drivers/nvme/host/fc.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (c) 2016 Avago Technologies.  All rights reserved.
   4  */
   5 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   6 #include <linux/module.h>
   7 #include <linux/parser.h>
   8 #include <uapi/scsi/fc/fc_fs.h>
   9 #include <uapi/scsi/fc/fc_els.h>
  10 #include <linux/delay.h>
  11 #include <linux/overflow.h>
  12
  13 #include "nvme.h"
  14 #include "fabrics.h"
  15 #include <linux/nvme-fc-driver.h>
  16 #include <linux/nvme-fc.h>
  17 #include "fc.h"
  18 #include <scsi/scsi_transport_fc.h>
  19
  20 /* *************************** Data Structures/Defines ****************** */
  21
  22
  23 enum nvme_fc_queue_flags {
  24         NVME_FC_Q_CONNECTED = 0,
  25         NVME_FC_Q_LIVE,
  26 };
  27
  28 #define NVME_FC_DEFAULT_DEV_LOSS_TMO    60      /* seconds */
  29 #define NVME_FC_DEFAULT_RECONNECT_TMO   2       /* delay between reconnects
  30                                                  * when connected and a
  31                                                  * connection failure.
  32                                                  */
  33
  34 struct nvme_fc_queue {
  35         struct nvme_fc_ctrl     *ctrl;
  36         struct device           *dev;
  37         struct blk_mq_hw_ctx    *hctx;
  38         void                    *lldd_handle;
  39         size_t                  cmnd_capsule_len;
  40         u32                     qnum;
  41         u32                     rqcnt;
  42         u32                     seqno;
  43
  44         u64                     connection_id;
  45         atomic_t                csn;
  46
  47         unsigned long           flags;
  48 } __aligned(sizeof(u64));       /* alignment for other things alloc'd with */
  49
  50 enum nvme_fcop_flags {
  51         FCOP_FLAGS_TERMIO       = (1 << 0),
  52         FCOP_FLAGS_AEN          = (1 << 1),
  53 };
  54
  55 struct nvmefc_ls_req_op {
  56         struct nvmefc_ls_req    ls_req;
  57
  58         struct nvme_fc_rport    *rport;
  59         struct nvme_fc_queue    *queue;
  60         struct request          *rq;
  61         u32                     flags;
  62
  63         int                     ls_error;
  64         struct completion       ls_done;
  65         struct list_head        lsreq_list;     /* rport->ls_req_list */
  66         bool                    req_queued;
  67 };
  68
  69 struct nvmefc_ls_rcv_op {
  70         struct nvme_fc_rport            *rport;
  71         struct nvmefc_ls_rsp            *lsrsp;
  72         union nvmefc_ls_requests        *rqstbuf;
  73         union nvmefc_ls_responses       *rspbuf;
  74         u16                             rqstdatalen;
  75         bool                            handled;
  76         dma_addr_t                      rspdma;
  77         struct list_head                lsrcv_list;     /* rport->ls_rcv_list */
  78 } __aligned(sizeof(u64));       /* alignment for other things alloc'd with */
  79
  80 enum nvme_fcpop_state {
  81         FCPOP_STATE_UNINIT      = 0,
  82         FCPOP_STATE_IDLE        = 1,
  83         FCPOP_STATE_ACTIVE      = 2,
  84         FCPOP_STATE_ABORTED     = 3,
  85         FCPOP_STATE_COMPLETE    = 4,
  86 };
  87
  88 struct nvme_fc_fcp_op {
  89         struct nvme_request     nreq;           /*
  90                                                  * nvme/host/core.c
  91                                                  * requires this to be
  92                                                  * the 1st element in the
  93                                                  * private structure
  94                                                  * associated with the
  95                                                  * request.
  96                                                  */
  97         struct nvmefc_fcp_req   fcp_req;
  98
  99         struct nvme_fc_ctrl     *ctrl;
 100         struct nvme_fc_queue    *queue;
 101         struct request          *rq;
 102
 103         atomic_t                state;
 104         u32                     flags;
 105         u32                     rqno;
 106         u32                     nents;
 107
 108         struct nvme_fc_cmd_iu   cmd_iu;
 109         struct nvme_fc_ersp_iu  rsp_iu;
 110 };
 111
 112 struct nvme_fcp_op_w_sgl {
 113         struct nvme_fc_fcp_op   op;
 114         struct scatterlist      sgl[NVME_INLINE_SG_CNT];
 115         uint8_t                 priv[];
 116 };
 117
 118 struct nvme_fc_lport {
 119         struct nvme_fc_local_port       localport;
 120
 121         struct ida                      endp_cnt;
 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;
 126         struct kref                     ref;
 127         atomic_t                        act_rport_cnt;
 128 } __aligned(sizeof(u64));       /* alignment for other things alloc'd with */
 129
 130 struct nvme_fc_rport {
 131         struct nvme_fc_remote_port      remoteport;
 132
 133         struct list_head                endp_list; /* for lport->endp_list */
 134         struct list_head                ctrl_list;
 135         struct list_head                ls_req_list;
 136         struct list_head                ls_rcv_list;
 137         struct list_head                disc_list;
 138         struct device                   *dev;   /* physical device for dma */
 139         struct nvme_fc_lport            *lport;
 140         spinlock_t                      lock;
 141         struct kref                     ref;
 142         atomic_t                        act_ctrl_cnt;
 143         unsigned long                   dev_loss_end;
 144         struct work_struct              lsrcv_work;
 145 } __aligned(sizeof(u64));       /* alignment for other things alloc'd with */
 146
 147 /* fc_ctrl flags values - specified as bit positions */
 148 #define ASSOC_ACTIVE            0
 149 #define ASSOC_FAILED            1
 150 #define FCCTRL_TERMIO           2
 151
 152 struct nvme_fc_ctrl {
 153         spinlock_t              lock;
 154         struct nvme_fc_queue    *queues;
 155         struct device           *dev;
 156         struct nvme_fc_lport    *lport;
 157         struct nvme_fc_rport    *rport;
 158         u32                     cnum;
 159
 160         bool                    ioq_live;
 161         u64                     association_id;
 162         struct nvmefc_ls_rcv_op *rcv_disconn;
 163
 164         struct list_head        ctrl_list;      /* rport->ctrl_list */
 165
 166         struct blk_mq_tag_set   admin_tag_set;
 167         struct blk_mq_tag_set   tag_set;
 168
 169         struct delayed_work     connect_work;
 170
 171         struct kref             ref;
 172         unsigned long           flags;
 173         u32                     iocnt;
 174         wait_queue_head_t       ioabort_wait;
 175
 176         struct nvme_fc_fcp_op   aen_ops[NVME_NR_AEN_COMMANDS];
 177
 178         struct nvme_ctrl        ctrl;
 179 };
 180
 181 static inline struct nvme_fc_ctrl *
 182 to_fc_ctrl(struct nvme_ctrl *ctrl)
 183 {
 184         return container_of(ctrl, struct nvme_fc_ctrl, ctrl);
 185 }
 186
 187 static inline struct nvme_fc_lport *
 188 localport_to_lport(struct nvme_fc_local_port *portptr)
 189 {
 190         return container_of(portptr, struct nvme_fc_lport, localport);
 191 }
 192
 193 static inline struct nvme_fc_rport *
 194 remoteport_to_rport(struct nvme_fc_remote_port *portptr)
 195 {
 196         return container_of(portptr, struct nvme_fc_rport, remoteport);
 197 }
 198
 199 static inline struct nvmefc_ls_req_op *
 200 ls_req_to_lsop(struct nvmefc_ls_req *lsreq)
 201 {
 202         return container_of(lsreq, struct nvmefc_ls_req_op, ls_req);
 203 }
 204
 205 static inline struct nvme_fc_fcp_op *
 206 fcp_req_to_fcp_op(struct nvmefc_fcp_req *fcpreq)
 207 {
 208         return container_of(fcpreq, struct nvme_fc_fcp_op, fcp_req);
 209 }
 210
 211
 212
 213 /* *************************** Globals **************************** */
 214
 215
 216 static DEFINE_SPINLOCK(nvme_fc_lock);
 217
 218 static LIST_HEAD(nvme_fc_lport_list);
 219 static DEFINE_IDA(nvme_fc_local_port_cnt);
 220 static DEFINE_IDA(nvme_fc_ctrl_cnt);
 221
 222 static struct workqueue_struct *nvme_fc_wq;
 223
 224 static bool nvme_fc_waiting_to_unload;
 225 static DECLARE_COMPLETION(nvme_fc_unload_proceed);
 226
 227 /*
 228  * These items are short-term. They will eventually be moved into
 229  * a generic FC class. See comments in module init.
 230  */
 231 static struct device *fc_udev_device;
 232
 233 static void nvme_fc_complete_rq(struct request *rq);
 234
 235 /* *********************** FC-NVME Port Management ************************ */
 236
 237 static void __nvme_fc_delete_hw_queue(struct nvme_fc_ctrl *,
 238                         struct nvme_fc_queue *, unsigned int);
 239
 240 static void nvme_fc_handle_ls_rqst_work(struct work_struct *work);
 241
 242
 243 static void
 244 nvme_fc_free_lport(struct kref *ref)
 245 {
 246         struct nvme_fc_lport *lport =
 247                 container_of(ref, struct nvme_fc_lport, ref);
 248         unsigned long flags;
 249
 250         WARN_ON(lport->localport.port_state != FC_OBJSTATE_DELETED);
 251         WARN_ON(!list_empty(&lport->endp_list));
 252
 253         /* remove from transport list */
 254         spin_lock_irqsave(&nvme_fc_lock, flags);
 255         list_del(&lport->port_list);
 256         if (nvme_fc_waiting_to_unload && list_empty(&nvme_fc_lport_list))
 257                 complete(&nvme_fc_unload_proceed);
 258         spin_unlock_irqrestore(&nvme_fc_lock, flags);
 259
 260         ida_simple_remove(&nvme_fc_local_port_cnt, lport->localport.port_num);
 261         ida_destroy(&lport->endp_cnt);
 262
 263         put_device(lport->dev);
 264
 265         kfree(lport);
 266 }
 267
 268 static void
 269 nvme_fc_lport_put(struct nvme_fc_lport *lport)
 270 {
 271         kref_put(&lport->ref, nvme_fc_free_lport);
 272 }
 273
 274 static int
 275 nvme_fc_lport_get(struct nvme_fc_lport *lport)
 276 {
 277         return kref_get_unless_zero(&lport->ref);
 278 }
 279
 280
 281 static struct nvme_fc_lport *
 282 nvme_fc_attach_to_unreg_lport(struct nvme_fc_port_info *pinfo,
 283                         struct nvme_fc_port_template *ops,
 284                         struct device *dev)
 285 {
 286         struct nvme_fc_lport *lport;
 287         unsigned long flags;
 288
 289         spin_lock_irqsave(&nvme_fc_lock, flags);
 290
 291         list_for_each_entry(lport, &nvme_fc_lport_list, port_list) {
 292                 if (lport->localport.node_name != pinfo->node_name ||
 293                     lport->localport.port_name != pinfo->port_name)
 294                         continue;
 295
 296                 if (lport->dev != dev) {
 297                         lport = ERR_PTR(-EXDEV);
 298                         goto out_done;
 299                 }
 300
 301                 if (lport->localport.port_state != FC_OBJSTATE_DELETED) {
 302                         lport = ERR_PTR(-EEXIST);
 303                         goto out_done;
 304                 }
 305
 306                 if (!nvme_fc_lport_get(lport)) {
 307                         /*
 308                          * fails if ref cnt already 0. If so,
 309                          * act as if lport already deleted
 310                          */
 311                         lport = NULL;
 312                         goto out_done;
 313                 }
 314
 315                 /* resume the lport */
 316
 317                 lport->ops = ops;
 318                 lport->localport.port_role = pinfo->port_role;
 319                 lport->localport.port_id = pinfo->port_id;
 320                 lport->localport.port_state = FC_OBJSTATE_ONLINE;
 321
 322                 spin_unlock_irqrestore(&nvme_fc_lock, flags);
 323
 324                 return lport;
 325         }
 326
 327         lport = NULL;
 328
 329 out_done:
 330         spin_unlock_irqrestore(&nvme_fc_lock, flags);
 331
 332         return lport;
 333 }
 334
 335 /**
 336  * nvme_fc_register_localport - transport entry point called by an
 337  *                              LLDD to register the existence of a NVME
 338  *                              host FC port.
 339  * @pinfo:     pointer to information about the port to be registered
 340  * @template:  LLDD entrypoints and operational parameters for the port
 341  * @dev:       physical hardware device node port corresponds to. Will be
 342  *             used for DMA mappings
 343  * @portptr:   pointer to a local port pointer. Upon success, the routine
 344  *             will allocate a nvme_fc_local_port structure and place its
 345  *             address in the local port pointer. Upon failure, local port
 346  *             pointer will be set to 0.
 347  *
 348  * Returns:
 349  * a completion status. Must be 0 upon success; a negative errno
 350  * (ex: -ENXIO) upon failure.
 351  */
 352 int
 353 nvme_fc_register_localport(struct nvme_fc_port_info *pinfo,
 354                         struct nvme_fc_port_template *template,
 355                         struct device *dev,
 356                         struct nvme_fc_local_port **portptr)
 357 {
 358         struct nvme_fc_lport *newrec;
 359         unsigned long flags;
 360         int ret, idx;
 361
 362         if (!template->localport_delete || !template->remoteport_delete ||
 363             !template->ls_req || !template->fcp_io ||
 364             !template->ls_abort || !template->fcp_abort ||
 365             !template->max_hw_queues || !template->max_sgl_segments ||
 366             !template->max_dif_sgl_segments || !template->dma_boundary) {
 367                 ret = -EINVAL;
 368                 goto out_reghost_failed;
 369         }
 370
 371         /*
 372          * look to see if there is already a localport that had been
 373          * deregistered and in the process of waiting for all the
 374          * references to fully be removed.  If the references haven't
 375          * expired, we can simply re-enable the localport. Remoteports
 376          * and controller reconnections should resume naturally.
 377          */
 378         newrec = nvme_fc_attach_to_unreg_lport(pinfo, template, dev);
 379
 380         /* found an lport, but something about its state is bad */
 381         if (IS_ERR(newrec)) {
 382                 ret = PTR_ERR(newrec);
 383                 goto out_reghost_failed;
 384
 385         /* found existing lport, which was resumed */
 386         } else if (newrec) {
 387                 *portptr = &newrec->localport;
 388                 return 0;
 389         }
 390
 391         /* nothing found - allocate a new localport struct */
 392
 393         newrec = kmalloc((sizeof(*newrec) + template->local_priv_sz),
 394                          GFP_KERNEL);
 395         if (!newrec) {
 396                 ret = -ENOMEM;
 397                 goto out_reghost_failed;
 398         }
 399
 400         idx = ida_simple_get(&nvme_fc_local_port_cnt, 0, 0, GFP_KERNEL);
 401         if (idx < 0) {
 402                 ret = -ENOSPC;
 403                 goto out_fail_kfree;
 404         }
 405
 406         if (!get_device(dev) && dev) {
 407                 ret = -ENODEV;
 408                 goto out_ida_put;
 409         }
 410
 411         INIT_LIST_HEAD(&newrec->port_list);
 412         INIT_LIST_HEAD(&newrec->endp_list);
 413         kref_init(&newrec->ref);
 414         atomic_set(&newrec->act_rport_cnt, 0);
 415         newrec->ops = template;
 416         newrec->dev = dev;
 417         ida_init(&newrec->endp_cnt);
 418         if (template->local_priv_sz)
 419                 newrec->localport.private = &newrec[1];
 420         else
 421                 newrec->localport.private = NULL;
 422         newrec->localport.node_name = pinfo->node_name;
 423         newrec->localport.port_name = pinfo->port_name;
 424         newrec->localport.port_role = pinfo->port_role;
 425         newrec->localport.port_id = pinfo->port_id;
 426         newrec->localport.port_state = FC_OBJSTATE_ONLINE;
 427         newrec->localport.port_num = idx;
 428
 429         spin_lock_irqsave(&nvme_fc_lock, flags);
 430         list_add_tail(&newrec->port_list, &nvme_fc_lport_list);
 431         spin_unlock_irqrestore(&nvme_fc_lock, flags);
 432
 433         if (dev)
 434                 dma_set_seg_boundary(dev, template->dma_boundary);
 435
 436         *portptr = &newrec->localport;
 437         return 0;
 438
 439 out_ida_put:
 440         ida_simple_remove(&nvme_fc_local_port_cnt, idx);
 441 out_fail_kfree:
 442         kfree(newrec);
 443 out_reghost_failed:
 444         *portptr = NULL;
 445
 446         return ret;
 447 }
 448 EXPORT_SYMBOL_GPL(nvme_fc_register_localport);
 449
 450 /**
 451  * nvme_fc_unregister_localport - transport entry point called by an
 452  *                              LLDD to deregister/remove a previously
 453  *                              registered a NVME host FC port.
 454  * @portptr: pointer to the (registered) local port that is to be deregistered.
 455  *
 456  * Returns:
 457  * a completion status. Must be 0 upon success; a negative errno
 458  * (ex: -ENXIO) upon failure.
 459  */
 460 int
 461 nvme_fc_unregister_localport(struct nvme_fc_local_port *portptr)
 462 {
 463         struct nvme_fc_lport *lport = localport_to_lport(portptr);
 464         unsigned long flags;
 465
 466         if (!portptr)
 467                 return -EINVAL;
 468
 469         spin_lock_irqsave(&nvme_fc_lock, flags);
 470
 471         if (portptr->port_state != FC_OBJSTATE_ONLINE) {
 472                 spin_unlock_irqrestore(&nvme_fc_lock, flags);
 473                 return -EINVAL;
 474         }
 475         portptr->port_state = FC_OBJSTATE_DELETED;
 476
 477         spin_unlock_irqrestore(&nvme_fc_lock, flags);
 478
 479         if (atomic_read(&lport->act_rport_cnt) == 0)
 480                 lport->ops->localport_delete(&lport->localport);
 481
 482         nvme_fc_lport_put(lport);
 483
 484         return 0;
 485 }
 486 EXPORT_SYMBOL_GPL(nvme_fc_unregister_localport);
 487
 488 /*
 489  * TRADDR strings, per FC-NVME are fixed format:
 490  *   "nn-0x<16hexdigits>:pn-0x<16hexdigits>" - 43 characters
 491  * udev event will only differ by prefix of what field is
 492  * being specified:
 493  *    "NVMEFC_HOST_TRADDR=" or "NVMEFC_TRADDR=" - 19 max characters
 494  *  19 + 43 + null_fudge = 64 characters
 495  */
 496 #define FCNVME_TRADDR_LENGTH            64
 497
 498 static void
 499 nvme_fc_signal_discovery_scan(struct nvme_fc_lport *lport,
 500                 struct nvme_fc_rport *rport)
 501 {
 502         char hostaddr[FCNVME_TRADDR_LENGTH];    /* NVMEFC_HOST_TRADDR=...*/
 503         char tgtaddr[FCNVME_TRADDR_LENGTH];     /* NVMEFC_TRADDR=...*/
 504         char *envp[4] = { "FC_EVENT=nvmediscovery", hostaddr, tgtaddr, NULL };
 505
 506         if (!(rport->remoteport.port_role & FC_PORT_ROLE_NVME_DISCOVERY))
 507                 return;
 508
 509         snprintf(hostaddr, sizeof(hostaddr),
 510                 "NVMEFC_HOST_TRADDR=nn-0x%016llx:pn-0x%016llx",
 511                 lport->localport.node_name, lport->localport.port_name);
 512         snprintf(tgtaddr, sizeof(tgtaddr),
 513                 "NVMEFC_TRADDR=nn-0x%016llx:pn-0x%016llx",
 514                 rport->remoteport.node_name, rport->remoteport.port_name);
 515         kobject_uevent_env(&fc_udev_device->kobj, KOBJ_CHANGE, envp);
 516 }
 517
 518 static void
 519 nvme_fc_free_rport(struct kref *ref)
 520 {
 521         struct nvme_fc_rport *rport =
 522                 container_of(ref, struct nvme_fc_rport, ref);
 523         struct nvme_fc_lport *lport =
 524                         localport_to_lport(rport->remoteport.localport);
 525         unsigned long flags;
 526
 527         WARN_ON(rport->remoteport.port_state != FC_OBJSTATE_DELETED);
 528         WARN_ON(!list_empty(&rport->ctrl_list));
 529
 530         /* remove from lport list */
 531         spin_lock_irqsave(&nvme_fc_lock, flags);
 532         list_del(&rport->endp_list);
 533         spin_unlock_irqrestore(&nvme_fc_lock, flags);
 534
 535         WARN_ON(!list_empty(&rport->disc_list));
 536         ida_simple_remove(&lport->endp_cnt, rport->remoteport.port_num);
 537
 538         kfree(rport);
 539
 540         nvme_fc_lport_put(lport);
 541 }
 542
 543 static void
 544 nvme_fc_rport_put(struct nvme_fc_rport *rport)
 545 {
 546         kref_put(&rport->ref, nvme_fc_free_rport);
 547 }
 548
 549 static int
 550 nvme_fc_rport_get(struct nvme_fc_rport *rport)
 551 {
 552         return kref_get_unless_zero(&rport->ref);
 553 }
 554
 555 static void
 556 nvme_fc_resume_controller(struct nvme_fc_ctrl *ctrl)
 557 {
 558         switch (ctrl->ctrl.state) {
 559         case NVME_CTRL_NEW:
 560         case NVME_CTRL_CONNECTING:
 561                 /*
 562                  * As all reconnects were suppressed, schedule a
 563                  * connect.
