0xffff);
1001 uuid_copy(&assoc_rqst->assoc_cmd.hostid, &ctrl->ctrl.opts->host->id);
1002 strncpy(assoc_rqst->assoc_cmd.hostnqn, ctrl->ctrl.opts->host->nqn,
1003 min(FCNVME_ASSOC_HOSTNQN_LEN, NVMF_NQN_SIZE));
1004 strncpy(assoc_rqst->assoc_cmd.subnqn, ctrl->ctrl.opts->subsysnqn,
1005 min(FCNVME_ASSOC_SUBNQN_LEN, NVMF_NQN_SIZE));
1007 lsop->queue = queue;
1008 lsreq->rqstaddr = assoc_rqst;
1009 lsreq->rqstlen = sizeof(*assoc_rqst);
1010 lsreq->rspaddr = assoc_acc;
1011 lsreq->rsplen = sizeof(*assoc_acc);
1012 lsreq->timeout = NVME_FC_CONNECT_TIMEOUT_SEC;
1014 ret = nvme_fc_send_ls_req(ctrl->rport, lsop);
1016 goto out_free_buffer;
1018 /* process connect LS completion */
1020 /* validate the ACC response */
1021 if (assoc_acc->hdr.w0.ls_cmd != FCNVME_LS_ACC)
1023 else if (assoc_acc->hdr.desc_list_len !=
1025 sizeof(struct fcnvme_ls_cr_assoc_acc)))
1026 fcret = VERR_CR_ASSOC_ACC_LEN;
1027 else if (assoc_acc->hdr.rqst.desc_tag !=
1028 cpu_to_be32(FCNVME_LSDESC_RQST))
1029 fcret = VERR_LSDESC_RQST;
1030 else if (assoc_acc->hdr.rqst.desc_len !=
1031 fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_rqst)))
1032 fcret = VERR_LSDESC_RQST_LEN;
1033 else if (assoc_acc->hdr.rqst.w0.ls_cmd != FCNVME_LS_CREATE_ASSOCIATION)
1034 fcret = VERR_CR_ASSOC;
1035 else if (assoc_acc->associd.desc_tag !=
1036 cpu_to_be32(FCNVME_LSDESC_ASSOC_ID))
1037 fcret = VERR_ASSOC_ID;
1038 else if (assoc_acc->associd.desc_len !=
1040 sizeof(struct fcnvme_lsdesc_assoc_id)))
1041 fcret = VERR_ASSOC_ID_LEN;
1042 else if (assoc_acc->connectid.desc_tag !=
1043 cpu_to_be32(FCNVME_LSDESC_CONN_ID))
1044 fcret = VERR_CONN_ID;
1045 else if (assoc_acc->connectid.desc_len !=
1046 fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_conn_id)))
1047 fcret = VERR_CONN_ID_LEN;
1052 "q %d connect failed: %s\n",
1053 queue->qnum, validation_errors[fcret]);
1055 ctrl->association_id =
1056 be64_to_cpu(assoc_acc->associd.association_id);
1057 queue->connection_id =
1058 be64_to_cpu(assoc_acc->connectid.connection_id);
1059 set_bit(NVME_FC_Q_CONNECTED, &queue->flags);
1067 "queue %d connect admin queue failed (%d).\n",
1073 nvme_fc_connect_queue(struct nvme_fc_ctrl *ctrl, struct nvme_fc_queue *queue,
1074 u16 qsize, u16 ersp_ratio)
1076 struct nvmefc_ls_req_op *lsop;
1077 struct nvmefc_ls_req *lsreq;
1078 struct fcnvme_ls_cr_conn_rqst *conn_rqst;
1079 struct fcnvme_ls_cr_conn_acc *conn_acc;
1082 lsop = kzalloc((sizeof(*lsop) +
1083 ctrl->lport->ops->lsrqst_priv_sz +
1084 sizeof(*conn_rqst) + sizeof(*conn_acc)), GFP_KERNEL);
1089 lsreq = &lsop->ls_req;
1091 lsreq->private = (void *)&lsop[1];
1092 conn_rqst = (struct fcnvme_ls_cr_conn_rqst *)
1093 (lsreq->private + ctrl->lport->ops->lsrqst_priv_sz);
1094 conn_acc = (struct fcnvme_ls_cr_conn_acc *)&conn_rqst[1];
1096 conn_rqst->w0.ls_cmd = FCNVME_LS_CREATE_CONNECTION;
1097 conn_rqst->desc_list_len = cpu_to_be32(
1098 sizeof(struct fcnvme_lsdesc_assoc_id) +
1099 sizeof(struct fcnvme_lsdesc_cr_conn_cmd));
1101 conn_rqst->associd.desc_tag = cpu_to_be32(FCNVME_LSDESC_ASSOC_ID);
1102 conn_rqst->associd.desc_len =
1104 sizeof(struct fcnvme_lsdesc_assoc_id));
1105 conn_rqst->associd.association_id = cpu_to_be64(ctrl->association_id);
1106 conn_rqst->connect_cmd.desc_tag =
1107 cpu_to_be32(FCNVME_LSDESC_CREATE_CONN_CMD);
1108 conn_rqst->connect_cmd.desc_len =
1110 sizeof(struct fcnvme_lsdesc_cr_conn_cmd));
1111 conn_rqst->connect_cmd.ersp_ratio = cpu_to_be16(ersp_ratio);
1112 conn_rqst->connect_cmd.qid = cpu_to_be16(queue->qnum);
1113 conn_rqst->connect_cmd.sqsize = cpu_to_be16(qsize);
1115 lsop->queue = queue;
1116 lsreq->rqstaddr = conn_rqst;
1117 lsreq->rqstlen = sizeof(*conn_rqst);
1118 lsreq->rspaddr = conn_acc;
1119 lsreq->rsplen = sizeof(*conn_acc);
1120 lsreq->timeout = NVME_FC_CONNECT_TIMEOUT_SEC;
1122 ret = nvme_fc_send_ls_req(ctrl->rport, lsop);
1124 goto out_free_buffer;
1126 /* process connect LS completion */
1128 /* validate the ACC response */
1129 if (conn_acc->hdr.w0.ls_cmd != FCNVME_LS_ACC)
1131 else if (conn_acc->hdr.desc_list_len !=
1132 fcnvme_lsdesc_len(sizeof(struct fcnvme_ls_cr_conn_acc)))
1133 fcret = VERR_CR_CONN_ACC_LEN;
1134 else if (conn_acc->hdr.rqst.desc_tag != cpu_to_be32(FCNVME_LSDESC_RQST))
1135 fcret = VERR_LSDESC_RQST;
1136 else if (conn_acc->hdr.rqst.desc_len !=
1137 fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_rqst)))
1138 fcret = VERR_LSDESC_RQST_LEN;
1139 else if (conn_acc->hdr.rqst.w0.ls_cmd != FCNVME_LS_CREATE_CONNECTION)
1140 fcret = VERR_CR_CONN;
1141 else if (conn_acc->connectid.desc_tag !=
1142 cpu_to_be32(FCNVME_LSDESC_CONN_ID))
1143 fcret = VERR_CONN_ID;
1144 else if (conn_acc->connectid.desc_len !=
1145 fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_conn_id)))
1146 fcret = VERR_CONN_ID_LEN;
1151 "q %d connect failed: %s\n",
1152 queue->qnum, validation_errors[fcret]);
1154 queue->connection_id =
1155 be64_to_cpu(conn_acc->connectid.connection_id);
1156 set_bit(NVME_FC_Q_CONNECTED, &queue->flags);
1164 "queue %d connect command failed (%d).\n",
1170 nvme_fc_disconnect_assoc_done(struct nvmefc_ls_req *lsreq, int status)
1172 struct nvmefc_ls_req_op *lsop = ls_req_to_lsop(lsreq);
1174 __nvme_fc_finish_ls_req(lsop);
1176 /* fc-nvme iniator doesn't care about success or failure of cmd */
1182 * This routine sends a FC-NVME LS to disconnect (aka terminate)
1183 * the FC-NVME Association. Terminating the association also
1184 * terminates the FC-NVME connections (per queue, both admin and io
1185 * queues) that are part of the association. E.g. things are torn
1186 * down, and the related FC-NVME Association ID and Connection IDs
1189 * The behavior of the fc-nvme initiator is such that it's
1190 * understanding of the association and connections will implicitly
1191 * be torn down. The action is implicit as it may be due to a loss of
1192 * connectivity with the fc-nvme target, so you may never get a
1193 * response even if you tried. As such, the action of this routine
1194 * is to asynchronously send the LS, ignore any results of the LS, and
1195 * continue on with terminating the association. If the fc-nvme target
1196 * is present and receives the LS, it too can tear down.
