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 ****************** */
33 enum nvme_fc_queue_flags {
34 NVME_FC_Q_CONNECTED = 0,
38 #define NVME_FC_DEFAULT_DEV_LOSS_TMO 60 /* seconds */
40 struct nvme_fc_queue {
41 struct nvme_fc_ctrl *ctrl;
43 struct blk_mq_hw_ctx *hctx;
45 size_t cmnd_capsule_len;
54 } __aligned(sizeof(u64)); /* alignment for other things alloc'd with */
56 enum nvme_fcop_flags {
57 FCOP_FLAGS_TERMIO = (1 << 0),
58 FCOP_FLAGS_AEN = (1 << 1),
61 struct nvmefc_ls_req_op {
62 struct nvmefc_ls_req ls_req;
64 struct nvme_fc_rport *rport;
65 struct nvme_fc_queue *queue;
70 struct completion ls_done;
71 struct list_head lsreq_list; /* rport->ls_req_list */
75 enum nvme_fcpop_state {
76 FCPOP_STATE_UNINIT = 0,
78 FCPOP_STATE_ACTIVE = 2,
79 FCPOP_STATE_ABORTED = 3,
80 FCPOP_STATE_COMPLETE = 4,
83 struct nvme_fc_fcp_op {
84 struct nvme_request nreq; /*
87 * the 1st element in the
92 struct nvmefc_fcp_req fcp_req;
94 struct nvme_fc_ctrl *ctrl;
95 struct nvme_fc_queue *queue;
103 struct nvme_fc_cmd_iu cmd_iu;
104 struct nvme_fc_ersp_iu rsp_iu;
107 struct nvme_fc_lport {
108 struct nvme_fc_local_port localport;
111 struct list_head port_list; /* nvme_fc_port_list */
112 struct list_head endp_list;
113 struct device *dev; /* physical device for dma */
114 struct nvme_fc_port_template *ops;
116 atomic_t act_rport_cnt;
117 } __aligned(sizeof(u64)); /* alignment for other things alloc'd with */
119 struct nvme_fc_rport {
120 struct nvme_fc_remote_port remoteport;
122 struct list_head endp_list; /* for lport->endp_list */
123 struct list_head ctrl_list;
124 struct list_head ls_req_list;
125 struct device *dev; /* physical device for dma */
126 struct nvme_fc_lport *lport;
129 atomic_t act_ctrl_cnt;
130 unsigned long dev_loss_end;
131 } __aligned(sizeof(u64)); /* alignment for other things alloc'd with */
133 enum nvme_fcctrl_flags {
134 FCCTRL_TERMIO = (1 << 0),
137 struct nvme_fc_ctrl {
139 struct nvme_fc_queue *queues;
141 struct nvme_fc_lport *lport;
142 struct nvme_fc_rport *rport;
148 struct list_head ctrl_list; /* rport->ctrl_list */
150 struct blk_mq_tag_set admin_tag_set;
151 struct blk_mq_tag_set tag_set;
153 struct delayed_work connect_work;
158 wait_queue_head_t ioabort_wait;
160 struct nvme_fc_fcp_op aen_ops[NVME_NR_AEN_COMMANDS];
162 struct nvme_ctrl ctrl;
165 static inline struct nvme_fc_ctrl *
166 to_fc_ctrl(struct nvme_ctrl *ctrl)
168 return container_of(ctrl, struct nvme_fc_ctrl, ctrl);
171 static inline struct nvme_fc_lport *
172 localport_to_lport(struct nvme_fc_local_port *portptr)
174 return container_of(portptr, struct nvme_fc_lport, localport);
177 static inline struct nvme_fc_rport *
178 remoteport_to_rport(struct nvme_fc_remote_port *portptr)
180 return container_of(portptr, struct nvme_fc_rport, remoteport);
183 static inline struct nvmefc_ls_req_op *
184 ls_req_to_lsop(struct nvmefc_ls_req *lsreq)
186 return container_of(lsreq, struct nvmefc_ls_req_op, ls_req);
189 static inline struct nvme_fc_fcp_op *
190 fcp_req_to_fcp_op(struct nvmefc_fcp_req *fcpreq)
192 return container_of(fcpreq, struct nvme_fc_fcp_op, fcp_req);
197 /* *************************** Globals **************************** */
200 static DEFINE_SPINLOCK(nvme_fc_lock);
202 static LIST_HEAD(nvme_fc_lport_list);
203 static DEFINE_IDA(nvme_fc_local_port_cnt);
204 static DEFINE_IDA(nvme_fc_ctrl_cnt);
209 * These items are short-term. They will eventually be moved into
210 * a generic FC class. See comments in module init.
212 static struct class *fc_class;
213 static struct device *fc_udev_device;
216 /* *********************** FC-NVME Port Management ************************ */
218 static void __nvme_fc_delete_hw_queue(struct nvme_fc_ctrl *,
219 struct nvme_fc_queue *, unsigned int);
222 nvme_fc_free_lport(struct kref *ref)
224 struct nvme_fc_lport *lport =
225 container_of(ref, struct nvme_fc_lport, ref);
228 WARN_ON(lport->localport.port_state != FC_OBJSTATE_DELETED);
229 WARN_ON(!list_empty(&lport->endp_list));
231 /* remove from transport list */
232 spin_lock_irqsave(&nvme_fc_lock, flags);
233 list_del(&lport->port_list);
234 spin_unlock_irqrestore(&nvme_fc_lock, flags);
236 ida_simple_remove(&nvme_fc_local_port_cnt, lport->localport.port_num);
237 ida_destroy(&lport->endp_cnt);
239 put_device(lport->dev);
245 nvme_fc_lport_put(struct nvme_fc_lport *lport)
247 kref_put(&lport->ref, nvme_fc_free_lport);
251 nvme_fc_lport_get(struct nvme_fc_lport *lport)
253 return kref_get_unless_zero(&lport->ref);
257 static struct nvme_fc_lport *
258 nvme_fc_attach_to_unreg_lport(struct nvme_fc_port_info *pinfo,
259 struct nvme_fc_port_template *ops,
262 struct nvme_fc_lport *lport;
265 spin_lock_irqsave(&nvme_fc_lock, flags);
267 list_for_each_entry(lport, &nvme_fc_lport_list, port_list) {
268 if (lport->localport.node_name != pinfo->node_name ||
269 lport->localport.port_name != pinfo->port_name)
272 if (lport->dev != dev) {
273 lport = ERR_PTR(-EXDEV);
277 if (lport->localport.port_state != FC_OBJSTATE_DELETED) {
278 lport = ERR_PTR(-EEXIST);
282 if (!nvme_fc_lport_get(lport)) {
284 * fails if ref cnt already 0. If so,
285 * act as if lport already deleted
291 /* resume the lport */
294 lport->localport.port_role = pinfo->port_role;
295 lport->localport.port_id = pinfo->port_id;
296 lport->localport.port_state = FC_OBJSTATE_ONLINE;
298 spin_unlock_irqrestore(&nvme_fc_lock, flags);
306 spin_unlock_irqrestore(&nvme_fc_lock, flags);
312 * nvme_fc_register_localport - transport entry point called by an
313 * LLDD to register the existence of a NVME
315 * @pinfo: pointer to information about the port to be registered
316 * @template: LLDD entrypoints and operational parameters for the port
317 * @dev: physical hardware device node port corresponds to. Will be
318 * used for DMA mappings
319 * @lport_p: pointer to a local port pointer. Upon success, the routine
320 * will allocate a nvme_fc_local_port structure and place its
321 * address in the local port pointer. Upon failure, local port
322 * pointer will be set to 0.
325 * a completion status. Must be 0 upon success; a negative errno
326 * (ex: -ENXIO) upon failure.
329 nvme_fc_register_localport(struct nvme_fc_port_info *pinfo,
330 struct nvme_fc_port_template *template,
332 struct nvme_fc_local_port **portptr)
334 struct nvme_fc_lport *newrec;
338 if (!template->localport_delete || !template->remoteport_delete ||
339 !template->ls_req || !template->fcp_io ||
340 !template->ls_abort || !template->fcp_abort ||
341 !template->max_hw_queues || !template->max_sgl_segments ||
342 !template->max_dif_sgl_segments || !template->dma_boundary) {
344 goto out_reghost_failed;
348 * look to see if there is already a localport that had been
349 * deregistered and in the process of waiting for all the
350 * references to fully be removed. If the references haven't
351 * expired, we can simply re-enable the localport. Remoteports
352 * and controller reconnections should resume naturally.
354 newrec = nvme_fc_attach_to_unreg_lport(pinfo, template, dev);
356 /* found an lport, but something about its state is bad */
357 if (IS_ERR(newrec)) {
358 ret = PTR_ERR(newrec);
359 goto out_reghost_failed;
361 /* found existing lport, which was resumed */
363 *portptr = &newrec->localport;
367 /* nothing found - allocate a new localport struct */
369 newrec = kmalloc((sizeof(*newrec) + template->local_priv_sz),
373 goto out_reghost_failed;
376 idx = ida_simple_get(&nvme_fc_local_port_cnt, 0, 0, GFP_KERNEL);
382 if (!get_device(dev) && dev) {
387 INIT_LIST_HEAD(&newrec->port_list);
388 INIT_LIST_HEAD(&newrec->endp_list);
389 kref_init(&newrec->ref);
390 atomic_set(&newrec->act_rport_cnt, 0);
391 newrec->ops = template;
393 ida_init(&newrec->endp_cnt);
394 newrec->localport.private = &newrec[1];
395 newrec->localport.node_name = pinfo->node_name;
396 newrec->localport.port_name = pinfo->port_name;
397 newrec->localport.port_role = pinfo->port_role;
398 newrec->localport.port_id = pinfo->port_id;
399 newrec->localport.port_state = FC_OBJSTATE_ONLINE;
400 newrec->localport.port_num = idx;
402 spin_lock_irqsave(&nvme_fc_lock, flags);
403 list_add_tail(&newrec->port_list, &nvme_fc_lport_list);
404 spin_unlock_irqrestore(&nvme_fc_lock, flags);
407 dma_set_seg_boundary(dev, template->dma_boundary);
409 *portptr = &newrec->localport;
413 ida_simple_remove(&nvme_fc_local_port_cnt, idx);
421 EXPORT_SYMBOL_GPL(nvme_fc_register_localport);
424 * nvme_fc_unregister_localport - transport entry point called by an
425 * LLDD to deregister/remove a previously
426 * registered a NVME host FC port.
427 * @localport: pointer to the (registered) local port that is to be
431 * a completion status. Must be 0 upon success; a negative errno
432 * (ex: -ENXIO) upon failure.
435 nvme_fc_unregister_localport(struct nvme_fc_local_port *portptr)
437 struct nvme_fc_lport *lport = localport_to_lport(portptr);
443 spin_lock_irqsave(&nvme_fc_lock, flags);
445 if (portptr->port_state != FC_OBJSTATE_ONLINE) {
446 spin_unlock_irqrestore(&nvme_fc_lock, flags);
449 portptr->port_state = FC_OBJSTATE_DELETED;
451 spin_unlock_irqrestore(&nvme_fc_lock, flags);
453 if (atomic_read(&lport->act_rport_cnt) == 0)
454 lport->ops->localport_delete(&lport->localport);
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 ida_simple_remove(&lport->endp_cnt, rport->remoteport.port_num);
513 nvme_fc_lport_put(lport);
517 nvme_fc_rport_put(struct nvme_fc_rport *rport)
519 kref_put(&rport->ref, nvme_fc_free_rport);
523 nvme_fc_rport_get(struct nvme_fc_rport *rport)
525 return kref_get_unless_zero(&rport->ref);
529 nvme_fc_resume_controller(struct nvme_fc_ctrl *ctrl)
531 switch (ctrl->ctrl.state) {
533 case NVME_CTRL_CONNECTING:
535 * As all reconnects were suppressed, schedule a
538 dev_info(ctrl->ctrl.device,
539 "NVME-FC{%d}: connectivity re-established. "
540 "Attempting reconnect\n", ctrl->cnum);
542 queue_delayed_work(nvme_wq, &ctrl->connect_work, 0);
545 case NVME_CTRL_RESETTING:
547 * Controller is already in the process of terminating the
548 * association. No need to do anything further. The reconnect
549 * step will naturally occur after the reset completes.
554 /* no action to take - let it delete */
559 static struct nvme_fc_rport *
560 nvme_fc_attach_to_suspended_rport(struct nvme_fc_lport *lport,
561 struct nvme_fc_port_info *pinfo)
563 struct nvme_fc_rport *rport;
564 struct nvme_fc_ctrl *ctrl;
567 spin_lock_irqsave(&nvme_fc_lock, flags);
569 list_for_each_entry(rport, &lport->endp_list, endp_list) {
570 if (rport->remoteport.node_name != pinfo->node_name ||
571 rport->remoteport.port_name != pinfo->port_name)
574 if (!nvme_fc_rport_get(rport)) {
575 rport = ERR_PTR(-ENOLCK);
579 spin_unlock_irqrestore(&nvme_fc_lock, flags);
581 spin_lock_irqsave(&rport->lock, flags);
583 /* has it been unregistered */
584 if (rport->remoteport.port_state != FC_OBJSTATE_DELETED) {
585 /* means lldd called us twice */
586 spin_unlock_irqrestore(&rport->lock, flags);
587 nvme_fc_rport_put(rport);
588 return ERR_PTR(-ESTALE);
591 rport->remoteport.port_state = FC_OBJSTATE_ONLINE;
592 rport->dev_loss_end = 0;
595 * kick off a reconnect attempt on all associations to the
596 * remote port. A successful reconnects will resume i/o.
