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 work_struct      ioerr_work;
 170         struct delayed_work     connect_work;
 171
 172         struct kref             ref;
 173         unsigned long           flags;
 174         u32                     iocnt;
 175         wait_queue_head_t       ioabort_wait;
 176
 177         struct nvme_fc_fcp_op   aen_ops[NVME_NR_AEN_COMMANDS];
 178
 179         struct nvme_ctrl        ctrl;
 180 };
 181
 182 static inline struct nvme_fc_ctrl *
 183 to_fc_ctrl(struct nvme_ctrl *ctrl)
 184 {
 185         return container_of(ctrl, struct nvme_fc_ctrl, ctrl);
 186 }
 187
 188 static inline struct nvme_fc_lport *
 189 localport_to_lport(struct nvme_fc_local_port *portptr)
 190 {
 191         return container_of(portptr, struct nvme_fc_lport, localport);
 192 }
 193
 194 static inline struct nvme_fc_rport *
 195 remoteport_to_rport(struct nvme_fc_remote_port *portptr)
 196 {
 197         return container_of(portptr, struct nvme_fc_rport, remoteport);
 198 }
 199
 200 static inline struct nvmefc_ls_req_op *
 201 ls_req_to_lsop(struct nvmefc_ls_req *lsreq)
 202 {
 203         return container_of(lsreq, struct nvmefc_ls_req_op, ls_req);
 204 }
 205
 206 static inline struct nvme_fc_fcp_op *
 207 fcp_req_to_fcp_op(struct nvmefc_fcp_req *fcpreq)
 208 {
 209         return container_of(fcpreq, struct nvme_fc_fcp_op, fcp_req);
 210 }
 211
 212
 213
 214 /* *************************** Globals **************************** */
 215
 216
 217 static DEFINE_SPINLOCK(nvme_fc_lock);
 218
 219 static LIST_HEAD(nvme_fc_lport_list);
 220 static DEFINE_IDA(nvme_fc_local_port_cnt);
 221 static DEFINE_IDA(nvme_fc_ctrl_cnt);
 222
 223 static struct workqueue_struct *nvme_fc_wq;
 224
 225 static bool nvme_fc_waiting_to_unload;
 226 static DECLARE_COMPLETION(nvme_fc_unload_proceed);
 227
 228 /*
 229  * These items are short-term. They will eventually be moved into
 230  * a generic FC class. See comments in module init.
 231  */
 232 static struct device *fc_udev_device;
 233
 234 static void nvme_fc_complete_rq(struct request *rq);
 235
 236 /* *********************** FC-NVME Port Management ************************ */
 237
 238 static void __nvme_fc_delete_hw_queue(struct nvme_fc_ctrl *,
 239                         struct nvme_fc_queue *, unsigned int);
 240
 241 static void nvme_fc_handle_ls_rqst_work(struct work_struct *work);
 242
 243
 244 static void
 245 nvme_fc_free_lport(struct kref *ref)
 246 {
 247         struct nvme_fc_lport *lport =
 248                 container_of(ref, struct nvme_fc_lport, ref);
 249         unsigned long flags;
 250
 251         WARN_ON(lport->localport.port_state != FC_OBJSTATE_DELETED);
 252         WARN_ON(!list_empty(&lport->endp_list));
 253
 254         /* remove from transport list */
 255         spin_lock_irqsave(&nvme_fc_lock, flags);
 256         list_del(&lport->port_list);
 257         if (nvme_fc_waiting_to_unload && list_empty(&nvme_fc_lport_list))
 258                 complete(&nvme_fc_unload_proceed);
 259         spin_unlock_irqrestore(&nvme_fc_lock, flags);
 260
 261         ida_simple_remove(&nvme_fc_local_port_cnt, lport->localport.port_num);
 262         ida_destroy(&lport->endp_cnt);
 263
 264         put_device(lport->dev);
 265
 266         kfree(lport);
 267 }
 268
 269 static void
 270 nvme_fc_lport_put(struct nvme_fc_lport *lport)
 271 {
 272         kref_put(&lport->ref, nvme_fc_free_lport);
 273 }
 274
 275 static int
 276 nvme_fc_lport_get(struct nvme_fc_lport *lport)
 277 {
 278         return kref_get_unless_zero(&lport->ref);
 279 }
 280
 281
 282 static struct nvme_fc_lport *
 283 nvme_fc_attach_to_unreg_lport(struct nvme_fc_port_info *pinfo,
 284                         struct nvme_fc_port_template *ops,
 285                         struct device *dev)
 286 {