1199 nvme_fc_xmt_disconnect_assoc(struct nvme_fc_ctrl *ctrl)
1201 struct fcnvme_ls_disconnect_rqst *discon_rqst;
1202 struct fcnvme_ls_disconnect_acc *discon_acc;
1203 struct nvmefc_ls_req_op *lsop;
1204 struct nvmefc_ls_req *lsreq;
1207 lsop = kzalloc((sizeof(*lsop) +
1208 ctrl->lport->ops->lsrqst_priv_sz +
1209 sizeof(*discon_rqst) + sizeof(*discon_acc)),
1212 /* couldn't sent it... too bad */
1215 lsreq = &lsop->ls_req;
1217 lsreq->private = (void *)&lsop[1];
1218 discon_rqst = (struct fcnvme_ls_disconnect_rqst *)
1219 (lsreq->private + ctrl->lport->ops->lsrqst_priv_sz);
1220 discon_acc = (struct fcnvme_ls_disconnect_acc *)&discon_rqst[1];
1222 discon_rqst->w0.ls_cmd = FCNVME_LS_DISCONNECT;
1223 discon_rqst->desc_list_len = cpu_to_be32(
1224 sizeof(struct fcnvme_lsdesc_assoc_id) +
1225 sizeof(struct fcnvme_lsdesc_disconn_cmd));
1227 discon_rqst->associd.desc_tag = cpu_to_be32(FCNVME_LSDESC_ASSOC_ID);
1228 discon_rqst->associd.desc_len =
1230 sizeof(struct fcnvme_lsdesc_assoc_id));
1232 discon_rqst->associd.association_id = cpu_to_be64(ctrl->association_id);
1234 discon_rqst->discon_cmd.desc_tag = cpu_to_be32(
1235 FCNVME_LSDESC_DISCONN_CMD);
1236 discon_rqst->discon_cmd.desc_len =
1238 sizeof(struct fcnvme_lsdesc_disconn_cmd));
1239 discon_rqst->discon_cmd.scope = FCNVME_DISCONN_ASSOCIATION;
1240 discon_rqst->discon_cmd.id = cpu_to_be64(ctrl->association_id);
1242 lsreq->rqstaddr = discon_rqst;
1243 lsreq->rqstlen = sizeof(*discon_rqst);
1244 lsreq->rspaddr = discon_acc;
1245 lsreq->rsplen = sizeof(*discon_acc);
1246 lsreq->timeout = NVME_FC_CONNECT_TIMEOUT_SEC;
1248 ret = nvme_fc_send_ls_req_async(ctrl->rport, lsop,
1249 nvme_fc_disconnect_assoc_done);
1253 /* only meaningful part to terminating the association */
1254 ctrl->association_id = 0;
1258 /* *********************** NVME Ctrl Routines **************************** */
1260 static void __nvme_fc_final_op_cleanup(struct request *rq);
1261 static void nvme_fc_error_recovery(struct nvme_fc_ctrl *ctrl, char *errmsg);
1264 nvme_fc_reinit_request(void *data, struct request *rq)
1266 struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
1267 struct nvme_fc_cmd_iu *cmdiu = &op->cmd_iu;
1269 memset(cmdiu, 0, sizeof(*cmdiu));
1270 cmdiu->scsi_id = NVME_CMD_SCSI_ID;
1271 cmdiu->fc_id = NVME_CMD_FC_ID;
1272 cmdiu->iu_len = cpu_to_be16(sizeof(*cmdiu) / sizeof(u32));
1273 memset(&op->rsp_iu, 0, sizeof(op->rsp_iu));
1279 __nvme_fc_exit_request(struct nvme_fc_ctrl *ctrl,
1280 struct nvme_fc_fcp_op *op)
1282 fc_dma_unmap_single(ctrl->lport->dev, op->fcp_req.rspdma,
1283 sizeof(op->rsp_iu), DMA_FROM_DEVICE);
1284 fc_dma_unmap_single(ctrl->lport->dev, op->fcp_req.cmddma,
1285 sizeof(op->cmd_iu), DMA_TO_DEVICE);
1287 atomic_set(&op->state, FCPOP_STATE_UNINIT);
1291 nvme_fc_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1292 unsigned int hctx_idx)
1294 struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
1296 return __nvme_fc_exit_request(set->driver_data, op);
1300 __nvme_fc_abort_op(struct nvme_fc_ctrl *ctrl, struct nvme_fc_fcp_op *op)
1304 state = atomic_xchg(&op->state, FCPOP_STATE_ABORTED);
1305 if (state != FCPOP_STATE_ACTIVE) {
1306 atomic_set(&op->state, state);
1310 ctrl->lport->ops->fcp_abort(&ctrl->lport->localport,
1311 &ctrl->rport->remoteport,
1312 op->queue->lldd_handle,
1319 nvme_fc_abort_aen_ops(struct nvme_fc_ctrl *ctrl)
1321 struct nvme_fc_fcp_op *aen_op = ctrl->aen_ops;
1322 unsigned long flags;
1325 for (i = 0; i < NVME_FC_NR_AEN_COMMANDS; i++, aen_op++) {
1326 if (atomic_read(&aen_op->state) != FCPOP_STATE_ACTIVE)
1329 spin_lock_irqsave(&ctrl->lock, flags);
1330 if (ctrl->flags & FCCTRL_TERMIO) {
1332 aen_op->flags |= FCOP_FLAGS_TERMIO;
1334 spin_unlock_irqrestore(&ctrl->lock, flags);
1336 ret = __nvme_fc_abort_op(ctrl, aen_op);
1339 * if __nvme_fc_abort_op failed the io wasn't
1340 * active. Thus this call path is running in
1341 * parallel to the io complete. Treat as non-error.
1344 /* back out the flags/counters */
1345 spin_lock_irqsave(&ctrl->lock, flags);
1346 if (ctrl->flags & FCCTRL_TERMIO)
1348 aen_op->flags &= ~FCOP_FLAGS_TERMIO;
1349 spin_unlock_irqrestore(&ctrl->lock, flags);
1356 __nvme_fc_fcpop_chk_teardowns(struct nvme_fc_ctrl *ctrl,
1357 struct nvme_fc_fcp_op *op)
1359 unsigned long flags;
1360 bool complete_rq = false;
1362 spin_lock_irqsave(&ctrl->lock, flags);
1363 if (unlikely(op->flags & FCOP_FLAGS_TERMIO)) {
1364 if (ctrl->flags & FCCTRL_TERMIO) {
1366 wake_up(&ctrl->ioabort_wait);
1369 if (op->flags & FCOP_FLAGS_RELEASED)
1372 op->flags |= FCOP_FLAGS_COMPLETE;
1373 spin_unlock_irqrestore(&ctrl->lock, flags);
1379 nvme_fc_fcpio_done(struct nvmefc_fcp_req *req)
1381 struct nvme_fc_fcp_op *op = fcp_req_to_fcp_op(req);
1382 struct request *rq = op->rq;
1383 struct nvmefc_fcp_req *freq = &op->fcp_req;
1384 struct nvme_fc_ctrl *ctrl = op->ctrl;
1385 struct nvme_fc_queue *queue = op->queue;
1386 struct nvme_completion *cqe = &op->rsp_iu.cqe;
1387 struct nvme_command *sqe = &op->cmd_iu.sqe;
1388 __le16 status = cpu_to_le16(NVME_SC_SUCCESS << 1);
1389 union nvme_result result;
1390 bool complete_rq, terminate_assoc = true;
1394 * The current linux implementation of a nvme controller
1395 * allocates a single tag set for all io queues and sizes
1396 * the io queues to fully hold all possible tags. Thus, the
1397 * implementation does not reference or care about the sqhd
1398 * value as it never needs to use the sqhd/sqtail pointers
1399 * for submission pacing.
1401 * This affects the FC-NVME implementation in two ways:
1402 * 1) As the value doesn't matter, we don't need to waste
1403 * cycles extracting it from ERSPs and stamping it in the
1404 * cases where the transport fabricates CQEs on successful
1406 * 2) The FC-NVME implementation requires that delivery of
1407 * ERSP completions are to go back to the nvme layer in order
1408 * relative to the rsn, such that the sqhd value will always
1409 * be "in order" for the nvme layer. As the nvme layer in
1410 * linux doesn't care about sqhd, there's no need to return
1414 * As the core nvme layer in linux currently does not look at
1415 * every field in the cqe - in cases where the FC transport must
1416 * fabricate a CQE, the following fields will not be set as they
1417 * are not referenced:
1418 * cqe.sqid, cqe.sqhd, cqe.command_id
1420 * Failure or error of an individual i/o, in a transport
1421 * detected fashion unrelated to the nvme completion status,
1422 * potentially cause the initiator and target sides to get out
1423 * of sync on SQ head/tail (aka outstanding io count allowed).
1424 * Per FC-NVME spec, failure of an individual command requires
1425 * the connection to be terminated, which in turn requires the
1426 * association to be terminated.
1429 fc_dma_sync_single_for_cpu(ctrl->lport->dev, op->fcp_req.rspdma,
1430 sizeof(op->rsp_iu), DMA_FROM_DEVICE);
1432 if (atomic_read(&op->state) == FCPOP_STATE_ABORTED)
1433 status = cpu_to_le16((NVME_SC_ABORT_REQ | NVME_SC_DNR) << 1);
1434 else if (freq->status)
1435 status = cpu_to_le16(NVME_SC_INTERNAL << 1);
1438 * For the linux implementation, if we have an unsuccesful
1439 * status, they blk-mq layer can typically be called with the
1440 * non-zero status and the content of the cqe isn't important.
1446 * command completed successfully relative to the wire
1447 * protocol. However, validate anything received and
1448 * extract the status and result from the cqe (create it
1452 switch (freq->rcv_rsplen) {
1455 case NVME_FC_SIZEOF_ZEROS_RSP:
1457 * No response payload or 12 bytes of payload (which
1458 * should all be zeros) are considered successful and
1459 * no payload in the CQE by the transport.