598 list_for_each_entry(ctrl, &rport->ctrl_list, ctrl_list)
599 nvme_fc_resume_controller(ctrl);
601 spin_unlock_irqrestore(&rport->lock, flags);
609 spin_unlock_irqrestore(&nvme_fc_lock, flags);
615 __nvme_fc_set_dev_loss_tmo(struct nvme_fc_rport *rport,
616 struct nvme_fc_port_info *pinfo)
618 if (pinfo->dev_loss_tmo)
619 rport->remoteport.dev_loss_tmo = pinfo->dev_loss_tmo;
621 rport->remoteport.dev_loss_tmo = NVME_FC_DEFAULT_DEV_LOSS_TMO;
625 * nvme_fc_register_remoteport - transport entry point called by an
626 * LLDD to register the existence of a NVME
627 * subsystem FC port on its fabric.
628 * @localport: pointer to the (registered) local port that the remote
629 * subsystem port is connected to.
630 * @pinfo: pointer to information about the port to be registered
631 * @rport_p: pointer to a remote port pointer. Upon success, the routine
632 * will allocate a nvme_fc_remote_port structure and place its
633 * address in the remote port pointer. Upon failure, remote port
634 * pointer will be set to 0.
637 * a completion status. Must be 0 upon success; a negative errno
638 * (ex: -ENXIO) upon failure.
641 nvme_fc_register_remoteport(struct nvme_fc_local_port *localport,
642 struct nvme_fc_port_info *pinfo,
643 struct nvme_fc_remote_port **portptr)
645 struct nvme_fc_lport *lport = localport_to_lport(localport);
646 struct nvme_fc_rport *newrec;
650 if (!nvme_fc_lport_get(lport)) {
652 goto out_reghost_failed;
656 * look to see if there is already a remoteport that is waiting
657 * for a reconnect (within dev_loss_tmo) with the same WWN's.
658 * If so, transition to it and reconnect.
660 newrec = nvme_fc_attach_to_suspended_rport(lport, pinfo);
662 /* found an rport, but something about its state is bad */
663 if (IS_ERR(newrec)) {
664 ret = PTR_ERR(newrec);
667 /* found existing rport, which was resumed */
669 nvme_fc_lport_put(lport);
670 __nvme_fc_set_dev_loss_tmo(newrec, pinfo);
671 nvme_fc_signal_discovery_scan(lport, newrec);
672 *portptr = &newrec->remoteport;
676 /* nothing found - allocate a new remoteport struct */
678 newrec = kmalloc((sizeof(*newrec) + lport->ops->remote_priv_sz),
685 idx = ida_simple_get(&lport->endp_cnt, 0, 0, GFP_KERNEL);
688 goto out_kfree_rport;
691 INIT_LIST_HEAD(&newrec->endp_list);
692 INIT_LIST_HEAD(&newrec->ctrl_list);
693 INIT_LIST_HEAD(&newrec->ls_req_list);
694 kref_init(&newrec->ref);
695 atomic_set(&newrec->act_ctrl_cnt, 0);
696 spin_lock_init(&newrec->lock);
697 newrec->remoteport.localport = &lport->localport;
698 newrec->dev = lport->dev;
699 newrec->lport = lport;
700 newrec->remoteport.private = &newrec[1];
701 newrec->remoteport.port_role = pinfo->port_role;
702 newrec->remoteport.node_name = pinfo->node_name;
703 newrec->remoteport.port_name = pinfo->port_name;
704 newrec->remoteport.port_id = pinfo->port_id;
705 newrec->remoteport.port_state = FC_OBJSTATE_ONLINE;
706 newrec->remoteport.port_num = idx;
707 __nvme_fc_set_dev_loss_tmo(newrec, pinfo);
709 spin_lock_irqsave(&nvme_fc_lock, flags);
710 list_add_tail(&newrec->endp_list, &lport->endp_list);
711 spin_unlock_irqrestore(&nvme_fc_lock, flags);
713 nvme_fc_signal_discovery_scan(lport, newrec);
715 *portptr = &newrec->remoteport;
721 nvme_fc_lport_put(lport);
726 EXPORT_SYMBOL_GPL(nvme_fc_register_remoteport);
729 nvme_fc_abort_lsops(struct nvme_fc_rport *rport)
731 struct nvmefc_ls_req_op *lsop;
735 spin_lock_irqsave(&rport->lock, flags);
737 list_for_each_entry(lsop, &rport->ls_req_list, lsreq_list) {
738 if (!(lsop->flags & FCOP_FLAGS_TERMIO)) {
739 lsop->flags |= FCOP_FLAGS_TERMIO;
740 spin_unlock_irqrestore(&rport->lock, flags);
741 rport->lport->ops->ls_abort(&rport->lport->localport,
747 spin_unlock_irqrestore(&rport->lock, flags);
753 nvme_fc_ctrl_connectivity_loss(struct nvme_fc_ctrl *ctrl)
755 dev_info(ctrl->ctrl.device,
756 "NVME-FC{%d}: controller connectivity lost. Awaiting "
757 "Reconnect", ctrl->cnum);
759 switch (ctrl->ctrl.state) {
763 * Schedule a controller reset. The reset will terminate the
764 * association and schedule the reconnect timer. Reconnects
765 * will be attempted until either the ctlr_loss_tmo
766 * (max_retries * connect_delay) expires or the remoteport's
767 * dev_loss_tmo expires.
769 if (nvme_reset_ctrl(&ctrl->ctrl)) {
770 dev_warn(ctrl->ctrl.device,
771 "NVME-FC{%d}: Couldn't schedule reset. "
772 "Deleting controller.\n",
774 nvme_delete_ctrl(&ctrl->ctrl);
778 case NVME_CTRL_CONNECTING:
780 * The association has already been terminated and the
781 * controller is attempting reconnects. No need to do anything
782 * futher. Reconnects will be attempted until either the
783 * ctlr_loss_tmo (max_retries * connect_delay) expires or the
784 * remoteport's dev_loss_tmo expires.
788 case NVME_CTRL_RESETTING:
790 * Controller is already in the process of terminating the
791 * association. No need to do anything further. The reconnect
792 * step will kick in naturally after the association is
797 case NVME_CTRL_DELETING:
799 /* no action to take - let it delete */
805 * nvme_fc_unregister_remoteport - transport entry point called by an
806 * LLDD to deregister/remove a previously
807 * registered a NVME subsystem FC port.
808 * @remoteport: pointer to the (registered) remote port that is to be
812 * a completion status. Must be 0 upon success; a negative errno
813 * (ex: -ENXIO) upon failure.
816 nvme_fc_unregister_remoteport(struct nvme_fc_remote_port *portptr)
818 struct nvme_fc_rport *rport = remoteport_to_rport(portptr);
819 struct nvme_fc_ctrl *ctrl;
825 spin_lock_irqsave(&rport->lock, flags);
827 if (portptr->port_state != FC_OBJSTATE_ONLINE) {
828 spin_unlock_irqrestore(&rport->lock, flags);
831 portptr->port_state = FC_OBJSTATE_DELETED;
833 rport->dev_loss_end = jiffies + (portptr->dev_loss_tmo * HZ);
835 list_for_each_entry(ctrl, &rport->ctrl_list, ctrl_list) {
836 /* if dev_loss_tmo==0, dev loss is immediate */
837 if (!portptr->dev_loss_tmo) {
838 dev_warn(ctrl->ctrl.device,
839 "NVME-FC{%d}: controller connectivity lost. "
840 "Deleting controller.\n",
842 nvme_delete_ctrl(&ctrl->ctrl);
844 nvme_fc_ctrl_connectivity_loss(ctrl);
847 spin_unlock_irqrestore(&rport->lock, flags);
849 nvme_fc_abort_lsops(rport);
851 if (atomic_read(&rport->act_ctrl_cnt) == 0)
852 rport->lport->ops->remoteport_delete(portptr);
855 * release the reference, which will allow, if all controllers
856 * go away, which should only occur after dev_loss_tmo occurs,
857 * for the rport to be torn down.
859 nvme_fc_rport_put(rport);
863 EXPORT_SYMBOL_GPL(nvme_fc_unregister_remoteport);
866 * nvme_fc_rescan_remoteport - transport entry point called by an
867 * LLDD to request a nvme device rescan.
868 * @remoteport: pointer to the (registered) remote port that is to be
874 nvme_fc_rescan_remoteport(struct nvme_fc_remote_port *remoteport)
876 struct nvme_fc_rport *rport = remoteport_to_rport(remoteport);
878 nvme_fc_signal_discovery_scan(rport->lport, rport);
880 EXPORT_SYMBOL_GPL(nvme_fc_rescan_remoteport);
883 nvme_fc_set_remoteport_devloss(struct nvme_fc_remote_port *portptr,
886 struct nvme_fc_rport *rport = remoteport_to_rport(portptr);
889 spin_lock_irqsave(&rport->lock, flags);
891 if (portptr->port_state != FC_OBJSTATE_ONLINE) {
892 spin_unlock_irqrestore(&rport->lock, flags);
896 /* a dev_loss_tmo of 0 (immediate) is allowed to be set */
897 rport->remoteport.dev_loss_tmo = dev_loss_tmo;
899 spin_unlock_irqrestore(&rport->lock, flags);
903 EXPORT_SYMBOL_GPL(nvme_fc_set_remoteport_devloss);
906 /* *********************** FC-NVME DMA Handling **************************** */
909 * The fcloop device passes in a NULL device pointer. Real LLD's will
910 * pass in a valid device pointer. If NULL is passed to the dma mapping
911 * routines, depending on the platform, it may or may not succeed, and
915 * Wrapper all the dma routines and check the dev pointer.
917 * If simple mappings (return just a dma address, we'll noop them,
918 * returning a dma address of 0.
920 * On more complex mappings (dma_map_sg), a pseudo routine fills
921 * in the scatter list, setting all dma addresses to 0.