 287         struct nvme_fc_lport *lport;
 288         unsigned long flags;
 289
 290         spin_lock_irqsave(&nvme_fc_lock, flags);
 291
 292         list_for_each_entry(lport, &nvme_fc_lport_list, port_list) {
 293                 if (lport->localport.node_name != pinfo->node_name ||
 294                     lport->localport.port_name != pinfo->port_name)
 295                         continue;
 296
 297                 if (lport->dev != dev) {
 298                         lport = ERR_PTR(-EXDEV);
 299                         goto out_done;
 300                 }
 301
 302                 if (lport->localport.port_state != FC_OBJSTATE_DELETED) {
 303                         lport = ERR_PTR(-EEXIST);
 304                         goto out_done;
 305                 }
 306
 307                 if (!nvme_fc_lport_get(lport)) {
 308                         /*
 309                          * fails if ref cnt already 0. If so,
 310                          * act as if lport already deleted
 311                          */
 312                         lport = NULL;
 313                         goto out_done;
 314                 }
 315
 316                 /* resume the lport */
 317
 318                 lport->ops = ops;
 319                 lport->localport.port_role = pinfo->port_role;
 320                 lport->localport.port_id = pinfo->port_id;
 321                 lport->localport.port_state = FC_OBJSTATE_ONLINE;
 322
 323                 spin_unlock_irqrestore(&nvme_fc_lock, flags);
 324
 325                 return lport;
 326         }
 327
 328         lport = NULL;
 329
 330 out_done:
 331         spin_unlock_irqrestore(&nvme_fc_lock, flags);
 332
 333         return lport;
 334 }
 335
 336 /**
 337  * nvme_fc_register_localport - transport entry point called by an
 338  *                              LLDD to register the existence of a NVME
 339  *                              host FC port.
 340  * @pinfo:     pointer to information about the port to be registered
 341  * @template:  LLDD entrypoints and operational parameters for the port
 342  * @dev:       physical hardware device node port corresponds to. Will be
 343  *             used for DMA mappings
 344  * @portptr:   pointer to a local port pointer. Upon success, the routine
 345  *             will allocate a nvme_fc_local_port structure and place its
 346  *             address in the local port pointer. Upon failure, local port
 347  *             pointer will be set to 0.
 348  *
 349  * Returns:
 350  * a completion status. Must be 0 upon success; a negative errno
 351  * (ex: -ENXIO) upon failure.
 352  */
 353 int
 354 nvme_fc_register_localport(struct nvme_fc_port_info *pinfo,
 355                         struct nvme_fc_port_template *template,
 356                         struct device *dev,
 357                         struct nvme_fc_local_port **portptr)
 358 {
 359         struct nvme_fc_lport *newrec;
 360         unsigned long flags;
 361         int ret, idx;
 362
 363         if (!template->localport_delete || !template->remoteport_delete ||
 364             !template->ls_req || !template->fcp_io ||
 365             !template->ls_abort || !template->fcp_abort ||
 366             !template->max_hw_queues || !template->max_sgl_segments ||
 367             !template->max_dif_sgl_segments || !template->dma_boundary) {
 368                 ret = -EINVAL;
 369                 goto out_reghost_failed;
 370         }
 371
 372         /*
 373          * look to see if there is already a localport that had been
 374          * deregistered and in the process of waiting for all the
 375          * references to fully be removed.  If the references haven't
 376          * expired, we can simply re-enable the localport. Remoteports
 377          * and controller reconnections should resume naturally.