 564                  */
 565                 dev_info(ctrl->ctrl.device,
 566                         "NVME-FC{%d}: connectivity re-established. "
 567                         "Attempting reconnect\n", ctrl->cnum);
 568
 569                 queue_delayed_work(nvme_wq, &ctrl->connect_work, 0);
 570                 break;
 571
 572         case NVME_CTRL_RESETTING:
 573                 /*
 574                  * Controller is already in the process of terminating the
 575                  * association. No need to do anything further. The reconnect
 576                  * step will naturally occur after the reset completes.
 577                  */
 578                 break;
 579
 580         default:
 581                 /* no action to take - let it delete */
 582                 break;
 583         }
 584 }
 585
 586 static struct nvme_fc_rport *
 587 nvme_fc_attach_to_suspended_rport(struct nvme_fc_lport *lport,
 588                                 struct nvme_fc_port_info *pinfo)
 589 {
 590         struct nvme_fc_rport *rport;
 591         struct nvme_fc_ctrl *ctrl;
 592         unsigned long flags;
 593
 594         spin_lock_irqsave(&nvme_fc_lock, flags);
 595
 596         list_for_each_entry(rport, &lport->endp_list, endp_list) {
 597                 if (rport->remoteport.node_name != pinfo->node_name ||
 598                     rport->remoteport.port_name != pinfo->port_name)
 599                         continue;
 600
 601                 if (!nvme_fc_rport_get(rport)) {
 602                         rport = ERR_PTR(-ENOLCK);
 603                         goto out_done;
 604                 }
 605
 606                 spin_unlock_irqrestore(&nvme_fc_lock, flags);
 607
 608                 spin_lock_irqsave(&rport->lock, flags);
 609
 610                 /* has it been unregistered */
 611                 if (rport->remoteport.port_state != FC_OBJSTATE_DELETED) {
 612                         /* means lldd called us twice */
 613                         spin_unlock_irqrestore(&rport->lock, flags);
 614                         nvme_fc_rport_put(rport);
 615                         return ERR_PTR(-ESTALE);
 616                 }
 617
 618                 rport->remoteport.port_role = pinfo->port_role;
 619                 rport->remoteport.port_id = pinfo->port_id;
 620                 rport->remoteport.port_state = FC_OBJSTATE_ONLINE;
 621                 rport->dev_loss_end = 0;
 622
 623                 /*
 624                  * kick off a reconnect attempt on all associations to the
 625                  * remote port. A successful reconnects will resume i/o.
 626                  */
 627                 list_for_each_entry(ctrl, &rport->ctrl_list, ctrl_list)
 628                         nvme_fc_resume_controller(ctrl);
 629
 630                 spin_unlock_irqrestore(&rport->lock, flags);
 631
 632                 return rport;
 633         }
 634
 635         rport = NULL;
 636
 637 out_done:
 638         spin_unlock_irqrestore(&nvme_fc_lock, flags);
 639
 640         return rport;
 641 }
 642
 643 static inline void
 644 __nvme_fc_set_dev_loss_tmo(struct nvme_fc_rport *rport,
 645                         struct nvme_fc_port_info *pinfo)
 646 {
 647         if (pinfo->dev_loss_tmo)
 648                 rport->remoteport.dev_loss_tmo = pinfo->dev_loss_tmo;
 649         else
 650                 rport->remoteport.dev_loss_tmo = NVME_FC_DEFAULT_DEV_LOSS_TMO;
 651 }
 652
 653 /**
 654  * nvme_fc_register_remoteport - transport entry point called by an
 655  *                              LLDD to register the existence of a NVME
 656  *                              subsystem FC port on its fabric.
 657  * @localport: pointer to the (registered) local port that the remote
 658  *             subsystem port is connected to.
 659  * @pinfo:     pointer to information about the port to be registered
 660  * @portptr:   pointer to a remote port pointer. Upon success, the routine
 661  *             will allocate a nvme_fc_remote_port structure and place its
 662  *             address in the remote port pointer. Upon failure, remote port
 663  *             pointer will be set to 0.
 664  *
 665  * Returns:
 666  * a completion status. Must be 0 upon success; a negative errno
 667  * (ex: -ENXIO) upon failure.
 668  */
 669 int
 670 nvme_fc_register_remoteport(struct nvme_fc_local_port *localport,
 671                                 struct nvme_fc_port_info *pinfo,
 672                                 struct nvme_fc_remote_port **portptr)
 673 {
 674         struct nvme_fc_lport *lport = localport_to_lport(localport);
 675         struct nvme_fc_rport *newrec;
 676         unsigned long flags;
 677         int ret, idx;
 678
 679         if (!nvme_fc_lport_get(lport)) {
 680                 ret = -ESHUTDOWN;
 681                 goto out_reghost_failed;
 682         }
 683
 684         /*
 685          * look to see if there is already a remoteport that is waiting
 686          * for a reconnect (within dev_loss_tmo) with the same WWN's.
 687          * If so, transition to it and reconnect.
 688          */
 689         newrec = nvme_fc_attach_to_suspended_rport(lport, pinfo);
 690
 691         /* found an rport, but something about its state is bad */
 692         if (IS_ERR(newrec)) {
 693                 ret = PTR_ERR(newrec);
 694                 goto out_lport_put;
 695
 696         /* found existing rport, which was resumed */
 697         } else if (newrec) {
 698                 nvme_fc_lport_put(lport);
 699                 __nvme_fc_set_dev_loss_tmo(newrec, pinfo);
 700                 nvme_fc_signal_discovery_scan(lport, newrec);
 701                 *portptr = &newrec->remoteport;
 702                 return 0;
 703         }
 704
 705         /* nothing found - allocate a new remoteport struct */
 706
 707         newrec = kmalloc((sizeof(*newrec) + lport->ops->remote_priv_sz),
 708                          GFP_KERNEL);
 709         if (!newrec) {
 710                 ret = -ENOMEM;
 711                 goto out_lport_put;
 712         }
 713
 714         idx = ida_simple_get(&lport->endp_cnt, 0, 0, GFP_KERNEL);
 715         if (idx < 0) {
 716                 ret = -ENOSPC;
 717                 goto out_kfree_rport;
 718         }
 719
 720         INIT_LIST_HEAD(&newrec->endp_list);
 721         INIT_LIST_HEAD(&newrec->ctrl_list);
 722         INIT_LIST_HEAD(&newrec->ls_req_list);
 723         INIT_LIST_HEAD(&newrec->disc_list);
 724         kref_init(&newrec->ref);
 725         atomic_set(&newrec->act_ctrl_cnt, 0);
 726         spin_lock_init(&newrec->lock);
 727         newrec->remoteport.localport = &lport->localport;
 728         INIT_LIST_HEAD(&newrec->ls_rcv_list);
 729         newrec->dev = lport->dev;
 730         newrec->lport = lport;
 731         if (lport->ops->remote_priv_sz)
 732                 newrec->remoteport.private = &newrec[1];
 733         else
 734                 newrec->remoteport.private = NULL;
 735         newrec->remoteport.port_role = pinfo->port_role;
 736         newrec->remoteport.node_name = pinfo->node_name;
 737         newrec->remoteport.port_name = pinfo->port_name;
 738         newrec->remoteport.port_id = pinfo->port_id;
 739         newrec->remoteport.port_state = FC_OBJSTATE_ONLINE;
 740         newrec->remoteport.port_num = idx;
 741         __nvme_fc_set_dev_loss_tmo(newrec, pinfo);
 742         INIT_WORK(&newrec->lsrcv_work, nvme_fc_handle_ls_rqst_work);
 743
 744         spin_lock_irqsave(&nvme_fc_lock, flags);
 745         list_add_tail(&newrec->endp_list, &lport->endp_list);
 746         spin_unlock_irqrestore(&nvme_fc_lock, flags);
 747
 748         nvme_fc_signal_discovery_scan(lport, newrec);
 749
 750         *portptr = &newrec->remoteport;
 751         return 0;
 752
 753 out_kfree_rport:
 754         kfree(newrec);
 755 out_lport_put:
 756         nvme_fc_lport_put(lport);
 757 out_reghost_failed:
 758         *portptr = NULL;
 759         return ret;
 760 }
 761 EXPORT_SYMBOL_GPL(nvme_fc_register_remoteport);
 762
 763 static int
 764 nvme_fc_abort_lsops(struct nvme_fc_rport *rport)
 765 {
 766         struct nvmefc_ls_req_op *lsop;
 767         unsigned long flags;
 768
 769 restart:
 770         spin_lock_irqsave(&rport->lock, flags);
 771
 772         list_for_each_entry(lsop, &rport->ls_req_list, lsreq_list) {
 773                 if (!(lsop->flags & FCOP_FLAGS_TERMIO)) {
 774                         lsop->flags |= FCOP_FLAGS_TERMIO;
 775                         spin_unlock_irqrestore(&rport->lock, flags);
 776                         rport->lport->ops->ls_abort(&rport->lport->localport,
 777                                                 &rport->remoteport,
 778                                                 &lsop->ls_req);
 779                         goto restart;
 780                 }
 781         }
 782         spin_unlock_irqrestore(&rport->lock, flags);
 783
 784         return 0;
 785 }
 786
 787 static void
 788 nvme_fc_ctrl_connectivity_loss(struct nvme_fc_ctrl *ctrl)
 789 {
 790         dev_info(ctrl->ctrl.device,
 791                 "NVME-FC{%d}: controller connectivity lost. Awaiting "
 792                 "Reconnect", ctrl->cnum);
 793
 794         switch (ctrl->ctrl.state) {
 795         case NVME_CTRL_NEW:
 796         case NVME_CTRL_LIVE:
 797                 /*
 798                  * Schedule a controller reset. The reset will terminate the
 799                  * association and schedule the reconnect timer.  Reconnects
 800                  * will be attempted until either the ctlr_loss_tmo
 801                  * (max_retries * connect_delay) expires or the remoteport's
 802                  * dev_loss_tmo expires.
 803                  */
 804                 if (nvme_reset_ctrl(&ctrl->ctrl)) {
 805                         dev_warn(ctrl->ctrl.device,
 806                                 "NVME-FC{%d}: Couldn't schedule reset.\n",
 807                                 ctrl->cnum);
 808                         nvme_delete_ctrl(&ctrl->ctrl);
 809                 }
 810                 break;
 811
 812         case NVME_CTRL_CONNECTING:
 813                 /*
 814                  * The association has already been terminated and the
 815                  * controller is attempting reconnects.  No need to do anything
 816                  * futher.  Reconnects will be attempted until either the
 817                  * ctlr_loss_tmo (max_retries * connect_delay) expires or the
 818                  * remoteport's dev_loss_tmo expires.
 819                  */
 820                 break;
 821
 822         case NVME_CTRL_RESETTING:
 823                 /*
 824                  * Controller is already in the process of terminating the
 825                  * association.  No need to do anything further. The reconnect
 826                  * step will kick in naturally after the association is
 827                  * terminated.
 828                  */
 829                 break;
 830
 831         case NVME_CTRL_DELETING:
 832         case NVME_CTRL_DELETING_NOIO:
 833         default:
 834                 /* no action to take - let it delete */
 835                 break;
 836         }
 837 }
 838
 839 /**
 840  * nvme_fc_unregister_remoteport - transport entry point called by an
 841  *                              LLDD to deregister/remove a previously
 842  *                              registered a NVME subsystem FC port.
 843  * @portptr: pointer to the (registered) remote port that is to be
 844  *           deregistered.
 845  *
 846  * Returns:
 847  * a completion status. Must be 0 upon success; a negative errno
 848  * (ex: -ENXIO) upon failure.
 849  */
 850 int
 851 nvme_fc_unregister_remoteport(struct nvme_fc_remote_port *portptr)
 852 {
 853         struct nvme_fc_rport *rport = remoteport_to_rport(portptr);
 854         struct nvme_fc_ctrl *ctrl;
 855         unsigned long flags;
 856
 857         if (!portptr)
 858                 return -EINVAL;
 859
 860         spin_lock_irqsave(&rport->lock, flags);
 861
 862         if (portptr->port_state != FC_OBJSTATE_ONLINE) {
 863                 spin_unlock_irqrestore(&rport->lock, flags);
 864                 return -EINVAL;
 865         }
 866         portptr->port_state = FC_OBJSTATE_DELETED;
 867
 868         rport->dev_loss_end = jiffies + (portptr->dev_loss_tmo * HZ);
 869
 870         list_for_each_entry(ctrl, &rport->ctrl_list, ctrl_list) {
 871                 /* if dev_loss_tmo==0, dev loss is immediate */
 872                 if (!portptr->dev_loss_tmo) {
 873                         dev_warn(ctrl->ctrl.device,
 874                                 "NVME-FC{%d}: controller connectivity lost.\n",
 875                                 ctrl->cnum);
 876                         nvme_delete_ctrl(&ctrl->ctrl);
 877                 } else
 878                         nvme_fc_ctrl_connectivity_loss(ctrl);
 879         }
 880
 881         spin_unlock_irqrestore(&rport->lock, flags);
 882
 883         nvme_fc_abort_lsops(rport);
 884
 885         if (atomic_read(&rport->act_ctrl_cnt) == 0)
 886                 rport->lport->ops->remoteport_delete(portptr);
 887
 888         /*
 889          * release the reference, which will allow, if all controllers
 890          * go away, which should only occur after dev_loss_tmo occurs,
 891          * for the rport to be torn down.
 892          */
 893         nvme_fc_rport_put(rport);
 894
 895         return 0;
 896 }
 897 EXPORT_SYMBOL_GPL(nvme_fc_unregister_remoteport);
 898
 899 /**
 900  * nvme_fc_rescan_remoteport - transport entry point called by an
 901  *                              LLDD to request a nvme device rescan.
 902  * @remoteport: pointer to the (registered) remote port that is to be
 903  *              rescanned.
 904  *
 905  * Returns: N/A
 906  */
 907 void
 908 nvme_fc_rescan_remoteport(struct nvme_fc_remote_port *remoteport)
 909 {
 910         struct nvme_fc_rport *rport = remoteport_to_rport(remoteport);
 911
 912         nvme_fc_signal_discovery_scan(rport->lport, rport);
 913 }
 914 EXPORT_SYMBOL_GPL(nvme_fc_rescan_remoteport);
 915
 916 int
 917 nvme_fc_set_remoteport_devloss(struct nvme_fc_remote_port *portptr,
 918                         u32 dev_loss_tmo)
 919 {
 920         struct nvme_fc_rport *rport = remoteport_to_rport(portptr);
 921         unsigned long flags;
 922
 923         spin_lock_irqsave(&rport->lock, flags);
 924
 925         if (portptr->port_state != FC_OBJSTATE_ONLINE) {
 926                 spin_unlock_irqrestore(&rport->lock, flags);
 927                 return -EINVAL;
 928         }
 929
 930         /* a dev_loss_tmo of 0 (immediate) is allowed to be set */
 931         rport->remoteport.dev_loss_tmo = dev_loss_tmo;
 932
 933         spin_unlock_irqrestore(&rport->lock, flags);
 934
 935         return 0;
 936 }
 937 EXPORT_SYMBOL_GPL(nvme_fc_set_remoteport_devloss);
 938
 939
 940 /* *********************** FC-NVME DMA Handling **************************** */
 941
 942 /*
 943  * The fcloop device passes in a NULL device pointer. Real LLD's will
 944  * pass in a valid device pointer. If NULL is passed to the dma mapping
 945  * routines, depending on the platform, it may or may not succeed, and
 946  * may crash.
 947  *
 948  * As such:
 949  * Wrapper all the dma routines and check the dev pointer.
 950  *
 951  * If simple mappings (return just a dma address, we'll noop them,
 952  * returning a dma address of 0.
 953  *
 954  * On more complex mappings (dma_map_sg), a pseudo routine fills
 955  * in the scatter list, setting all dma addresses to 0.