924 static inline dma_addr_t
925 fc_dma_map_single(struct device *dev, void *ptr, size_t size,
926 enum dma_data_direction dir)
928 return dev ? dma_map_single(dev, ptr, size, dir) : (dma_addr_t)0L;
932 fc_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
934 return dev ? dma_mapping_error(dev, dma_addr) : 0;
938 fc_dma_unmap_single(struct device *dev, dma_addr_t addr, size_t size,
939 enum dma_data_direction dir)
942 dma_unmap_single(dev, addr, size, dir);
946 fc_dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr, size_t size,
947 enum dma_data_direction dir)
950 dma_sync_single_for_cpu(dev, addr, size, dir);
954 fc_dma_sync_single_for_device(struct device *dev, dma_addr_t addr, size_t size,
955 enum dma_data_direction dir)
958 dma_sync_single_for_device(dev, addr, size, dir);
961 /* pseudo dma_map_sg call */
963 fc_map_sg(struct scatterlist *sg, int nents)
965 struct scatterlist *s;
968 WARN_ON(nents == 0 || sg[0].length == 0);
970 for_each_sg(sg, s, nents, i) {
972 #ifdef CONFIG_NEED_SG_DMA_LENGTH
973 s->dma_length = s->length;
980 fc_dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
981 enum dma_data_direction dir)
983 return dev ? dma_map_sg(dev, sg, nents, dir) : fc_map_sg(sg, nents);
987 fc_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
988 enum dma_data_direction dir)
991 dma_unmap_sg(dev, sg, nents, dir);
994 /* *********************** FC-NVME LS Handling **************************** */
996 static void nvme_fc_ctrl_put(struct nvme_fc_ctrl *);
997 static int nvme_fc_ctrl_get(struct nvme_fc_ctrl *);
1001 __nvme_fc_finish_ls_req(struct nvmefc_ls_req_op *lsop)
1003 struct nvme_fc_rport *rport = lsop->rport;
1004 struct nvmefc_ls_req *lsreq = &lsop->ls_req;
1005 unsigned long flags;
1007 spin_lock_irqsave(&rport->lock, flags);
1009 if (!lsop->req_queued) {
1010 spin_unlock_irqrestore(&rport->lock, flags);
1014 list_del(&lsop->lsreq_list);
1016 lsop->req_queued = false;
1018 spin_unlock_irqrestore(&rport->lock, flags);
1020 fc_dma_unmap_single(rport->dev, lsreq->rqstdma,
1021 (lsreq->rqstlen + lsreq->rsplen),
1024 nvme_fc_rport_put(rport);
1028 __nvme_fc_send_ls_req(struct nvme_fc_rport *rport,
1029 struct nvmefc_ls_req_op *lsop,
1030 void (*done)(struct nvmefc_ls_req *req, int status))
1032 struct nvmefc_ls_req *lsreq = &lsop->ls_req;
1033 unsigned long flags;
1036 if (rport->remoteport.port_state != FC_OBJSTATE_ONLINE)
1037 return -ECONNREFUSED;
1039 if (!nvme_fc_rport_get(rport))
1043 lsop->rport = rport;
1044 lsop->req_queued = false;
1045 INIT_LIST_HEAD(&lsop->lsreq_list);
1046 init_completion(&lsop->ls_done);
1048 lsreq->rqstdma = fc_dma_map_single(rport->dev, lsreq->rqstaddr,
1049 lsreq->rqstlen + lsreq->rsplen,
1051 if (fc_dma_mapping_error(rport->dev, lsreq->rqstdma)) {
1055 lsreq->rspdma = lsreq->rqstdma + lsreq->rqstlen;
1057 spin_lock_irqsave(&rport->lock, flags);
1059 list_add_tail(&lsop->lsreq_list, &rport->ls_req_list);
1061 lsop->req_queued = true;
1063 spin_unlock_irqrestore(&rport->lock, flags);
1065 ret = rport->lport->ops->ls_req(&rport->lport->localport,
1066 &rport->remoteport, lsreq);
1073 lsop->ls_error = ret;
1074 spin_lock_irqsave(&rport->lock, flags);
1075 lsop->req_queued = false;
1076 list_del(&lsop->lsreq_list);
1077 spin_unlock_irqrestore(&rport->lock, flags);
1078 fc_dma_unmap_single(rport->dev, lsreq->rqstdma,
1079 (lsreq->rqstlen + lsreq->rsplen),
1082 nvme_fc_rport_put(rport);
1088 nvme_fc_send_ls_req_done(struct nvmefc_ls_req *lsreq, int status)
1090 struct nvmefc_ls_req_op *lsop = ls_req_to_lsop(lsreq);
1092 lsop->ls_error = status;
1093 complete(&lsop->ls_done);
1097 nvme_fc_send_ls_req(struct nvme_fc_rport *rport, struct nvmefc_ls_req_op *lsop)
1099 struct nvmefc_ls_req *lsreq = &lsop->ls_req;
1100 struct fcnvme_ls_rjt *rjt = lsreq->rspaddr;
1103 ret = __nvme_fc_send_ls_req(rport, lsop, nvme_fc_send_ls_req_done);
1107 * No timeout/not interruptible as we need the struct
1108 * to exist until the lldd calls us back. Thus mandate
1109 * wait until driver calls back. lldd responsible for
1110 * the timeout action
1112 wait_for_completion(&lsop->ls_done);
1114 __nvme_fc_finish_ls_req(lsop);
1116 ret = lsop->ls_error;
1122 /* ACC or RJT payload ? */
1123 if (rjt->w0.ls_cmd == FCNVME_LS_RJT)
1130 nvme_fc_send_ls_req_async(struct nvme_fc_rport *rport,
1131 struct nvmefc_ls_req_op *lsop,
1132 void (*done)(struct nvmefc_ls_req *req, int status))
1134 /* don't wait for completion */
1136 return __nvme_fc_send_ls_req(rport, lsop, done);
1139 /* Validation Error indexes into the string table below */
1143 VERR_LSDESC_RQST = 2,
1144 VERR_LSDESC_RQST_LEN = 3,
1146 VERR_ASSOC_ID_LEN = 5,
1148 VERR_CONN_ID_LEN = 7,
1150 VERR_CR_ASSOC_ACC_LEN = 9,
1152 VERR_CR_CONN_ACC_LEN = 11,
1154 VERR_DISCONN_ACC_LEN = 13,
1157 static char *validation_errors[] = {
1161 "Bad LSDESC_RQST Length",
1162 "Not Association ID",
1163 "Bad Association ID Length",
1164 "Not Connection ID",
1165 "Bad Connection ID Length",
1166 "Not CR_ASSOC Rqst",
1167 "Bad CR_ASSOC ACC Length",
1169 "Bad CR_CONN ACC Length",
1170 "Not Disconnect Rqst",
1171 "Bad Disconnect ACC Length",
1175 nvme_fc_connect_admin_queue(struct nvme_fc_ctrl *ctrl,
1176 struct nvme_fc_queue *queue, u16 qsize, u16 ersp_ratio)
1178 struct nvmefc_ls_req_op *lsop;
1179 struct nvmefc_ls_req *lsreq;
1180 struct fcnvme_ls_cr_assoc_rqst *assoc_rqst;
1181 struct fcnvme_ls_cr_assoc_acc *assoc_acc;
1184 lsop = kzalloc((sizeof(*lsop) +
1185 ctrl->lport->ops->lsrqst_priv_sz +
1186 sizeof(*assoc_rqst) + sizeof(*assoc_acc)), GFP_KERNEL);
1191 lsreq = &lsop->ls_req;
1193 lsreq->private = (void *)&lsop[1];
1194 assoc_rqst = (struct fcnvme_ls_cr_assoc_rqst *)
1195 (lsreq->private + ctrl->lport->ops->lsrqst_priv_sz);
1196 assoc_acc = (struct fcnvme_ls_cr_assoc_acc *)&assoc_rqst[1];
1198 assoc_rqst->w0.ls_cmd = FCNVME_LS_CREATE_ASSOCIATION;
1199 assoc_rqst->desc_list_len =
1200 cpu_to_be32(sizeof(struct fcnvme_lsdesc_cr_assoc_cmd));
1202 assoc_rqst->assoc_cmd.desc_tag =
1203 cpu_to_be32(FCNVME_LSDESC_CREATE_ASSOC_CMD);
1204 assoc_rqst->assoc_cmd.desc_len =
1206 sizeof(struct fcnvme_lsdesc_cr_assoc_cmd));
1208 assoc_rqst->assoc_cmd.ersp_ratio = cpu_to_be16(ersp_ratio);
1209 assoc_rqst->assoc_cmd.sqsize = cpu_to_be16(qsize - 1);
1210 /* Linux supports only Dynamic controllers */
1211 assoc_rqst->assoc_cmd.cntlid = cpu_to_be16(0xffff);
1212 uuid_copy(&assoc_rqst->assoc_cmd.hostid, &ctrl->ctrl.opts->host->id);
1213 strncpy(assoc_rqst->assoc_cmd.hostnqn, ctrl->ctrl.opts->host->nqn,
1214 min(FCNVME_ASSOC_HOSTNQN_LEN, NVMF_NQN_SIZE));
1215 strncpy(assoc_rqst->assoc_cmd.subnqn, ctrl->ctrl.opts->subsysnqn,
1216 min(FCNVME_ASSOC_SUBNQN_LEN, NVMF_NQN_SIZE));
1218 lsop->queue = queue;
1219 lsreq->rqstaddr = assoc_rqst;
1220 lsreq->rqstlen = sizeof(*assoc_rqst);
1221 lsreq->rspaddr = assoc_acc;
1222 lsreq->rsplen = sizeof(*assoc_acc);
1223 lsreq->timeout = NVME_FC_CONNECT_TIMEOUT_SEC;
1225 ret = nvme_fc_send_ls_req(ctrl->rport, lsop);
1227 goto out_free_buffer;
1229 /* process connect LS completion */
1231 /* validate the ACC response */
1232 if (assoc_acc->hdr.w0.ls_cmd != FCNVME_LS_ACC)
1234 else if (assoc_acc->hdr.desc_list_len !=
1236 sizeof(struct fcnvme_ls_cr_assoc_acc)))
1237 fcret = VERR_CR_ASSOC_ACC_LEN;
1238 else if (assoc_acc->hdr.rqst.desc_tag !=
1239 cpu_to_be32(FCNVME_LSDESC_RQST))
1240 fcret = VERR_LSDESC_RQST;
1241 else if (assoc_acc->hdr.rqst.desc_len !=
1242 fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_rqst)))
1243 fcret = VERR_LSDESC_RQST_LEN;
1244 else if (assoc_acc->hdr.rqst.w0.ls_cmd != FCNVME_LS_CREATE_ASSOCIATION)
1245 fcret = VERR_CR_ASSOC;
1246 else if (assoc_acc->associd.desc_tag !=
1247 cpu_to_be32(FCNVME_LSDESC_ASSOC_ID))
1248 fcret = VERR_ASSOC_ID;
1249 else if (assoc_acc->associd.desc_len !=
1251 sizeof(struct fcnvme_lsdesc_assoc_id)))
1252 fcret = VERR_ASSOC_ID_LEN;
1253 else if (assoc_acc->connectid.desc_tag !=
1254 cpu_to_be32(FCNVME_LSDESC_CONN_ID))
1255 fcret = VERR_CONN_ID;
1256 else if (assoc_acc->connectid.desc_len !=
1257 fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_conn_id)))
1258 fcret = VERR_CONN_ID_LEN;
1263 "q %d connect failed: %s\n",
1264 queue->qnum, validation_errors[fcret]);
1266 ctrl->association_id =
1267 be64_to_cpu(assoc_acc->associd.association_id);
1268 queue->connection_id =
1269 be64_to_cpu(assoc_acc->connectid.connection_id);
1270 set_bit(NVME_FC_Q_CONNECTED, &queue->flags);
1278 "queue %d connect admin queue failed (%d).\n",
1284 nvme_fc_connect_queue(struct nvme_fc_ctrl *ctrl, struct nvme_fc_queue *queue,
1285 u16 qsize, u16 ersp_ratio)
1287 struct nvmefc_ls_req_op *lsop;
1288 struct nvmefc_ls_req *lsreq;
1289 struct fcnvme_ls_cr_conn_rqst *conn_rqst;
1290 struct fcnvme_ls_cr_conn_acc *conn_acc;
1293 lsop = kzalloc((sizeof(*lsop) +
1294 ctrl->lport->ops->lsrqst_priv_sz +
1295 sizeof(*conn_rqst) + sizeof(*conn_acc)), GFP_KERNEL);
1300 lsreq = &lsop->ls_req;
1302 lsreq->private = (void *)&lsop[1];
1303 conn_rqst = (struct fcnvme_ls_cr_conn_rqst *)
1304 (lsreq->private + ctrl->lport->ops->lsrqst_priv_sz);
1305 conn_acc = (struct fcnvme_ls_cr_conn_acc *)&conn_rqst[1];
1307 conn_rqst->w0.ls_cmd = FCNVME_LS_CREATE_CONNECTION;
1308 conn_rqst->desc_list_len = cpu_to_be32(
1309 sizeof(struct fcnvme_lsdesc_assoc_id) +
1310 sizeof(struct fcnvme_lsdesc_cr_conn_cmd));
1312 conn_rqst->associd.desc_tag = cpu_to_be32(FCNVME_LSDESC_ASSOC_ID);
1313 conn_rqst->associd.desc_len =
1315 sizeof(struct fcnvme_lsdesc_assoc_id));
1316 conn_rqst->associd.association_id = cpu_to_be64(ctrl->association_id);
1317 conn_rqst->connect_cmd.desc_tag =
1318 cpu_to_be32(FCNVME_LSDESC_CREATE_CONN_CMD);
1319 conn_rqst->connect_cmd.desc_len =
1321 sizeof(struct fcnvme_lsdesc_cr_conn_cmd));
1322 conn_rqst->connect_cmd.ersp_ratio = cpu_to_be16(ersp_ratio);
1323 conn_rqst->connect_cmd.qid = cpu_to_be16(queue->qnum);
1324 conn_rqst->connect_cmd.sqsize = cpu_to_be16(qsize - 1);
1326 lsop->queue = queue;
1327 lsreq->rqstaddr = conn_rqst;
1328 lsreq->rqstlen = sizeof(*conn_rqst);
1329 lsreq->rspaddr = conn_acc;
1330 lsreq->rsplen = sizeof(*conn_acc);
1331 lsreq->timeout = NVME_FC_CONNECT_TIMEOUT_SEC;
1333 ret = nvme_fc_send_ls_req(ctrl->rport, lsop);
1335 goto out_free_buffer;
1337 /* process connect LS completion */
1339 /* validate the ACC response */
1340 if (conn_acc->hdr.w0.ls_cmd != FCNVME_LS_ACC)
1342 else if (conn_acc->hdr.desc_list_len !=
1343 fcnvme_lsdesc_len(sizeof(struct fcnvme_ls_cr_conn_acc)))
1344 fcret = VERR_CR_CONN_ACC_LEN;
1345 else if (conn_acc->hdr.rqst.desc_tag != cpu_to_be32(FCNVME_LSDESC_RQST))
1346 fcret = VERR_LSDESC_RQST;
1347 else if (conn_acc->hdr.rqst.desc_len !=
1348 fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_rqst)))
1349 fcret = VERR_LSDESC_RQST_LEN;
1350 else if (conn_acc->hdr.rqst.w0.ls_cmd != FCNVME_LS_CREATE_CONNECTION)
1351 fcret = VERR_CR_CONN;
1352 else if (conn_acc->connectid.desc_tag !=
1353 cpu_to_be32(FCNVME_LSDESC_CONN_ID))
1354 fcret = VERR_CONN_ID;
1355 else if (conn_acc->connectid.desc_len !=
1356 fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_conn_id)))
1357 fcret = VERR_CONN_ID_LEN;
1362 "q %d connect failed: %s\n",
1363 queue->qnum, validation_errors[fcret]);
1365 queue->connection_id =
1366 be64_to_cpu(conn_acc->connectid.connection_id);
1367 set_bit(NVME_FC_Q_CONNECTED, &queue->flags);
1375 "queue %d connect command failed (%d).\n",
1381 nvme_fc_disconnect_assoc_done(struct nvmefc_ls_req *lsreq, int status)
1383 struct nvmefc_ls_req_op *lsop = ls_req_to_lsop(lsreq);
1385 __nvme_fc_finish_ls_req(lsop);
1387 /* fc-nvme iniator doesn't care about success or failure of cmd */
1393 * This routine sends a FC-NVME LS to disconnect (aka terminate)
1394 * the FC-NVME Association. Terminating the association also
1395 * terminates the FC-NVME connections (per queue, both admin and io
1396 * queues) that are part of the association. E.g. things are torn
1397 * down, and the related FC-NVME Association ID and Connection IDs
1400 * The behavior of the fc-nvme initiator is such that it's
1401 * understanding of the association and connections will implicitly
1402 * be torn down. The action is implicit as it may be due to a loss of
1403 * connectivity with the fc-nvme target, so you may never get a
1404 * response even if you tried. As such, the action of this routine
1405 * is to asynchronously send the LS, ignore any results of the LS, and
1406 * continue on with terminating the association. If the fc-nvme target
1407 * is present and receives the LS, it too can tear down.