 378          */
 379         newrec = nvme_fc_attach_to_unreg_lport(pinfo, template, dev);
 380
 381         /* found an lport, but something about its state is bad */
 382         if (IS_ERR(newrec)) {
 383                 ret = PTR_ERR(newrec);
 384                 goto out_reghost_failed;
 385
 386         /* found existing lport, which was resumed */
 387         } else if (newrec) {
 388                 *portptr = &newrec->localport;
 389                 return 0;
 390         }
 391
 392         /* nothing found - allocate a new localport struct */
 393
 394         newrec = kmalloc((sizeof(*newrec) + template->local_priv_sz),
 395                          GFP_KERNEL);
 396         if (!newrec) {
 397                 ret = -ENOMEM;
 398                 goto out_reghost_failed;
 399         }
 400
 401         idx = ida_simple_get(&nvme_fc_local_port_cnt, 0, 0, GFP_KERNEL);
 402         if (idx < 0) {
 403                 ret = -ENOSPC;
 404                 goto out_fail_kfree;
 405         }
 406
 407         if (!get_device(dev) && dev) {
 408                 ret = -ENODEV;
 409                 goto out_ida_put;
 410         }
 411
 412         INIT_LIST_HEAD(&newrec->port_list);
 413         INIT_LIST_HEAD(&newrec->endp_list);
 414         kref_init(&newrec->ref);
 415         atomic_set(&newrec->act_rport_cnt, 0);
 416         newrec->ops = template;
 417         newrec->dev = dev;
 418         ida_init(&newrec->endp_cnt);
 419         if (template->local_priv_sz)
 420                 newrec->localport.private = &newrec[1];
 421         else
 422                 newrec->localport.private = NULL;
 423         newrec->localport.node_name = pinfo->node_name;
 424         newrec->localport.port_name = pinfo->port_name;
 425         newrec->localport.port_role = pinfo->port_role;
 426         newrec->localport.port_id = pinfo->port_id;
 427         newrec->localport.port_state = FC_OBJSTATE_ONLINE;
 428         newrec->localport.port_num = idx;
 429
 430         spin_lock_irqsave(&nvme_fc_lock, flags);
 431         list_add_tail(&newrec->port_list, &nvme_fc_lport_list);
 432         spin_unlock_irqrestore(&nvme_fc_lock, flags);
 433
 434         if (dev)
 435                 dma_set_seg_boundary(dev, template->dma_boundary);
 436
 437         *portptr = &newrec->localport;
 438         return 0;
 439
 440 out_ida_put:
 441         ida_simple_remove(&nvme_fc_local_port_cnt, idx);
 442 out_fail_kfree:
 443         kfree(newrec);
 444 out_reghost_failed:
 445         *portptr = NULL;
 446
 447         return ret;
 448 }
 449 EXPORT_SYMBOL_GPL(nvme_fc_register_localport);
 450
 451 /**
 452  * nvme_fc_unregister_localport - transport entry point called by an
 453  *                              LLDD to deregister/remove a previously
 454  *                              registered a NVME host FC port.
 455  * @portptr: pointer to the (registered) local port that is to be deregistered.
 456  *
 457  * Returns:
 458  * a completion status. Must be 0 upon success; a negative errno
 459  * (ex: -ENXIO) upon failure.