 956  */
 957
 958 static inline dma_addr_t
 959 fc_dma_map_single(struct device *dev, void *ptr, size_t size,
 960                 enum dma_data_direction dir)
 961 {
 962         return dev ? dma_map_single(dev, ptr, size, dir) : (dma_addr_t)0L;
 963 }
 964
 965 static inline int
 966 fc_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
 967 {
 968         return dev ? dma_mapping_error(dev, dma_addr) : 0;
 969 }
 970
 971 static inline void
 972 fc_dma_unmap_single(struct device *dev, dma_addr_t addr, size_t size,
 973         enum dma_data_direction dir)
 974 {
 975         if (dev)
 976                 dma_unmap_single(dev, addr, size, dir);
 977 }
 978
 979 static inline void
 980 fc_dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr, size_t size,
 981                 enum dma_data_direction dir)
 982 {
 983         if (dev)
 984                 dma_sync_single_for_cpu(dev, addr, size, dir);
 985 }
 986
 987 static inline void
 988 fc_dma_sync_single_for_device(struct device *dev, dma_addr_t addr, size_t size,
 989                 enum dma_data_direction dir)
 990 {
 991         if (dev)
 992                 dma_sync_single_for_device(dev, addr, size, dir);
 993 }
 994
 995 /* pseudo dma_map_sg call */
 996 static int
 997 fc_map_sg(struct scatterlist *sg, int nents)
 998 {
 999         struct scatterlist *s;
1000         int i;
1001
1002         WARN_ON(nents == 0 || sg[0].length == 0);
1003
1004         for_each_sg(sg, s, nents, i) {
1005                 s->dma_address = 0L;
1006 #ifdef CONFIG_NEED_SG_DMA_LENGTH
1007                 s->dma_length = s->length;
1008 #endif
1009         }
1010         return nents;
1011 }
1012
1013 static inline int
1014 fc_dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
1015                 enum dma_data_direction dir)
1016 {
1017         return dev ? dma_map_sg(dev, sg, nents, dir) : fc_map_sg(sg, nents);
1018 }
1019
1020 static inline void
1021 fc_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
1022                 enum dma_data_direction dir)
1023 {
1024         if (dev)
1025                 dma_unmap_sg(dev, sg, nents, dir);
1026 }
1027
1028 /* *********************** FC-NVME LS Handling **************************** */
1029
1030 static void nvme_fc_ctrl_put(struct nvme_fc_ctrl *);
1031 static int nvme_fc_ctrl_get(struct nvme_fc_ctrl *);
1032
1033 static void nvme_fc_error_recovery(struct nvme_fc_ctrl *ctrl, char *errmsg);
1034
1035 static void
1036 __nvme_fc_finish_ls_req(struct nvmefc_ls_req_op *lsop)
1037 {
1038         struct nvme_fc_rport *rport = lsop->rport;
1039         struct nvmefc_ls_req *lsreq = &lsop->ls_req;
1040         unsigned long flags;
1041
1042         spin_lock_irqsave(&rport->lock, flags);
1043
1044         if (!lsop->req_queued) {
1045                 spin_unlock_irqrestore(&rport->lock, flags);
1046                 return;
1047         }
1048
1049         list_del(&lsop->lsreq_list);
1050
1051         lsop->req_queued = false;
1052
1053         spin_unlock_irqrestore(&rport->lock, flags);
1054
1055         fc_dma_unmap_single(rport->dev, lsreq->rqstdma,
1056                                   (lsreq->rqstlen + lsreq->rsplen),
1057                                   DMA_BIDIRECTIONAL);
1058
1059         nvme_fc_rport_put(rport);
1060 }
1061
1062 static int
1063 __nvme_fc_send_ls_req(struct nvme_fc_rport *rport,
1064                 struct nvmefc_ls_req_op *lsop,
1065                 void (*done)(struct nvmefc_ls_req *req, int status))
1066 {
1067         struct nvmefc_ls_req *lsreq = &lsop->ls_req;
1068         unsigned long flags;
1069         int ret = 0;
1070
1071         if (rport->remoteport.port_state != FC_OBJSTATE_ONLINE)
1072                 return -ECONNREFUSED;
1073
1074         if (!nvme_fc_rport_get(rport))
1075                 return -ESHUTDOWN;
1076
1077         lsreq->done = done;
1078         lsop->rport = rport;
1079         lsop->req_queued = false;
1080         INIT_LIST_HEAD(&lsop->lsreq_list);
1081         init_completion(&lsop->ls_done);
1082
1083         lsreq->rqstdma = fc_dma_map_single(rport->dev, lsreq->rqstaddr,
1084                                   lsreq->rqstlen + lsreq->rsplen,
1085                                   DMA_BIDIRECTIONAL);
1086         if (fc_dma_mapping_error(rport->dev, lsreq->rqstdma)) {
1087                 ret = -EFAULT;
1088                 goto out_putrport;
1089         }
1090         lsreq->rspdma = lsreq->rqstdma + lsreq->rqstlen;
1091
1092         spin_lock_irqsave(&rport->lock, flags);
1093
1094         list_add_tail(&lsop->lsreq_list, &rport->ls_req_list);
1095
1096         lsop->req_queued = true;
1097
1098         spin_unlock_irqrestore(&rport->lock, flags);
1099
1100         ret = rport->lport->ops->ls_req(&rport->lport->localport,
1101                                         &rport->remoteport, lsreq);
1102         if (ret)
1103                 goto out_unlink;
1104
1105         return 0;
1106
1107 out_unlink:
1108         lsop->ls_error = ret;
1109         spin_lock_irqsave(&rport->lock, flags);
1110         lsop->req_queued = false;
1111         list_del(&lsop->lsreq_list);
1112         spin_unlock_irqrestore(&rport->lock, flags);
1113         fc_dma_unmap_single(rport->dev, lsreq->rqstdma,
1114                                   (lsreq->rqstlen + lsreq->rsplen),
1115                                   DMA_BIDIRECTIONAL);
1116 out_putrport:
1117         nvme_fc_rport_put(rport);
1118
1119         return ret;
1120 }
1121
1122 static void
1123 nvme_fc_send_ls_req_done(struct nvmefc_ls_req *lsreq, int status)
1124 {
1125         struct nvmefc_ls_req_op *lsop = ls_req_to_lsop(lsreq);
1126
1127         lsop->ls_error = status;
1128         complete(&lsop->ls_done);
1129 }
1130
1131 static int
1132 nvme_fc_send_ls_req(struct nvme_fc_rport *rport, struct nvmefc_ls_req_op *lsop)
1133 {
1134         struct nvmefc_ls_req *lsreq = &lsop->ls_req;
1135         struct fcnvme_ls_rjt *rjt = lsreq->rspaddr;
1136         int ret;
1137
1138         ret = __nvme_fc_send_ls_req(rport, lsop, nvme_fc_send_ls_req_done);
1139
1140         if (!ret) {
1141                 /*
1142                  * No timeout/not interruptible as we need the struct
1143                  * to exist until the lldd calls us back. Thus mandate
1144                  * wait until driver calls back. lldd responsible for
1145                  * the timeout action
1146                  */
1147                 wait_for_completion(&lsop->ls_done);
1148
1149                 __nvme_fc_finish_ls_req(lsop);
1150
1151                 ret = lsop->ls_error;
1152         }
1153
1154         if (ret)
1155                 return ret;
1156
1157         /* ACC or RJT payload ? */
1158         if (rjt->w0.ls_cmd == FCNVME_LS_RJT)
1159                 return -ENXIO;
1160
1161         return 0;
1162 }
1163
1164 static int
1165 nvme_fc_send_ls_req_async(struct nvme_fc_rport *rport,
1166                 struct nvmefc_ls_req_op *lsop,
1167                 void (*done)(struct nvmefc_ls_req *req, int status))
1168 {
1169         /* don't wait for completion */
1170
1171         return __nvme_fc_send_ls_req(rport, lsop, done);
1172 }
1173
1174 static int
1175 nvme_fc_connect_admin_queue(struct nvme_fc_ctrl *ctrl,
1176         struct nvme_fc_queue *queue, u16 qsize, u16 ersp_ratio)
1177 {
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;
1182         unsigned long flags;
1183         int ret, fcret = 0;
1184
1185         lsop = kzalloc((sizeof(*lsop) +
1186                          sizeof(*assoc_rqst) + sizeof(*assoc_acc) +
1187                          ctrl->lport->ops->lsrqst_priv_sz), GFP_KERNEL);
1188         if (!lsop) {
1189                 dev_info(ctrl->ctrl.device,
1190                         "NVME-FC{%d}: send Create Association failed: ENOMEM\n",
1191                         ctrl->cnum);
1192                 ret = -ENOMEM;
1193                 goto out_no_memory;
1194         }
1195
1196         assoc_rqst = (struct fcnvme_ls_cr_assoc_rqst *)&lsop[1];
1197         assoc_acc = (struct fcnvme_ls_cr_assoc_acc *)&assoc_rqst[1];
1198         lsreq = &lsop->ls_req;
1199         if (ctrl->lport->ops->lsrqst_priv_sz)
1200                 lsreq->private = &assoc_acc[1];
1201         else
1202                 lsreq->private = NULL;
1203
1204         assoc_rqst->w0.ls_cmd = FCNVME_LS_CREATE_ASSOCIATION;
1205         assoc_rqst->desc_list_len =
1206                         cpu_to_be32(sizeof(struct fcnvme_lsdesc_cr_assoc_cmd));
1207
1208         assoc_rqst->assoc_cmd.desc_tag =
1209                         cpu_to_be32(FCNVME_LSDESC_CREATE_ASSOC_CMD);
1210         assoc_rqst->assoc_cmd.desc_len =
1211                         fcnvme_lsdesc_len(
1212                                 sizeof(struct fcnvme_lsdesc_cr_assoc_cmd));
1213
1214         assoc_rqst->assoc_cmd.ersp_ratio = cpu_to_be16(ersp_ratio);
1215         assoc_rqst->assoc_cmd.sqsize = cpu_to_be16(qsize - 1);
1216         /* Linux supports only Dynamic controllers */
1217         assoc_rqst->assoc_cmd.cntlid = cpu_to_be16(0xffff);
1218         uuid_copy(&assoc_rqst->assoc_cmd.hostid, &ctrl->ctrl.opts->host->id);
1219         strncpy(assoc_rqst->assoc_cmd.hostnqn, ctrl->ctrl.opts->host->nqn,
1220                 min(FCNVME_ASSOC_HOSTNQN_LEN, NVMF_NQN_SIZE));
1221         strncpy(assoc_rqst->assoc_cmd.subnqn, ctrl->ctrl.opts->subsysnqn,
1222                 min(FCNVME_ASSOC_SUBNQN_LEN, NVMF_NQN_SIZE));
1223
1224         lsop->queue = queue;
1225         lsreq->rqstaddr = assoc_rqst;
1226         lsreq->rqstlen = sizeof(*assoc_rqst);
1227         lsreq->rspaddr = assoc_acc;
1228         lsreq->rsplen = sizeof(*assoc_acc);
1229         lsreq->timeout = NVME_FC_LS_TIMEOUT_SEC;
1230
1231         ret = nvme_fc_send_ls_req(ctrl->rport, lsop);
1232         if (ret)
1233                 goto out_free_buffer;
1234
1235         /* process connect LS completion */
1236
1237         /* validate the ACC response */
1238         if (assoc_acc->hdr.w0.ls_cmd != FCNVME_LS_ACC)
1239                 fcret = VERR_LSACC;
1240         else if (assoc_acc->hdr.desc_list_len !=
1241                         fcnvme_lsdesc_len(
1242                                 sizeof(struct fcnvme_ls_cr_assoc_acc)))
1243                 fcret = VERR_CR_ASSOC_ACC_LEN;
1244         else if (assoc_acc->hdr.rqst.desc_tag !=
1245                         cpu_to_be32(FCNVME_LSDESC_RQST))
1246                 fcret = VERR_LSDESC_RQST;
1247         else if (assoc_acc->hdr.rqst.desc_len !=
1248                         fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_rqst)))
1249                 fcret = VERR_LSDESC_RQST_LEN;
1250         else if (assoc_acc->hdr.rqst.w0.ls_cmd != FCNVME_LS_CREATE_ASSOCIATION)
1251                 fcret = VERR_CR_ASSOC;
1252         else if (assoc_acc->associd.desc_tag !=
1253                         cpu_to_be32(FCNVME_LSDESC_ASSOC_ID))
1254                 fcret = VERR_ASSOC_ID;
1255         else if (assoc_acc->associd.desc_len !=
1256                         fcnvme_lsdesc_len(
1257                                 sizeof(struct fcnvme_lsdesc_assoc_id)))
1258                 fcret = VERR_ASSOC_ID_LEN;
1259         else if (assoc_acc->connectid.desc_tag !=
1260                         cpu_to_be32(FCNVME_LSDESC_CONN_ID))
1261                 fcret = VERR_CONN_ID;
1262         else if (assoc_acc->connectid.desc_len !=
1263                         fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_conn_id)))
1264                 fcret = VERR_CONN_ID_LEN;
1265
1266         if (fcret) {
1267                 ret = -EBADF;
1268                 dev_err(ctrl->dev,
1269                         "q %d Create Association LS failed: %s\n",
1270                         queue->qnum, validation_errors[fcret]);
1271         } else {
1272                 spin_lock_irqsave(&ctrl->lock, flags);
1273                 ctrl->association_id =
1274                         be64_to_cpu(assoc_acc->associd.association_id);
1275                 queue->connection_id =
1276                         be64_to_cpu(assoc_acc->connectid.connection_id);
1277                 set_bit(NVME_FC_Q_CONNECTED, &queue->flags);
1278                 spin_unlock_irqrestore(&ctrl->lock, flags);
1279         }
1280
1281 out_free_buffer:
1282         kfree(lsop);
1283 out_no_memory:
1284         if (ret)
1285                 dev_err(ctrl->dev,
1286                         "queue %d connect admin queue failed (%d).\n",
1287                         queue->qnum, ret);
1288         return ret;
1289 }
1290
1291 static int
1292 nvme_fc_connect_queue(struct nvme_fc_ctrl *ctrl, struct nvme_fc_queue *queue,
1293                         u16 qsize, u16 ersp_ratio)
1294 {
1295         struct nvmefc_ls_req_op *lsop;
1296         struct nvmefc_ls_req *lsreq;
1297         struct fcnvme_ls_cr_conn_rqst *conn_rqst;
1298         struct fcnvme_ls_cr_conn_acc *conn_acc;
1299         int ret, fcret = 0;
1300
1301         lsop = kzalloc((sizeof(*lsop) +
1302                          sizeof(*conn_rqst) + sizeof(*conn_acc) +
1303                          ctrl->lport->ops->lsrqst_priv_sz), GFP_KERNEL);
1304         if (!lsop) {
1305                 dev_info(ctrl->ctrl.device,
1306                         "NVME-FC{%d}: send Create Connection failed: ENOMEM\n",
1307                         ctrl->cnum);
1308                 ret = -ENOMEM;
1309                 goto out_no_memory;
1310         }
1311
1312         conn_rqst = (struct fcnvme_ls_cr_conn_rqst *)&lsop[1];
1313         conn_acc = (struct fcnvme_ls_cr_conn_acc *)&conn_rqst[1];
1314         lsreq = &lsop->ls_req;
1315         if (ctrl->lport->ops->lsrqst_priv_sz)
1316                 lsreq->private = (void *)&conn_acc[1];
1317         else
1318                 lsreq->private = NULL;
1319
1320         conn_rqst->w0.ls_cmd = FCNVME_LS_CREATE_CONNECTION;
1321         conn_rqst->desc_list_len = cpu_to_be32(
1322                                 sizeof(struct fcnvme_lsdesc_assoc_id) +
1323                                 sizeof(struct fcnvme_lsdesc_cr_conn_cmd));
1324
1325         conn_rqst->associd.desc_tag = cpu_to_be32(FCNVME_LSDESC_ASSOC_ID);
1326         conn_rqst->associd.desc_len =
1327                         fcnvme_lsdesc_len(
1328                                 sizeof(struct fcnvme_lsdesc_assoc_id));
1329         conn_rqst->associd.association_id = cpu_to_be64(ctrl->association_id);
1330         conn_rqst->connect_cmd.desc_tag =
1331                         cpu_to_be32(FCNVME_LSDESC_CREATE_CONN_CMD);
1332         conn_rqst->connect_cmd.desc_len =
1333                         fcnvme_lsdesc_len(
1334                                 sizeof(struct fcnvme_lsdesc_cr_conn_cmd));
1335         conn_rqst->connect_cmd.ersp_ratio = cpu_to_be16(ersp_ratio);
1336         conn_rqst->connect_cmd.qid  = cpu_to_be16(queue->qnum);
1337         conn_rqst->connect_cmd.sqsize = cpu_to_be16(qsize - 1);
1338
1339         lsop->queue = queue;
1340         lsreq->rqstaddr = conn_rqst;
1341         lsreq->rqstlen = sizeof(*conn_rqst);
1342         lsreq->rspaddr = conn_acc;
1343         lsreq->rsplen = sizeof(*conn_acc);
1344         lsreq->timeout = NVME_FC_LS_TIMEOUT_SEC;
1345
1346         ret = nvme_fc_send_ls_req(ctrl->rport, lsop);
1347         if (ret)
1348                 goto out_free_buffer;
1349
1350         /* process connect LS completion */
1351
1352         /* validate the ACC response */
1353         if (conn_acc->hdr.w0.ls_cmd != FCNVME_LS_ACC)
1354                 fcret = VERR_LSACC;
1355         else if (conn_acc->hdr.desc_list_len !=
1356                         fcnvme_lsdesc_len(sizeof(struct fcnvme_ls_cr_conn_acc)))
1357                 fcret = VERR_CR_CONN_ACC_LEN;
1358         else if (conn_acc->hdr.rqst.desc_tag != cpu_to_be32(FCNVME_LSDESC_RQST))
1359                 fcret = VERR_LSDESC_RQST;
1360         else if (conn_acc->hdr.rqst.desc_len !=
1361                         fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_rqst)))
1362                 fcret = VERR_LSDESC_RQST_LEN;
1363         else if (conn_acc->hdr.rqst.w0.ls_cmd != FCNVME_LS_CREATE_CONNECTION)
1364                 fcret = VERR_CR_CONN;
1365         else if (conn_acc->connectid.desc_tag !=
1366                         cpu_to_be32(FCNVME_LSDESC_CONN_ID))
1367                 fcret = VERR_CONN_ID;
1368         else if (conn_acc->connectid.desc_len !=
1369                         fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_conn_id)))
1370                 fcret = VERR_CONN_ID_LEN;
1371
1372         if (fcret) {
1373                 ret = -EBADF;
1374                 dev_err(ctrl->dev,
1375                         "q %d Create I/O Connection LS failed: %s\n",
1376                         queue->qnum, validation_errors[fcret]);
1377         } else {
1378                 queue->connection_id =
1379                         be64_to_cpu(conn_acc->connectid.connection_id);
1380                 set_bit(NVME_FC_Q_CONNECTED, &queue->flags);
1381         }
1382
1383 out_free_buffer:
1384         kfree(lsop);
1385 out_no_memory:
1386         if (ret)
1387                 dev_err(ctrl->dev,
1388                         "queue %d connect I/O queue failed (%d).\n",
1389                         queue->qnum, ret);
1390         return ret;
1391 }
1392
1393 static void
1394 nvme_fc_disconnect_assoc_done(struct nvmefc_ls_req *lsreq, int status)
1395 {
1396         struct nvmefc_ls_req_op *lsop = ls_req_to_lsop(lsreq);
1397
1398         __nvme_fc_finish_ls_req(lsop);
1399
1400         /* fc-nvme initiator doesn't care about success or failure of cmd */
1401
1402         kfree(lsop);
1403 }
1404
1405 /*
1406  * This routine sends a FC-NVME LS to disconnect (aka terminate)
1407  * the FC-NVME Association.  Terminating the association also
1408  * terminates the FC-NVME connections (per queue, both admin and io
1409  * queues) that are part of the association. E.g. things are torn
1410  * down, and the related FC-NVME Association ID and Connection IDs
1411  * become invalid.
1412  *
1413  * The behavior of the fc-nvme initiator is such that it's
1414  * understanding of the association and connections will implicitly
1415  * be torn down. The action is implicit as it may be due to a loss of
1416  * connectivity with the fc-nvme target, so you may never get a
1417  * response even if you tried.  As such, the action of this routine
1418  * is to asynchronously send the LS, ignore any results of the LS, and
1419  * continue on with terminating the association. If the fc-nvme target
1420  * is present and receives the LS, it too can tear down.