1461 if (freq->transferred_length !=
1462 be32_to_cpu(op->cmd_iu.data_len)) {
1463 status = cpu_to_le16(NVME_SC_INTERNAL << 1);
1469 case sizeof(struct nvme_fc_ersp_iu):
1471 * The ERSP IU contains a full completion with CQE.
1472 * Validate ERSP IU and look at cqe.
1474 if (unlikely(be16_to_cpu(op->rsp_iu.iu_len) !=
1475 (freq->rcv_rsplen / 4) ||
1476 be32_to_cpu(op->rsp_iu.xfrd_len) !=
1477 freq->transferred_length ||
1478 op->rsp_iu.status_code ||
1479 sqe->common.command_id != cqe->command_id)) {
1480 status = cpu_to_le16(NVME_SC_INTERNAL << 1);
1483 result = cqe->result;
1484 status = cqe->status;
1488 status = cpu_to_le16(NVME_SC_INTERNAL << 1);
1492 terminate_assoc = false;
1495 if (op->flags & FCOP_FLAGS_AEN) {
1496 nvme_complete_async_event(&queue->ctrl->ctrl, status, &result);
1497 complete_rq = __nvme_fc_fcpop_chk_teardowns(ctrl, op);
1498 atomic_set(&op->state, FCPOP_STATE_IDLE);
1499 op->flags = FCOP_FLAGS_AEN; /* clear other flags */
1500 nvme_fc_ctrl_put(ctrl);
1504 complete_rq = __nvme_fc_fcpop_chk_teardowns(ctrl, op);
1506 if (unlikely(op->flags & FCOP_FLAGS_TERMIO)) {
1507 status = cpu_to_le16(NVME_SC_ABORT_REQ << 1);
1508 if (blk_queue_dying(rq->q))
1509 status |= cpu_to_le16(NVME_SC_DNR << 1);
1511 nvme_end_request(rq, status, result);
1513 __nvme_fc_final_op_cleanup(rq);
1516 if (terminate_assoc)
1517 nvme_fc_error_recovery(ctrl, "transport detected io error");
1521 __nvme_fc_init_request(struct nvme_fc_ctrl *ctrl,
1522 struct nvme_fc_queue *queue, struct nvme_fc_fcp_op *op,
1523 struct request *rq, u32 rqno)
1525 struct nvme_fc_cmd_iu *cmdiu = &op->cmd_iu;
1528 memset(op, 0, sizeof(*op));
1529 op->fcp_req.cmdaddr = &op->cmd_iu;
1530 op->fcp_req.cmdlen = sizeof(op->cmd_iu);
1531 op->fcp_req.rspaddr = &op->rsp_iu;
1532 op->fcp_req.rsplen = sizeof(op->rsp_iu);
1533 op->fcp_req.done = nvme_fc_fcpio_done;
1534 op->fcp_req.first_sgl = (struct scatterlist *)&op[1];
1535 op->fcp_req.private = &op->fcp_req.first_sgl[SG_CHUNK_SIZE];
1541 cmdiu->scsi_id = NVME_CMD_SCSI_ID;
1542 cmdiu->fc_id = NVME_CMD_FC_ID;
1543 cmdiu->iu_len = cpu_to_be16(sizeof(*cmdiu) / sizeof(u32));
1545 op->fcp_req.cmddma = fc_dma_map_single(ctrl->lport->dev,
1546 &op->cmd_iu, sizeof(op->cmd_iu), DMA_TO_DEVICE);
1547 if (fc_dma_mapping_error(ctrl->lport->dev, op->fcp_req.cmddma)) {
1549 "FCP Op failed - cmdiu dma mapping failed.\n");
1554 op->fcp_req.rspdma = fc_dma_map_single(ctrl->lport->dev,
1555 &op->rsp_iu, sizeof(op->rsp_iu),
1557 if (fc_dma_mapping_error(ctrl->lport->dev, op->fcp_req.rspdma)) {
1559 "FCP Op failed - rspiu dma mapping failed.\n");
1563 atomic_set(&op->state, FCPOP_STATE_IDLE);
1569 nvme_fc_init_request(struct blk_mq_tag_set *set, struct request *rq,
1570 unsigned int hctx_idx, unsigned int numa_node)
1572 struct nvme_fc_ctrl *ctrl = set->driver_data;
1573 struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
1574 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
1575 struct nvme_fc_queue *queue = &ctrl->queues[queue_idx];
1577 return __nvme_fc_init_request(ctrl, queue, op, rq, queue->rqcnt++);
1581 nvme_fc_init_aen_ops(struct nvme_fc_ctrl *ctrl)
1583 struct nvme_fc_fcp_op *aen_op;
1584 struct nvme_fc_cmd_iu *cmdiu;
1585 struct nvme_command *sqe;
1589 aen_op = ctrl->aen_ops;
1590 for (i = 0; i < NVME_FC_NR_AEN_COMMANDS; i++, aen_op++) {
1591 private = kzalloc(ctrl->lport->ops->fcprqst_priv_sz,
1596 cmdiu = &aen_op->cmd_iu;
1598 ret = __nvme_fc_init_request(ctrl, &ctrl->queues[0],
1599 aen_op, (struct request *)NULL,
1600 (AEN_CMDID_BASE + i));
1606 aen_op->flags = FCOP_FLAGS_AEN;
1607 aen_op->fcp_req.first_sgl = NULL; /* no sg list */
1608 aen_op->fcp_req.private = private;
1610 memset(sqe, 0, sizeof(*sqe));
1611 sqe->common.opcode = nvme_admin_async_event;
1612 /* Note: core layer may overwrite the sqe.command_id value */
1613 sqe->common.command_id = AEN_CMDID_BASE + i;
1619 nvme_fc_term_aen_ops(struct nvme_fc_ctrl *ctrl)
1621 struct nvme_fc_fcp_op *aen_op;
1624 aen_op = ctrl->aen_ops;
1625 for (i = 0; i < NVME_FC_NR_AEN_COMMANDS; i++, aen_op++) {
1626 if (!aen_op->fcp_req.private)
1629 __nvme_fc_exit_request(ctrl, aen_op);
1631 kfree(aen_op->fcp_req.private);
1632 aen_op->fcp_req.private = NULL;
1637 __nvme_fc_init_hctx(struct blk_mq_hw_ctx *hctx, struct nvme_fc_ctrl *ctrl,
1640 struct nvme_fc_queue *queue = &ctrl->queues[qidx];
1642 hctx->driver_data = queue;
1647 nvme_fc_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
1648 unsigned int hctx_idx)
1650 struct nvme_fc_ctrl *ctrl = data;
1652 __nvme_fc_init_hctx(hctx, ctrl, hctx_idx + 1);
1658 nvme_fc_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
1659 unsigned int hctx_idx)
1661 struct nvme_fc_ctrl *ctrl = data;
1663 __nvme_fc_init_hctx(hctx, ctrl, hctx_idx);
1669 nvme_fc_init_queue(struct nvme_fc_ctrl *ctrl, int idx, size_t queue_size)
1671 struct nvme_fc_queue *queue;
1673 queue = &ctrl->queues[idx];
1674 memset(queue, 0, sizeof(*queue));
1677 atomic_set(&queue->csn, 1);
1678 queue->dev = ctrl->dev;
1681 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
1683 queue->cmnd_capsule_len = sizeof(struct nvme_command);
1685 queue->queue_size = queue_size;
1688 * Considered whether we should allocate buffers for all SQEs
1689 * and CQEs and dma map them - mapping their respective entries
1690 * into the request structures (kernel vm addr and dma address)
1691 * thus the driver could use the buffers/mappings directly.
1692 * It only makes sense if the LLDD would use them for its
1693 * messaging api. It's very unlikely most adapter api's would use
1694 * a native NVME sqe/cqe. More reasonable if FC-NVME IU payload
1695 * structures were used instead.
1700 * This routine terminates a queue at the transport level.
1701 * The transport has already ensured that all outstanding ios on
1702 * the queue have been terminated.
1703 * The transport will send a Disconnect LS request to terminate
1704 * the queue's connection. Termination of the admin queue will also
1705 * terminate the association at the target.
1708 nvme_fc_free_queue(struct nvme_fc_queue *queue)
1710 if (!test_and_clear_bit(NVME_FC_Q_CONNECTED, &queue->flags))
1714 * Current implementation never disconnects a single queue.
1715 * It always terminates a whole association. So there is never
1716 * a disconnect(queue) LS sent to the target.