 460  */
 461 int
 462 nvme_fc_unregister_localport(struct nvme_fc_local_port *portptr)
 463 {
 464         struct nvme_fc_lport *lport = localport_to_lport(portptr);
 465         unsigned long flags;
 466
 467         if (!portptr)
 468                 return -EINVAL;
 469
 470         spin_lock_irqsave(&nvme_fc_lock, flags);
 471
 472         if (portptr->port_state != FC_OBJSTATE_ONLINE) {
 473                 spin_unlock_irqrestore(&nvme_fc_lock, flags);
 474                 return -EINVAL;
 475         }
 476         portptr->port_state = FC_OBJSTATE_DELETED;
 477
 478         spin_unlock_irqrestore(&nvme_fc_lock, flags);
 479
 480         if (atomic_read(&lport->act_rport_cnt) == 0)
 481                 lport->ops->localport_delete(&lport->localport);
 482
 483         nvme_fc_lport_put(lport);
 484
 485         return 0;
 486 }
 487 EXPORT_SYMBOL_GPL(nvme_fc_unregister_localport);
 488
 489 /*
 490  * TRADDR strings, per FC-NVME are fixed format:
 491  *   "nn-0x<16hexdigits>:pn-0x<16hexdigits>" - 43 characters
 492  * udev event will only differ by prefix of what field is
 493  * being specified:
 494  *    "NVMEFC_HOST_TRADDR=" or "NVMEFC_TRADDR=" - 19 max characters
 495  *  19 + 43 + null_fudge = 64 characters
 496  */
 497 #define FCNVME_TRADDR_LENGTH            64
 498
 499 static void
 500 nvme_fc_signal_discovery_scan(struct nvme_fc_lport *lport,
 501                 struct nvme_fc_rport *rport)
 502 {
 503         char hostaddr[FCNVME_TRADDR_LENGTH];    /* NVMEFC_HOST_TRADDR=...*/
 504         char tgtaddr[FCNVME_TRADDR_LENGTH];     /* NVMEFC_TRADDR=...*/
 505         char *envp[4] = { "FC_EVENT=nvmediscovery", hostaddr, tgtaddr, NULL };
 506
 507         if (!(rport->remoteport.port_role & FC_PORT_ROLE_NVME_DISCOVERY))
 508                 return;
 509
 510         snprintf(hostaddr, sizeof(hostaddr),
 511                 "NVMEFC_HOST_TRADDR=nn-0x%016llx:pn-0x%016llx",
 512                 lport->localport.node_name, lport->localport.port_name);
 513         snprintf(tgtaddr, sizeof(tgtaddr),
 514                 "NVMEFC_TRADDR=nn-0x%016llx:pn-0x%016llx",
 515                 rport->remoteport.node_name, rport->remoteport.port_name);
 516         kobject_uevent_env(&fc_udev_device->kobj, KOBJ_CHANGE, envp);
 517 }
 518
 519 static void
 520 nvme_fc_free_rport(struct kref *ref)
 521 {
 522         struct nvme_fc_rport *rport =
 523                 container_of(ref, struct nvme_fc_rport, ref);
 524         struct nvme_fc_lport *lport =
 525                         localport_to_lport(rport->remoteport.localport);
 526         unsigned long flags;
 527
 528         WARN_ON(rport->remoteport.port_state != FC_OBJSTATE_DELETED);
 529         WARN_ON(!list_empty(&rport->ctrl_list));
 530
 531         /* remove from lport list */
 532         spin_lock_irqsave(&nvme_fc_lock, flags);
 533         list_del(&rport->endp_list);
 534         spin_unlock_irqrestore(&nvme_fc_lock, flags);
 535
 536         WARN_ON(!list_empty(&rport->disc_list));
 537         ida_simple_remove(&lport->endp_cnt, rport->remoteport.port_num);
 538
 539         kfree(rport);
 540
 541         nvme_fc_lport_put(lport);
 542 }
 543
 544 static void
 545 nvme_fc_rport_put(struct nvme_fc_rport *rport)
 546 {
 547         kref_put(&rport->ref, nvme_fc_free_rport);
 548 }
 549
 550 static int
 551 nvme_fc_rport_get(struct nvme_fc_rport *rport)
 552 {
 553         return kref_get_unless_zero(&rport->ref);
 554 }
 555
 556 static void
 557 nvme_fc_resume_controller(struct nvme_fc_ctrl *ctrl)
 558 {
 559         switch (ctrl->ctrl.state) {
 560         case NVME_CTRL_NEW:
 561         case NVME_CTRL_CONNECTING:
 562                 /*
 563                  * As all reconnects were suppressed, schedule a
 564                  * connect.