1421  */
1422 static void
1423 nvme_fc_xmt_disconnect_assoc(struct nvme_fc_ctrl *ctrl)
1424 {
1425         struct fcnvme_ls_disconnect_assoc_rqst *discon_rqst;
1426         struct fcnvme_ls_disconnect_assoc_acc *discon_acc;
1427         struct nvmefc_ls_req_op *lsop;
1428         struct nvmefc_ls_req *lsreq;
1429         int ret;
1430
1431         lsop = kzalloc((sizeof(*lsop) +
1432                         sizeof(*discon_rqst) + sizeof(*discon_acc) +
1433                         ctrl->lport->ops->lsrqst_priv_sz), GFP_KERNEL);
1434         if (!lsop) {
1435                 dev_info(ctrl->ctrl.device,
1436                         "NVME-FC{%d}: send Disconnect Association "
1437                         "failed: ENOMEM\n",
1438                         ctrl->cnum);
1439                 return;
1440         }
1441
1442         discon_rqst = (struct fcnvme_ls_disconnect_assoc_rqst *)&lsop[1];
1443         discon_acc = (struct fcnvme_ls_disconnect_assoc_acc *)&discon_rqst[1];
1444         lsreq = &lsop->ls_req;
1445         if (ctrl->lport->ops->lsrqst_priv_sz)
1446                 lsreq->private = (void *)&discon_acc[1];
1447         else
1448                 lsreq->private = NULL;
1449
1450         nvmefc_fmt_lsreq_discon_assoc(lsreq, discon_rqst, discon_acc,
1451                                 ctrl->association_id);
1452
1453         ret = nvme_fc_send_ls_req_async(ctrl->rport, lsop,
1454                                 nvme_fc_disconnect_assoc_done);
1455         if (ret)
1456                 kfree(lsop);
1457 }
1458
1459 static void
1460 nvme_fc_xmt_ls_rsp_done(struct nvmefc_ls_rsp *lsrsp)
1461 {
1462         struct nvmefc_ls_rcv_op *lsop = lsrsp->nvme_fc_private;
1463         struct nvme_fc_rport *rport = lsop->rport;
1464         struct nvme_fc_lport *lport = rport->lport;
1465         unsigned long flags;
1466
1467         spin_lock_irqsave(&rport->lock, flags);
1468         list_del(&lsop->lsrcv_list);
1469         spin_unlock_irqrestore(&rport->lock, flags);
1470
1471         fc_dma_sync_single_for_cpu(lport->dev, lsop->rspdma,
1472                                 sizeof(*lsop->rspbuf), DMA_TO_DEVICE);
1473         fc_dma_unmap_single(lport->dev, lsop->rspdma,
1474                         sizeof(*lsop->rspbuf), DMA_TO_DEVICE);
1475
1476         kfree(lsop);
1477
1478         nvme_fc_rport_put(rport);
1479 }
1480
1481 static void
1482 nvme_fc_xmt_ls_rsp(struct nvmefc_ls_rcv_op *lsop)
1483 {
1484         struct nvme_fc_rport *rport = lsop->rport;
1485         struct nvme_fc_lport *lport = rport->lport;
1486         struct fcnvme_ls_rqst_w0 *w0 = &lsop->rqstbuf->w0;
1487         int ret;
1488
1489         fc_dma_sync_single_for_device(lport->dev, lsop->rspdma,
1490                                   sizeof(*lsop->rspbuf), DMA_TO_DEVICE);
1491
1492         ret = lport->ops->xmt_ls_rsp(&lport->localport, &rport->remoteport,
1493                                      lsop->lsrsp);
1494         if (ret) {
1495                 dev_warn(lport->dev,
1496                         "LLDD rejected LS RSP xmt: LS %d status %d\n",
1497                         w0->ls_cmd, ret);
1498                 nvme_fc_xmt_ls_rsp_done(lsop->lsrsp);
1499                 return;
1500         }
1501 }
1502
1503 static struct nvme_fc_ctrl *
1504 nvme_fc_match_disconn_ls(struct nvme_fc_rport *rport,
1505                       struct nvmefc_ls_rcv_op *lsop)
1506 {
1507         struct fcnvme_ls_disconnect_assoc_rqst *rqst =
1508                                         &lsop->rqstbuf->rq_dis_assoc;
1509         struct nvme_fc_ctrl *ctrl, *ret = NULL;
1510         struct nvmefc_ls_rcv_op *oldls = NULL;
1511         u64 association_id = be64_to_cpu(rqst->associd.association_id);
1512         unsigned long flags;
1513
1514         spin_lock_irqsave(&rport->lock, flags);
1515
1516         list_for_each_entry(ctrl, &rport->ctrl_list, ctrl_list) {
1517                 if (!nvme_fc_ctrl_get(ctrl))
1518                         continue;
1519                 spin_lock(&ctrl->lock);
1520                 if (association_id == ctrl->association_id) {
1521                         oldls = ctrl->rcv_disconn;
1522                         ctrl->rcv_disconn = lsop;
1523                         ret = ctrl;
1524                 }
1525                 spin_unlock(&ctrl->lock);
1526                 if (ret)
1527                         /* leave the ctrl get reference */
1528                         break;
1529                 nvme_fc_ctrl_put(ctrl);
1530         }
1531
1532         spin_unlock_irqrestore(&rport->lock, flags);
1533
1534         /* transmit a response for anything that was pending */
1535         if (oldls) {
1536                 dev_info(rport->lport->dev,
1537                         "NVME-FC{%d}: Multiple Disconnect Association "
1538                         "LS's received\n", ctrl->cnum);
1539                 /* overwrite good response with bogus failure */
1540                 oldls->lsrsp->rsplen = nvme_fc_format_rjt(oldls->rspbuf,
1541                                                 sizeof(*oldls->rspbuf),
1542                                                 rqst->w0.ls_cmd,
1543                                                 FCNVME_RJT_RC_UNAB,
1544                                                 FCNVME_RJT_EXP_NONE, 0);
1545                 nvme_fc_xmt_ls_rsp(oldls);
1546         }
1547
1548         return ret;
1549 }
1550
1551 /*
1552  * returns true to mean LS handled and ls_rsp can be sent
1553  * returns false to defer ls_rsp xmt (will be done as part of
1554  *     association termination)
1555  */
1556 static bool
1557 nvme_fc_ls_disconnect_assoc(struct nvmefc_ls_rcv_op *lsop)
1558 {
1559         struct nvme_fc_rport *rport = lsop->rport;
1560         struct fcnvme_ls_disconnect_assoc_rqst *rqst =
1561                                         &lsop->rqstbuf->rq_dis_assoc;
1562         struct fcnvme_ls_disconnect_assoc_acc *acc =
1563                                         &lsop->rspbuf->rsp_dis_assoc;
1564         struct nvme_fc_ctrl *ctrl = NULL;
1565         int ret = 0;
1566
1567         memset(acc, 0, sizeof(*acc));
1568
1569         ret = nvmefc_vldt_lsreq_discon_assoc(lsop->rqstdatalen, rqst);
1570         if (!ret) {
1571                 /* match an active association */
1572                 ctrl = nvme_fc_match_disconn_ls(rport, lsop);
1573                 if (!ctrl)
1574                         ret = VERR_NO_ASSOC;
1575         }
1576
1577         if (ret) {
1578                 dev_info(rport->lport->dev,
1579                         "Disconnect LS failed: %s\n",
1580                         validation_errors[ret]);
1581                 lsop->lsrsp->rsplen = nvme_fc_format_rjt(acc,
1582                                         sizeof(*acc), rqst->w0.ls_cmd,
1583                                         (ret == VERR_NO_ASSOC) ?
1584                                                 FCNVME_RJT_RC_INV_ASSOC :
1585                                                 FCNVME_RJT_RC_LOGIC,
1586                                         FCNVME_RJT_EXP_NONE, 0);
1587                 return true;
1588         }
1589
1590         /* format an ACCept response */
1591
1592         lsop->lsrsp->rsplen = sizeof(*acc);
1593
1594         nvme_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1595                         fcnvme_lsdesc_len(
1596                                 sizeof(struct fcnvme_ls_disconnect_assoc_acc)),
1597                         FCNVME_LS_DISCONNECT_ASSOC);
1598
1599         /*
1600          * the transmit of the response will occur after the exchanges
1601          * for the association have been ABTS'd by
1602          * nvme_fc_delete_association().
1603          */
1604
1605         /* fail the association */
1606         nvme_fc_error_recovery(ctrl, "Disconnect Association LS received");
1607
1608         /* release the reference taken by nvme_fc_match_disconn_ls() */
1609         nvme_fc_ctrl_put(ctrl);
1610
1611         return false;
1612 }
1613
1614 /*
1615  * Actual Processing routine for received FC-NVME LS Requests from the LLD
1616  * returns true if a response should be sent afterward, false if rsp will
1617  * be sent asynchronously.
1618  */
1619 static bool
1620 nvme_fc_handle_ls_rqst(struct nvmefc_ls_rcv_op *lsop)
1621 {
1622         struct fcnvme_ls_rqst_w0 *w0 = &lsop->rqstbuf->w0;
1623         bool ret = true;
1624
1625         lsop->lsrsp->nvme_fc_private = lsop;
1626         lsop->lsrsp->rspbuf = lsop->rspbuf;
1627         lsop->lsrsp->rspdma = lsop->rspdma;
1628         lsop->lsrsp->done = nvme_fc_xmt_ls_rsp_done;
1629         /* Be preventative. handlers will later set to valid length */
1630         lsop->lsrsp->rsplen = 0;
1631
1632         /*
1633          * handlers:
1634          *   parse request input, execute the request, and format the
1635          *   LS response
1636          */
1637         switch (w0->ls_cmd) {
1638         case FCNVME_LS_DISCONNECT_ASSOC:
1639                 ret = nvme_fc_ls_disconnect_assoc(lsop);
1640                 break;
1641         case FCNVME_LS_DISCONNECT_CONN:
1642                 lsop->lsrsp->rsplen = nvme_fc_format_rjt(lsop->rspbuf,
1643                                 sizeof(*lsop->rspbuf), w0->ls_cmd,
1644                                 FCNVME_RJT_RC_UNSUP, FCNVME_RJT_EXP_NONE, 0);
1645                 break;
1646         case FCNVME_LS_CREATE_ASSOCIATION:
1647         case FCNVME_LS_CREATE_CONNECTION:
1648                 lsop->lsrsp->rsplen = nvme_fc_format_rjt(lsop->rspbuf,
1649                                 sizeof(*lsop->rspbuf), w0->ls_cmd,
1650                                 FCNVME_RJT_RC_LOGIC, FCNVME_RJT_EXP_NONE, 0);
1651                 break;
1652         default:
1653                 lsop->lsrsp->rsplen = nvme_fc_format_rjt(lsop->rspbuf,
1654                                 sizeof(*lsop->rspbuf), w0->ls_cmd,
1655                                 FCNVME_RJT_RC_INVAL, FCNVME_RJT_EXP_NONE, 0);
1656                 break;
1657         }
1658
1659         return(ret);
1660 }
1661
1662 static void
1663 nvme_fc_handle_ls_rqst_work(struct work_struct *work)
1664 {
1665         struct nvme_fc_rport *rport =
1666                 container_of(work, struct nvme_fc_rport, lsrcv_work);
1667         struct fcnvme_ls_rqst_w0 *w0;
1668         struct nvmefc_ls_rcv_op *lsop;
1669         unsigned long flags;
1670         bool sendrsp;
1671
1672 restart:
1673         sendrsp = true;
1674         spin_lock_irqsave(&rport->lock, flags);
1675         list_for_each_entry(lsop, &rport->ls_rcv_list, lsrcv_list) {
1676                 if (lsop->handled)
1677                         continue;
1678
1679                 lsop->handled = true;
1680                 if (rport->remoteport.port_state == FC_OBJSTATE_ONLINE) {
1681                         spin_unlock_irqrestore(&rport->lock, flags);
1682                         sendrsp = nvme_fc_handle_ls_rqst(lsop);
1683                 } else {
1684                         spin_unlock_irqrestore(&rport->lock, flags);
1685                         w0 = &lsop->rqstbuf->w0;
1686                         lsop->lsrsp->rsplen = nvme_fc_format_rjt(
1687                                                 lsop->rspbuf,
1688                                                 sizeof(*lsop->rspbuf),
1689                                                 w0->ls_cmd,
1690                                                 FCNVME_RJT_RC_UNAB,
1691                                                 FCNVME_RJT_EXP_NONE, 0);
1692                 }
1693                 if (sendrsp)
1694                         nvme_fc_xmt_ls_rsp(lsop);
1695                 goto restart;
1696         }
1697         spin_unlock_irqrestore(&rport->lock, flags);
1698 }
1699
1700 /**
1701  * nvme_fc_rcv_ls_req - transport entry point called by an LLDD
1702  *                       upon the reception of a NVME LS request.
1703  *
1704  * The nvme-fc layer will copy payload to an internal structure for
1705  * processing.  As such, upon completion of the routine, the LLDD may
1706  * immediately free/reuse the LS request buffer passed in the call.
1707  *
1708  * If this routine returns error, the LLDD should abort the exchange.
1709  *
1710  * @remoteport: pointer to the (registered) remote port that the LS
1711  *              was received from. The remoteport is associated with
1712  *              a specific localport.
1713  * @lsrsp:      pointer to a nvmefc_ls_rsp response structure to be
1714  *              used to reference the exchange corresponding to the LS
1715  *              when issuing an ls response.
1716  * @lsreqbuf:   pointer to the buffer containing the LS Request
1717  * @lsreqbuf_len: length, in bytes, of the received LS request
1718  */
1719 int
1720 nvme_fc_rcv_ls_req(struct nvme_fc_remote_port *portptr,
1721                         struct nvmefc_ls_rsp *lsrsp,
1722                         void *lsreqbuf, u32 lsreqbuf_len)
1723 {
1724         struct nvme_fc_rport *rport = remoteport_to_rport(portptr);
1725         struct nvme_fc_lport *lport = rport->lport;
1726         struct fcnvme_ls_rqst_w0 *w0 = (struct fcnvme_ls_rqst_w0 *)lsreqbuf;
1727         struct nvmefc_ls_rcv_op *lsop;
1728         unsigned long flags;
1729         int ret;
1730
1731         nvme_fc_rport_get(rport);
1732
1733         /* validate there's a routine to transmit a response */
1734         if (!lport->ops->xmt_ls_rsp) {
1735                 dev_info(lport->dev,
1736                         "RCV %s LS failed: no LLDD xmt_ls_rsp\n",
1737                         (w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
1738                                 nvmefc_ls_names[w0->ls_cmd] : "");
1739                 ret = -EINVAL;
1740                 goto out_put;
1741         }
1742
1743         if (lsreqbuf_len > sizeof(union nvmefc_ls_requests)) {
1744                 dev_info(lport->dev,
1745                         "RCV %s LS failed: payload too large\n",
1746                         (w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
1747                                 nvmefc_ls_names[w0->ls_cmd] : "");
1748                 ret = -E2BIG;
1749                 goto out_put;
1750         }
1751
1752         lsop = kzalloc(sizeof(*lsop) +
1753                         sizeof(union nvmefc_ls_requests) +
1754                         sizeof(union nvmefc_ls_responses),
1755                         GFP_KERNEL);
1756         if (!lsop) {
1757                 dev_info(lport->dev,
1758                         "RCV %s LS failed: No memory\n",
1759                         (w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
1760                                 nvmefc_ls_names[w0->ls_cmd] : "");
1761                 ret = -ENOMEM;
1762                 goto out_put;
1763         }
1764         lsop->rqstbuf = (union nvmefc_ls_requests *)&lsop[1];
1765         lsop->rspbuf = (union nvmefc_ls_responses *)&lsop->rqstbuf[1];
1766
1767         lsop->rspdma = fc_dma_map_single(lport->dev, lsop->rspbuf,
1768                                         sizeof(*lsop->rspbuf),
1769                                         DMA_TO_DEVICE);
1770         if (fc_dma_mapping_error(lport->dev, lsop->rspdma)) {
1771                 dev_info(lport->dev,
1772                         "RCV %s LS failed: DMA mapping failure\n",
1773                         (w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
1774                                 nvmefc_ls_names[w0->ls_cmd] : "");
1775                 ret = -EFAULT;
1776                 goto out_free;
1777         }
1778
1779         lsop->rport = rport;
1780         lsop->lsrsp = lsrsp;
1781
1782         memcpy(lsop->rqstbuf, lsreqbuf, lsreqbuf_len);
1783         lsop->rqstdatalen = lsreqbuf_len;
1784
1785         spin_lock_irqsave(&rport->lock, flags);
1786         if (rport->remoteport.port_state != FC_OBJSTATE_ONLINE) {
1787                 spin_unlock_irqrestore(&rport->lock, flags);
1788                 ret = -ENOTCONN;
1789                 goto out_unmap;
1790         }
1791         list_add_tail(&lsop->lsrcv_list, &rport->ls_rcv_list);
1792         spin_unlock_irqrestore(&rport->lock, flags);
1793
1794         schedule_work(&rport->lsrcv_work);
1795
1796         return 0;
1797
1798 out_unmap:
1799         fc_dma_unmap_single(lport->dev, lsop->rspdma,
1800                         sizeof(*lsop->rspbuf), DMA_TO_DEVICE);
1801 out_free:
1802         kfree(lsop);
1803 out_put:
1804         nvme_fc_rport_put(rport);
1805         return ret;
1806 }
1807 EXPORT_SYMBOL_GPL(nvme_fc_rcv_ls_req);
1808
1809
1810 /* *********************** NVME Ctrl Routines **************************** */
1811
1812 static void
1813 __nvme_fc_exit_request(struct nvme_fc_ctrl *ctrl,
1814                 struct nvme_fc_fcp_op *op)
1815 {
1816         fc_dma_unmap_single(ctrl->lport->dev, op->fcp_req.rspdma,
1817                                 sizeof(op->rsp_iu), DMA_FROM_DEVICE);
1818         fc_dma_unmap_single(ctrl->lport->dev, op->fcp_req.cmddma,
1819                                 sizeof(op->cmd_iu), DMA_TO_DEVICE);
1820
1821         atomic_set(&op->state, FCPOP_STATE_UNINIT);
1822 }
1823
1824 static void
1825 nvme_fc_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1826                 unsigned int hctx_idx)
1827 {
1828         struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
1829
1830         return __nvme_fc_exit_request(set->driver_data, op);
1831 }
1832
1833 static int
1834 __nvme_fc_abort_op(struct nvme_fc_ctrl *ctrl, struct nvme_fc_fcp_op *op)
1835 {
1836         unsigned long flags;
1837         int opstate;
1838
1839         spin_lock_irqsave(&ctrl->lock, flags);
1840         opstate = atomic_xchg(&op->state, FCPOP_STATE_ABORTED);
1841         if (opstate != FCPOP_STATE_ACTIVE)
1842                 atomic_set(&op->state, opstate);
1843         else if (test_bit(FCCTRL_TERMIO, &ctrl->flags)) {
1844                 op->flags |= FCOP_FLAGS_TERMIO;
1845                 ctrl->iocnt++;
1846         }
1847         spin_unlock_irqrestore(&ctrl->lock, flags);
1848
1849         if (opstate != FCPOP_STATE_ACTIVE)
1850                 return -ECANCELED;
1851
1852         ctrl->lport->ops->fcp_abort(&ctrl->lport->localport,
1853                                         &ctrl->rport->remoteport,
1854                                         op->queue->lldd_handle,
1855                                         &op->fcp_req);
1856
1857         return 0;
1858 }
1859
1860 static void
1861 nvme_fc_abort_aen_ops(struct nvme_fc_ctrl *ctrl)
1862 {
1863         struct nvme_fc_fcp_op *aen_op = ctrl->aen_ops;
1864         int i;
1865
1866         /* ensure we've initialized the ops once */
1867         if (!(aen_op->flags & FCOP_FLAGS_AEN))
1868                 return;
1869
1870         for (i = 0; i < NVME_NR_AEN_COMMANDS; i++, aen_op++)
1871                 __nvme_fc_abort_op(ctrl, aen_op);
1872 }
1873
1874 static inline void
1875 __nvme_fc_fcpop_chk_teardowns(struct nvme_fc_ctrl *ctrl,
1876                 struct nvme_fc_fcp_op *op, int opstate)
1877 {
1878         unsigned long flags;
1879
1880         if (opstate == FCPOP_STATE_ABORTED) {
1881                 spin_lock_irqsave(&ctrl->lock, flags);
1882                 if (test_bit(FCCTRL_TERMIO, &ctrl->flags) &&
1883                     op->flags & FCOP_FLAGS_TERMIO) {
1884                         if (!--ctrl->iocnt)
1885                                 wake_up(&ctrl->ioabort_wait);
1886                 }
1887                 spin_unlock_irqrestore(&ctrl->lock, flags);
1888         }
1889 }
1890
1891 static void
1892 nvme_fc_fcpio_done(struct nvmefc_fcp_req *req)
1893 {
1894         struct nvme_fc_fcp_op *op = fcp_req_to_fcp_op(req);
1895         struct request *rq = op->rq;
1896         struct nvmefc_fcp_req *freq = &op->fcp_req;
1897         struct nvme_fc_ctrl *ctrl = op->ctrl;
1898         struct nvme_fc_queue *queue = op->queue;
1899         struct nvme_completion *cqe = &op->rsp_iu.cqe;
1900         struct nvme_command *sqe = &op->cmd_iu.sqe;
1901         __le16 status = cpu_to_le16(NVME_SC_SUCCESS << 1);
1902         union nvme_result result;
1903         bool terminate_assoc = true;
1904         int opstate;
1905
1906         /*
1907          * WARNING:
1908          * The current linux implementation of a nvme controller
1909          * allocates a single tag set for all io queues and sizes
1910          * the io queues to fully hold all possible tags. Thus, the
1911          * implementation does not reference or care about the sqhd
1912          * value as it never needs to use the sqhd/sqtail pointers
1913          * for submission pacing.
1914          *
1915          * This affects the FC-NVME implementation in two ways:
1916          * 1) As the value doesn't matter, we don't need to waste
1917          *    cycles extracting it from ERSPs and stamping it in the
1918          *    cases where the transport fabricates CQEs on successful
1919          *    completions.
1920          * 2) The FC-NVME implementation requires that delivery of
1921          *    ERSP completions are to go back to the nvme layer in order
1922          *    relative to the rsn, such that the sqhd value will always
1923          *    be "in order" for the nvme layer. As the nvme layer in
1924          *    linux doesn't care about sqhd, there's no need to return
1925          *    them in order.
1926          *
1927          * Additionally:
1928          * As the core nvme layer in linux currently does not look at
1929          * every field in the cqe - in cases where the FC transport must
1930          * fabricate a CQE, the following fields will not be set as they
1931          * are not referenced:
1932          *      cqe.sqid,  cqe.sqhd,  cqe.command_id
1933          *
1934          * Failure or error of an individual i/o, in a transport
1935          * detected fashion unrelated to the nvme completion status,
1936          * potentially cause the initiator and target sides to get out
1937          * of sync on SQ head/tail (aka outstanding io count allowed).
1938          * Per FC-NVME spec, failure of an individual command requires
1939          * the connection to be terminated, which in turn requires the
1940          * association to be terminated.
1941          */
1942
1943         opstate = atomic_xchg(&op->state, FCPOP_STATE_COMPLETE);
1944
1945         fc_dma_sync_single_for_cpu(ctrl->lport->dev, op->fcp_req.rspdma,
1946                                 sizeof(op->rsp_iu), DMA_FROM_DEVICE);
1947
1948         if (opstate == FCPOP_STATE_ABORTED)
1949                 status = cpu_to_le16(NVME_SC_HOST_PATH_ERROR << 1);
1950         else if (freq->status) {
1951                 status = cpu_to_le16(NVME_SC_HOST_PATH_ERROR << 1);
1952                 dev_info(ctrl->ctrl.device,
1953                         "NVME-FC{%d}: io failed due to lldd error %d\n",
1954                         ctrl->cnum, freq->status);
1955         }
1956
1957         /*
1958          * For the linux implementation, if we have an unsuccesful
1959          * status, they blk-mq layer can typically be called with the
1960          * non-zero status and the content of the cqe isn't important.