1719 queue->connection_id = 0;
1720 clear_bit(NVME_FC_Q_CONNECTED, &queue->flags);
1724 __nvme_fc_delete_hw_queue(struct nvme_fc_ctrl *ctrl,
1725 struct nvme_fc_queue *queue, unsigned int qidx)
1727 if (ctrl->lport->ops->delete_queue)
1728 ctrl->lport->ops->delete_queue(&ctrl->lport->localport, qidx,
1729 queue->lldd_handle);
1730 queue->lldd_handle = NULL;
1734 nvme_fc_free_io_queues(struct nvme_fc_ctrl *ctrl)
1738 for (i = 1; i < ctrl->ctrl.queue_count; i++)
1739 nvme_fc_free_queue(&ctrl->queues[i]);
1743 __nvme_fc_create_hw_queue(struct nvme_fc_ctrl *ctrl,
1744 struct nvme_fc_queue *queue, unsigned int qidx, u16 qsize)
1748 queue->lldd_handle = NULL;
1749 if (ctrl->lport->ops->create_queue)
1750 ret = ctrl->lport->ops->create_queue(&ctrl->lport->localport,
1751 qidx, qsize, &queue->lldd_handle);
1757 nvme_fc_delete_hw_io_queues(struct nvme_fc_ctrl *ctrl)
1759 struct nvme_fc_queue *queue = &ctrl->queues[ctrl->ctrl.queue_count - 1];
1762 for (i = ctrl->ctrl.queue_count - 1; i >= 1; i--, queue--)
1763 __nvme_fc_delete_hw_queue(ctrl, queue, i);
1767 nvme_fc_create_hw_io_queues(struct nvme_fc_ctrl *ctrl, u16 qsize)
1769 struct nvme_fc_queue *queue = &ctrl->queues[1];
1772 for (i = 1; i < ctrl->ctrl.queue_count; i++, queue++) {
1773 ret = __nvme_fc_create_hw_queue(ctrl, queue, i, qsize);
1782 __nvme_fc_delete_hw_queue(ctrl, &ctrl->queues[i], i);
1787 nvme_fc_connect_io_queues(struct nvme_fc_ctrl *ctrl, u16 qsize)
1791 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
1792 ret = nvme_fc_connect_queue(ctrl, &ctrl->queues[i], qsize,
1796 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
1805 nvme_fc_init_io_queues(struct nvme_fc_ctrl *ctrl)
1809 for (i = 1; i < ctrl->ctrl.queue_count; i++)
1810 nvme_fc_init_queue(ctrl, i, ctrl->ctrl.sqsize);
1814 nvme_fc_ctrl_free(struct kref *ref)
1816 struct nvme_fc_ctrl *ctrl =
1817 container_of(ref, struct nvme_fc_ctrl, ref);
1818 unsigned long flags;
1820 if (ctrl->ctrl.tagset) {
1821 blk_cleanup_queue(ctrl->ctrl.connect_q);
1822 blk_mq_free_tag_set(&ctrl->tag_set);
1825 /* remove from rport list */
1826 spin_lock_irqsave(&ctrl->rport->lock, flags);
1827 list_del(&ctrl->ctrl_list);
1828 spin_unlock_irqrestore(&ctrl->rport->lock, flags);
1830 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1831 blk_cleanup_queue(ctrl->ctrl.admin_q);
1832 blk_mq_free_tag_set(&ctrl->admin_tag_set);
1834 kfree(ctrl->queues);
1836 put_device(ctrl->dev);
1837 nvme_fc_rport_put(ctrl->rport);
1839 ida_simple_remove(&nvme_fc_ctrl_cnt, ctrl->cnum);
1840 if (ctrl->ctrl.opts)
1841 nvmf_free_options(ctrl->ctrl.opts);
1846 nvme_fc_ctrl_put(struct nvme_fc_ctrl *ctrl)
1848 kref_put(&ctrl->ref, nvme_fc_ctrl_free);
1852 nvme_fc_ctrl_get(struct nvme_fc_ctrl *ctrl)
1854 return kref_get_unless_zero(&ctrl->ref);
1858 * All accesses from nvme core layer done - can now free the
1859 * controller. Called after last nvme_put_ctrl() call
1862 nvme_fc_nvme_ctrl_freed(struct nvme_ctrl *nctrl)
1864 struct nvme_fc_ctrl *ctrl = to_fc_ctrl(nctrl);
1866 WARN_ON(nctrl != &ctrl->ctrl);
1868 nvme_fc_ctrl_put(ctrl);
1872 nvme_fc_error_recovery(struct nvme_fc_ctrl *ctrl, char *errmsg)
1874 /* only proceed if in LIVE state - e.g. on first error */
1875 if (ctrl->ctrl.state != NVME_CTRL_LIVE)
1878 dev_warn(ctrl->ctrl.device,
1879 "NVME-FC{%d}: transport association error detected: %s\n",
1880 ctrl->cnum, errmsg);
1881 dev_warn(ctrl->ctrl.device,
1882 "NVME-FC{%d}: resetting controller\n", ctrl->cnum);
1884 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING)) {
1885 dev_err(ctrl->ctrl.device,
1886 "NVME-FC{%d}: error_recovery: Couldn't change state "
1887 "to RECONNECTING\n", ctrl->cnum);
1891 nvme_reset_ctrl(&ctrl->ctrl);
1894 static enum blk_eh_timer_return
1895 nvme_fc_timeout(struct request *rq, bool reserved)
1897 struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
1898 struct nvme_fc_ctrl *ctrl = op->ctrl;
1902 return BLK_EH_RESET_TIMER;
1904 ret = __nvme_fc_abort_op(ctrl, op);
1906 /* io wasn't active to abort consider it done */
1907 return BLK_EH_HANDLED;
1910 * we can't individually ABTS an io without affecting the queue,
1911 * thus killing the queue, adn thus the association.
1912 * So resolve by performing a controller reset, which will stop
1913 * the host/io stack, terminate the association on the link,
1914 * and recreate an association on the link.
1916 nvme_fc_error_recovery(ctrl, "io timeout error");
1918 return BLK_EH_HANDLED;
1922 nvme_fc_map_data(struct nvme_fc_ctrl *ctrl, struct request *rq,
1923 struct nvme_fc_fcp_op *op)
1925 struct nvmefc_fcp_req *freq = &op->fcp_req;
1926 enum dma_data_direction dir;
1931 if (!blk_rq_payload_bytes(rq))
1934 freq->sg_table.sgl = freq->first_sgl;
1935 ret = sg_alloc_table_chained(&freq->sg_table,
1936 blk_rq_nr_phys_segments(rq), freq->sg_table.sgl);
1940 op->nents = blk_rq_map_sg(rq->q, rq, freq->sg_table.sgl);
1941 WARN_ON(op->nents > blk_rq_nr_phys_segments(rq));
1942 dir = (rq_data_dir(rq) == WRITE) ? DMA_TO_DEVICE : DMA_FROM_DEVICE;
1943 freq->sg_cnt = fc_dma_map_sg(ctrl->lport->dev, freq->sg_table.sgl,
1945 if (unlikely(freq->sg_cnt <= 0)) {
1946 sg_free_table_chained(&freq->sg_table, true);
1952 * TODO: blk_integrity_rq(rq) for DIF
1958 nvme_fc_unmap_data(struct nvme_fc_ctrl *ctrl, struct request *rq,
1959 struct nvme_fc_fcp_op *op)
1961 struct nvmefc_fcp_req *freq = &op->fcp_req;
1966 fc_dma_unmap_sg(ctrl->lport->dev, freq->sg_table.sgl, op->nents,
1967 ((rq_data_dir(rq) == WRITE) ?
1968 DMA_TO_DEVICE : DMA_FROM_DEVICE));
1970 nvme_cleanup_cmd(rq);
1972 sg_free_table_chained(&freq->sg_table, true);
1978 * In FC, the queue is a logical thing. At transport connect, the target
1979 * creates its "queue" and returns a handle that is to be given to the
1980 * target whenever it posts something to the corresponding SQ. When an
1981 * SQE is sent on a SQ, FC effectively considers the SQE, or rather the
1982 * command contained within the SQE, an io, and assigns a FC exchange
1983 * to it. The SQE and the associated SQ handle are sent in the initial
1984 * CMD IU sents on the exchange. All transfers relative to the io occur
1985 * as part of the exchange. The CQE is the last thing for the io,
1986 * which is transferred (explicitly or implicitly) with the RSP IU
1987 * sent on the exchange. After the CQE is received, the FC exchange is
1988 * terminaed and the Exchange may be used on a different io.
1990 * The transport to LLDD api has the transport making a request for a
1991 * new fcp io request to the LLDD. The LLDD then allocates a FC exchange
1992 * resource and transfers the command. The LLDD will then process all
1993 * steps to complete the io. Upon completion, the transport done routine
1996 * So - while the operation is outstanding to the LLDD, there is a link
1997 * level FC exchange resource that is also outstanding. This must be
1998 * considered in all cleanup operations.
2001 nvme_fc_start_fcp_op(struct nvme_fc_ctrl *ctrl, struct nvme_fc_queue *queue,
2002 struct nvme_fc_fcp_op *op, u32 data_len,
2003 enum nvmefc_fcp_datadir io_dir)
2005 struct nvme_fc_cmd_iu *cmdiu = &op->cmd_iu;
2006 struct nvme_command *sqe = &cmdiu->sqe;
2011 * before attempting to send the io, check to see if we believe
2012 * the target device is present
2014 if (ctrl->rport->remoteport.port_state != FC_OBJSTATE_ONLINE)
2017 if (!nvme_fc_ctrl_get(ctrl))
2018 return BLK_STS_IOERR;
2020 /* format the FC-NVME CMD IU and fcp_req */
2021 cmdiu->connection_id = cpu_to_be64(queue->connection_id);
2022 csn = atomic_inc_return(&queue->csn);
2023 cmdiu->csn = cpu_to_be32(csn);
2024 cmdiu->data_len = cpu_to_be32(data_len);
2026 case NVMEFC_FCP_WRITE:
2027 cmdiu->flags = FCNVME_CMD_FLAGS_WRITE;
2029 case NVMEFC_FCP_READ:
2030 cmdiu->flags = FCNVME_CMD_FLAGS_READ;
2032 case NVMEFC_FCP_NODATA:
2036 op->fcp_req.payload_length = data_len;
2037 op->fcp_req.io_dir = io_dir;
2038 op->fcp_req.transferred_length = 0;
2039 op->fcp_req.rcv_rsplen = 0;
2040 op->fcp_req.status = NVME_SC_SUCCESS;
2041 op->fcp_req.sqid = cpu_to_le16(queue->qnum);
2044 * validate per fabric rules, set fields mandated by fabric spec
2045 * as well as those by FC-NVME spec.