1410 nvme_fc_xmt_disconnect_assoc(struct nvme_fc_ctrl *ctrl)
1412 struct fcnvme_ls_disconnect_rqst *discon_rqst;
1413 struct fcnvme_ls_disconnect_acc *discon_acc;
1414 struct nvmefc_ls_req_op *lsop;
1415 struct nvmefc_ls_req *lsreq;
1418 lsop = kzalloc((sizeof(*lsop) +
1419 ctrl->lport->ops->lsrqst_priv_sz +
1420 sizeof(*discon_rqst) + sizeof(*discon_acc)),
1423 /* couldn't sent it... too bad */
1426 lsreq = &lsop->ls_req;
1428 lsreq->private = (void *)&lsop[1];
1429 discon_rqst = (struct fcnvme_ls_disconnect_rqst *)
1430 (lsreq->private + ctrl->lport->ops->lsrqst_priv_sz);
1431 discon_acc = (struct fcnvme_ls_disconnect_acc *)&discon_rqst[1];
1433 discon_rqst->w0.ls_cmd = FCNVME_LS_DISCONNECT;
1434 discon_rqst->desc_list_len = cpu_to_be32(
1435 sizeof(struct fcnvme_lsdesc_assoc_id) +
1436 sizeof(struct fcnvme_lsdesc_disconn_cmd));
1438 discon_rqst->associd.desc_tag = cpu_to_be32(FCNVME_LSDESC_ASSOC_ID);
1439 discon_rqst->associd.desc_len =
1441 sizeof(struct fcnvme_lsdesc_assoc_id));
1443 discon_rqst->associd.association_id = cpu_to_be64(ctrl->association_id);
1445 discon_rqst->discon_cmd.desc_tag = cpu_to_be32(
1446 FCNVME_LSDESC_DISCONN_CMD);
1447 discon_rqst->discon_cmd.desc_len =
1449 sizeof(struct fcnvme_lsdesc_disconn_cmd));
1450 discon_rqst->discon_cmd.scope = FCNVME_DISCONN_ASSOCIATION;
1451 discon_rqst->discon_cmd.id = cpu_to_be64(ctrl->association_id);
1453 lsreq->rqstaddr = discon_rqst;
1454 lsreq->rqstlen = sizeof(*discon_rqst);
1455 lsreq->rspaddr = discon_acc;
1456 lsreq->rsplen = sizeof(*discon_acc);
1457 lsreq->timeout = NVME_FC_CONNECT_TIMEOUT_SEC;
1459 ret = nvme_fc_send_ls_req_async(ctrl->rport, lsop,
1460 nvme_fc_disconnect_assoc_done);
1464 /* only meaningful part to terminating the association */
1465 ctrl->association_id = 0;
1469 /* *********************** NVME Ctrl Routines **************************** */
1471 static void nvme_fc_error_recovery(struct nvme_fc_ctrl *ctrl, char *errmsg);
1474 nvme_fc_reinit_request(void *data, struct request *rq)
1476 struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
1477 struct nvme_fc_cmd_iu *cmdiu = &op->cmd_iu;
1479 memset(cmdiu, 0, sizeof(*cmdiu));
1480 cmdiu->scsi_id = NVME_CMD_SCSI_ID;
1481 cmdiu->fc_id = NVME_CMD_FC_ID;
1482 cmdiu->iu_len = cpu_to_be16(sizeof(*cmdiu) / sizeof(u32));
1483 memset(&op->rsp_iu, 0, sizeof(op->rsp_iu));
1489 __nvme_fc_exit_request(struct nvme_fc_ctrl *ctrl,
1490 struct nvme_fc_fcp_op *op)
1492 fc_dma_unmap_single(ctrl->lport->dev, op->fcp_req.rspdma,
1493 sizeof(op->rsp_iu), DMA_FROM_DEVICE);
1494 fc_dma_unmap_single(ctrl->lport->dev, op->fcp_req.cmddma,
1495 sizeof(op->cmd_iu), DMA_TO_DEVICE);
1497 atomic_set(&op->state, FCPOP_STATE_UNINIT);
1501 nvme_fc_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1502 unsigned int hctx_idx)
1504 struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
1506 return __nvme_fc_exit_request(set->driver_data, op);
1510 __nvme_fc_abort_op(struct nvme_fc_ctrl *ctrl, struct nvme_fc_fcp_op *op)
1512 unsigned long flags;
1515 spin_lock_irqsave(&ctrl->lock, flags);
1516 opstate = atomic_xchg(&op->state, FCPOP_STATE_ABORTED);
1517 if (opstate != FCPOP_STATE_ACTIVE)
1518 atomic_set(&op->state, opstate);
1519 else if (ctrl->flags & FCCTRL_TERMIO)
1521 spin_unlock_irqrestore(&ctrl->lock, flags);
1523 if (opstate != FCPOP_STATE_ACTIVE)
1526 ctrl->lport->ops->fcp_abort(&ctrl->lport->localport,
1527 &ctrl->rport->remoteport,
1528 op->queue->lldd_handle,
1535 nvme_fc_abort_aen_ops(struct nvme_fc_ctrl *ctrl)
1537 struct nvme_fc_fcp_op *aen_op = ctrl->aen_ops;
1540 for (i = 0; i < NVME_NR_AEN_COMMANDS; i++, aen_op++)
1541 __nvme_fc_abort_op(ctrl, aen_op);
1545 __nvme_fc_fcpop_chk_teardowns(struct nvme_fc_ctrl *ctrl,
1546 struct nvme_fc_fcp_op *op, int opstate)
1548 unsigned long flags;
1550 if (opstate == FCPOP_STATE_ABORTED) {
1551 spin_lock_irqsave(&ctrl->lock, flags);
1552 if (ctrl->flags & FCCTRL_TERMIO) {
1554 wake_up(&ctrl->ioabort_wait);
1556 spin_unlock_irqrestore(&ctrl->lock, flags);
1561 nvme_fc_fcpio_done(struct nvmefc_fcp_req *req)
1563 struct nvme_fc_fcp_op *op = fcp_req_to_fcp_op(req);
1564 struct request *rq = op->rq;
1565 struct nvmefc_fcp_req *freq = &op->fcp_req;
1566 struct nvme_fc_ctrl *ctrl = op->ctrl;
1567 struct nvme_fc_queue *queue = op->queue;
1568 struct nvme_completion *cqe = &op->rsp_iu.cqe;
1569 struct nvme_command *sqe = &op->cmd_iu.sqe;
1570 __le16 status = cpu_to_le16(NVME_SC_SUCCESS << 1);
1571 union nvme_result result;
1572 bool terminate_assoc = true;
1577 * The current linux implementation of a nvme controller
1578 * allocates a single tag set for all io queues and sizes
1579 * the io queues to fully hold all possible tags. Thus, the
1580 * implementation does not reference or care about the sqhd
1581 * value as it never needs to use the sqhd/sqtail pointers
1582 * for submission pacing.
1584 * This affects the FC-NVME implementation in two ways:
1585 * 1) As the value doesn't matter, we don't need to waste
1586 * cycles extracting it from ERSPs and stamping it in the
1587 * cases where the transport fabricates CQEs on successful
1589 * 2) The FC-NVME implementation requires that delivery of
1590 * ERSP completions are to go back to the nvme layer in order
1591 * relative to the rsn, such that the sqhd value will always
1592 * be "in order" for the nvme layer. As the nvme layer in
1593 * linux doesn't care about sqhd, there's no need to return
1597 * As the core nvme layer in linux currently does not look at
1598 * every field in the cqe - in cases where the FC transport must
1599 * fabricate a CQE, the following fields will not be set as they
1600 * are not referenced:
1601 * cqe.sqid, cqe.sqhd, cqe.command_id
1603 * Failure or error of an individual i/o, in a transport
1604 * detected fashion unrelated to the nvme completion status,
1605 * potentially cause the initiator and target sides to get out
1606 * of sync on SQ head/tail (aka outstanding io count allowed).
1607 * Per FC-NVME spec, failure of an individual command requires
1608 * the connection to be terminated, which in turn requires the
1609 * association to be terminated.
1612 opstate = atomic_xchg(&op->state, FCPOP_STATE_COMPLETE);
1614 fc_dma_sync_single_for_cpu(ctrl->lport->dev, op->fcp_req.rspdma,
1615 sizeof(op->rsp_iu), DMA_FROM_DEVICE);
1617 if (opstate == FCPOP_STATE_ABORTED)
1618 status = cpu_to_le16(NVME_SC_ABORT_REQ << 1);
1619 else if (freq->status)
1620 status = cpu_to_le16(NVME_SC_INTERNAL << 1);
1623 * For the linux implementation, if we have an unsuccesful
1624 * status, they blk-mq layer can typically be called with the
1625 * non-zero status and the content of the cqe isn't important.
1631 * command completed successfully relative to the wire
1632 * protocol. However, validate anything received and
1633 * extract the status and result from the cqe (create it
1637 switch (freq->rcv_rsplen) {
1640 case NVME_FC_SIZEOF_ZEROS_RSP:
1642 * No response payload or 12 bytes of payload (which
1643 * should all be zeros) are considered successful and
1644 * no payload in the CQE by the transport.
1646 if (freq->transferred_length !=
1647 be32_to_cpu(op->cmd_iu.data_len)) {
1648 status = cpu_to_le16(NVME_SC_INTERNAL << 1);
1654 case sizeof(struct nvme_fc_ersp_iu):
1656 * The ERSP IU contains a full completion with CQE.
1657 * Validate ERSP IU and look at cqe.