1961          */
1962         if (status)
1963                 goto done;
1964
1965         /*
1966          * command completed successfully relative to the wire
1967          * protocol. However, validate anything received and
1968          * extract the status and result from the cqe (create it
1969          * where necessary).
1970          */
1971
1972         switch (freq->rcv_rsplen) {
1973
1974         case 0:
1975         case NVME_FC_SIZEOF_ZEROS_RSP:
1976                 /*
1977                  * No response payload or 12 bytes of payload (which
1978                  * should all be zeros) are considered successful and
1979                  * no payload in the CQE by the transport.
1980                  */
1981                 if (freq->transferred_length !=
1982                     be32_to_cpu(op->cmd_iu.data_len)) {
1983                         status = cpu_to_le16(NVME_SC_HOST_PATH_ERROR << 1);
1984                         dev_info(ctrl->ctrl.device,
1985                                 "NVME-FC{%d}: io failed due to bad transfer "
1986                                 "length: %d vs expected %d\n",
1987                                 ctrl->cnum, freq->transferred_length,
1988                                 be32_to_cpu(op->cmd_iu.data_len));
1989                         goto done;
1990                 }
1991                 result.u64 = 0;
1992                 break;
1993
1994         case sizeof(struct nvme_fc_ersp_iu):
1995                 /*
1996                  * The ERSP IU contains a full completion with CQE.
1997                  * Validate ERSP IU and look at cqe.
1998                  */
1999                 if (unlikely(be16_to_cpu(op->rsp_iu.iu_len) !=
2000                                         (freq->rcv_rsplen / 4) ||
2001                              be32_to_cpu(op->rsp_iu.xfrd_len) !=
2002                                         freq->transferred_length ||
2003                              op->rsp_iu.ersp_result ||
2004                              sqe->common.command_id != cqe->command_id)) {
2005                         status = cpu_to_le16(NVME_SC_HOST_PATH_ERROR << 1);
2006                         dev_info(ctrl->ctrl.device,
2007                                 "NVME-FC{%d}: io failed due to bad NVMe_ERSP: "
2008                                 "iu len %d, xfr len %d vs %d, status code "
2009                                 "%d, cmdid %d vs %d\n",
2010                                 ctrl->cnum, be16_to_cpu(op->rsp_iu.iu_len),
2011                                 be32_to_cpu(op->rsp_iu.xfrd_len),
2012                                 freq->transferred_length,
2013                                 op->rsp_iu.ersp_result,
2014                                 sqe->common.command_id,
2015                                 cqe->command_id);
2016                         goto done;
2017                 }
2018                 result = cqe->result;
2019                 status = cqe->status;
2020                 break;
2021
2022         default:
2023                 status = cpu_to_le16(NVME_SC_HOST_PATH_ERROR << 1);
2024                 dev_info(ctrl->ctrl.device,
2025                         "NVME-FC{%d}: io failed due to odd NVMe_xRSP iu "
2026                         "len %d\n",
2027                         ctrl->cnum, freq->rcv_rsplen);
2028                 goto done;
2029         }
2030
2031         terminate_assoc = false;
2032
2033 done:
2034         if (op->flags & FCOP_FLAGS_AEN) {
2035                 nvme_complete_async_event(&queue->ctrl->ctrl, status, &result);
2036                 __nvme_fc_fcpop_chk_teardowns(ctrl, op, opstate);
2037                 atomic_set(&op->state, FCPOP_STATE_IDLE);
2038                 op->flags = FCOP_FLAGS_AEN;     /* clear other flags */
2039                 nvme_fc_ctrl_put(ctrl);
2040                 goto check_error;
2041         }
2042
2043         __nvme_fc_fcpop_chk_teardowns(ctrl, op, opstate);
2044         if (!nvme_try_complete_req(rq, status, result))
2045                 nvme_fc_complete_rq(rq);
2046
2047 check_error:
2048         if (terminate_assoc)
2049                 nvme_fc_error_recovery(ctrl, "transport detected io error");
2050 }
2051
2052 static int
2053 __nvme_fc_init_request(struct nvme_fc_ctrl *ctrl,
2054                 struct nvme_fc_queue *queue, struct nvme_fc_fcp_op *op,
2055                 struct request *rq, u32 rqno)
2056 {
2057         struct nvme_fcp_op_w_sgl *op_w_sgl =
2058                 container_of(op, typeof(*op_w_sgl), op);
2059         struct nvme_fc_cmd_iu *cmdiu = &op->cmd_iu;
2060         int ret = 0;
2061
2062         memset(op, 0, sizeof(*op));
2063         op->fcp_req.cmdaddr = &op->cmd_iu;
2064         op->fcp_req.cmdlen = sizeof(op->cmd_iu);
2065         op->fcp_req.rspaddr = &op->rsp_iu;
2066         op->fcp_req.rsplen = sizeof(op->rsp_iu);
2067         op->fcp_req.done = nvme_fc_fcpio_done;
2068         op->ctrl = ctrl;
2069         op->queue = queue;
2070         op->rq = rq;
2071         op->rqno = rqno;
2072
2073         cmdiu->format_id = NVME_CMD_FORMAT_ID;
2074         cmdiu->fc_id = NVME_CMD_FC_ID;
2075         cmdiu->iu_len = cpu_to_be16(sizeof(*cmdiu) / sizeof(u32));
2076         if (queue->qnum)
2077                 cmdiu->rsv_cat = fccmnd_set_cat_css(0,
2078                                         (NVME_CC_CSS_NVM >> NVME_CC_CSS_SHIFT));
2079         else
2080                 cmdiu->rsv_cat = fccmnd_set_cat_admin(0);
2081
2082         op->fcp_req.cmddma = fc_dma_map_single(ctrl->lport->dev,
2083                                 &op->cmd_iu, sizeof(op->cmd_iu), DMA_TO_DEVICE);
2084         if (fc_dma_mapping_error(ctrl->lport->dev, op->fcp_req.cmddma)) {
2085                 dev_err(ctrl->dev,
2086                         "FCP Op failed - cmdiu dma mapping failed.\n");
2087                 ret = -EFAULT;
2088                 goto out_on_error;
2089         }
2090
2091         op->fcp_req.rspdma = fc_dma_map_single(ctrl->lport->dev,
2092                                 &op->rsp_iu, sizeof(op->rsp_iu),
2093                                 DMA_FROM_DEVICE);
2094         if (fc_dma_mapping_error(ctrl->lport->dev, op->fcp_req.rspdma)) {
2095                 dev_err(ctrl->dev,
2096                         "FCP Op failed - rspiu dma mapping failed.\n");
2097                 ret = -EFAULT;
2098         }
2099
2100         atomic_set(&op->state, FCPOP_STATE_IDLE);
2101 out_on_error:
2102         return ret;
2103 }
2104
2105 static int
2106 nvme_fc_init_request(struct blk_mq_tag_set *set, struct request *rq,
2107                 unsigned int hctx_idx, unsigned int numa_node)
2108 {
2109         struct nvme_fc_ctrl *ctrl = set->driver_data;
2110         struct nvme_fcp_op_w_sgl *op = blk_mq_rq_to_pdu(rq);
2111         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
2112         struct nvme_fc_queue *queue = &ctrl->queues[queue_idx];
2113         int res;
2114
2115         res = __nvme_fc_init_request(ctrl, queue, &op->op, rq, queue->rqcnt++);
2116         if (res)
2117                 return res;
2118         op->op.fcp_req.first_sgl = op->sgl;
2119         op->op.fcp_req.private = &op->priv[0];
2120         nvme_req(rq)->ctrl = &ctrl->ctrl;
2121         return res;
2122 }
2123
2124 static int
2125 nvme_fc_init_aen_ops(struct nvme_fc_ctrl *ctrl)
2126 {
2127         struct nvme_fc_fcp_op *aen_op;
2128         struct nvme_fc_cmd_iu *cmdiu;
2129         struct nvme_command *sqe;
2130         void *private = NULL;
2131         int i, ret;
2132
2133         aen_op = ctrl->aen_ops;
2134         for (i = 0; i < NVME_NR_AEN_COMMANDS; i++, aen_op++) {
2135                 if (ctrl->lport->ops->fcprqst_priv_sz) {
2136                         private = kzalloc(ctrl->lport->ops->fcprqst_priv_sz,
2137                                                 GFP_KERNEL);
2138                         if (!private)
2139                                 return -ENOMEM;
2140                 }
2141
2142                 cmdiu = &aen_op->cmd_iu;
2143                 sqe = &cmdiu->sqe;
2144                 ret = __nvme_fc_init_request(ctrl, &ctrl->queues[0],
2145                                 aen_op, (struct request *)NULL,
2146                                 (NVME_AQ_BLK_MQ_DEPTH + i));
2147                 if (ret) {
2148                         kfree(private);
2149                         return ret;
2150                 }
2151
2152                 aen_op->flags = FCOP_FLAGS_AEN;
2153                 aen_op->fcp_req.private = private;
2154
2155                 memset(sqe, 0, sizeof(*sqe));
2156                 sqe->common.opcode = nvme_admin_async_event;
2157                 /* Note: core layer may overwrite the sqe.command_id value */
2158                 sqe->common.command_id = NVME_AQ_BLK_MQ_DEPTH + i;
2159         }
2160         return 0;
2161 }
2162
2163 static void
2164 nvme_fc_term_aen_ops(struct nvme_fc_ctrl *ctrl)
2165 {
2166         struct nvme_fc_fcp_op *aen_op;
2167         int i;
2168
2169         cancel_work_sync(&ctrl->ctrl.async_event_work);
2170         aen_op = ctrl->aen_ops;
2171         for (i = 0; i < NVME_NR_AEN_COMMANDS; i++, aen_op++) {
2172                 __nvme_fc_exit_request(ctrl, aen_op);
2173
2174                 kfree(aen_op->fcp_req.private);
2175                 aen_op->fcp_req.private = NULL;
2176         }
2177 }
2178
2179 static inline void
2180 __nvme_fc_init_hctx(struct blk_mq_hw_ctx *hctx, struct nvme_fc_ctrl *ctrl,
2181                 unsigned int qidx)
2182 {
2183         struct nvme_fc_queue *queue = &ctrl->queues[qidx];
2184
2185         hctx->driver_data = queue;
2186         queue->hctx = hctx;
2187 }
2188
2189 static int
2190 nvme_fc_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
2191                 unsigned int hctx_idx)
2192 {
2193         struct nvme_fc_ctrl *ctrl = data;
2194
2195         __nvme_fc_init_hctx(hctx, ctrl, hctx_idx + 1);
2196
2197         return 0;
2198 }
2199
2200 static int
2201 nvme_fc_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
2202                 unsigned int hctx_idx)
2203 {
2204         struct nvme_fc_ctrl *ctrl = data;
2205
2206         __nvme_fc_init_hctx(hctx, ctrl, hctx_idx);
2207
2208         return 0;
2209 }
2210
2211 static void
2212 nvme_fc_init_queue(struct nvme_fc_ctrl *ctrl, int idx)
2213 {
2214         struct nvme_fc_queue *queue;
2215
2216         queue = &ctrl->queues[idx];
2217         memset(queue, 0, sizeof(*queue));
2218         queue->ctrl = ctrl;
2219         queue->qnum = idx;
2220         atomic_set(&queue->csn, 0);
2221         queue->dev = ctrl->dev;
2222
2223         if (idx > 0)
2224                 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
2225         else
2226                 queue->cmnd_capsule_len = sizeof(struct nvme_command);
2227
2228         /*
2229          * Considered whether we should allocate buffers for all SQEs
2230          * and CQEs and dma map them - mapping their respective entries
2231          * into the request structures (kernel vm addr and dma address)
2232          * thus the driver could use the buffers/mappings directly.
2233          * It only makes sense if the LLDD would use them for its
2234          * messaging api. It's very unlikely most adapter api's would use
2235          * a native NVME sqe/cqe. More reasonable if FC-NVME IU payload
2236          * structures were used instead.
2237          */
2238 }
2239
2240 /*
2241  * This routine terminates a queue at the transport level.
2242  * The transport has already ensured that all outstanding ios on
2243  * the queue have been terminated.
2244  * The transport will send a Disconnect LS request to terminate
2245  * the queue's connection. Termination of the admin queue will also
2246  * terminate the association at the target.
2247  */
2248 static void
2249 nvme_fc_free_queue(struct nvme_fc_queue *queue)
2250 {
2251         if (!test_and_clear_bit(NVME_FC_Q_CONNECTED, &queue->flags))
2252                 return;
2253
2254         clear_bit(NVME_FC_Q_LIVE, &queue->flags);
2255         /*
2256          * Current implementation never disconnects a single queue.
2257          * It always terminates a whole association. So there is never
2258          * a disconnect(queue) LS sent to the target.
2259          */
2260
2261         queue->connection_id = 0;
2262         atomic_set(&queue->csn, 0);
2263 }
2264
2265 static void
2266 __nvme_fc_delete_hw_queue(struct nvme_fc_ctrl *ctrl,
2267         struct nvme_fc_queue *queue, unsigned int qidx)
2268 {
2269         if (ctrl->lport->ops->delete_queue)
2270                 ctrl->lport->ops->delete_queue(&ctrl->lport->localport, qidx,
2271                                 queue->lldd_handle);
2272         queue->lldd_handle = NULL;
2273 }
2274
2275 static void
2276 nvme_fc_free_io_queues(struct nvme_fc_ctrl *ctrl)
2277 {
2278         int i;
2279
2280         for (i = 1; i < ctrl->ctrl.queue_count; i++)
2281                 nvme_fc_free_queue(&ctrl->queues[i]);
2282 }
2283
2284 static int
2285 __nvme_fc_create_hw_queue(struct nvme_fc_ctrl *ctrl,
2286         struct nvme_fc_queue *queue, unsigned int qidx, u16 qsize)
2287 {
2288         int ret = 0;
2289
2290         queue->lldd_handle = NULL;
2291         if (ctrl->lport->ops->create_queue)
2292                 ret = ctrl->lport->ops->create_queue(&ctrl->lport->localport,
2293                                 qidx, qsize, &queue->lldd_handle);
2294
2295         return ret;
2296 }
2297
2298 static void
2299 nvme_fc_delete_hw_io_queues(struct nvme_fc_ctrl *ctrl)
2300 {
2301         struct nvme_fc_queue *queue = &ctrl->queues[ctrl->ctrl.queue_count - 1];
2302         int i;
2303
2304         for (i = ctrl->ctrl.queue_count - 1; i >= 1; i--, queue--)
2305                 __nvme_fc_delete_hw_queue(ctrl, queue, i);
2306 }
2307
2308 static int
2309 nvme_fc_create_hw_io_queues(struct nvme_fc_ctrl *ctrl, u16 qsize)
2310 {
2311         struct nvme_fc_queue *queue = &ctrl->queues[1];
2312         int i, ret;
2313
2314         for (i = 1; i < ctrl->ctrl.queue_count; i++, queue++) {
2315                 ret = __nvme_fc_create_hw_queue(ctrl, queue, i, qsize);
2316                 if (ret)
2317                         goto delete_queues;
2318         }
2319
2320         return 0;
2321
2322 delete_queues:
2323         for (; i > 0; i--)
2324                 __nvme_fc_delete_hw_queue(ctrl, &ctrl->queues[i], i);
2325         return ret;
2326 }
2327
2328 static int
2329 nvme_fc_connect_io_queues(struct nvme_fc_ctrl *ctrl, u16 qsize)
2330 {
2331         int i, ret = 0;
2332
2333         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
2334                 ret = nvme_fc_connect_queue(ctrl, &ctrl->queues[i], qsize,
2335                                         (qsize / 5));
2336                 if (ret)
2337                         break;
2338                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i, false);
2339                 if (ret)
2340                         break;
2341
2342                 set_bit(NVME_FC_Q_LIVE, &ctrl->queues[i].flags);
2343         }
2344
2345         return ret;
2346 }
2347
2348 static void
2349 nvme_fc_init_io_queues(struct nvme_fc_ctrl *ctrl)
2350 {
2351         int i;
2352
2353         for (i = 1; i < ctrl->ctrl.queue_count; i++)
2354                 nvme_fc_init_queue(ctrl, i);
2355 }
2356
2357 static void
2358 nvme_fc_ctrl_free(struct kref *ref)
2359 {
2360         struct nvme_fc_ctrl *ctrl =
2361                 container_of(ref, struct nvme_fc_ctrl, ref);
2362         unsigned long flags;
2363
2364         if (ctrl->ctrl.tagset) {
2365                 blk_cleanup_queue(ctrl->ctrl.connect_q);
2366                 blk_mq_free_tag_set(&ctrl->tag_set);
2367         }
2368
2369         /* remove from rport list */
2370         spin_lock_irqsave(&ctrl->rport->lock, flags);
2371         list_del(&ctrl->ctrl_list);
2372         spin_unlock_irqrestore(&ctrl->rport->lock, flags);
2373
2374         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
2375         blk_cleanup_queue(ctrl->ctrl.admin_q);
2376         blk_cleanup_queue(ctrl->ctrl.fabrics_q);
2377         blk_mq_free_tag_set(&ctrl->admin_tag_set);
2378
2379         kfree(ctrl->queues);
2380
2381         put_device(ctrl->dev);
2382         nvme_fc_rport_put(ctrl->rport);
2383
2384         ida_simple_remove(&nvme_fc_ctrl_cnt, ctrl->cnum);
2385         if (ctrl->ctrl.opts)
2386                 nvmf_free_options(ctrl->ctrl.opts);
2387         kfree(ctrl);
2388 }
2389
2390 static void
2391 nvme_fc_ctrl_put(struct nvme_fc_ctrl *ctrl)
2392 {
2393         kref_put(&ctrl->ref, nvme_fc_ctrl_free);
2394 }
2395
2396 static int
2397 nvme_fc_ctrl_get(struct nvme_fc_ctrl *ctrl)
2398 {
2399         return kref_get_unless_zero(&ctrl->ref);
2400 }
2401
2402 /*
2403  * All accesses from nvme core layer done - can now free the
2404  * controller. Called after last nvme_put_ctrl() call
2405  */
2406 static void
2407 nvme_fc_nvme_ctrl_freed(struct nvme_ctrl *nctrl)
2408 {
2409         struct nvme_fc_ctrl *ctrl = to_fc_ctrl(nctrl);
2410
2411         WARN_ON(nctrl != &ctrl->ctrl);
2412
2413         nvme_fc_ctrl_put(ctrl);
2414 }
2415
2416 /*
2417  * This routine is used by the transport when it needs to find active
2418  * io on a queue that is to be terminated. The transport uses
2419  * blk_mq_tagset_busy_itr() to find the busy requests, which then invoke
2420  * this routine to kill them on a 1 by 1 basis.
2421  *
2422  * As FC allocates FC exchange for each io, the transport must contact
2423  * the LLDD to terminate the exchange, thus releasing the FC exchange.
2424  * After terminating the exchange the LLDD will call the transport's
2425  * normal io done path for the request, but it will have an aborted
2426  * status. The done path will return the io request back to the block
2427  * layer with an error status.
2428  */
2429 static bool
2430 nvme_fc_terminate_exchange(struct request *req, void *data, bool reserved)
2431 {
2432         struct nvme_ctrl *nctrl = data;
2433         struct nvme_fc_ctrl *ctrl = to_fc_ctrl(nctrl);
2434         struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(req);
2435
2436         __nvme_fc_abort_op(ctrl, op);
2437         return true;
2438 }
2439
2440 /*
2441  * This routine runs through all outstanding commands on the association
2442  * and aborts them.  This routine is typically be called by the
2443  * delete_association routine. It is also called due to an error during
2444  * reconnect. In that scenario, it is most likely a command that initializes
2445  * the controller, including fabric Connect commands on io queues, that
2446  * may have timed out or failed thus the io must be killed for the connect
2447  * thread to see the error.