2047 WARN_ON_ONCE(sqe->common.metadata);
2048 sqe->common.flags |= NVME_CMD_SGL_METABUF;
2051 * format SQE DPTR field per FC-NVME rules:
2052 * type=0x5 Transport SGL Data Block Descriptor
2053 * subtype=0xA Transport-specific value
2055 * length=length of the data series
2057 sqe->rw.dptr.sgl.type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2058 NVME_SGL_FMT_TRANSPORT_A;
2059 sqe->rw.dptr.sgl.length = cpu_to_le32(data_len);
2060 sqe->rw.dptr.sgl.addr = 0;
2062 if (!(op->flags & FCOP_FLAGS_AEN)) {
2063 ret = nvme_fc_map_data(ctrl, op->rq, op);
2065 nvme_cleanup_cmd(op->rq);
2066 nvme_fc_ctrl_put(ctrl);
2067 if (ret == -ENOMEM || ret == -EAGAIN)
2068 return BLK_STS_RESOURCE;
2069 return BLK_STS_IOERR;
2073 fc_dma_sync_single_for_device(ctrl->lport->dev, op->fcp_req.cmddma,
2074 sizeof(op->cmd_iu), DMA_TO_DEVICE);
2076 atomic_set(&op->state, FCPOP_STATE_ACTIVE);
2078 if (!(op->flags & FCOP_FLAGS_AEN))
2079 blk_mq_start_request(op->rq);
2081 ret = ctrl->lport->ops->fcp_io(&ctrl->lport->localport,
2082 &ctrl->rport->remoteport,
2083 queue->lldd_handle, &op->fcp_req);
2086 if (!(op->flags & FCOP_FLAGS_AEN))
2087 nvme_fc_unmap_data(ctrl, op->rq, op);
2089 nvme_fc_ctrl_put(ctrl);
2091 if (ctrl->rport->remoteport.port_state == FC_OBJSTATE_ONLINE &&
2093 return BLK_STS_IOERR;
2101 if (!(op->flags & FCOP_FLAGS_AEN) && queue->hctx)
2102 blk_mq_delay_run_hw_queue(queue->hctx, NVMEFC_QUEUE_DELAY);
2104 return BLK_STS_RESOURCE;
2108 nvme_fc_queue_rq(struct blk_mq_hw_ctx *hctx,
2109 const struct blk_mq_queue_data *bd)
2111 struct nvme_ns *ns = hctx->queue->queuedata;
2112 struct nvme_fc_queue *queue = hctx->driver_data;
2113 struct nvme_fc_ctrl *ctrl = queue->ctrl;
2114 struct request *rq = bd->rq;
2115 struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
2116 struct nvme_fc_cmd_iu *cmdiu = &op->cmd_iu;
2117 struct nvme_command *sqe = &cmdiu->sqe;
2118 enum nvmefc_fcp_datadir io_dir;
2122 ret = nvme_setup_cmd(ns, rq, sqe);
2126 data_len = blk_rq_payload_bytes(rq);
2128 io_dir = ((rq_data_dir(rq) == WRITE) ?
2129 NVMEFC_FCP_WRITE : NVMEFC_FCP_READ);
2131 io_dir = NVMEFC_FCP_NODATA;
2133 return nvme_fc_start_fcp_op(ctrl, queue, op, data_len, io_dir);
2136 static struct blk_mq_tags *
2137 nvme_fc_tagset(struct nvme_fc_queue *queue)
2139 if (queue->qnum == 0)
2140 return queue->ctrl->admin_tag_set.tags[queue->qnum];
2142 return queue->ctrl->tag_set.tags[queue->qnum - 1];
2146 nvme_fc_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
2149 struct nvme_fc_queue *queue = hctx->driver_data;
2150 struct nvme_fc_ctrl *ctrl = queue->ctrl;
2151 struct request *req;
2152 struct nvme_fc_fcp_op *op;
2154 req = blk_mq_tag_to_rq(nvme_fc_tagset(queue), tag);
2158 op = blk_mq_rq_to_pdu(req);
2160 if ((atomic_read(&op->state) == FCPOP_STATE_ACTIVE) &&
2161 (ctrl->lport->ops->poll_queue))
2162 ctrl->lport->ops->poll_queue(&ctrl->lport->localport,
2163 queue->lldd_handle);
2165 return ((atomic_read(&op->state) != FCPOP_STATE_ACTIVE));
2169 nvme_fc_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
2171 struct nvme_fc_ctrl *ctrl = to_fc_ctrl(arg);
2172 struct nvme_fc_fcp_op *aen_op;
2173 unsigned long flags;
2174 bool terminating = false;
2177 if (aer_idx > NVME_FC_NR_AEN_COMMANDS)
2180 spin_lock_irqsave(&ctrl->lock, flags);
2181 if (ctrl->flags & FCCTRL_TERMIO)
2183 spin_unlock_irqrestore(&ctrl->lock, flags);
2188 aen_op = &ctrl->aen_ops[aer_idx];
2190 ret = nvme_fc_start_fcp_op(ctrl, aen_op->queue, aen_op, 0,
2193 dev_err(ctrl->ctrl.device,
2194 "failed async event work [%d]\n", aer_idx);
2198 __nvme_fc_final_op_cleanup(struct request *rq)
2200 struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
2201 struct nvme_fc_ctrl *ctrl = op->ctrl;
2203 atomic_set(&op->state, FCPOP_STATE_IDLE);
2204 op->flags &= ~(FCOP_FLAGS_TERMIO | FCOP_FLAGS_RELEASED |
2205 FCOP_FLAGS_COMPLETE);
2207 nvme_fc_unmap_data(ctrl, rq, op);
2208 nvme_complete_rq(rq);
2209 nvme_fc_ctrl_put(ctrl);
2214 nvme_fc_complete_rq(struct request *rq)
2216 struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
2217 struct nvme_fc_ctrl *ctrl = op->ctrl;
2218 unsigned long flags;
2219 bool completed = false;
2222 * the core layer, on controller resets after calling
2223 * nvme_shutdown_ctrl(), calls complete_rq without our
2224 * calling blk_mq_complete_request(), thus there may still
2225 * be live i/o outstanding with the LLDD. Means transport has
2226 * to track complete calls vs fcpio_done calls to know what
2227 * path to take on completes and dones.
2229 spin_lock_irqsave(&ctrl->lock, flags);
2230 if (op->flags & FCOP_FLAGS_COMPLETE)
2233 op->flags |= FCOP_FLAGS_RELEASED;
2234 spin_unlock_irqrestore(&ctrl->lock, flags);
2237 __nvme_fc_final_op_cleanup(rq);
2241 * This routine is used by the transport when it needs to find active
2242 * io on a queue that is to be terminated. The transport uses
2243 * blk_mq_tagset_busy_itr() to find the busy requests, which then invoke
2244 * this routine to kill them on a 1 by 1 basis.
2246 * As FC allocates FC exchange for each io, the transport must contact
2247 * the LLDD to terminate the exchange, thus releasing the FC exchange.
2248 * After terminating the exchange the LLDD will call the transport's
2249 * normal io done path for the request, but it will have an aborted
2250 * status. The done path will return the io request back to the block
2251 * layer with an error status.
2254 nvme_fc_terminate_exchange(struct request *req, void *data, bool reserved)
2256 struct nvme_ctrl *nctrl = data;
2257 struct nvme_fc_ctrl *ctrl = to_fc_ctrl(nctrl);
2258 struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(req);
2259 unsigned long flags;
2262 if (!blk_mq_request_started(req))
2265 spin_lock_irqsave(&ctrl->lock, flags);
2266 if (ctrl->flags & FCCTRL_TERMIO) {
2268 op->flags |= FCOP_FLAGS_TERMIO;
2270 spin_unlock_irqrestore(&ctrl->lock, flags);
2272 status = __nvme_fc_abort_op(ctrl, op);
2275 * if __nvme_fc_abort_op failed the io wasn't
2276 * active. Thus this call path is running in
2277 * parallel to the io complete. Treat as non-error.