1659 if (unlikely(be16_to_cpu(op->rsp_iu.iu_len) !=
1660 (freq->rcv_rsplen / 4) ||
1661 be32_to_cpu(op->rsp_iu.xfrd_len) !=
1662 freq->transferred_length ||
1663 op->rsp_iu.status_code ||
1664 sqe->common.command_id != cqe->command_id)) {
1665 status = cpu_to_le16(NVME_SC_INTERNAL << 1);
1668 result = cqe->result;
1669 status = cqe->status;
1673 status = cpu_to_le16(NVME_SC_INTERNAL << 1);
1677 terminate_assoc = false;
1680 if (op->flags & FCOP_FLAGS_AEN) {
1681 nvme_complete_async_event(&queue->ctrl->ctrl, status, &result);
1682 __nvme_fc_fcpop_chk_teardowns(ctrl, op, opstate);
1683 atomic_set(&op->state, FCPOP_STATE_IDLE);
1684 op->flags = FCOP_FLAGS_AEN; /* clear other flags */
1685 nvme_fc_ctrl_put(ctrl);
1690 * Force failures of commands if we're killing the controller
1691 * or have an error on a command used to create an new association
1694 (blk_queue_dying(rq->q) ||
1695 ctrl->ctrl.state == NVME_CTRL_NEW ||
1696 ctrl->ctrl.state == NVME_CTRL_CONNECTING))
1697 status |= cpu_to_le16(NVME_SC_DNR << 1);
1699 __nvme_fc_fcpop_chk_teardowns(ctrl, op, opstate);
1700 nvme_end_request(rq, status, result);
1703 if (terminate_assoc)
1704 nvme_fc_error_recovery(ctrl, "transport detected io error");
1708 __nvme_fc_init_request(struct nvme_fc_ctrl *ctrl,
1709 struct nvme_fc_queue *queue, struct nvme_fc_fcp_op *op,
1710 struct request *rq, u32 rqno)
1712 struct nvme_fc_cmd_iu *cmdiu = &op->cmd_iu;
1715 memset(op, 0, sizeof(*op));
1716 op->fcp_req.cmdaddr = &op->cmd_iu;
1717 op->fcp_req.cmdlen = sizeof(op->cmd_iu);
1718 op->fcp_req.rspaddr = &op->rsp_iu;
1719 op->fcp_req.rsplen = sizeof(op->rsp_iu);
1720 op->fcp_req.done = nvme_fc_fcpio_done;
1721 op->fcp_req.first_sgl = (struct scatterlist *)&op[1];
1722 op->fcp_req.private = &op->fcp_req.first_sgl[SG_CHUNK_SIZE];
1728 cmdiu->scsi_id = NVME_CMD_SCSI_ID;
1729 cmdiu->fc_id = NVME_CMD_FC_ID;
1730 cmdiu->iu_len = cpu_to_be16(sizeof(*cmdiu) / sizeof(u32));
1732 op->fcp_req.cmddma = fc_dma_map_single(ctrl->lport->dev,
1733 &op->cmd_iu, sizeof(op->cmd_iu), DMA_TO_DEVICE);
1734 if (fc_dma_mapping_error(ctrl->lport->dev, op->fcp_req.cmddma)) {
1736 "FCP Op failed - cmdiu dma mapping failed.\n");
1741 op->fcp_req.rspdma = fc_dma_map_single(ctrl->lport->dev,
1742 &op->rsp_iu, sizeof(op->rsp_iu),
1744 if (fc_dma_mapping_error(ctrl->lport->dev, op->fcp_req.rspdma)) {
1746 "FCP Op failed - rspiu dma mapping failed.\n");
1750 atomic_set(&op->state, FCPOP_STATE_IDLE);
1756 nvme_fc_init_request(struct blk_mq_tag_set *set, struct request *rq,
1757 unsigned int hctx_idx, unsigned int numa_node)
1759 struct nvme_fc_ctrl *ctrl = set->driver_data;
1760 struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
1761 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
1762 struct nvme_fc_queue *queue = &ctrl->queues[queue_idx];
1764 return __nvme_fc_init_request(ctrl, queue, op, rq, queue->rqcnt++);
1768 nvme_fc_init_aen_ops(struct nvme_fc_ctrl *ctrl)
1770 struct nvme_fc_fcp_op *aen_op;
1771 struct nvme_fc_cmd_iu *cmdiu;
1772 struct nvme_command *sqe;
1776 aen_op = ctrl->aen_ops;
1777 for (i = 0; i < NVME_NR_AEN_COMMANDS; i++, aen_op++) {
1778 private = kzalloc(ctrl->lport->ops->fcprqst_priv_sz,
1783 cmdiu = &aen_op->cmd_iu;
1785 ret = __nvme_fc_init_request(ctrl, &ctrl->queues[0],
1786 aen_op, (struct request *)NULL,
1787 (NVME_AQ_BLK_MQ_DEPTH + i));
1793 aen_op->flags = FCOP_FLAGS_AEN;
1794 aen_op->fcp_req.first_sgl = NULL; /* no sg list */
1795 aen_op->fcp_req.private = private;
1797 memset(sqe, 0, sizeof(*sqe));
1798 sqe->common.opcode = nvme_admin_async_event;
1799 /* Note: core layer may overwrite the sqe.command_id value */
1800 sqe->common.command_id = NVME_AQ_BLK_MQ_DEPTH + i;
1806 nvme_fc_term_aen_ops(struct nvme_fc_ctrl *ctrl)
1808 struct nvme_fc_fcp_op *aen_op;
1811 aen_op = ctrl->aen_ops;
1812 for (i = 0; i < NVME_NR_AEN_COMMANDS; i++, aen_op++) {
1813 if (!aen_op->fcp_req.private)
1816 __nvme_fc_exit_request(ctrl, aen_op);
1818 kfree(aen_op->fcp_req.private);
1819 aen_op->fcp_req.private = NULL;
1824 __nvme_fc_init_hctx(struct blk_mq_hw_ctx *hctx, struct nvme_fc_ctrl *ctrl,
1827 struct nvme_fc_queue *queue = &ctrl->queues[qidx];
1829 hctx->driver_data = queue;
1834 nvme_fc_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
1835 unsigned int hctx_idx)
1837 struct nvme_fc_ctrl *ctrl = data;
1839 __nvme_fc_init_hctx(hctx, ctrl, hctx_idx + 1);
1845 nvme_fc_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
1846 unsigned int hctx_idx)
1848 struct nvme_fc_ctrl *ctrl = data;
1850 __nvme_fc_init_hctx(hctx, ctrl, hctx_idx);
1856 nvme_fc_init_queue(struct nvme_fc_ctrl *ctrl, int idx)
1858 struct nvme_fc_queue *queue;
1860 queue = &ctrl->queues[idx];
1861 memset(queue, 0, sizeof(*queue));
1864 atomic_set(&queue->csn, 1);
1865 queue->dev = ctrl->dev;
1868 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
1870 queue->cmnd_capsule_len = sizeof(struct nvme_command);
1873 * Considered whether we should allocate buffers for all SQEs
1874 * and CQEs and dma map them - mapping their respective entries
1875 * into the request structures (kernel vm addr and dma address)
1876 * thus the driver could use the buffers/mappings directly.
1877 * It only makes sense if the LLDD would use them for its
1878 * messaging api. It's very unlikely most adapter api's would use
1879 * a native NVME sqe/cqe. More reasonable if FC-NVME IU payload
1880 * structures were used instead.
1885 * This routine terminates a queue at the transport level.
1886 * The transport has already ensured that all outstanding ios on
1887 * the queue have been terminated.
1888 * The transport will send a Disconnect LS request to terminate
1889 * the queue's connection. Termination of the admin queue will also
1890 * terminate the association at the target.
1893 nvme_fc_free_queue(struct nvme_fc_queue *queue)
1895 if (!test_and_clear_bit(NVME_FC_Q_CONNECTED, &queue->flags))
1898 clear_bit(NVME_FC_Q_LIVE, &queue->flags);
1900 * Current implementation never disconnects a single queue.
1901 * It always terminates a whole association. So there is never
1902 * a disconnect(queue) LS sent to the target.
1905 queue->connection_id = 0;
1909 __nvme_fc_delete_hw_queue(struct nvme_fc_ctrl *ctrl,
1910 struct nvme_fc_queue *queue, unsigned int qidx)
1912 if (ctrl->lport->ops->delete_queue)
1913 ctrl->lport->ops->delete_queue(&ctrl->lport->localport, qidx,
1914 queue->lldd_handle);
1915 queue->lldd_handle = NULL;
1919 nvme_fc_free_io_queues(struct nvme_fc_ctrl *ctrl)
1923 for (i = 1; i < ctrl->ctrl.queue_count; i++)
1924 nvme_fc_free_queue(&ctrl->queues[i]);
1928 __nvme_fc_create_hw_queue(struct nvme_fc_ctrl *ctrl,
1929 struct nvme_fc_queue *queue, unsigned int qidx, u16 qsize)
1933 queue->lldd_handle = NULL;
1934 if (ctrl->lport->ops->create_queue)
1935 ret = ctrl->lport->ops->create_queue(&ctrl->lport->localport,
1936 qidx, qsize, &queue->lldd_handle);
1942 nvme_fc_delete_hw_io_queues(struct nvme_fc_ctrl *ctrl)
1944 struct nvme_fc_queue *queue = &ctrl->queues[ctrl->ctrl.queue_count - 1];
1947 for (i = ctrl->ctrl.queue_count - 1; i >= 1; i--, queue--)
1948 __nvme_fc_delete_hw_queue(ctrl, queue, i);
1952 nvme_fc_create_hw_io_queues(struct nvme_fc_ctrl *ctrl, u16 qsize)
1954 struct nvme_fc_queue *queue = &ctrl->queues[1];
1957 for (i = 1; i < ctrl->ctrl.queue_count; i++, queue++) {
1958 ret = __nvme_fc_create_hw_queue(ctrl, queue, i, qsize);
1967 __nvme_fc_delete_hw_queue(ctrl, &ctrl->queues[i], i);
1972 nvme_fc_connect_io_queues(struct nvme_fc_ctrl *ctrl, u16 qsize)
1976 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
1977 ret = nvme_fc_connect_queue(ctrl, &ctrl->queues[i], qsize,
1981 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
1985 set_bit(NVME_FC_Q_LIVE, &ctrl->queues[i].flags);
1992 nvme_fc_init_io_queues(struct nvme_fc_ctrl *ctrl)
1996 for (i = 1; i < ctrl->ctrl.queue_count; i++)
1997 nvme_fc_init_queue(ctrl, i);
2001 nvme_fc_ctrl_free(struct kref *ref)
2003 struct nvme_fc_ctrl *ctrl =
2004 container_of(ref, struct nvme_fc_ctrl, ref);
2005 unsigned long flags;
2007 if (ctrl->ctrl.tagset) {
2008 blk_cleanup_queue(ctrl->ctrl.connect_q);
2009 blk_mq_free_tag_set(&ctrl->tag_set);
2012 /* remove from rport list */
2013 spin_lock_irqsave(&ctrl->rport->lock, flags);
2014 list_del(&ctrl->ctrl_list);
2015 spin_unlock_irqrestore(&ctrl->rport->lock, flags);
2017 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
2018 blk_cleanup_queue(ctrl->ctrl.admin_q);
2019 blk_mq_free_tag_set(&ctrl->admin_tag_set);
2021 kfree(ctrl->queues);
2023 put_device(ctrl->dev);
2024 nvme_fc_rport_put(ctrl->rport);
2026 ida_simple_remove(&nvme_fc_ctrl_cnt, ctrl->cnum);
2027 if (ctrl->ctrl.opts)
2028 nvmf_free_options(ctrl->ctrl.opts);
2033 nvme_fc_ctrl_put(struct nvme_fc_ctrl *ctrl)
2035 kref_put(&ctrl->ref, nvme_fc_ctrl_free);
2039 nvme_fc_ctrl_get(struct nvme_fc_ctrl *ctrl)
2041 return kref_get_unless_zero(&ctrl->ref);
2045 * All accesses from nvme core layer done - can now free the
2046 * controller. Called after last nvme_put_ctrl() call
2049 nvme_fc_nvme_ctrl_freed(struct nvme_ctrl *nctrl)
2051 struct nvme_fc_ctrl *ctrl = to_fc_ctrl(nctrl);
2053 WARN_ON(nctrl != &ctrl->ctrl);
2055 nvme_fc_ctrl_put(ctrl);
2059 nvme_fc_error_recovery(struct nvme_fc_ctrl *ctrl, char *errmsg)
2061 /* only proceed if in LIVE state - e.g. on first error */
2062 if (ctrl->ctrl.state != NVME_CTRL_LIVE)
2065 dev_warn(ctrl->ctrl.device,
2066 "NVME-FC{%d}: transport association error detected: %s\n",
2067 ctrl->cnum, errmsg);
2068 dev_warn(ctrl->ctrl.device,
2069 "NVME-FC{%d}: resetting controller\n", ctrl->cnum);
2071 nvme_reset_ctrl(&ctrl->ctrl);
2074 static enum blk_eh_timer_return
2075 nvme_fc_timeout(struct request *rq, bool reserved)
2077 struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
2078 struct nvme_fc_ctrl *ctrl = op->ctrl;
2081 if (ctrl->rport->remoteport.port_state != FC_OBJSTATE_ONLINE ||
2082 atomic_read(&op->state) == FCPOP_STATE_ABORTED)
2083 return BLK_EH_RESET_TIMER;
2085 ret = __nvme_fc_abort_op(ctrl, op);
2087 /* io wasn't active to abort */
2088 return BLK_EH_NOT_HANDLED;
2091 * we can't individually ABTS an io without affecting the queue,
2092 * thus killing the queue, adn thus the association.
2093 * So resolve by performing a controller reset, which will stop
2094 * the host/io stack, terminate the association on the link,
2095 * and recreate an association on the link.
2097 nvme_fc_error_recovery(ctrl, "io timeout error");
2100 * the io abort has been initiated. Have the reset timer
2101 * restarted and the abort completion will complete the io
2102 * shortly. Avoids a synchronous wait while the abort finishes.
2104 return BLK_EH_RESET_TIMER;
2108 nvme_fc_map_data(struct nvme_fc_ctrl *ctrl, struct request *rq,
2109 struct nvme_fc_fcp_op *op)
2111 struct nvmefc_fcp_req *freq = &op->fcp_req;
2112 enum dma_data_direction dir;
2117 if (!blk_rq_payload_bytes(rq))
2120 freq->sg_table.sgl = freq->first_sgl;
2121 ret = sg_alloc_table_chained(&freq->sg_table,
2122 blk_rq_nr_phys_segments(rq), freq->sg_table.sgl);
2126 op->nents = blk_rq_map_sg(rq->q, rq, freq->sg_table.sgl);
2127 WARN_ON(op->nents > blk_rq_nr_phys_segments(rq));
2128 dir = (rq_data_dir(rq) == WRITE) ? DMA_TO_DEVICE : DMA_FROM_DEVICE;
2129 freq->sg_cnt = fc_dma_map_sg(ctrl->lport->dev, freq->sg_table.sgl,
2131 if (unlikely(freq->sg_cnt <= 0)) {
2132 sg_free_table_chained(&freq->sg_table, true);
2138 * TODO: blk_integrity_rq(rq) for DIF
2144 nvme_fc_unmap_data(struct nvme_fc_ctrl *ctrl, struct request *rq,
2145 struct nvme_fc_fcp_op *op)
2147 struct nvmefc_fcp_req *freq = &op->fcp_req;
2152 fc_dma_unmap_sg(ctrl->lport->dev, freq->sg_table.sgl, op->nents,
2153 ((rq_data_dir(rq) == WRITE) ?
2154 DMA_TO_DEVICE : DMA_FROM_DEVICE));
2156 nvme_cleanup_cmd(rq);
2158 sg_free_table_chained(&freq->sg_table, true);
2164 * In FC, the queue is a logical thing. At transport connect, the target
2165 * creates its "queue" and returns a handle that is to be given to the
2166 * target whenever it posts something to the corresponding SQ. When an
2167 * SQE is sent on a SQ, FC effectively considers the SQE, or rather the
2168 * command contained within the SQE, an io, and assigns a FC exchange
2169 * to it. The SQE and the associated SQ handle are sent in the initial
2170 * CMD IU sents on the exchange. All transfers relative to the io occur
2171 * as part of the exchange. The CQE is the last thing for the io,
2172 * which is transferred (explicitly or implicitly) with the RSP IU
2173 * sent on the exchange. After the CQE is received, the FC exchange is
2174 * terminaed and the Exchange may be used on a different io.