2448  */
2449 static void
2450 __nvme_fc_abort_outstanding_ios(struct nvme_fc_ctrl *ctrl, bool start_queues)
2451 {
2452         /*
2453          * If io queues are present, stop them and terminate all outstanding
2454          * ios on them. As FC allocates FC exchange for each io, the
2455          * transport must contact the LLDD to terminate the exchange,
2456          * thus releasing the FC exchange. We use blk_mq_tagset_busy_itr()
2457          * to tell us what io's are busy and invoke a transport routine
2458          * to kill them with the LLDD.  After terminating the exchange
2459          * the LLDD will call the transport's normal io done path, but it
2460          * will have an aborted status. The done path will return the
2461          * io requests back to the block layer as part of normal completions
2462          * (but with error status).
2463          */
2464         if (ctrl->ctrl.queue_count > 1) {
2465                 nvme_stop_queues(&ctrl->ctrl);
2466                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
2467                                 nvme_fc_terminate_exchange, &ctrl->ctrl);
2468                 blk_mq_tagset_wait_completed_request(&ctrl->tag_set);
2469                 if (start_queues)
2470                         nvme_start_queues(&ctrl->ctrl);
2471         }
2472
2473         /*
2474          * Other transports, which don't have link-level contexts bound
2475          * to sqe's, would try to gracefully shutdown the controller by
2476          * writing the registers for shutdown and polling (call
2477          * nvme_shutdown_ctrl()). Given a bunch of i/o was potentially
2478          * just aborted and we will wait on those contexts, and given
2479          * there was no indication of how live the controlelr is on the
2480          * link, don't send more io to create more contexts for the
2481          * shutdown. Let the controller fail via keepalive failure if
2482          * its still present.
2483          */
2484
2485         /*
2486          * clean up the admin queue. Same thing as above.
2487          */
2488         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
2489         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
2490                                 nvme_fc_terminate_exchange, &ctrl->ctrl);
2491         blk_mq_tagset_wait_completed_request(&ctrl->admin_tag_set);
2492 }
2493
2494 static void
2495 nvme_fc_error_recovery(struct nvme_fc_ctrl *ctrl, char *errmsg)
2496 {
2497         /*
2498          * if an error (io timeout, etc) while (re)connecting, the remote
2499          * port requested terminating of the association (disconnect_ls)
2500          * or an error (timeout or abort) occurred on an io while creating
2501          * the controller.  Abort any ios on the association and let the
2502          * create_association error path resolve things.
2503          */
2504         if (ctrl->ctrl.state == NVME_CTRL_CONNECTING) {
2505                 __nvme_fc_abort_outstanding_ios(ctrl, true);
2506                 set_bit(ASSOC_FAILED, &ctrl->flags);
2507                 return;
2508         }
2509
2510         /* Otherwise, only proceed if in LIVE state - e.g. on first error */
2511         if (ctrl->ctrl.state != NVME_CTRL_LIVE)
2512                 return;
2513
2514         dev_warn(ctrl->ctrl.device,
2515                 "NVME-FC{%d}: transport association event: %s\n",
2516                 ctrl->cnum, errmsg);
2517         dev_warn(ctrl->ctrl.device,
2518                 "NVME-FC{%d}: resetting controller\n", ctrl->cnum);
2519
2520         nvme_reset_ctrl(&ctrl->ctrl);
2521 }
2522
2523 static enum blk_eh_timer_return
2524 nvme_fc_timeout(struct request *rq, bool reserved)
2525 {
2526         struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
2527         struct nvme_fc_ctrl *ctrl = op->ctrl;
2528         struct nvme_fc_cmd_iu *cmdiu = &op->cmd_iu;
2529         struct nvme_command *sqe = &cmdiu->sqe;
2530
2531         /*
2532          * Attempt to abort the offending command. Command completion
2533          * will detect the aborted io and will fail the connection.
2534          */
2535         dev_info(ctrl->ctrl.device,
2536                 "NVME-FC{%d.%d}: io timeout: opcode %d fctype %d w10/11: "
2537                 "x%08x/x%08x\n",
2538                 ctrl->cnum, op->queue->qnum, sqe->common.opcode,
2539                 sqe->connect.fctype, sqe->common.cdw10, sqe->common.cdw11);
2540         if (__nvme_fc_abort_op(ctrl, op))
2541                 nvme_fc_error_recovery(ctrl, "io timeout abort failed");
2542
2543         /*
2544          * the io abort has been initiated. Have the reset timer
2545          * restarted and the abort completion will complete the io
2546          * shortly. Avoids a synchronous wait while the abort finishes.
2547          */
2548         return BLK_EH_RESET_TIMER;
2549 }
2550
2551 static int
2552 nvme_fc_map_data(struct nvme_fc_ctrl *ctrl, struct request *rq,
2553                 struct nvme_fc_fcp_op *op)
2554 {
2555         struct nvmefc_fcp_req *freq = &op->fcp_req;
2556         int ret;
2557
2558         freq->sg_cnt = 0;
2559
2560         if (!blk_rq_nr_phys_segments(rq))
2561                 return 0;
2562
2563         freq->sg_table.sgl = freq->first_sgl;
2564         ret = sg_alloc_table_chained(&freq->sg_table,
2565                         blk_rq_nr_phys_segments(rq), freq->sg_table.sgl,
2566                         NVME_INLINE_SG_CNT);
2567         if (ret)
2568                 return -ENOMEM;
2569
2570         op->nents = blk_rq_map_sg(rq->q, rq, freq->sg_table.sgl);
2571         WARN_ON(op->nents > blk_rq_nr_phys_segments(rq));
2572         freq->sg_cnt = fc_dma_map_sg(ctrl->lport->dev, freq->sg_table.sgl,
2573                                 op->nents, rq_dma_dir(rq));
2574         if (unlikely(freq->sg_cnt <= 0)) {
2575                 sg_free_table_chained(&freq->sg_table, NVME_INLINE_SG_CNT);
2576                 freq->sg_cnt = 0;
2577                 return -EFAULT;
2578         }
2579
2580         /*
2581          * TODO: blk_integrity_rq(rq)  for DIF
2582          */
2583         return 0;
2584 }
2585
2586 static void
2587 nvme_fc_unmap_data(struct nvme_fc_ctrl *ctrl, struct request *rq,
2588                 struct nvme_fc_fcp_op *op)
2589 {
2590         struct nvmefc_fcp_req *freq = &op->fcp_req;
2591
2592         if (!freq->sg_cnt)
2593                 return;
2594
2595         fc_dma_unmap_sg(ctrl->lport->dev, freq->sg_table.sgl, op->nents,
2596                         rq_dma_dir(rq));
2597
2598         sg_free_table_chained(&freq->sg_table, NVME_INLINE_SG_CNT);
2599
2600         freq->sg_cnt = 0;
2601 }
2602
2603 /*
2604  * In FC, the queue is a logical thing. At transport connect, the target
2605  * creates its "queue" and returns a handle that is to be given to the
2606  * target whenever it posts something to the corresponding SQ.  When an
2607  * SQE is sent on a SQ, FC effectively considers the SQE, or rather the
2608  * command contained within the SQE, an io, and assigns a FC exchange
2609  * to it. The SQE and the associated SQ handle are sent in the initial
2610  * CMD IU sents on the exchange. All transfers relative to the io occur
2611  * as part of the exchange.  The CQE is the last thing for the io,
2612  * which is transferred (explicitly or implicitly) with the RSP IU
2613  * sent on the exchange. After the CQE is received, the FC exchange is
2614  * terminaed and the Exchange may be used on a different io.
2615  *
2616  * The transport to LLDD api has the transport making a request for a
2617  * new fcp io request to the LLDD. The LLDD then allocates a FC exchange
2618  * resource and transfers the command. The LLDD will then process all
2619  * steps to complete the io. Upon completion, the transport done routine
2620  * is called.
2621  *
2622  * So - while the operation is outstanding to the LLDD, there is a link
2623  * level FC exchange resource that is also outstanding. This must be
2624  * considered in all cleanup operations.
2625  */
2626 static blk_status_t
2627 nvme_fc_start_fcp_op(struct nvme_fc_ctrl *ctrl, struct nvme_fc_queue *queue,
2628         struct nvme_fc_fcp_op *op, u32 data_len,
2629         enum nvmefc_fcp_datadir io_dir)
2630 {
2631         struct nvme_fc_cmd_iu *cmdiu = &op->cmd_iu;
2632         struct nvme_command *sqe = &cmdiu->sqe;
2633         int ret, opstate;
2634
2635         /*
2636          * before attempting to send the io, check to see if we believe
2637          * the target device is present
2638          */
2639         if (ctrl->rport->remoteport.port_state != FC_OBJSTATE_ONLINE)
2640                 return BLK_STS_RESOURCE;
2641
2642         if (!nvme_fc_ctrl_get(ctrl))
2643                 return BLK_STS_IOERR;
2644
2645         /* format the FC-NVME CMD IU and fcp_req */
2646         cmdiu->connection_id = cpu_to_be64(queue->connection_id);
2647         cmdiu->data_len = cpu_to_be32(data_len);
2648         switch (io_dir) {
2649         case NVMEFC_FCP_WRITE:
2650                 cmdiu->flags = FCNVME_CMD_FLAGS_WRITE;
2651                 break;
2652         case NVMEFC_FCP_READ:
2653                 cmdiu->flags = FCNVME_CMD_FLAGS_READ;
2654                 break;
2655         case NVMEFC_FCP_NODATA:
2656                 cmdiu->flags = 0;
2657                 break;
2658         }
2659         op->fcp_req.payload_length = data_len;
2660         op->fcp_req.io_dir = io_dir;
2661         op->fcp_req.transferred_length = 0;
2662         op->fcp_req.rcv_rsplen = 0;
2663         op->fcp_req.status = NVME_SC_SUCCESS;
2664         op->fcp_req.sqid = cpu_to_le16(queue->qnum);
2665
2666         /*
2667          * validate per fabric rules, set fields mandated by fabric spec
2668          * as well as those by FC-NVME spec.
2669          */
2670         WARN_ON_ONCE(sqe->common.metadata);
2671         sqe->common.flags |= NVME_CMD_SGL_METABUF;
2672
2673         /*
2674          * format SQE DPTR field per FC-NVME rules:
2675          *    type=0x5     Transport SGL Data Block Descriptor
2676          *    subtype=0xA  Transport-specific value
2677          *    address=0
2678          *    length=length of the data series
2679          */
2680         sqe->rw.dptr.sgl.type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2681                                         NVME_SGL_FMT_TRANSPORT_A;
2682         sqe->rw.dptr.sgl.length = cpu_to_le32(data_len);
2683         sqe->rw.dptr.sgl.addr = 0;
2684
2685         if (!(op->flags & FCOP_FLAGS_AEN)) {
2686                 ret = nvme_fc_map_data(ctrl, op->rq, op);
2687                 if (ret < 0) {
2688                         nvme_cleanup_cmd(op->rq);
2689                         nvme_fc_ctrl_put(ctrl);
2690                         if (ret == -ENOMEM || ret == -EAGAIN)
2691                                 return BLK_STS_RESOURCE;
2692                         return BLK_STS_IOERR;
2693                 }
2694         }
2695
2696         fc_dma_sync_single_for_device(ctrl->lport->dev, op->fcp_req.cmddma,
2697                                   sizeof(op->cmd_iu), DMA_TO_DEVICE);
2698
2699         atomic_set(&op->state, FCPOP_STATE_ACTIVE);
2700
2701         if (!(op->flags & FCOP_FLAGS_AEN))
2702                 blk_mq_start_request(op->rq);
2703
2704         cmdiu->csn = cpu_to_be32(atomic_inc_return(&queue->csn));
2705         ret = ctrl->lport->ops->fcp_io(&ctrl->lport->localport,
2706                                         &ctrl->rport->remoteport,
2707                                         queue->lldd_handle, &op->fcp_req);
2708
2709         if (ret) {
2710                 /*
2711                  * If the lld fails to send the command is there an issue with
2712                  * the csn value?  If the command that fails is the Connect,
2713                  * no - as the connection won't be live.  If it is a command
2714                  * post-connect, it's possible a gap in csn may be created.
2715                  * Does this matter?  As Linux initiators don't send fused
2716                  * commands, no.  The gap would exist, but as there's nothing
2717                  * that depends on csn order to be delivered on the target
2718                  * side, it shouldn't hurt.  It would be difficult for a
2719                  * target to even detect the csn gap as it has no idea when the
2720                  * cmd with the csn was supposed to arrive.
2721                  */
2722                 opstate = atomic_xchg(&op->state, FCPOP_STATE_COMPLETE);
2723                 __nvme_fc_fcpop_chk_teardowns(ctrl, op, opstate);
2724
2725                 if (!(op->flags & FCOP_FLAGS_AEN)) {
2726                         nvme_fc_unmap_data(ctrl, op->rq, op);
2727                         nvme_cleanup_cmd(op->rq);
2728                 }
2729
2730                 nvme_fc_ctrl_put(ctrl);
2731
2732                 if (ctrl->rport->remoteport.port_state == FC_OBJSTATE_ONLINE &&
2733                                 ret != -EBUSY)
2734                         return BLK_STS_IOERR;
2735
2736                 return BLK_STS_RESOURCE;
2737         }
2738
2739         return BLK_STS_OK;
2740 }
2741
2742 static blk_status_t
2743 nvme_fc_queue_rq(struct blk_mq_hw_ctx *hctx,
2744                         const struct blk_mq_queue_data *bd)
2745 {
2746         struct nvme_ns *ns = hctx->queue->queuedata;
2747         struct nvme_fc_queue *queue = hctx->driver_data;
2748         struct nvme_fc_ctrl *ctrl = queue->ctrl;
2749         struct request *rq = bd->rq;
2750         struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
2751         struct nvme_fc_cmd_iu *cmdiu = &op->cmd_iu;
2752         struct nvme_command *sqe = &cmdiu->sqe;
2753         enum nvmefc_fcp_datadir io_dir;
2754         bool queue_ready = test_bit(NVME_FC_Q_LIVE, &queue->flags);
2755         u32 data_len;
2756         blk_status_t ret;
2757
2758         if (ctrl->rport->remoteport.port_state != FC_OBJSTATE_ONLINE ||
2759             !nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2760                 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
2761
2762         ret = nvme_setup_cmd(ns, rq, sqe);
2763         if (ret)
2764                 return ret;
2765
2766         /*
2767          * nvme core doesn't quite treat the rq opaquely. Commands such
2768          * as WRITE ZEROES will return a non-zero rq payload_bytes yet
2769          * there is no actual payload to be transferred.
2770          * To get it right, key data transmission on there being 1 or
2771          * more physical segments in the sg list. If there is no
2772          * physical segments, there is no payload.
2773          */
2774         if (blk_rq_nr_phys_segments(rq)) {
2775                 data_len = blk_rq_payload_bytes(rq);
2776                 io_dir = ((rq_data_dir(rq) == WRITE) ?
2777                                         NVMEFC_FCP_WRITE : NVMEFC_FCP_READ);
2778         } else {
2779                 data_len = 0;
2780                 io_dir = NVMEFC_FCP_NODATA;
2781         }
2782
2783
2784         return nvme_fc_start_fcp_op(ctrl, queue, op, data_len, io_dir);
2785 }
2786
2787 static void
2788 nvme_fc_submit_async_event(struct nvme_ctrl *arg)
2789 {
2790         struct nvme_fc_ctrl *ctrl = to_fc_ctrl(arg);
2791         struct nvme_fc_fcp_op *aen_op;
2792         blk_status_t ret;
2793
2794         if (test_bit(FCCTRL_TERMIO, &ctrl->flags))
2795                 return;
2796
2797         aen_op = &ctrl->aen_ops[0];
2798
2799         ret = nvme_fc_start_fcp_op(ctrl, aen_op->queue, aen_op, 0,
2800                                         NVMEFC_FCP_NODATA);
2801         if (ret)
2802                 dev_err(ctrl->ctrl.device,
2803                         "failed async event work\n");
2804 }
2805
2806 static void
2807 nvme_fc_complete_rq(struct request *rq)
2808 {
2809         struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
2810         struct nvme_fc_ctrl *ctrl = op->ctrl;
2811
2812         atomic_set(&op->state, FCPOP_STATE_IDLE);
2813         op->flags &= ~FCOP_FLAGS_TERMIO;
2814
2815         nvme_fc_unmap_data(ctrl, rq, op);
2816         nvme_complete_rq(rq);
2817         nvme_fc_ctrl_put(ctrl);
2818 }
2819
2820
2821 static const struct blk_mq_ops nvme_fc_mq_ops = {
2822         .queue_rq       = nvme_fc_queue_rq,
2823         .complete       = nvme_fc_complete_rq,
2824         .init_request   = nvme_fc_init_request,
2825         .exit_request   = nvme_fc_exit_request,
2826         .init_hctx      = nvme_fc_init_hctx,
2827         .timeout        = nvme_fc_timeout,
2828 };
2829
2830 static int
2831 nvme_fc_create_io_queues(struct nvme_fc_ctrl *ctrl)
2832 {
2833         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2834         unsigned int nr_io_queues;
2835         int ret;
2836
2837         nr_io_queues = min(min(opts->nr_io_queues, num_online_cpus()),
2838                                 ctrl->lport->ops->max_hw_queues);
2839         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
2840         if (ret) {
2841                 dev_info(ctrl->ctrl.device,
2842                         "set_queue_count failed: %d\n", ret);
2843                 return ret;
2844         }
2845
2846         ctrl->ctrl.queue_count = nr_io_queues + 1;
2847         if (!nr_io_queues)
2848                 return 0;
2849
2850         nvme_fc_init_io_queues(ctrl);
2851
2852         memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
2853         ctrl->tag_set.ops = &nvme_fc_mq_ops;
2854         ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
2855         ctrl->tag_set.reserved_tags = 1; /* fabric connect */
2856         ctrl->tag_set.numa_node = ctrl->ctrl.numa_node;
2857         ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
2858         ctrl->tag_set.cmd_size =
2859                 struct_size((struct nvme_fcp_op_w_sgl *)NULL, priv,
2860                             ctrl->lport->ops->fcprqst_priv_sz);
2861         ctrl->tag_set.driver_data = ctrl;
2862         ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
2863         ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
2864
2865         ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
2866         if (ret)
2867                 return ret;
2868
2869         ctrl->ctrl.tagset = &ctrl->tag_set;
2870
2871         ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
2872         if (IS_ERR(ctrl->ctrl.connect_q)) {
2873                 ret = PTR_ERR(ctrl->ctrl.connect_q);
2874                 goto out_free_tag_set;
2875         }
2876
2877         ret = nvme_fc_create_hw_io_queues(ctrl, ctrl->ctrl.sqsize + 1);
2878         if (ret)
2879                 goto out_cleanup_blk_queue;
2880
2881         ret = nvme_fc_connect_io_queues(ctrl, ctrl->ctrl.sqsize + 1);
2882         if (ret)
2883                 goto out_delete_hw_queues;
2884
2885         ctrl->ioq_live = true;
2886
2887         return 0;
2888
2889 out_delete_hw_queues:
2890         nvme_fc_delete_hw_io_queues(ctrl);
2891 out_cleanup_blk_queue:
2892         blk_cleanup_queue(ctrl->ctrl.connect_q);
2893 out_free_tag_set:
2894         blk_mq_free_tag_set(&ctrl->tag_set);
2895         nvme_fc_free_io_queues(ctrl);
2896
2897         /* force put free routine to ignore io queues */
2898         ctrl->ctrl.tagset = NULL;
2899
2900         return ret;
2901 }
2902
2903 static int
2904 nvme_fc_recreate_io_queues(struct nvme_fc_ctrl *ctrl)
2905 {
2906         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2907         u32 prior_ioq_cnt = ctrl->ctrl.queue_count - 1;
2908         unsigned int nr_io_queues;
2909         int ret;
2910
2911         nr_io_queues = min(min(opts->nr_io_queues, num_online_cpus()),
2912                                 ctrl->lport->ops->max_hw_queues);
2913         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
2914         if (ret) {
2915                 dev_info(ctrl->ctrl.device,
2916                         "set_queue_count failed: %d\n", ret);
2917                 return ret;
2918         }
2919
2920         if (!nr_io_queues && prior_ioq_cnt) {
2921                 dev_info(ctrl->ctrl.device,
2922                         "Fail Reconnect: At least 1 io queue "
2923                         "required (was %d)\n", prior_ioq_cnt);
2924                 return -ENOSPC;
2925         }
2926
2927         ctrl->ctrl.queue_count = nr_io_queues + 1;
2928         /* check for io queues existing */
2929         if (ctrl->ctrl.queue_count == 1)
2930                 return 0;
2931
2932         ret = nvme_fc_create_hw_io_queues(ctrl, ctrl->ctrl.sqsize + 1);
2933         if (ret)
2934                 goto out_free_io_queues;
2935
2936         ret = nvme_fc_connect_io_queues(ctrl, ctrl->ctrl.sqsize + 1);
2937         if (ret)
2938                 goto out_delete_hw_queues;
2939
2940         if (prior_ioq_cnt != nr_io_queues) {
2941                 dev_info(ctrl->ctrl.device,
2942                         "reconnect: revising io queue count from %d to %d\n",
2943                         prior_ioq_cnt, nr_io_queues);
2944                 nvme_wait_freeze(&ctrl->ctrl);
2945                 blk_mq_update_nr_hw_queues(&ctrl->tag_set, nr_io_queues);
2946                 nvme_unfreeze(&ctrl->ctrl);
2947         }
2948
2949         return 0;
2950
2951 out_delete_hw_queues:
2952         nvme_fc_delete_hw_io_queues(ctrl);
2953 out_free_io_queues:
2954         nvme_fc_free_io_queues(ctrl);
2955         return ret;
2956 }
2957
2958 static void
2959 nvme_fc_rport_active_on_lport(struct nvme_fc_rport *rport)
2960 {
2961         struct nvme_fc_lport *lport = rport->lport;
2962
2963         atomic_inc(&lport->act_rport_cnt);
2964 }
2965
2966 static void
2967 nvme_fc_rport_inactive_on_lport(struct nvme_fc_rport *rport)
2968 {
2969         struct nvme_fc_lport *lport = rport->lport;
2970         u32 cnt;
2971
2972         cnt = atomic_dec_return(&lport->act_rport_cnt);
2973         if (cnt == 0 && lport->localport.port_state == FC_OBJSTATE_DELETED)
2974                 lport->ops->localport_delete(&lport->localport);
2975 }
2976
2977 static int
2978 nvme_fc_ctlr_active_on_rport(struct nvme_fc_ctrl *ctrl)
2979 {
2980         struct nvme_fc_rport *rport = ctrl->rport;
2981         u32 cnt;
2982
2983         if (test_and_set_bit(ASSOC_ACTIVE, &ctrl->flags))
2984                 return 1;
2985
2986         cnt = atomic_inc_return(&rport->act_ctrl_cnt);
2987         if (cnt == 1)
2988                 nvme_fc_rport_active_on_lport(rport);
2989
2990         return 0;
2991 }
2992
2993 static int
2994 nvme_fc_ctlr_inactive_on_rport(struct nvme_fc_ctrl *ctrl)
2995 {
2996         struct nvme_fc_rport *rport = ctrl->rport;
2997         struct nvme_fc_lport *lport = rport->lport;
2998         u32 cnt;
2999
3000         /* clearing of ctrl->flags ASSOC_ACTIVE bit is in association delete */
3001
3002         cnt = atomic_dec_return(&rport->act_ctrl_cnt);
3003         if (cnt == 0) {
3004                 if (rport->remoteport.port_state == FC_OBJSTATE_DELETED)
3005                         lport->ops->remoteport_delete(&rport->remoteport);
3006                 nvme_fc_rport_inactive_on_lport(rport);
3007         }
3008
3009         return 0;
3010 }
3011
3012 /*
3013  * This routine restarts the controller on the host side, and
3014  * on the link side, recreates the controller association.