2280 /* back out the flags/counters */
2281 spin_lock_irqsave(&ctrl->lock, flags);
2282 if (ctrl->flags & FCCTRL_TERMIO)
2284 op->flags &= ~FCOP_FLAGS_TERMIO;
2285 spin_unlock_irqrestore(&ctrl->lock, flags);
2291 static const struct blk_mq_ops nvme_fc_mq_ops = {
2292 .queue_rq = nvme_fc_queue_rq,
2293 .complete = nvme_fc_complete_rq,
2294 .init_request = nvme_fc_init_request,
2295 .exit_request = nvme_fc_exit_request,
2296 .init_hctx = nvme_fc_init_hctx,
2297 .poll = nvme_fc_poll,
2298 .timeout = nvme_fc_timeout,
2302 nvme_fc_create_io_queues(struct nvme_fc_ctrl *ctrl)
2304 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2305 unsigned int nr_io_queues;
2308 nr_io_queues = min(min(opts->nr_io_queues, num_online_cpus()),
2309 ctrl->lport->ops->max_hw_queues);
2310 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
2312 dev_info(ctrl->ctrl.device,
2313 "set_queue_count failed: %d\n", ret);
2317 ctrl->ctrl.queue_count = nr_io_queues + 1;
2321 nvme_fc_init_io_queues(ctrl);
2323 memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
2324 ctrl->tag_set.ops = &nvme_fc_mq_ops;
2325 ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
2326 ctrl->tag_set.reserved_tags = 1; /* fabric connect */
2327 ctrl->tag_set.numa_node = NUMA_NO_NODE;
2328 ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
2329 ctrl->tag_set.cmd_size = sizeof(struct nvme_fc_fcp_op) +
2331 sizeof(struct scatterlist)) +
2332 ctrl->lport->ops->fcprqst_priv_sz;
2333 ctrl->tag_set.driver_data = ctrl;
2334 ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
2335 ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
2337 ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
2341 ctrl->ctrl.tagset = &ctrl->tag_set;
2343 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
2344 if (IS_ERR(ctrl->ctrl.connect_q)) {
2345 ret = PTR_ERR(ctrl->ctrl.connect_q);
2346 goto out_free_tag_set;
2349 ret = nvme_fc_create_hw_io_queues(ctrl, ctrl->ctrl.opts->queue_size);
2351 goto out_cleanup_blk_queue;
2353 ret = nvme_fc_connect_io_queues(ctrl, ctrl->ctrl.opts->queue_size);
2355 goto out_delete_hw_queues;
2359 out_delete_hw_queues:
2360 nvme_fc_delete_hw_io_queues(ctrl);
2361 out_cleanup_blk_queue:
2362 blk_cleanup_queue(ctrl->ctrl.connect_q);
2364 blk_mq_free_tag_set(&ctrl->tag_set);
2365 nvme_fc_free_io_queues(ctrl);
2367 /* force put free routine to ignore io queues */
2368 ctrl->ctrl.tagset = NULL;
2374 nvme_fc_reinit_io_queues(struct nvme_fc_ctrl *ctrl)
2376 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2377 unsigned int nr_io_queues;
2380 nr_io_queues = min(min(opts->nr_io_queues, num_online_cpus()),
2381 ctrl->lport->ops->max_hw_queues);
2382 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
2384 dev_info(ctrl->ctrl.device,
2385 "set_queue_count failed: %d\n", ret);
2389 ctrl->ctrl.queue_count = nr_io_queues + 1;
2390 /* check for io queues existing */
2391 if (ctrl->ctrl.queue_count == 1)
2394 nvme_fc_init_io_queues(ctrl);
2396 ret = nvme_reinit_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
2398 goto out_free_io_queues;
2400 ret = nvme_fc_create_hw_io_queues(ctrl, ctrl->ctrl.opts->queue_size);
2402 goto out_free_io_queues;
2404 ret = nvme_fc_connect_io_queues(ctrl, ctrl->ctrl.opts->queue_size);
2406 goto out_delete_hw_queues;
2408 blk_mq_update_nr_hw_queues(&ctrl->tag_set, nr_io_queues);
2412 out_delete_hw_queues:
2413 nvme_fc_delete_hw_io_queues(ctrl);
2415 nvme_fc_free_io_queues(ctrl);
2420 * This routine restarts the controller on the host side, and
2421 * on the link side, recreates the controller association.
2424 nvme_fc_create_association(struct nvme_fc_ctrl *ctrl)
2426 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2431 ++ctrl->ctrl.nr_reconnects;
2434 * Create the admin queue
2437 nvme_fc_init_queue(ctrl, 0, NVME_FC_AQ_BLKMQ_DEPTH);
2439 ret = __nvme_fc_create_hw_queue(ctrl, &ctrl->queues[0], 0,
2440 NVME_FC_AQ_BLKMQ_DEPTH);
2442 goto out_free_queue;
2444 ret = nvme_fc_connect_admin_queue(ctrl, &ctrl->queues[0],
2445 NVME_FC_AQ_BLKMQ_DEPTH,
2446 (NVME_FC_AQ_BLKMQ_DEPTH / 4));
2448 goto out_delete_hw_queue;
2450 if (ctrl->ctrl.state != NVME_CTRL_NEW)
2451 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
2453 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
2455 goto out_disconnect_admin_queue;
2458 * Check controller capabilities
2460 * todo:- add code to check if ctrl attributes changed from
2461 * prior connection values
2464 ret = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->ctrl.cap);
2466 dev_err(ctrl->ctrl.device,
2467 "prop_get NVME_REG_CAP failed\n");
2468 goto out_disconnect_admin_queue;
2472 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap) + 1, ctrl->ctrl.sqsize);
2474 ret = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
2476 goto out_disconnect_admin_queue;
2478 segs = min_t(u32, NVME_FC_MAX_SEGMENTS,
2479 ctrl->lport->ops->max_sgl_segments);
2480 ctrl->ctrl.max_hw_sectors = (segs - 1) << (PAGE_SHIFT - 9);
2482 ret = nvme_init_identify(&ctrl->ctrl);
2484 goto out_disconnect_admin_queue;
2488 /* FC-NVME does not have other data in the capsule */
2489 if (ctrl->ctrl.icdoff) {
2490 dev_err(ctrl->ctrl.device, "icdoff %d is not supported!\n",
2492 goto out_disconnect_admin_queue;
2495 /* FC-NVME supports normal SGL Data Block Descriptors */
2497 if (opts->queue_size > ctrl->ctrl.maxcmd) {
2498 /* warn if maxcmd is lower than queue_size */
2499 dev_warn(ctrl->ctrl.device,
2500 "queue_size %zu > ctrl maxcmd %u, reducing "
2502 opts->queue_size, ctrl->ctrl.maxcmd);
2503 opts->queue_size = ctrl->ctrl.maxcmd;
2506 ret = nvme_fc_init_aen_ops(ctrl);
2508 goto out_term_aen_ops;
2511 * Create the io queues
2514 if (ctrl->ctrl.queue_count > 1) {
2515 if (ctrl->ctrl.state == NVME_CTRL_NEW)
2516 ret = nvme_fc_create_io_queues(ctrl);
2518 ret = nvme_fc_reinit_io_queues(ctrl);
2520 goto out_term_aen_ops;
2523 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
2524 WARN_ON_ONCE(!changed);
2526 ctrl->ctrl.nr_reconnects = 0;
2528 nvme_start_ctrl(&ctrl->ctrl);
2530 return 0; /* Success */
2533 nvme_fc_term_aen_ops(ctrl);
2534 out_disconnect_admin_queue:
2535 /* send a Disconnect(association) LS to fc-nvme target */
2536 nvme_fc_xmt_disconnect_assoc(ctrl);
2537 out_delete_hw_queue:
2538 __nvme_fc_delete_hw_queue(ctrl, &ctrl->queues[0], 0);
2540 nvme_fc_free_queue(&ctrl->queues[0]);
2546 * This routine stops operation of the controller on the host side.
2547 * On the host os stack side: Admin and IO queues are stopped,
2548 * outstanding ios on them terminated via FC ABTS.
2549 * On the link side: the association is terminated.
2552 nvme_fc_delete_association(struct nvme_fc_ctrl *ctrl)
2554 unsigned long flags;
2556 spin_lock_irqsave(&ctrl->lock, flags);
2557 ctrl->flags |= FCCTRL_TERMIO;
2559 spin_unlock_irqrestore(&ctrl->lock, flags);
2562 * If io queues are present, stop them and terminate all outstanding
2563 * ios on them. As FC allocates FC exchange for each io, the
2564 * transport must contact the LLDD to terminate the exchange,
2565 * thus releasing the FC exchange. We use blk_mq_tagset_busy_itr()
2566 * to tell us what io's are busy and invoke a transport routine
2567 * to kill them with the LLDD. After terminating the exchange
2568 * the LLDD will call the transport's normal io done path, but it
2569 * will have an aborted status. The done path will return the
2570 * io requests back to the block layer as part of normal completions
2571 * (but with error status).
2573 if (ctrl->ctrl.queue_count > 1) {
2574 nvme_stop_queues(&ctrl->ctrl);
2575 blk_mq_tagset_busy_iter(&ctrl->tag_set,
2576 nvme_fc_terminate_exchange, &ctrl->ctrl);
2580 * Other transports, which don't have link-level contexts bound
2581 * to sqe's, would try to gracefully shutdown the controller by
2582 * writing the registers for shutdown and polling (call
2583 * nvme_shutdown_ctrl()). Given a bunch of i/o was potentially
2584 * just aborted and we will wait on those contexts, and given
2585 * there was no indication of how live the controlelr is on the
2586 * link, don't send more io to create more contexts for the
2587 * shutdown. Let the controller fail via keepalive failure if
2588 * its still present.