2176 * The transport to LLDD api has the transport making a request for a
2177 * new fcp io request to the LLDD. The LLDD then allocates a FC exchange
2178 * resource and transfers the command. The LLDD will then process all
2179 * steps to complete the io. Upon completion, the transport done routine
2182 * So - while the operation is outstanding to the LLDD, there is a link
2183 * level FC exchange resource that is also outstanding. This must be
2184 * considered in all cleanup operations.
2187 nvme_fc_start_fcp_op(struct nvme_fc_ctrl *ctrl, struct nvme_fc_queue *queue,
2188 struct nvme_fc_fcp_op *op, u32 data_len,
2189 enum nvmefc_fcp_datadir io_dir)
2191 struct nvme_fc_cmd_iu *cmdiu = &op->cmd_iu;
2192 struct nvme_command *sqe = &cmdiu->sqe;
2197 * before attempting to send the io, check to see if we believe
2198 * the target device is present
2200 if (ctrl->rport->remoteport.port_state != FC_OBJSTATE_ONLINE)
2201 return BLK_STS_RESOURCE;
2203 if (!nvme_fc_ctrl_get(ctrl))
2204 return BLK_STS_IOERR;
2206 /* format the FC-NVME CMD IU and fcp_req */
2207 cmdiu->connection_id = cpu_to_be64(queue->connection_id);
2208 csn = atomic_inc_return(&queue->csn);
2209 cmdiu->csn = cpu_to_be32(csn);
2210 cmdiu->data_len = cpu_to_be32(data_len);
2212 case NVMEFC_FCP_WRITE:
2213 cmdiu->flags = FCNVME_CMD_FLAGS_WRITE;
2215 case NVMEFC_FCP_READ:
2216 cmdiu->flags = FCNVME_CMD_FLAGS_READ;
2218 case NVMEFC_FCP_NODATA:
2222 op->fcp_req.payload_length = data_len;
2223 op->fcp_req.io_dir = io_dir;
2224 op->fcp_req.transferred_length = 0;
2225 op->fcp_req.rcv_rsplen = 0;
2226 op->fcp_req.status = NVME_SC_SUCCESS;
2227 op->fcp_req.sqid = cpu_to_le16(queue->qnum);
2230 * validate per fabric rules, set fields mandated by fabric spec
2231 * as well as those by FC-NVME spec.
2233 WARN_ON_ONCE(sqe->common.metadata);
2234 sqe->common.flags |= NVME_CMD_SGL_METABUF;
2237 * format SQE DPTR field per FC-NVME rules:
2238 * type=0x5 Transport SGL Data Block Descriptor
2239 * subtype=0xA Transport-specific value
2241 * length=length of the data series
2243 sqe->rw.dptr.sgl.type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2244 NVME_SGL_FMT_TRANSPORT_A;
2245 sqe->rw.dptr.sgl.length = cpu_to_le32(data_len);
2246 sqe->rw.dptr.sgl.addr = 0;
2248 if (!(op->flags & FCOP_FLAGS_AEN)) {
2249 ret = nvme_fc_map_data(ctrl, op->rq, op);
2251 nvme_cleanup_cmd(op->rq);
2252 nvme_fc_ctrl_put(ctrl);
2253 if (ret == -ENOMEM || ret == -EAGAIN)
2254 return BLK_STS_RESOURCE;
2255 return BLK_STS_IOERR;
2259 fc_dma_sync_single_for_device(ctrl->lport->dev, op->fcp_req.cmddma,
2260 sizeof(op->cmd_iu), DMA_TO_DEVICE);
2262 atomic_set(&op->state, FCPOP_STATE_ACTIVE);
2264 if (!(op->flags & FCOP_FLAGS_AEN))
2265 blk_mq_start_request(op->rq);
2267 ret = ctrl->lport->ops->fcp_io(&ctrl->lport->localport,
2268 &ctrl->rport->remoteport,
2269 queue->lldd_handle, &op->fcp_req);
2272 if (!(op->flags & FCOP_FLAGS_AEN))
2273 nvme_fc_unmap_data(ctrl, op->rq, op);
2275 nvme_fc_ctrl_put(ctrl);
2277 if (ctrl->rport->remoteport.port_state == FC_OBJSTATE_ONLINE &&
2279 return BLK_STS_IOERR;
2281 return BLK_STS_RESOURCE;
2287 static inline blk_status_t nvme_fc_is_ready(struct nvme_fc_queue *queue,
2290 if (unlikely(!test_bit(NVME_FC_Q_LIVE, &queue->flags)))
2291 return nvmf_check_init_req(&queue->ctrl->ctrl, rq);
2296 nvme_fc_queue_rq(struct blk_mq_hw_ctx *hctx,
2297 const struct blk_mq_queue_data *bd)
2299 struct nvme_ns *ns = hctx->queue->queuedata;
2300 struct nvme_fc_queue *queue = hctx->driver_data;
2301 struct nvme_fc_ctrl *ctrl = queue->ctrl;
2302 struct request *rq = bd->rq;
2303 struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
2304 struct nvme_fc_cmd_iu *cmdiu = &op->cmd_iu;
2305 struct nvme_command *sqe = &cmdiu->sqe;
2306 enum nvmefc_fcp_datadir io_dir;
2310 ret = nvme_fc_is_ready(queue, rq);
2314 ret = nvme_setup_cmd(ns, rq, sqe);
2318 data_len = blk_rq_payload_bytes(rq);
2320 io_dir = ((rq_data_dir(rq) == WRITE) ?
2321 NVMEFC_FCP_WRITE : NVMEFC_FCP_READ);
2323 io_dir = NVMEFC_FCP_NODATA;
2325 return nvme_fc_start_fcp_op(ctrl, queue, op, data_len, io_dir);
2328 static struct blk_mq_tags *
2329 nvme_fc_tagset(struct nvme_fc_queue *queue)
2331 if (queue->qnum == 0)
2332 return queue->ctrl->admin_tag_set.tags[queue->qnum];
2334 return queue->ctrl->tag_set.tags[queue->qnum - 1];
2338 nvme_fc_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
2341 struct nvme_fc_queue *queue = hctx->driver_data;
2342 struct nvme_fc_ctrl *ctrl = queue->ctrl;
2343 struct request *req;
2344 struct nvme_fc_fcp_op *op;
2346 req = blk_mq_tag_to_rq(nvme_fc_tagset(queue), tag);
2350 op = blk_mq_rq_to_pdu(req);
2352 if ((atomic_read(&op->state) == FCPOP_STATE_ACTIVE) &&
2353 (ctrl->lport->ops->poll_queue))
2354 ctrl->lport->ops->poll_queue(&ctrl->lport->localport,
2355 queue->lldd_handle);
2357 return ((atomic_read(&op->state) != FCPOP_STATE_ACTIVE));
2361 nvme_fc_submit_async_event(struct nvme_ctrl *arg)
2363 struct nvme_fc_ctrl *ctrl = to_fc_ctrl(arg);
2364 struct nvme_fc_fcp_op *aen_op;
2365 unsigned long flags;
2366 bool terminating = false;
2369 spin_lock_irqsave(&ctrl->lock, flags);
2370 if (ctrl->flags & FCCTRL_TERMIO)
2372 spin_unlock_irqrestore(&ctrl->lock, flags);
2377 aen_op = &ctrl->aen_ops[0];
2379 ret = nvme_fc_start_fcp_op(ctrl, aen_op->queue, aen_op, 0,
2382 dev_err(ctrl->ctrl.device,
2383 "failed async event work\n");
2387 nvme_fc_complete_rq(struct request *rq)
2389 struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
2390 struct nvme_fc_ctrl *ctrl = op->ctrl;
2392 atomic_set(&op->state, FCPOP_STATE_IDLE);
2394 nvme_fc_unmap_data(ctrl, rq, op);
2395 nvme_complete_rq(rq);
2396 nvme_fc_ctrl_put(ctrl);
2400 * This routine is used by the transport when it needs to find active
2401 * io on a queue that is to be terminated. The transport uses
2402 * blk_mq_tagset_busy_itr() to find the busy requests, which then invoke
2403 * this routine to kill them on a 1 by 1 basis.
2405 * As FC allocates FC exchange for each io, the transport must contact
2406 * the LLDD to terminate the exchange, thus releasing the FC exchange.
2407 * After terminating the exchange the LLDD will call the transport's
2408 * normal io done path for the request, but it will have an aborted
2409 * status. The done path will return the io request back to the block
2410 * layer with an error status.
2413 nvme_fc_terminate_exchange(struct request *req, void *data, bool reserved)
2415 struct nvme_ctrl *nctrl = data;
2416 struct nvme_fc_ctrl *ctrl = to_fc_ctrl(nctrl);
2417 struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(req);
2419 if (!blk_mq_request_started(req))
2422 __nvme_fc_abort_op(ctrl, op);
2426 static const struct blk_mq_ops nvme_fc_mq_ops = {
2427 .queue_rq = nvme_fc_queue_rq,
2428 .complete = nvme_fc_complete_rq,
2429 .init_request = nvme_fc_init_request,
2430 .exit_request = nvme_fc_exit_request,
2431 .init_hctx = nvme_fc_init_hctx,
2432 .poll = nvme_fc_poll,
2433 .timeout = nvme_fc_timeout,
2437 nvme_fc_create_io_queues(struct nvme_fc_ctrl *ctrl)
2439 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2440 unsigned int nr_io_queues;
2443 nr_io_queues = min(min(opts->nr_io_queues, num_online_cpus()),
2444 ctrl->lport->ops->max_hw_queues);
2445 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
2447 dev_info(ctrl->ctrl.device,
2448 "set_queue_count failed: %d\n", ret);
2452 ctrl->ctrl.queue_count = nr_io_queues + 1;
2456 nvme_fc_init_io_queues(ctrl);
2458 memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
2459 ctrl->tag_set.ops = &nvme_fc_mq_ops;
2460 ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
2461 ctrl->tag_set.reserved_tags = 1; /* fabric connect */
2462 ctrl->tag_set.numa_node = NUMA_NO_NODE;
2463 ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
2464 ctrl->tag_set.cmd_size = sizeof(struct nvme_fc_fcp_op) +
2466 sizeof(struct scatterlist)) +
2467 ctrl->lport->ops->fcprqst_priv_sz;
2468 ctrl->tag_set.driver_data = ctrl;
2469 ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
2470 ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
2472 ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
2476 ctrl->ctrl.tagset = &ctrl->tag_set;
2478 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
2479 if (IS_ERR(ctrl->ctrl.connect_q)) {
2480 ret = PTR_ERR(ctrl->ctrl.connect_q);
2481 goto out_free_tag_set;
2484 ret = nvme_fc_create_hw_io_queues(ctrl, ctrl->ctrl.sqsize + 1);
2486 goto out_cleanup_blk_queue;
2488 ret = nvme_fc_connect_io_queues(ctrl, ctrl->ctrl.sqsize + 1);
2490 goto out_delete_hw_queues;
2494 out_delete_hw_queues:
2495 nvme_fc_delete_hw_io_queues(ctrl);
2496 out_cleanup_blk_queue:
2497 blk_cleanup_queue(ctrl->ctrl.connect_q);
2499 blk_mq_free_tag_set(&ctrl->tag_set);
2500 nvme_fc_free_io_queues(ctrl);
2502 /* force put free routine to ignore io queues */
2503 ctrl->ctrl.tagset = NULL;
2509 nvme_fc_reinit_io_queues(struct nvme_fc_ctrl *ctrl)
2511 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2512 unsigned int nr_io_queues;
2515 nr_io_queues = min(min(opts->nr_io_queues, num_online_cpus()),
2516 ctrl->lport->ops->max_hw_queues);
2517 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
2519 dev_info(ctrl->ctrl.device,
2520 "set_queue_count failed: %d\n", ret);
2524 ctrl->ctrl.queue_count = nr_io_queues + 1;
2525 /* check for io queues existing */
2526 if (ctrl->ctrl.queue_count == 1)
2529 nvme_fc_init_io_queues(ctrl);
2531 ret = nvme_reinit_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
2533 goto out_free_io_queues;
2535 ret = nvme_fc_create_hw_io_queues(ctrl, ctrl->ctrl.sqsize + 1);
2537 goto out_free_io_queues;
2539 ret = nvme_fc_connect_io_queues(ctrl, ctrl->ctrl.sqsize + 1);
2541 goto out_delete_hw_queues;
2543 blk_mq_update_nr_hw_queues(&ctrl->tag_set, nr_io_queues);
2547 out_delete_hw_queues:
2548 nvme_fc_delete_hw_io_queues(ctrl);
2550 nvme_fc_free_io_queues(ctrl);
2555 nvme_fc_rport_active_on_lport(struct nvme_fc_rport *rport)
2557 struct nvme_fc_lport *lport = rport->lport;
2559 atomic_inc(&lport->act_rport_cnt);
2563 nvme_fc_rport_inactive_on_lport(struct nvme_fc_rport *rport)
2565 struct nvme_fc_lport *lport = rport->lport;
2568 cnt = atomic_dec_return(&lport->act_rport_cnt);
2569 if (cnt == 0 && lport->localport.port_state == FC_OBJSTATE_DELETED)
2570 lport->ops->localport_delete(&lport->localport);
2574 nvme_fc_ctlr_active_on_rport(struct nvme_fc_ctrl *ctrl)
2576 struct nvme_fc_rport *rport = ctrl->rport;
2579 if (ctrl->assoc_active)
2582 ctrl->assoc_active = true;
2583 cnt = atomic_inc_return(&rport->act_ctrl_cnt);
2585 nvme_fc_rport_active_on_lport(rport);
2591 nvme_fc_ctlr_inactive_on_rport(struct nvme_fc_ctrl *ctrl)
2593 struct nvme_fc_rport *rport = ctrl->rport;
2594 struct nvme_fc_lport *lport = rport->lport;
2597 /* ctrl->assoc_active=false will be set independently */
2599 cnt = atomic_dec_return(&rport->act_ctrl_cnt);
2601 if (rport->remoteport.port_state == FC_OBJSTATE_DELETED)
2602 lport->ops->remoteport_delete(&rport->remoteport);
2603 nvme_fc_rport_inactive_on_lport(rport);
2610 * This routine restarts the controller on the host side, and
2611 * on the link side, recreates the controller association.