3015  */
3016 static int
3017 nvme_fc_create_association(struct nvme_fc_ctrl *ctrl)
3018 {
3019         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
3020         struct nvmefc_ls_rcv_op *disls = NULL;
3021         unsigned long flags;
3022         int ret;
3023         bool changed;
3024
3025         ++ctrl->ctrl.nr_reconnects;
3026
3027         if (ctrl->rport->remoteport.port_state != FC_OBJSTATE_ONLINE)
3028                 return -ENODEV;
3029
3030         if (nvme_fc_ctlr_active_on_rport(ctrl))
3031                 return -ENOTUNIQ;
3032
3033         dev_info(ctrl->ctrl.device,
3034                 "NVME-FC{%d}: create association : host wwpn 0x%016llx "
3035                 " rport wwpn 0x%016llx: NQN \"%s\"\n",
3036                 ctrl->cnum, ctrl->lport->localport.port_name,
3037                 ctrl->rport->remoteport.port_name, ctrl->ctrl.opts->subsysnqn);
3038
3039         clear_bit(ASSOC_FAILED, &ctrl->flags);
3040
3041         /*
3042          * Create the admin queue
3043          */
3044
3045         ret = __nvme_fc_create_hw_queue(ctrl, &ctrl->queues[0], 0,
3046                                 NVME_AQ_DEPTH);
3047         if (ret)
3048                 goto out_free_queue;
3049
3050         ret = nvme_fc_connect_admin_queue(ctrl, &ctrl->queues[0],
3051                                 NVME_AQ_DEPTH, (NVME_AQ_DEPTH / 4));
3052         if (ret)
3053                 goto out_delete_hw_queue;
3054
3055         ret = nvmf_connect_admin_queue(&ctrl->ctrl);
3056         if (ret)
3057                 goto out_disconnect_admin_queue;
3058
3059         set_bit(NVME_FC_Q_LIVE, &ctrl->queues[0].flags);
3060
3061         /*
3062          * Check controller capabilities
3063          *
3064          * todo:- add code to check if ctrl attributes changed from
3065          * prior connection values
3066          */
3067
3068         ret = nvme_enable_ctrl(&ctrl->ctrl);
3069         if (ret || test_bit(ASSOC_FAILED, &ctrl->flags))
3070                 goto out_disconnect_admin_queue;
3071
3072         ctrl->ctrl.max_segments = ctrl->lport->ops->max_sgl_segments;
3073         ctrl->ctrl.max_hw_sectors = ctrl->ctrl.max_segments <<
3074                                                 (ilog2(SZ_4K) - 9);
3075
3076         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
3077
3078         ret = nvme_init_identify(&ctrl->ctrl);
3079         if (ret || test_bit(ASSOC_FAILED, &ctrl->flags))
3080                 goto out_disconnect_admin_queue;
3081
3082         /* sanity checks */
3083
3084         /* FC-NVME does not have other data in the capsule */
3085         if (ctrl->ctrl.icdoff) {
3086                 dev_err(ctrl->ctrl.device, "icdoff %d is not supported!\n",
3087                                 ctrl->ctrl.icdoff);
3088                 goto out_disconnect_admin_queue;
3089         }
3090
3091         /* FC-NVME supports normal SGL Data Block Descriptors */
3092
3093         if (opts->queue_size > ctrl->ctrl.maxcmd) {
3094                 /* warn if maxcmd is lower than queue_size */
3095                 dev_warn(ctrl->ctrl.device,
3096                         "queue_size %zu > ctrl maxcmd %u, reducing "
3097                         "to maxcmd\n",
3098                         opts->queue_size, ctrl->ctrl.maxcmd);
3099                 opts->queue_size = ctrl->ctrl.maxcmd;
3100         }
3101
3102         if (opts->queue_size > ctrl->ctrl.sqsize + 1) {
3103                 /* warn if sqsize is lower than queue_size */
3104                 dev_warn(ctrl->ctrl.device,
3105                         "queue_size %zu > ctrl sqsize %u, reducing "
3106                         "to sqsize\n",
3107                         opts->queue_size, ctrl->ctrl.sqsize + 1);
3108                 opts->queue_size = ctrl->ctrl.sqsize + 1;
3109         }
3110
3111         ret = nvme_fc_init_aen_ops(ctrl);
3112         if (ret)
3113                 goto out_term_aen_ops;
3114
3115         /*
3116          * Create the io queues
3117          */
3118
3119         if (ctrl->ctrl.queue_count > 1) {
3120                 if (!ctrl->ioq_live)
3121                         ret = nvme_fc_create_io_queues(ctrl);
3122                 else
3123                         ret = nvme_fc_recreate_io_queues(ctrl);
3124         }
3125         if (ret || test_bit(ASSOC_FAILED, &ctrl->flags))
3126                 goto out_term_aen_ops;
3127
3128         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
3129
3130         ctrl->ctrl.nr_reconnects = 0;
3131
3132         if (changed)
3133                 nvme_start_ctrl(&ctrl->ctrl);
3134
3135         return 0;       /* Success */
3136
3137 out_term_aen_ops:
3138         nvme_fc_term_aen_ops(ctrl);
3139 out_disconnect_admin_queue:
3140         /* send a Disconnect(association) LS to fc-nvme target */
3141         nvme_fc_xmt_disconnect_assoc(ctrl);
3142         spin_lock_irqsave(&ctrl->lock, flags);
3143         ctrl->association_id = 0;
3144         disls = ctrl->rcv_disconn;
3145         ctrl->rcv_disconn = NULL;
3146         spin_unlock_irqrestore(&ctrl->lock, flags);
3147         if (disls)
3148                 nvme_fc_xmt_ls_rsp(disls);
3149 out_delete_hw_queue:
3150         __nvme_fc_delete_hw_queue(ctrl, &ctrl->queues[0], 0);
3151 out_free_queue:
3152         nvme_fc_free_queue(&ctrl->queues[0]);
3153         clear_bit(ASSOC_ACTIVE, &ctrl->flags);
3154         nvme_fc_ctlr_inactive_on_rport(ctrl);
3155
3156         return ret;
3157 }
3158
3159
3160 /*
3161  * This routine stops operation of the controller on the host side.
3162  * On the host os stack side: Admin and IO queues are stopped,
3163  *   outstanding ios on them terminated via FC ABTS.
3164  * On the link side: the association is terminated.
3165  */
3166 static void
3167 nvme_fc_delete_association(struct nvme_fc_ctrl *ctrl)
3168 {
3169         struct nvmefc_ls_rcv_op *disls = NULL;
3170         unsigned long flags;
3171
3172         if (!test_and_clear_bit(ASSOC_ACTIVE, &ctrl->flags))
3173                 return;
3174
3175         spin_lock_irqsave(&ctrl->lock, flags);
3176         set_bit(FCCTRL_TERMIO, &ctrl->flags);
3177         ctrl->iocnt = 0;
3178         spin_unlock_irqrestore(&ctrl->lock, flags);
3179
3180         __nvme_fc_abort_outstanding_ios(ctrl, false);
3181
3182         /* kill the aens as they are a separate path */
3183         nvme_fc_abort_aen_ops(ctrl);
3184
3185         /* wait for all io that had to be aborted */
3186         spin_lock_irq(&ctrl->lock);
3187         wait_event_lock_irq(ctrl->ioabort_wait, ctrl->iocnt == 0, ctrl->lock);
3188         clear_bit(FCCTRL_TERMIO, &ctrl->flags);
3189         spin_unlock_irq(&ctrl->lock);
3190
3191         nvme_fc_term_aen_ops(ctrl);
3192
3193         /*
3194          * send a Disconnect(association) LS to fc-nvme target
3195          * Note: could have been sent at top of process, but
3196          * cleaner on link traffic if after the aborts complete.
3197          * Note: if association doesn't exist, association_id will be 0
3198          */
3199         if (ctrl->association_id)
3200                 nvme_fc_xmt_disconnect_assoc(ctrl);
3201
3202         spin_lock_irqsave(&ctrl->lock, flags);
3203         ctrl->association_id = 0;
3204         disls = ctrl->rcv_disconn;
3205         ctrl->rcv_disconn = NULL;
3206         spin_unlock_irqrestore(&ctrl->lock, flags);
3207         if (disls)
3208                 /*
3209                  * if a Disconnect Request was waiting for a response, send
3210                  * now that all ABTS's have been issued (and are complete).
3211                  */
3212                 nvme_fc_xmt_ls_rsp(disls);
3213
3214         if (ctrl->ctrl.tagset) {
3215                 nvme_fc_delete_hw_io_queues(ctrl);
3216                 nvme_fc_free_io_queues(ctrl);
3217         }
3218
3219         __nvme_fc_delete_hw_queue(ctrl, &ctrl->queues[0], 0);
3220         nvme_fc_free_queue(&ctrl->queues[0]);
3221
3222         /* re-enable the admin_q so anything new can fast fail */
3223         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
3224
3225         /* resume the io queues so that things will fast fail */
3226         nvme_start_queues(&ctrl->ctrl);
3227
3228         nvme_fc_ctlr_inactive_on_rport(ctrl);
3229 }
3230
3231 static void
3232 nvme_fc_delete_ctrl(struct nvme_ctrl *nctrl)
3233 {
3234         struct nvme_fc_ctrl *ctrl = to_fc_ctrl(nctrl);
3235
3236         cancel_delayed_work_sync(&ctrl->connect_work);
3237         /*
3238          * kill the association on the link side.  this will block
3239          * waiting for io to terminate
3240          */
3241         nvme_fc_delete_association(ctrl);
3242 }
3243
3244 static void
3245 nvme_fc_reconnect_or_delete(struct nvme_fc_ctrl *ctrl, int status)
3246 {
3247         struct nvme_fc_rport *rport = ctrl->rport;
3248         struct nvme_fc_remote_port *portptr = &rport->remoteport;
3249         unsigned long recon_delay = ctrl->ctrl.opts->reconnect_delay * HZ;
3250         bool recon = true;
3251
3252         if (ctrl->ctrl.state != NVME_CTRL_CONNECTING)
3253                 return;
3254
3255         if (portptr->port_state == FC_OBJSTATE_ONLINE)
3256                 dev_info(ctrl->ctrl.device,
3257                         "NVME-FC{%d}: reset: Reconnect attempt failed (%d)\n",
3258                         ctrl->cnum, status);
3259         else if (time_after_eq(jiffies, rport->dev_loss_end))
3260                 recon = false;
3261
3262         if (recon && nvmf_should_reconnect(&ctrl->ctrl)) {
3263                 if (portptr->port_state == FC_OBJSTATE_ONLINE)
3264                         dev_info(ctrl->ctrl.device,
3265                                 "NVME-FC{%d}: Reconnect attempt in %ld "
3266                                 "seconds\n",
3267                                 ctrl->cnum, recon_delay / HZ);
3268                 else if (time_after(jiffies + recon_delay, rport->dev_loss_end))
3269                         recon_delay = rport->dev_loss_end - jiffies;
3270
3271                 queue_delayed_work(nvme_wq, &ctrl->connect_work, recon_delay);
3272         } else {
3273                 if (portptr->port_state == FC_OBJSTATE_ONLINE)
3274                         dev_warn(ctrl->ctrl.device,
3275                                 "NVME-FC{%d}: Max reconnect attempts (%d) "
3276                                 "reached.\n",
3277                                 ctrl->cnum, ctrl->ctrl.nr_reconnects);
3278                 else
3279                         dev_warn(ctrl->ctrl.device,
3280                                 "NVME-FC{%d}: dev_loss_tmo (%d) expired "
3281                                 "while waiting for remoteport connectivity.\n",
3282                                 ctrl->cnum, min_t(int, portptr->dev_loss_tmo,
3283                                         (ctrl->ctrl.opts->max_reconnects *
3284                                          ctrl->ctrl.opts->reconnect_delay)));
3285                 WARN_ON(nvme_delete_ctrl(&ctrl->ctrl));
3286         }
3287 }
3288
3289 static void
3290 nvme_fc_reset_ctrl_work(struct work_struct *work)
3291 {
3292         struct nvme_fc_ctrl *ctrl =
3293                 container_of(work, struct nvme_fc_ctrl, ctrl.reset_work);
3294
3295         nvme_stop_ctrl(&ctrl->ctrl);
3296
3297         /* will block will waiting for io to terminate */
3298         nvme_fc_delete_association(ctrl);
3299
3300         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING))
3301                 dev_err(ctrl->ctrl.device,
3302                         "NVME-FC{%d}: error_recovery: Couldn't change state "
3303                         "to CONNECTING\n", ctrl->cnum);
3304
3305         if (ctrl->rport->remoteport.port_state == FC_OBJSTATE_ONLINE) {
3306                 if (!queue_delayed_work(nvme_wq, &ctrl->connect_work, 0)) {
3307                         dev_err(ctrl->ctrl.device,
3308                                 "NVME-FC{%d}: failed to schedule connect "
3309                                 "after reset\n", ctrl->cnum);
3310                 } else {
3311                         flush_delayed_work(&ctrl->connect_work);
3312                 }
3313         } else {
3314                 nvme_fc_reconnect_or_delete(ctrl, -ENOTCONN);
3315         }
3316 }
3317
3318
3319 static const struct nvme_ctrl_ops nvme_fc_ctrl_ops = {
3320         .name                   = "fc",
3321         .module                 = THIS_MODULE,
3322         .flags                  = NVME_F_FABRICS,
3323         .reg_read32             = nvmf_reg_read32,
3324         .reg_read64             = nvmf_reg_read64,
3325         .reg_write32            = nvmf_reg_write32,
3326         .free_ctrl              = nvme_fc_nvme_ctrl_freed,
3327         .submit_async_event     = nvme_fc_submit_async_event,
3328         .delete_ctrl            = nvme_fc_delete_ctrl,
3329         .get_address            = nvmf_get_address,
3330 };
3331
3332 static void
3333 nvme_fc_connect_ctrl_work(struct work_struct *work)
3334 {
3335         int ret;
3336
3337         struct nvme_fc_ctrl *ctrl =
3338                         container_of(to_delayed_work(work),
3339                                 struct nvme_fc_ctrl, connect_work);
3340
3341         ret = nvme_fc_create_association(ctrl);
3342         if (ret)
3343                 nvme_fc_reconnect_or_delete(ctrl, ret);
3344         else
3345                 dev_info(ctrl->ctrl.device,
3346                         "NVME-FC{%d}: controller connect complete\n",
3347                         ctrl->cnum);
3348 }
3349
3350
3351 static const struct blk_mq_ops nvme_fc_admin_mq_ops = {
3352         .queue_rq       = nvme_fc_queue_rq,
3353         .complete       = nvme_fc_complete_rq,
3354         .init_request   = nvme_fc_init_request,
3355         .exit_request   = nvme_fc_exit_request,
3356         .init_hctx      = nvme_fc_init_admin_hctx,
3357         .timeout        = nvme_fc_timeout,
3358 };
3359
3360
3361 /*
3362  * Fails a controller request if it matches an existing controller
3363  * (association) with the same tuple:
3364  * <Host NQN, Host ID, local FC port, remote FC port, SUBSYS NQN>
3365  *
3366  * The ports don't need to be compared as they are intrinsically
3367  * already matched by the port pointers supplied.
3368  */
3369 static bool
3370 nvme_fc_existing_controller(struct nvme_fc_rport *rport,
3371                 struct nvmf_ctrl_options *opts)
3372 {
3373         struct nvme_fc_ctrl *ctrl;
3374         unsigned long flags;
3375         bool found = false;
3376
3377         spin_lock_irqsave(&rport->lock, flags);
3378         list_for_each_entry(ctrl, &rport->ctrl_list, ctrl_list) {
3379                 found = nvmf_ctlr_matches_baseopts(&ctrl->ctrl, opts);
3380                 if (found)
3381                         break;
3382         }
3383         spin_unlock_irqrestore(&rport->lock, flags);
3384
3385         return found;
3386 }
3387
3388 static struct nvme_ctrl *
3389 nvme_fc_init_ctrl(struct device *dev, struct nvmf_ctrl_options *opts,
3390         struct nvme_fc_lport *lport, struct nvme_fc_rport *rport)
3391 {
3392         struct nvme_fc_ctrl *ctrl;
3393         unsigned long flags;
3394         int ret, idx, ctrl_loss_tmo;
3395
3396         if (!(rport->remoteport.port_role &
3397             (FC_PORT_ROLE_NVME_DISCOVERY | FC_PORT_ROLE_NVME_TARGET))) {
3398                 ret = -EBADR;
3399                 goto out_fail;
3400         }
3401
3402         if (!opts->duplicate_connect &&
3403             nvme_fc_existing_controller(rport, opts)) {
3404                 ret = -EALREADY;
3405                 goto out_fail;
3406         }
3407
3408         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
3409         if (!ctrl) {
3410                 ret = -ENOMEM;
3411                 goto out_fail;
3412         }
3413
3414         idx = ida_simple_get(&nvme_fc_ctrl_cnt, 0, 0, GFP_KERNEL);
3415         if (idx < 0) {
3416                 ret = -ENOSPC;
3417                 goto out_free_ctrl;
3418         }
3419
3420         /*
3421          * if ctrl_loss_tmo is being enforced and the default reconnect delay
3422          * is being used, change to a shorter reconnect delay for FC.