2592 * clean up the admin queue. Same thing as above.
2593 * use blk_mq_tagset_busy_itr() and the transport routine to
2594 * terminate the exchanges.
2596 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
2597 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
2598 nvme_fc_terminate_exchange, &ctrl->ctrl);
2600 /* kill the aens as they are a separate path */
2601 nvme_fc_abort_aen_ops(ctrl);
2603 /* wait for all io that had to be aborted */
2604 spin_lock_irqsave(&ctrl->lock, flags);
2605 wait_event_lock_irq(ctrl->ioabort_wait, ctrl->iocnt == 0, ctrl->lock);
2606 ctrl->flags &= ~FCCTRL_TERMIO;
2607 spin_unlock_irqrestore(&ctrl->lock, flags);
2609 nvme_fc_term_aen_ops(ctrl);
2612 * send a Disconnect(association) LS to fc-nvme target
2613 * Note: could have been sent at top of process, but
2614 * cleaner on link traffic if after the aborts complete.
2615 * Note: if association doesn't exist, association_id will be 0
2617 if (ctrl->association_id)
2618 nvme_fc_xmt_disconnect_assoc(ctrl);
2620 if (ctrl->ctrl.tagset) {
2621 nvme_fc_delete_hw_io_queues(ctrl);
2622 nvme_fc_free_io_queues(ctrl);
2625 __nvme_fc_delete_hw_queue(ctrl, &ctrl->queues[0], 0);
2626 nvme_fc_free_queue(&ctrl->queues[0]);
2630 nvme_fc_delete_ctrl_work(struct work_struct *work)
2632 struct nvme_fc_ctrl *ctrl =
2633 container_of(work, struct nvme_fc_ctrl, delete_work);
2635 cancel_work_sync(&ctrl->ctrl.reset_work);
2636 cancel_delayed_work_sync(&ctrl->connect_work);
2637 nvme_stop_ctrl(&ctrl->ctrl);
2638 nvme_remove_namespaces(&ctrl->ctrl);
2640 * kill the association on the link side. this will block
2641 * waiting for io to terminate
2643 nvme_fc_delete_association(ctrl);
2646 * tear down the controller
2647 * After the last reference on the nvme ctrl is removed,
2648 * the transport nvme_fc_nvme_ctrl_freed() callback will be
2649 * invoked. From there, the transport will tear down it's
2650 * logical queues and association.
2652 nvme_uninit_ctrl(&ctrl->ctrl);
2654 nvme_put_ctrl(&ctrl->ctrl);
2658 __nvme_fc_schedule_delete_work(struct nvme_fc_ctrl *ctrl)
2660 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
2663 if (!queue_work(nvme_wq, &ctrl->delete_work))
2670 __nvme_fc_del_ctrl(struct nvme_fc_ctrl *ctrl)
2672 return __nvme_fc_schedule_delete_work(ctrl) ? -EBUSY : 0;
2676 * Request from nvme core layer to delete the controller
2679 nvme_fc_del_nvme_ctrl(struct nvme_ctrl *nctrl)
2681 struct nvme_fc_ctrl *ctrl = to_fc_ctrl(nctrl);
2684 if (!kref_get_unless_zero(&ctrl->ctrl.kref))
2687 ret = __nvme_fc_del_ctrl(ctrl);
2690 flush_workqueue(nvme_wq);
2692 nvme_put_ctrl(&ctrl->ctrl);
2698 nvme_fc_reconnect_or_delete(struct nvme_fc_ctrl *ctrl, int status)
2700 /* If we are resetting/deleting then do nothing */
2701 if (ctrl->ctrl.state != NVME_CTRL_RECONNECTING) {
2702 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
2703 ctrl->ctrl.state == NVME_CTRL_LIVE);
2707 dev_info(ctrl->ctrl.device,
2708 "NVME-FC{%d}: reset: Reconnect attempt failed (%d)\n",
2709 ctrl->cnum, status);
2711 if (nvmf_should_reconnect(&ctrl->ctrl)) {
2712 dev_info(ctrl->ctrl.device,
2713 "NVME-FC{%d}: Reconnect attempt in %d seconds.\n",
2714 ctrl->cnum, ctrl->ctrl.opts->reconnect_delay);
2715 queue_delayed_work(nvme_wq, &ctrl->connect_work,
2716 ctrl->ctrl.opts->reconnect_delay * HZ);
2718 dev_warn(ctrl->ctrl.device,
2719 "NVME-FC{%d}: Max reconnect attempts (%d) "
2720 "reached. Removing controller\n",
2721 ctrl->cnum, ctrl->ctrl.nr_reconnects);
2722 WARN_ON(__nvme_fc_schedule_delete_work(ctrl));
2727 nvme_fc_reset_ctrl_work(struct work_struct *work)
2729 struct nvme_fc_ctrl *ctrl =
2730 container_of(work, struct nvme_fc_ctrl, ctrl.reset_work);
2733 nvme_stop_ctrl(&ctrl->ctrl);
2734 /* will block will waiting for io to terminate */
2735 nvme_fc_delete_association(ctrl);
2737 ret = nvme_fc_create_association(ctrl);
2739 nvme_fc_reconnect_or_delete(ctrl, ret);
2741 dev_info(ctrl->ctrl.device,
2742 "NVME-FC{%d}: controller reset complete\n", ctrl->cnum);
2745 static const struct nvme_ctrl_ops nvme_fc_ctrl_ops = {
2747 .module = THIS_MODULE,
2748 .flags = NVME_F_FABRICS,
2749 .reg_read32 = nvmf_reg_read32,
2750 .reg_read64 = nvmf_reg_read64,
2751 .reg_write32 = nvmf_reg_write32,
2752 .free_ctrl = nvme_fc_nvme_ctrl_freed,
2753 .submit_async_event = nvme_fc_submit_async_event,
2754 .delete_ctrl = nvme_fc_del_nvme_ctrl,
2755 .get_address = nvmf_get_address,
2756 .reinit_request = nvme_fc_reinit_request,
2760 nvme_fc_connect_ctrl_work(struct work_struct *work)
2764 struct nvme_fc_ctrl *ctrl =
2765 container_of(to_delayed_work(work),
2766 struct nvme_fc_ctrl, connect_work);
2768 ret = nvme_fc_create_association(ctrl);
2770 nvme_fc_reconnect_or_delete(ctrl, ret);
2772 dev_info(ctrl->ctrl.device,
2773 "NVME-FC{%d}: controller reconnect complete\n",
2778 static const struct blk_mq_ops nvme_fc_admin_mq_ops = {
2779 .queue_rq = nvme_fc_queue_rq,
2780 .complete = nvme_fc_complete_rq,
2781 .init_request = nvme_fc_init_request,
2782 .exit_request = nvme_fc_exit_request,
2783 .init_hctx = nvme_fc_init_admin_hctx,
2784 .timeout = nvme_fc_timeout,
2788 static struct nvme_ctrl *
2789 nvme_fc_init_ctrl(struct device *dev, struct nvmf_ctrl_options *opts,
2790 struct nvme_fc_lport *lport, struct nvme_fc_rport *rport)
2792 struct nvme_fc_ctrl *ctrl;
2793 unsigned long flags;
2796 if (!(rport->remoteport.port_role &
2797 (FC_PORT_ROLE_NVME_DISCOVERY | FC_PORT_ROLE_NVME_TARGET))) {
2802 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2808 idx = ida_simple_get(&nvme_fc_ctrl_cnt, 0, 0, GFP_KERNEL);
2814 ctrl->ctrl.opts = opts;
2815 INIT_LIST_HEAD(&ctrl->ctrl_list);
2816 ctrl->lport = lport;
2817 ctrl->rport = rport;
2818 ctrl->dev = lport->dev;
2821 get_device(ctrl->dev);
2822 kref_init(&ctrl->ref);
2824 INIT_WORK(&ctrl->delete_work, nvme_fc_delete_ctrl_work);
2825 INIT_WORK(&ctrl->ctrl.reset_work, nvme_fc_reset_ctrl_work);
2826 INIT_DELAYED_WORK(&ctrl->connect_work, nvme_fc_connect_ctrl_work);
2827 spin_lock_init(&ctrl->lock);
2829 /* io queue count */
2830 ctrl->ctrl.queue_count = min_t(unsigned int,
2832 lport->ops->max_hw_queues);
2833 ctrl->ctrl.queue_count++; /* +1 for admin queue */
2835 ctrl->ctrl.sqsize = opts->queue_size - 1;
2836 ctrl->ctrl.kato = opts->kato;
2839 ctrl->queues = kcalloc(ctrl->ctrl.queue_count,
2840 sizeof(struct nvme_fc_queue), GFP_KERNEL);
2844 memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
2845 ctrl->admin_tag_set.ops = &nvme_fc_admin_mq_ops;
2846 ctrl->admin_tag_set.queue_depth = NVME_FC_AQ_BLKMQ_DEPTH;
2847 ctrl->admin_tag_set.reserved_tags = 2; /* fabric connect + Keep-Alive */
2848 ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
2849 ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_fc_fcp_op) +
2851 sizeof(struct scatterlist)) +
2852 ctrl->lport->ops->fcprqst_priv_sz;
2853 ctrl->admin_tag_set.driver_data = ctrl;
2854 ctrl->admin_tag_set.nr_hw_queues = 1;
2855 ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
2857 ret = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
2859 goto out_free_queues;
2860 ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
2862 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
2863 if (IS_ERR(ctrl->ctrl.admin_q)) {
2864 ret = PTR_ERR(ctrl->ctrl.admin_q);
2865 goto out_free_admin_tag_set;
2869 * Would have been nice to init io queues tag set as well.