2614 nvme_fc_create_association(struct nvme_fc_ctrl *ctrl)
2616 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2620 ++ctrl->ctrl.nr_reconnects;
2622 if (ctrl->rport->remoteport.port_state != FC_OBJSTATE_ONLINE)
2625 if (nvme_fc_ctlr_active_on_rport(ctrl))
2629 * Create the admin queue
2632 nvme_fc_init_queue(ctrl, 0);
2634 ret = __nvme_fc_create_hw_queue(ctrl, &ctrl->queues[0], 0,
2637 goto out_free_queue;
2639 ret = nvme_fc_connect_admin_queue(ctrl, &ctrl->queues[0],
2640 NVME_AQ_DEPTH, (NVME_AQ_DEPTH / 4));
2642 goto out_delete_hw_queue;
2644 if (ctrl->ctrl.state != NVME_CTRL_NEW)
2645 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
2647 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
2649 goto out_disconnect_admin_queue;
2651 set_bit(NVME_FC_Q_LIVE, &ctrl->queues[0].flags);
2654 * Check controller capabilities
2656 * todo:- add code to check if ctrl attributes changed from
2657 * prior connection values
2660 ret = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->ctrl.cap);
2662 dev_err(ctrl->ctrl.device,
2663 "prop_get NVME_REG_CAP failed\n");
2664 goto out_disconnect_admin_queue;
2668 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
2670 ret = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
2672 goto out_disconnect_admin_queue;
2674 ctrl->ctrl.max_hw_sectors =
2675 (ctrl->lport->ops->max_sgl_segments - 1) << (PAGE_SHIFT - 9);
2677 ret = nvme_init_identify(&ctrl->ctrl);
2679 goto out_disconnect_admin_queue;
2683 /* FC-NVME does not have other data in the capsule */
2684 if (ctrl->ctrl.icdoff) {
2685 dev_err(ctrl->ctrl.device, "icdoff %d is not supported!\n",
2687 goto out_disconnect_admin_queue;
2690 /* FC-NVME supports normal SGL Data Block Descriptors */
2692 if (opts->queue_size > ctrl->ctrl.maxcmd) {
2693 /* warn if maxcmd is lower than queue_size */
2694 dev_warn(ctrl->ctrl.device,
2695 "queue_size %zu > ctrl maxcmd %u, reducing "
2697 opts->queue_size, ctrl->ctrl.maxcmd);
2698 opts->queue_size = ctrl->ctrl.maxcmd;
2701 if (opts->queue_size > ctrl->ctrl.sqsize + 1) {
2702 /* warn if sqsize is lower than queue_size */
2703 dev_warn(ctrl->ctrl.device,
2704 "queue_size %zu > ctrl sqsize %u, clamping down\n",
2705 opts->queue_size, ctrl->ctrl.sqsize + 1);
2706 opts->queue_size = ctrl->ctrl.sqsize + 1;
2709 ret = nvme_fc_init_aen_ops(ctrl);
2711 goto out_term_aen_ops;
2714 * Create the io queues
2717 if (ctrl->ctrl.queue_count > 1) {
2718 if (ctrl->ctrl.state == NVME_CTRL_NEW)
2719 ret = nvme_fc_create_io_queues(ctrl);
2721 ret = nvme_fc_reinit_io_queues(ctrl);
2723 goto out_term_aen_ops;
2726 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
2728 ctrl->ctrl.nr_reconnects = 0;
2731 nvme_start_ctrl(&ctrl->ctrl);
2733 return 0; /* Success */
2736 nvme_fc_term_aen_ops(ctrl);
2737 out_disconnect_admin_queue:
2738 /* send a Disconnect(association) LS to fc-nvme target */
2739 nvme_fc_xmt_disconnect_assoc(ctrl);
2740 out_delete_hw_queue:
2741 __nvme_fc_delete_hw_queue(ctrl, &ctrl->queues[0], 0);
2743 nvme_fc_free_queue(&ctrl->queues[0]);
2744 ctrl->assoc_active = false;
2745 nvme_fc_ctlr_inactive_on_rport(ctrl);
2751 * This routine stops operation of the controller on the host side.
2752 * On the host os stack side: Admin and IO queues are stopped,
2753 * outstanding ios on them terminated via FC ABTS.
2754 * On the link side: the association is terminated.
2757 nvme_fc_delete_association(struct nvme_fc_ctrl *ctrl)
2759 unsigned long flags;
2761 if (!ctrl->assoc_active)
2763 ctrl->assoc_active = false;
2765 spin_lock_irqsave(&ctrl->lock, flags);
2766 ctrl->flags |= FCCTRL_TERMIO;
2768 spin_unlock_irqrestore(&ctrl->lock, flags);
2771 * If io queues are present, stop them and terminate all outstanding
2772 * ios on them. As FC allocates FC exchange for each io, the
2773 * transport must contact the LLDD to terminate the exchange,
2774 * thus releasing the FC exchange. We use blk_mq_tagset_busy_itr()
2775 * to tell us what io's are busy and invoke a transport routine
2776 * to kill them with the LLDD. After terminating the exchange
2777 * the LLDD will call the transport's normal io done path, but it
2778 * will have an aborted status. The done path will return the
2779 * io requests back to the block layer as part of normal completions
2780 * (but with error status).
2782 if (ctrl->ctrl.queue_count > 1) {
2783 nvme_stop_queues(&ctrl->ctrl);
2784 blk_mq_tagset_busy_iter(&ctrl->tag_set,
2785 nvme_fc_terminate_exchange, &ctrl->ctrl);
2789 * Other transports, which don't have link-level contexts bound
2790 * to sqe's, would try to gracefully shutdown the controller by
2791 * writing the registers for shutdown and polling (call
2792 * nvme_shutdown_ctrl()). Given a bunch of i/o was potentially
2793 * just aborted and we will wait on those contexts, and given
2794 * there was no indication of how live the controlelr is on the
2795 * link, don't send more io to create more contexts for the
2796 * shutdown. Let the controller fail via keepalive failure if
2797 * its still present.
2801 * clean up the admin queue. Same thing as above.
2802 * use blk_mq_tagset_busy_itr() and the transport routine to
2803 * terminate the exchanges.
2805 if (ctrl->ctrl.state != NVME_CTRL_NEW)
2806 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
2807 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
2808 nvme_fc_terminate_exchange, &ctrl->ctrl);
2810 /* kill the aens as they are a separate path */
2811 nvme_fc_abort_aen_ops(ctrl);
2813 /* wait for all io that had to be aborted */
2814 spin_lock_irq(&ctrl->lock);
2815 wait_event_lock_irq(ctrl->ioabort_wait, ctrl->iocnt == 0, ctrl->lock);
2816 ctrl->flags &= ~FCCTRL_TERMIO;
2817 spin_unlock_irq(&ctrl->lock);
2819 nvme_fc_term_aen_ops(ctrl);
2822 * send a Disconnect(association) LS to fc-nvme target
2823 * Note: could have been sent at top of process, but
2824 * cleaner on link traffic if after the aborts complete.
2825 * Note: if association doesn't exist, association_id will be 0
2827 if (ctrl->association_id)
2828 nvme_fc_xmt_disconnect_assoc(ctrl);
2830 if (ctrl->ctrl.tagset) {
2831 nvme_fc_delete_hw_io_queues(ctrl);
2832 nvme_fc_free_io_queues(ctrl);
2835 __nvme_fc_delete_hw_queue(ctrl, &ctrl->queues[0], 0);
2836 nvme_fc_free_queue(&ctrl->queues[0]);
2838 /* re-enable the admin_q so anything new can fast fail */
2839 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
2841 nvme_fc_ctlr_inactive_on_rport(ctrl);
2845 nvme_fc_delete_ctrl(struct nvme_ctrl *nctrl)
2847 struct nvme_fc_ctrl *ctrl = to_fc_ctrl(nctrl);
2849 cancel_delayed_work_sync(&ctrl->connect_work);
2851 * kill the association on the link side. this will block
2852 * waiting for io to terminate
2854 nvme_fc_delete_association(ctrl);
2856 /* resume the io queues so that things will fast fail */
2857 nvme_start_queues(nctrl);
2861 nvme_fc_reconnect_or_delete(struct nvme_fc_ctrl *ctrl, int status)
2863 struct nvme_fc_rport *rport = ctrl->rport;
2864 struct nvme_fc_remote_port *portptr = &rport->remoteport;
2865 unsigned long recon_delay = ctrl->ctrl.opts->reconnect_delay * HZ;
2868 if (ctrl->ctrl.state != NVME_CTRL_CONNECTING)
2871 if (portptr->port_state == FC_OBJSTATE_ONLINE)
2872 dev_info(ctrl->ctrl.device,
2873 "NVME-FC{%d}: reset: Reconnect attempt failed (%d)\n",
2874 ctrl->cnum, status);
2875 else if (time_after_eq(jiffies, rport->dev_loss_end))
2878 if (recon && nvmf_should_reconnect(&ctrl->ctrl)) {
2879 if (portptr->port_state == FC_OBJSTATE_ONLINE)
2880 dev_info(ctrl->ctrl.device,
2881 "NVME-FC{%d}: Reconnect attempt in %ld "
2883 ctrl->cnum, recon_delay / HZ);
2884 else if (time_after(jiffies + recon_delay, rport->dev_loss_end))
2885 recon_delay = rport->dev_loss_end - jiffies;
2887 queue_delayed_work(nvme_wq, &ctrl->connect_work, recon_delay);
2889 if (portptr->port_state == FC_OBJSTATE_ONLINE)
2890 dev_warn(ctrl->ctrl.device,
2891 "NVME-FC{%d}: Max reconnect attempts (%d) "
2892 "reached. Removing controller\n",
2893 ctrl->cnum, ctrl->ctrl.nr_reconnects);
2895 dev_warn(ctrl->ctrl.device,
2896 "NVME-FC{%d}: dev_loss_tmo (%d) expired "
2897 "while waiting for remoteport connectivity. "
2898 "Removing controller\n", ctrl->cnum,
2899 portptr->dev_loss_tmo);
2900 WARN_ON(nvme_delete_ctrl(&ctrl->ctrl));
2905 nvme_fc_reset_ctrl_work(struct work_struct *work)
2907 struct nvme_fc_ctrl *ctrl =
2908 container_of(work, struct nvme_fc_ctrl, ctrl.reset_work);
2911 nvme_stop_ctrl(&ctrl->ctrl);
2913 /* will block will waiting for io to terminate */
2914 nvme_fc_delete_association(ctrl);
2916 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2917 dev_err(ctrl->ctrl.device,
2918 "NVME-FC{%d}: error_recovery: Couldn't change state "
2919 "to CONNECTING\n", ctrl->cnum);
2923 if (ctrl->rport->remoteport.port_state == FC_OBJSTATE_ONLINE)
2924 ret = nvme_fc_create_association(ctrl);
2929 nvme_fc_reconnect_or_delete(ctrl, ret);
2931 dev_info(ctrl->ctrl.device,
2932 "NVME-FC{%d}: controller reset complete\n",
2936 static const struct nvme_ctrl_ops nvme_fc_ctrl_ops = {
2938 .module = THIS_MODULE,
2939 .flags = NVME_F_FABRICS,
2940 .reg_read32 = nvmf_reg_read32,
2941 .reg_read64 = nvmf_reg_read64,
2942 .reg_write32 = nvmf_reg_write32,
2943 .free_ctrl = nvme_fc_nvme_ctrl_freed,
2944 .submit_async_event = nvme_fc_submit_async_event,
2945 .delete_ctrl = nvme_fc_delete_ctrl,
2946 .get_address = nvmf_get_address,
2947 .reinit_request = nvme_fc_reinit_request,
2951 nvme_fc_connect_ctrl_work(struct work_struct *work)
2955 struct nvme_fc_ctrl *ctrl =
2956 container_of(to_delayed_work(work),
2957 struct nvme_fc_ctrl, connect_work);
2959 ret = nvme_fc_create_association(ctrl);
2961 nvme_fc_reconnect_or_delete(ctrl, ret);
2963 dev_info(ctrl->ctrl.device,
2964 "NVME-FC{%d}: controller reconnect complete\n",
2969 static const struct blk_mq_ops nvme_fc_admin_mq_ops = {
2970 .queue_rq = nvme_fc_queue_rq,
2971 .complete = nvme_fc_complete_rq,
2972 .init_request = nvme_fc_init_request,
2973 .exit_request = nvme_fc_exit_request,
2974 .init_hctx = nvme_fc_init_admin_hctx,
2975 .timeout = nvme_fc_timeout,
2980 * Fails a controller request if it matches an existing controller
2981 * (association) with the same tuple:
2982 * <Host NQN, Host ID, local FC port, remote FC port, SUBSYS NQN>
2984 * The ports don't need to be compared as they are intrinsically
2985 * already matched by the port pointers supplied.