3423          */
3424         if (opts->max_reconnects != -1 &&
3425             opts->reconnect_delay == NVMF_DEF_RECONNECT_DELAY &&
3426             opts->reconnect_delay > NVME_FC_DEFAULT_RECONNECT_TMO) {
3427                 ctrl_loss_tmo = opts->max_reconnects * opts->reconnect_delay;
3428                 opts->reconnect_delay = NVME_FC_DEFAULT_RECONNECT_TMO;
3429                 opts->max_reconnects = DIV_ROUND_UP(ctrl_loss_tmo,
3430                                                 opts->reconnect_delay);
3431         }
3432
3433         ctrl->ctrl.opts = opts;
3434         ctrl->ctrl.nr_reconnects = 0;
3435         if (lport->dev)
3436                 ctrl->ctrl.numa_node = dev_to_node(lport->dev);
3437         else
3438                 ctrl->ctrl.numa_node = NUMA_NO_NODE;
3439         INIT_LIST_HEAD(&ctrl->ctrl_list);
3440         ctrl->lport = lport;
3441         ctrl->rport = rport;
3442         ctrl->dev = lport->dev;
3443         ctrl->cnum = idx;
3444         ctrl->ioq_live = false;
3445         init_waitqueue_head(&ctrl->ioabort_wait);
3446
3447         get_device(ctrl->dev);
3448         kref_init(&ctrl->ref);
3449
3450         INIT_WORK(&ctrl->ctrl.reset_work, nvme_fc_reset_ctrl_work);
3451         INIT_DELAYED_WORK(&ctrl->connect_work, nvme_fc_connect_ctrl_work);
3452         spin_lock_init(&ctrl->lock);
3453
3454         /* io queue count */
3455         ctrl->ctrl.queue_count = min_t(unsigned int,
3456                                 opts->nr_io_queues,
3457                                 lport->ops->max_hw_queues);
3458         ctrl->ctrl.queue_count++;       /* +1 for admin queue */
3459
3460         ctrl->ctrl.sqsize = opts->queue_size - 1;
3461         ctrl->ctrl.kato = opts->kato;
3462         ctrl->ctrl.cntlid = 0xffff;
3463
3464         ret = -ENOMEM;
3465         ctrl->queues = kcalloc(ctrl->ctrl.queue_count,
3466                                 sizeof(struct nvme_fc_queue), GFP_KERNEL);
3467         if (!ctrl->queues)
3468                 goto out_free_ida;
3469
3470         nvme_fc_init_queue(ctrl, 0);
3471
3472         memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
3473         ctrl->admin_tag_set.ops = &nvme_fc_admin_mq_ops;
3474         ctrl->admin_tag_set.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
3475         ctrl->admin_tag_set.reserved_tags = 2; /* fabric connect + Keep-Alive */
3476         ctrl->admin_tag_set.numa_node = ctrl->ctrl.numa_node;
3477         ctrl->admin_tag_set.cmd_size =
3478                 struct_size((struct nvme_fcp_op_w_sgl *)NULL, priv,
3479                             ctrl->lport->ops->fcprqst_priv_sz);
3480         ctrl->admin_tag_set.driver_data = ctrl;
3481         ctrl->admin_tag_set.nr_hw_queues = 1;
3482         ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
3483         ctrl->admin_tag_set.flags = BLK_MQ_F_NO_SCHED;
3484
3485         ret = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
3486         if (ret)
3487                 goto out_free_queues;
3488         ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
3489
3490         ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
3491         if (IS_ERR(ctrl->ctrl.fabrics_q)) {
3492                 ret = PTR_ERR(ctrl->ctrl.fabrics_q);
3493                 goto out_free_admin_tag_set;
3494         }
3495
3496         ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
3497         if (IS_ERR(ctrl->ctrl.admin_q)) {
3498                 ret = PTR_ERR(ctrl->ctrl.admin_q);
3499                 goto out_cleanup_fabrics_q;
3500         }
3501
3502         /*
3503          * Would have been nice to init io queues tag set as well.
3504          * However, we require interaction from the controller
3505          * for max io queue count before we can do so.
3506          * Defer this to the connect path.
3507          */
3508
3509         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_fc_ctrl_ops, 0);
3510         if (ret)
3511                 goto out_cleanup_admin_q;
3512
3513         /* at this point, teardown path changes to ref counting on nvme ctrl */
3514
3515         spin_lock_irqsave(&rport->lock, flags);
3516         list_add_tail(&ctrl->ctrl_list, &rport->ctrl_list);
3517         spin_unlock_irqrestore(&rport->lock, flags);
3518
3519         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING) ||
3520             !nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
3521                 dev_err(ctrl->ctrl.device,
3522                         "NVME-FC{%d}: failed to init ctrl state\n", ctrl->cnum);
3523                 goto fail_ctrl;
3524         }
3525
3526         if (!queue_delayed_work(nvme_wq, &ctrl->connect_work, 0)) {
3527                 dev_err(ctrl->ctrl.device,
3528                         "NVME-FC{%d}: failed to schedule initial connect\n",
3529                         ctrl->cnum);
3530                 goto fail_ctrl;
3531         }
3532
3533         flush_delayed_work(&ctrl->connect_work);
3534
3535         dev_info(ctrl->ctrl.device,
3536                 "NVME-FC{%d}: new ctrl: NQN \"%s\"\n",
3537                 ctrl->cnum, ctrl->ctrl.opts->subsysnqn);
3538
3539         return &ctrl->ctrl;
3540
3541 fail_ctrl:
3542         nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING);
3543         cancel_work_sync(&ctrl->ctrl.reset_work);
3544         cancel_delayed_work_sync(&ctrl->connect_work);
3545
3546         ctrl->ctrl.opts = NULL;
3547
3548         /* initiate nvme ctrl ref counting teardown */
3549         nvme_uninit_ctrl(&ctrl->ctrl);
3550
3551         /* Remove core ctrl ref. */
3552         nvme_put_ctrl(&ctrl->ctrl);
3553
3554         /* as we're past the point where we transition to the ref
3555          * counting teardown path, if we return a bad pointer here,
3556          * the calling routine, thinking it's prior to the
3557          * transition, will do an rport put. Since the teardown
3558          * path also does a rport put, we do an extra get here to
3559          * so proper order/teardown happens.
3560          */
3561         nvme_fc_rport_get(rport);
3562
3563         return ERR_PTR(-EIO);
3564
3565 out_cleanup_admin_q:
3566         blk_cleanup_queue(ctrl->ctrl.admin_q);
3567 out_cleanup_fabrics_q:
3568         blk_cleanup_queue(ctrl->ctrl.fabrics_q);
3569 out_free_admin_tag_set:
3570         blk_mq_free_tag_set(&ctrl->admin_tag_set);
3571 out_free_queues:
3572         kfree(ctrl->queues);
3573 out_free_ida:
3574         put_device(ctrl->dev);
3575         ida_simple_remove(&nvme_fc_ctrl_cnt, ctrl->cnum);
3576 out_free_ctrl:
3577         kfree(ctrl);
3578 out_fail:
3579         /* exit via here doesn't follow ctlr ref points */
3580         return ERR_PTR(ret);
3581 }
3582
3583
3584 struct nvmet_fc_traddr {
3585         u64     nn;
3586         u64     pn;
3587 };
3588
3589 static int
3590 __nvme_fc_parse_u64(substring_t *sstr, u64 *val)
3591 {
3592         u64 token64;
3593
3594         if (match_u64(sstr, &token64))
3595                 return -EINVAL;
3596         *val = token64;
3597
3598         return 0;
3599 }
3600
3601 /*
3602  * This routine validates and extracts the WWN's from the TRADDR string.
3603  * As kernel parsers need the 0x to determine number base, universally
3604  * build string to parse with 0x prefix before parsing name strings.
3605  */
3606 static int
3607 nvme_fc_parse_traddr(struct nvmet_fc_traddr *traddr, char *buf, size_t blen)
3608 {
3609         char name[2 + NVME_FC_TRADDR_HEXNAMELEN + 1];
3610         substring_t wwn = { name, &name[sizeof(name)-1] };
3611         int nnoffset, pnoffset;
3612
3613         /* validate if string is one of the 2 allowed formats */
3614         if (strnlen(buf, blen) == NVME_FC_TRADDR_MAXLENGTH &&
3615                         !strncmp(buf, "nn-0x", NVME_FC_TRADDR_OXNNLEN) &&
3616                         !strncmp(&buf[NVME_FC_TRADDR_MAX_PN_OFFSET],
3617                                 "pn-0x", NVME_FC_TRADDR_OXNNLEN)) {
3618                 nnoffset = NVME_FC_TRADDR_OXNNLEN;
3619                 pnoffset = NVME_FC_TRADDR_MAX_PN_OFFSET +
3620                                                 NVME_FC_TRADDR_OXNNLEN;
3621         } else if ((strnlen(buf, blen) == NVME_FC_TRADDR_MINLENGTH &&
3622                         !strncmp(buf, "nn-", NVME_FC_TRADDR_NNLEN) &&
3623                         !strncmp(&buf[NVME_FC_TRADDR_MIN_PN_OFFSET],
3624                                 "pn-", NVME_FC_TRADDR_NNLEN))) {
3625                 nnoffset = NVME_FC_TRADDR_NNLEN;
3626                 pnoffset = NVME_FC_TRADDR_MIN_PN_OFFSET + NVME_FC_TRADDR_NNLEN;
3627         } else
3628                 goto out_einval;
3629
3630         name[0] = '0';
3631         name[1] = 'x';
3632         name[2 + NVME_FC_TRADDR_HEXNAMELEN] = 0;
3633
3634         memcpy(&name[2], &buf[nnoffset], NVME_FC_TRADDR_HEXNAMELEN);
3635         if (__nvme_fc_parse_u64(&wwn, &traddr->nn))
3636                 goto out_einval;
3637
3638         memcpy(&name[2], &buf[pnoffset], NVME_FC_TRADDR_HEXNAMELEN);
3639         if (__nvme_fc_parse_u64(&wwn, &traddr->pn))
3640                 goto out_einval;
3641
3642         return 0;
3643
3644 out_einval:
3645         pr_warn("%s: bad traddr string\n", __func__);
3646         return -EINVAL;
3647 }
3648
3649 static struct nvme_ctrl *
3650 nvme_fc_create_ctrl(struct device *dev, struct nvmf_ctrl_options *opts)
3651 {
3652         struct nvme_fc_lport *lport;
3653         struct nvme_fc_rport *rport;
3654         struct nvme_ctrl *ctrl;
3655         struct nvmet_fc_traddr laddr = { 0L, 0L };
3656         struct nvmet_fc_traddr raddr = { 0L, 0L };
3657         unsigned long flags;
3658         int ret;
3659
3660         ret = nvme_fc_parse_traddr(&raddr, opts->traddr, NVMF_TRADDR_SIZE);
3661         if (ret || !raddr.nn || !raddr.pn)
3662                 return ERR_PTR(-EINVAL);
3663
3664         ret = nvme_fc_parse_traddr(&laddr, opts->host_traddr, NVMF_TRADDR_SIZE);
3665         if (ret || !laddr.nn || !laddr.pn)
3666                 return ERR_PTR(-EINVAL);
3667
3668         /* find the host and remote ports to connect together */
3669         spin_lock_irqsave(&nvme_fc_lock, flags);
3670         list_for_each_entry(lport, &nvme_fc_lport_list, port_list) {
3671                 if (lport->localport.node_name != laddr.nn ||
3672                     lport->localport.port_name != laddr.pn ||
3673                     lport->localport.port_state != FC_OBJSTATE_ONLINE)
3674                         continue;
3675
3676                 list_for_each_entry(rport, &lport->endp_list, endp_list) {
3677                         if (rport->remoteport.node_name != raddr.nn ||
3678                             rport->remoteport.port_name != raddr.pn ||
3679                             rport->remoteport.port_state != FC_OBJSTATE_ONLINE)
3680                                 continue;
3681
3682                         /* if fail to get reference fall through. Will error */
3683                         if (!nvme_fc_rport_get(rport))
3684                                 break;
3685
3686                         spin_unlock_irqrestore(&nvme_fc_lock, flags);
3687
3688                         ctrl = nvme_fc_init_ctrl(dev, opts, lport, rport);
3689                         if (IS_ERR(ctrl))
3690                                 nvme_fc_rport_put(rport);
3691                         return ctrl;
3692                 }
3693         }
3694         spin_unlock_irqrestore(&nvme_fc_lock, flags);
3695
3696         pr_warn("%s: %s - %s combination not found\n",
3697                 __func__, opts->traddr, opts->host_traddr);
3698         return ERR_PTR(-ENOENT);
3699 }
3700
3701
3702 static struct nvmf_transport_ops nvme_fc_transport = {
3703         .name           = "fc",
3704         .module         = THIS_MODULE,
3705         .required_opts  = NVMF_OPT_TRADDR | NVMF_OPT_HOST_TRADDR,
3706         .allowed_opts   = NVMF_OPT_RECONNECT_DELAY | NVMF_OPT_CTRL_LOSS_TMO,
3707         .create_ctrl    = nvme_fc_create_ctrl,
3708 };
3709
3710 /* Arbitrary successive failures max. With lots of subsystems could be high */
3711 #define DISCOVERY_MAX_FAIL      20
3712
3713 static ssize_t nvme_fc_nvme_discovery_store(struct device *dev,
3714                 struct device_attribute *attr, const char *buf, size_t count)
3715 {
3716         unsigned long flags;
3717         LIST_HEAD(local_disc_list);
3718         struct nvme_fc_lport *lport;
3719         struct nvme_fc_rport *rport;
3720         int failcnt = 0;
3721
3722         spin_lock_irqsave(&nvme_fc_lock, flags);
3723 restart:
3724         list_for_each_entry(lport, &nvme_fc_lport_list, port_list) {
3725                 list_for_each_entry(rport, &lport->endp_list, endp_list) {
3726                         if (!nvme_fc_lport_get(lport))
3727                                 continue;
3728                         if (!nvme_fc_rport_get(rport)) {
3729                                 /*
3730                                  * This is a temporary condition. Upon restart
3731                                  * this rport will be gone from the list.
3732                                  *
3733                                  * Revert the lport put and retry.  Anything
3734                                  * added to the list already will be skipped (as
3735                                  * they are no longer list_empty).  Loops should
3736                                  * resume at rports that were not yet seen.
3737                                  */
3738                                 nvme_fc_lport_put(lport);
3739
3740                                 if (failcnt++ < DISCOVERY_MAX_FAIL)
3741                                         goto restart;
3742
3743                                 pr_err("nvme_discovery: too many reference "
3744                                        "failures\n");
3745                                 goto process_local_list;
3746                         }
3747                         if (list_empty(&rport->disc_list))
3748                                 list_add_tail(&rport->disc_list,
3749                                               &local_disc_list);
3750                 }
3751         }
3752
3753 process_local_list:
3754         while (!list_empty(&local_disc_list)) {
3755                 rport = list_first_entry(&local_disc_list,
3756                                          struct nvme_fc_rport, disc_list);
3757                 list_del_init(&rport->disc_list);
3758                 spin_unlock_irqrestore(&nvme_fc_lock, flags);
3759
3760                 lport = rport->lport;
3761                 /* signal discovery. Won't hurt if it repeats */
3762                 nvme_fc_signal_discovery_scan(lport, rport);
3763                 nvme_fc_rport_put(rport);
3764                 nvme_fc_lport_put(lport);
3765
3766                 spin_lock_irqsave(&nvme_fc_lock, flags);
3767         }
3768         spin_unlock_irqrestore(&nvme_fc_lock, flags);
3769
3770         return count;
3771 }
3772 static DEVICE_ATTR(nvme_discovery, 0200, NULL, nvme_fc_nvme_discovery_store);
3773
3774 static struct attribute *nvme_fc_attrs[] = {
3775         &dev_attr_nvme_discovery.attr,
3776         NULL
3777 };
3778
3779 static struct attribute_group nvme_fc_attr_group = {
3780         .attrs = nvme_fc_attrs,
3781 };
3782
3783 static const struct attribute_group *nvme_fc_attr_groups[] = {
3784         &nvme_fc_attr_group,
3785         NULL
3786 };
3787
3788 static struct class fc_class = {
3789         .name = "fc",
3790         .dev_groups = nvme_fc_attr_groups,
3791         .owner = THIS_MODULE,
3792 };
3793
3794 static int __init nvme_fc_init_module(void)
3795 {
3796         int ret;
3797
3798         nvme_fc_wq = alloc_workqueue("nvme_fc_wq", WQ_MEM_RECLAIM, 0);
3799         if (!nvme_fc_wq)
3800                 return -ENOMEM;
3801
3802         /*
3803          * NOTE:
3804          * It is expected that in the future the kernel will combine
3805          * the FC-isms that are currently under scsi and now being
3806          * added to by NVME into a new standalone FC class. The SCSI
3807          * and NVME protocols and their devices would be under this
3808          * new FC class.
3809          *
3810          * As we need something to post FC-specific udev events to,
3811          * specifically for nvme probe events, start by creating the
3812          * new device class.  When the new standalone FC class is
3813          * put in place, this code will move to a more generic
3814          * location for the class.
3815          */
3816         ret = class_register(&fc_class);
3817         if (ret) {
3818                 pr_err("couldn't register class fc\n");
3819                 goto out_destroy_wq;
3820         }
3821
3822         /*
3823          * Create a device for the FC-centric udev events
3824          */
3825         fc_udev_device = device_create(&fc_class, NULL, MKDEV(0, 0), NULL,
3826                                 "fc_udev_device");
3827         if (IS_ERR(fc_udev_device)) {
3828                 pr_err("couldn't create fc_udev device!\n");
3829                 ret = PTR_ERR(fc_udev_device);
3830                 goto out_destroy_class;
3831         }
3832
3833         ret = nvmf_register_transport(&nvme_fc_transport);
3834         if (ret)
3835                 goto out_destroy_device;
3836
3837         return 0;
3838
3839 out_destroy_device:
3840         device_destroy(&fc_class, MKDEV(0, 0));
3841 out_destroy_class:
3842         class_unregister(&fc_class);
3843 out_destroy_wq:
3844         destroy_workqueue(nvme_fc_wq);
3845
3846         return ret;
3847 }
3848
3849 static void
3850 nvme_fc_delete_controllers(struct nvme_fc_rport *rport)
3851 {
3852         struct nvme_fc_ctrl *ctrl;
3853
3854         spin_lock(&rport->lock);
3855         list_for_each_entry(ctrl, &rport->ctrl_list, ctrl_list) {
3856                 dev_warn(ctrl->ctrl.device,
3857                         "NVME-FC{%d}: transport unloading: deleting ctrl\n",
3858                         ctrl->cnum);
3859                 nvme_delete_ctrl(&ctrl->ctrl);
3860         }
3861         spin_unlock(&rport->lock);
3862 }
3863
3864 static void
3865 nvme_fc_cleanup_for_unload(void)
3866 {
3867         struct nvme_fc_lport *lport;
3868         struct nvme_fc_rport *rport;
3869
3870         list_for_each_entry(lport, &nvme_fc_lport_list, port_list) {
3871                 list_for_each_entry(rport, &lport->endp_list, endp_list) {
3872                         nvme_fc_delete_controllers(rport);
3873                 }
3874         }
3875 }
3876
3877 static void __exit nvme_fc_exit_module(void)
3878 {
3879         unsigned long flags;
3880         bool need_cleanup = false;
3881
3882         spin_lock_irqsave(&nvme_fc_lock, flags);
3883         nvme_fc_waiting_to_unload = true;
3884         if (!list_empty(&nvme_fc_lport_list)) {
3885                 need_cleanup = true;
3886                 nvme_fc_cleanup_for_unload();
3887         }
3888         spin_unlock_irqrestore(&nvme_fc_lock, flags);
3889         if (need_cleanup) {
3890                 pr_info("%s: waiting for ctlr deletes\n", __func__);
3891                 wait_for_completion(&nvme_fc_unload_proceed);
3892                 pr_info("%s: ctrl deletes complete\n", __func__);
3893         }
3894
3895         nvmf_unregister_transport(&nvme_fc_transport);
3896
3897         ida_destroy(&nvme_fc_local_port_cnt);
3898         ida_destroy(&nvme_fc_ctrl_cnt);
3899
3900         device_destroy(&fc_class, MKDEV(0, 0));
3901         class_unregister(&fc_class);
3902         destroy_workqueue(nvme_fc_wq);
3903 }
3904
3905 module_init(nvme_fc_init_module);
3906 module_exit(nvme_fc_exit_module);
3907
3908 MODULE_LICENSE("GPL v2");