2870 * However, we require interaction from the controller
2871 * for max io queue count before we can do so.
2872 * Defer this to the connect path.
2875 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_fc_ctrl_ops, 0);
2877 goto out_cleanup_admin_q;
2879 /* at this point, teardown path changes to ref counting on nvme ctrl */
2881 spin_lock_irqsave(&rport->lock, flags);
2882 list_add_tail(&ctrl->ctrl_list, &rport->ctrl_list);
2883 spin_unlock_irqrestore(&rport->lock, flags);
2885 ret = nvme_fc_create_association(ctrl);
2887 ctrl->ctrl.opts = NULL;
2888 /* initiate nvme ctrl ref counting teardown */
2889 nvme_uninit_ctrl(&ctrl->ctrl);
2890 nvme_put_ctrl(&ctrl->ctrl);
2892 /* Remove core ctrl ref. */
2893 nvme_put_ctrl(&ctrl->ctrl);
2895 /* as we're past the point where we transition to the ref
2896 * counting teardown path, if we return a bad pointer here,
2897 * the calling routine, thinking it's prior to the
2898 * transition, will do an rport put. Since the teardown
2899 * path also does a rport put, we do an extra get here to
2900 * so proper order/teardown happens.
2902 nvme_fc_rport_get(rport);
2906 return ERR_PTR(ret);
2909 kref_get(&ctrl->ctrl.kref);
2911 dev_info(ctrl->ctrl.device,
2912 "NVME-FC{%d}: new ctrl: NQN \"%s\"\n",
2913 ctrl->cnum, ctrl->ctrl.opts->subsysnqn);
2917 out_cleanup_admin_q:
2918 blk_cleanup_queue(ctrl->ctrl.admin_q);
2919 out_free_admin_tag_set:
2920 blk_mq_free_tag_set(&ctrl->admin_tag_set);
2922 kfree(ctrl->queues);
2924 put_device(ctrl->dev);
2925 ida_simple_remove(&nvme_fc_ctrl_cnt, ctrl->cnum);
2929 /* exit via here doesn't follow ctlr ref points */
2930 return ERR_PTR(ret);
2934 struct nvmet_fc_traddr {
2940 __nvme_fc_parse_u64(substring_t *sstr, u64 *val)
2944 if (match_u64(sstr, &token64))
2952 * This routine validates and extracts the WWN's from the TRADDR string.
2953 * As kernel parsers need the 0x to determine number base, universally
2954 * build string to parse with 0x prefix before parsing name strings.
2957 nvme_fc_parse_traddr(struct nvmet_fc_traddr *traddr, char *buf, size_t blen)
2959 char name[2 + NVME_FC_TRADDR_HEXNAMELEN + 1];
2960 substring_t wwn = { name, &name[sizeof(name)-1] };
2961 int nnoffset, pnoffset;
2963 /* validate it string one of the 2 allowed formats */
2964 if (strnlen(buf, blen) == NVME_FC_TRADDR_MAXLENGTH &&
2965 !strncmp(buf, "nn-0x", NVME_FC_TRADDR_OXNNLEN) &&
2966 !strncmp(&buf[NVME_FC_TRADDR_MAX_PN_OFFSET],
2967 "pn-0x", NVME_FC_TRADDR_OXNNLEN)) {
2968 nnoffset = NVME_FC_TRADDR_OXNNLEN;
2969 pnoffset = NVME_FC_TRADDR_MAX_PN_OFFSET +
2970 NVME_FC_TRADDR_OXNNLEN;
2971 } else if ((strnlen(buf, blen) == NVME_FC_TRADDR_MINLENGTH &&
2972 !strncmp(buf, "nn-", NVME_FC_TRADDR_NNLEN) &&
2973 !strncmp(&buf[NVME_FC_TRADDR_MIN_PN_OFFSET],
2974 "pn-", NVME_FC_TRADDR_NNLEN))) {
2975 nnoffset = NVME_FC_TRADDR_NNLEN;
2976 pnoffset = NVME_FC_TRADDR_MIN_PN_OFFSET + NVME_FC_TRADDR_NNLEN;
2982 name[2 + NVME_FC_TRADDR_HEXNAMELEN] = 0;
2984 memcpy(&name[2], &buf[nnoffset], NVME_FC_TRADDR_HEXNAMELEN);
2985 if (__nvme_fc_parse_u64(&wwn, &traddr->nn))
2988 memcpy(&name[2], &buf[pnoffset], NVME_FC_TRADDR_HEXNAMELEN);
2989 if (__nvme_fc_parse_u64(&wwn, &traddr->pn))
2995 pr_warn("%s: bad traddr string\n", __func__);
2999 static struct nvme_ctrl *
3000 nvme_fc_create_ctrl(struct device *dev, struct nvmf_ctrl_options *opts)
3002 struct nvme_fc_lport *lport;
3003 struct nvme_fc_rport *rport;
3004 struct nvme_ctrl *ctrl;
3005 struct nvmet_fc_traddr laddr = { 0L, 0L };
3006 struct nvmet_fc_traddr raddr = { 0L, 0L };
3007 unsigned long flags;
3010 ret = nvme_fc_parse_traddr(&raddr, opts->traddr, NVMF_TRADDR_SIZE);
3011 if (ret || !raddr.nn || !raddr.pn)
3012 return ERR_PTR(-EINVAL);
3014 ret = nvme_fc_parse_traddr(&laddr, opts->host_traddr, NVMF_TRADDR_SIZE);
3015 if (ret || !laddr.nn || !laddr.pn)
3016 return ERR_PTR(-EINVAL);
3018 /* find the host and remote ports to connect together */
3019 spin_lock_irqsave(&nvme_fc_lock, flags);
3020 list_for_each_entry(lport, &nvme_fc_lport_list, port_list) {
3021 if (lport->localport.node_name != laddr.nn ||
3022 lport->localport.port_name != laddr.pn)
3025 list_for_each_entry(rport, &lport->endp_list, endp_list) {
3026 if (rport->remoteport.node_name != raddr.nn ||
3027 rport->remoteport.port_name != raddr.pn)
3030 /* if fail to get reference fall through. Will error */
3031 if (!nvme_fc_rport_get(rport))
3034 spin_unlock_irqrestore(&nvme_fc_lock, flags);
3036 ctrl = nvme_fc_init_ctrl(dev, opts, lport, rport);
3038 nvme_fc_rport_put(rport);
3042 spin_unlock_irqrestore(&nvme_fc_lock, flags);
3044 return ERR_PTR(-ENOENT);
3048 static struct nvmf_transport_ops nvme_fc_transport = {
3050 .required_opts = NVMF_OPT_TRADDR | NVMF_OPT_HOST_TRADDR,
3051 .allowed_opts = NVMF_OPT_RECONNECT_DELAY | NVMF_OPT_CTRL_LOSS_TMO,
3052 .create_ctrl = nvme_fc_create_ctrl,
3055 static int __init nvme_fc_init_module(void)
3061 * It is expected that in the future the kernel will combine
3062 * the FC-isms that are currently under scsi and now being
3063 * added to by NVME into a new standalone FC class. The SCSI
3064 * and NVME protocols and their devices would be under this
3067 * As we need something to post FC-specific udev events to,
3068 * specifically for nvme probe events, start by creating the
3069 * new device class. When the new standalone FC class is
3070 * put in place, this code will move to a more generic
3071 * location for the class.
3073 fc_class = class_create(THIS_MODULE, "fc");
3074 if (IS_ERR(fc_class)) {
3075 pr_err("couldn't register class fc\n");
3076 return PTR_ERR(fc_class);
3080 * Create a device for the FC-centric udev events
3082 fc_udev_device = device_create(fc_class, NULL, MKDEV(0, 0), NULL,
3084 if (IS_ERR(fc_udev_device)) {
3085 pr_err("couldn't create fc_udev device!\n");
3086 ret = PTR_ERR(fc_udev_device);
3087 goto out_destroy_class;
3090 ret = nvmf_register_transport(&nvme_fc_transport);
3092 goto out_destroy_device;
3097 device_destroy(fc_class, MKDEV(0, 0));
3099 class_destroy(fc_class);
3103 static void __exit nvme_fc_exit_module(void)
3105 /* sanity check - all lports should be removed */
3106 if (!list_empty(&nvme_fc_lport_list))
3107 pr_warn("%s: localport list not empty\n", __func__);
3109 nvmf_unregister_transport(&nvme_fc_transport);
3111 ida_destroy(&nvme_fc_local_port_cnt);
3112 ida_destroy(&nvme_fc_ctrl_cnt);
3114 device_destroy(fc_class, MKDEV(0, 0));
3115 class_destroy(fc_class);
3118 module_init(nvme_fc_init_module);
3119 module_exit(nvme_fc_exit_module);
3121 MODULE_LICENSE("GPL v2");