2988 nvme_fc_existing_controller(struct nvme_fc_rport *rport,
2989 struct nvmf_ctrl_options *opts)
2991 struct nvme_fc_ctrl *ctrl;
2992 unsigned long flags;
2995 spin_lock_irqsave(&rport->lock, flags);
2996 list_for_each_entry(ctrl, &rport->ctrl_list, ctrl_list) {
2997 found = nvmf_ctlr_matches_baseopts(&ctrl->ctrl, opts);
3001 spin_unlock_irqrestore(&rport->lock, flags);
3006 static struct nvme_ctrl *
3007 nvme_fc_init_ctrl(struct device *dev, struct nvmf_ctrl_options *opts,
3008 struct nvme_fc_lport *lport, struct nvme_fc_rport *rport)
3010 struct nvme_fc_ctrl *ctrl;
3011 unsigned long flags;
3012 int ret, idx, retry;
3014 if (!(rport->remoteport.port_role &
3015 (FC_PORT_ROLE_NVME_DISCOVERY | FC_PORT_ROLE_NVME_TARGET))) {
3020 if (!opts->duplicate_connect &&
3021 nvme_fc_existing_controller(rport, opts)) {
3026 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
3032 idx = ida_simple_get(&nvme_fc_ctrl_cnt, 0, 0, GFP_KERNEL);
3038 ctrl->ctrl.opts = opts;
3039 INIT_LIST_HEAD(&ctrl->ctrl_list);
3040 ctrl->lport = lport;
3041 ctrl->rport = rport;
3042 ctrl->dev = lport->dev;
3044 ctrl->assoc_active = false;
3045 init_waitqueue_head(&ctrl->ioabort_wait);
3047 get_device(ctrl->dev);
3048 kref_init(&ctrl->ref);
3050 INIT_WORK(&ctrl->ctrl.reset_work, nvme_fc_reset_ctrl_work);
3051 INIT_DELAYED_WORK(&ctrl->connect_work, nvme_fc_connect_ctrl_work);
3052 spin_lock_init(&ctrl->lock);
3054 /* io queue count */
3055 ctrl->ctrl.queue_count = min_t(unsigned int,
3057 lport->ops->max_hw_queues);
3058 ctrl->ctrl.queue_count++; /* +1 for admin queue */
3060 ctrl->ctrl.sqsize = opts->queue_size - 1;
3061 ctrl->ctrl.kato = opts->kato;
3064 ctrl->queues = kcalloc(ctrl->ctrl.queue_count,
3065 sizeof(struct nvme_fc_queue), GFP_KERNEL);
3069 memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
3070 ctrl->admin_tag_set.ops = &nvme_fc_admin_mq_ops;
3071 ctrl->admin_tag_set.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
3072 ctrl->admin_tag_set.reserved_tags = 2; /* fabric connect + Keep-Alive */
3073 ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
3074 ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_fc_fcp_op) +
3076 sizeof(struct scatterlist)) +
3077 ctrl->lport->ops->fcprqst_priv_sz;
3078 ctrl->admin_tag_set.driver_data = ctrl;
3079 ctrl->admin_tag_set.nr_hw_queues = 1;
3080 ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
3081 ctrl->admin_tag_set.flags = BLK_MQ_F_NO_SCHED;
3083 ret = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
3085 goto out_free_queues;
3086 ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
3088 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
3089 if (IS_ERR(ctrl->ctrl.admin_q)) {
3090 ret = PTR_ERR(ctrl->ctrl.admin_q);
3091 goto out_free_admin_tag_set;
3095 * Would have been nice to init io queues tag set as well.
3096 * However, we require interaction from the controller
3097 * for max io queue count before we can do so.
3098 * Defer this to the connect path.
3101 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_fc_ctrl_ops, 0);
3103 goto out_cleanup_admin_q;
3105 /* at this point, teardown path changes to ref counting on nvme ctrl */
3107 spin_lock_irqsave(&rport->lock, flags);
3108 list_add_tail(&ctrl->ctrl_list, &rport->ctrl_list);
3109 spin_unlock_irqrestore(&rport->lock, flags);
3112 * It's possible that transactions used to create the association
3113 * may fail. Examples: CreateAssociation LS or CreateIOConnection
3114 * LS gets dropped/corrupted/fails; or a frame gets dropped or a
3115 * command times out for one of the actions to init the controller
3116 * (Connect, Get/Set_Property, Set_Features, etc). Many of these
3117 * transport errors (frame drop, LS failure) inherently must kill
3118 * the association. The transport is coded so that any command used
3119 * to create the association (prior to a LIVE state transition
3120 * while NEW or CONNECTING) will fail if it completes in error or
3123 * As such: as the connect request was mostly likely due to a
3124 * udev event that discovered the remote port, meaning there is
3125 * not an admin or script there to restart if the connect
3126 * request fails, retry the initial connection creation up to
3127 * three times before giving up and declaring failure.
3129 for (retry = 0; retry < 3; retry++) {
3130 ret = nvme_fc_create_association(ctrl);
3136 /* couldn't schedule retry - fail out */
3137 dev_err(ctrl->ctrl.device,
3138 "NVME-FC{%d}: Connect retry failed\n", ctrl->cnum);
3140 ctrl->ctrl.opts = NULL;
3142 /* initiate nvme ctrl ref counting teardown */
3143 nvme_uninit_ctrl(&ctrl->ctrl);
3145 /* Remove core ctrl ref. */
3146 nvme_put_ctrl(&ctrl->ctrl);
3148 /* as we're past the point where we transition to the ref
3149 * counting teardown path, if we return a bad pointer here,
3150 * the calling routine, thinking it's prior to the
3151 * transition, will do an rport put. Since the teardown
3152 * path also does a rport put, we do an extra get here to
3153 * so proper order/teardown happens.
3155 nvme_fc_rport_get(rport);
3159 return ERR_PTR(ret);
3162 nvme_get_ctrl(&ctrl->ctrl);
3164 dev_info(ctrl->ctrl.device,
3165 "NVME-FC{%d}: new ctrl: NQN \"%s\"\n",
3166 ctrl->cnum, ctrl->ctrl.opts->subsysnqn);
3170 out_cleanup_admin_q:
3171 blk_cleanup_queue(ctrl->ctrl.admin_q);
3172 out_free_admin_tag_set:
3173 blk_mq_free_tag_set(&ctrl->admin_tag_set);
3175 kfree(ctrl->queues);
3177 put_device(ctrl->dev);
3178 ida_simple_remove(&nvme_fc_ctrl_cnt, ctrl->cnum);
3182 /* exit via here doesn't follow ctlr ref points */
3183 return ERR_PTR(ret);
3187 struct nvmet_fc_traddr {
3193 __nvme_fc_parse_u64(substring_t *sstr, u64 *val)
3197 if (match_u64(sstr, &token64))
3205 * This routine validates and extracts the WWN's from the TRADDR string.
3206 * As kernel parsers need the 0x to determine number base, universally
3207 * build string to parse with 0x prefix before parsing name strings.
3210 nvme_fc_parse_traddr(struct nvmet_fc_traddr *traddr, char *buf, size_t blen)
3212 char name[2 + NVME_FC_TRADDR_HEXNAMELEN + 1];
3213 substring_t wwn = { name, &name[sizeof(name)-1] };
3214 int nnoffset, pnoffset;
3216 /* validate it string one of the 2 allowed formats */
3217 if (strnlen(buf, blen) == NVME_FC_TRADDR_MAXLENGTH &&
3218 !strncmp(buf, "nn-0x", NVME_FC_TRADDR_OXNNLEN) &&
3219 !strncmp(&buf[NVME_FC_TRADDR_MAX_PN_OFFSET],
3220 "pn-0x", NVME_FC_TRADDR_OXNNLEN)) {
3221 nnoffset = NVME_FC_TRADDR_OXNNLEN;
3222 pnoffset = NVME_FC_TRADDR_MAX_PN_OFFSET +
3223 NVME_FC_TRADDR_OXNNLEN;
3224 } else if ((strnlen(buf, blen) == NVME_FC_TRADDR_MINLENGTH &&
3225 !strncmp(buf, "nn-", NVME_FC_TRADDR_NNLEN) &&
3226 !strncmp(&buf[NVME_FC_TRADDR_MIN_PN_OFFSET],
3227 "pn-", NVME_FC_TRADDR_NNLEN))) {
3228 nnoffset = NVME_FC_TRADDR_NNLEN;
3229 pnoffset = NVME_FC_TRADDR_MIN_PN_OFFSET + NVME_FC_TRADDR_NNLEN;
3235 name[2 + NVME_FC_TRADDR_HEXNAMELEN] = 0;
3237 memcpy(&name[2], &buf[nnoffset], NVME_FC_TRADDR_HEXNAMELEN);
3238 if (__nvme_fc_parse_u64(&wwn, &traddr->nn))
3241 memcpy(&name[2], &buf[pnoffset], NVME_FC_TRADDR_HEXNAMELEN);
3242 if (__nvme_fc_parse_u64(&wwn, &traddr->pn))
3248 pr_warn("%s: bad traddr string\n", __func__);
3252 static struct nvme_ctrl *
3253 nvme_fc_create_ctrl(struct device *dev, struct nvmf_ctrl_options *opts)
3255 struct nvme_fc_lport *lport;
3256 struct nvme_fc_rport *rport;
3257 struct nvme_ctrl *ctrl;
3258 struct nvmet_fc_traddr laddr = { 0L, 0L };
3259 struct nvmet_fc_traddr raddr = { 0L, 0L };
3260 unsigned long flags;
3263 ret = nvme_fc_parse_traddr(&raddr, opts->traddr, NVMF_TRADDR_SIZE);
3264 if (ret || !raddr.nn || !raddr.pn)
3265 return ERR_PTR(-EINVAL);
3267 ret = nvme_fc_parse_traddr(&laddr, opts->host_traddr, NVMF_TRADDR_SIZE);
3268 if (ret || !laddr.nn || !laddr.pn)
3269 return ERR_PTR(-EINVAL);
3271 /* find the host and remote ports to connect together */
3272 spin_lock_irqsave(&nvme_fc_lock, flags);
3273 list_for_each_entry(lport, &nvme_fc_lport_list, port_list) {
3274 if (lport->localport.node_name != laddr.nn ||
3275 lport->localport.port_name != laddr.pn)
3278 list_for_each_entry(rport, &lport->endp_list, endp_list) {
3279 if (rport->remoteport.node_name != raddr.nn ||
3280 rport->remoteport.port_name != raddr.pn)
3283 /* if fail to get reference fall through. Will error */
3284 if (!nvme_fc_rport_get(rport))
3287 spin_unlock_irqrestore(&nvme_fc_lock, flags);
3289 ctrl = nvme_fc_init_ctrl(dev, opts, lport, rport);
3291 nvme_fc_rport_put(rport);
3295 spin_unlock_irqrestore(&nvme_fc_lock, flags);
3297 return ERR_PTR(-ENOENT);
3301 static struct nvmf_transport_ops nvme_fc_transport = {
3303 .module = THIS_MODULE,
3304 .required_opts = NVMF_OPT_TRADDR | NVMF_OPT_HOST_TRADDR,
3305 .allowed_opts = NVMF_OPT_RECONNECT_DELAY | NVMF_OPT_CTRL_LOSS_TMO,
3306 .create_ctrl = nvme_fc_create_ctrl,
3309 static int __init nvme_fc_init_module(void)
3315 * It is expected that in the future the kernel will combine
3316 * the FC-isms that are currently under scsi and now being
3317 * added to by NVME into a new standalone FC class. The SCSI
3318 * and NVME protocols and their devices would be under this
3321 * As we need something to post FC-specific udev events to,
3322 * specifically for nvme probe events, start by creating the
3323 * new device class. When the new standalone FC class is
3324 * put in place, this code will move to a more generic
3325 * location for the class.
3327 fc_class = class_create(THIS_MODULE, "fc");
3328 if (IS_ERR(fc_class)) {
3329 pr_err("couldn't register class fc\n");
3330 return PTR_ERR(fc_class);
3334 * Create a device for the FC-centric udev events
3336 fc_udev_device = device_create(fc_class, NULL, MKDEV(0, 0), NULL,
3338 if (IS_ERR(fc_udev_device)) {
3339 pr_err("couldn't create fc_udev device!\n");
3340 ret = PTR_ERR(fc_udev_device);
3341 goto out_destroy_class;
3344 ret = nvmf_register_transport(&nvme_fc_transport);
3346 goto out_destroy_device;
3351 device_destroy(fc_class, MKDEV(0, 0));
3353 class_destroy(fc_class);
3357 static void __exit nvme_fc_exit_module(void)
3359 /* sanity check - all lports should be removed */
3360 if (!list_empty(&nvme_fc_lport_list))
3361 pr_warn("%s: localport list not empty\n", __func__);
3363 nvmf_unregister_transport(&nvme_fc_transport);
3365 ida_destroy(&nvme_fc_local_port_cnt);
3366 ida_destroy(&nvme_fc_ctrl_cnt);
3368 device_destroy(fc_class, MKDEV(0, 0));
3369 class_destroy(fc_class);
3372 module_init(nvme_fc_init_module);
3373 module_exit(nvme_fc_exit_module);
3375 MODULE_LICENSE("GPL v2");