 565                  */
 566                 dev_info(ctrl->ctrl.device,
 567                         "NVME-FC{%d}: connectivity re-established. "
 568                         "Attempting reconnect\n", ctrl->cnum);
 569
 570                 queue_delayed_work(nvme_wq, &ctrl->connect_work, 0);
 571                 break;
 572
 573         case NVME_CTRL_RESETTING:
 574                 /*
 575                  * Controller is already in the process of terminating the
 576                  * association. No need to do anything further. The reconnect
 577                  * step will naturally occur after the reset completes.
 578                  */
 579                 break;
 580
 581         default:
 582                 /* no action to take - let it delete */
 583                 break;
 584         }
 585 }
 586
 587 static struct nvme_fc_rport *
 588 nvme_fc_attach_to_suspended_rport(struct nvme_fc_lport *lport,
 589                                 struct nvme_fc_port_info *pinfo)
 590 {
 591         struct nvme_fc_rport *rport;
 592         struct nvme_fc_ctrl *ctrl;
 593         unsigned long flags;
 594
 595         spin_lock_irqsave(&nvme_fc_lock, flags);
 596
 597         list_for_each_entry(rport, &lport->endp_list, endp_list) {
 598                 if (rport->remoteport.node_name != pinfo->node_name ||
 599                     rport->remoteport.port_name != pinfo->port_name)
 600                         continue;
 601
 602                 if (!nvme_fc_rport_get(rport)) {
 603                         rport = ERR_PTR(-ENOLCK);
 604                         goto out_done;
 605                 }
 606
 607                 spin_unlock_irqrestore(&nvme_fc_lock, flags);
 608
 609                 spin_lock_irqsave(&rport->lock, flags);
 610
 611                 /* has it been unregistered */
 612                 if (rport->remoteport.port_state != FC_OBJSTATE_DELETED) {
 613                         /* means lldd called us twice */
 614                         spin_unlock_irqrestore(&rport->lock, flags);
 615                         nvme_fc_rport_put(rport);
 616                         return ERR_PTR(-ESTALE);
 617                 }
 618
 619                 rport->remoteport.port_role = pinfo->port_role;
 620                 rport->remoteport.port_id = pinfo->port_id;
 621                 rport->remoteport.port_state = FC_OBJSTATE_ONLINE;
 622                 rport->dev_loss_end = 0;
 623
 624                 /*
 625                  * kick off a reconnect attempt on all associations to the
 626                  * remote port. A successful reconnects will resume i/o.
 627                  */
 628                 list_for_each_entry(ctrl, &rport->ctrl_list, ctrl_list)
 629                         nvme_fc_resume_controller(ctrl);
 630
 631                 spin_unlock_irqrestore(&rport->lock, flags);
 632
 633                 return rport;
 634         }
 635
 636         rport = NULL;
 637
 638 out_done:
 639         spin_unlock_irqrestore(&nvme_fc_lock, flags);
 640
 641         return rport;
 642 }
 643
 644 static inline void
 645 __nvme_fc_set_dev_loss_tmo(struct nvme_fc_rport *rport,
 646                         struct nvme_fc_port_info *pinfo)
 647 {
 648         if (pinfo->dev_loss_tmo)
 649                 rport->remoteport.dev_loss_tmo = pinfo->dev_loss_tmo;
 650         else
 651                 rport->remoteport.dev_loss_tmo = NVME_FC_DEFAULT_DEV_LOSS_TMO;
 652 }
 653
 654 /**
 655  * nvme_fc_register_remoteport - transport entry point called by an
 656  *                              LLDD to register the existence of a NVME
 657  *                              subsystem FC port on its fabric.
 658  * @localport: pointer to the (registered) local port that the remote
 659  *             subsystem port is connected to.
 660  * @pinfo:     pointer to information about the port to be registered
 661  * @portptr:   pointer to a remote port pointer. Upon success, the routine
 662  *             will allocate a nvme_fc_remote_port structure and place its
 663  *             address in the remote port pointer. Upon failure, remote port
 664  *             pointer will be set to 0.
 665  *
 666  * Returns:
 667  * a completion status. Must be 0 upon success; a negative errno
 668  * (ex: -ENXIO) upon failure.
 669  */
 670 int
 671 nvme_fc_register_remoteport(struct nvme_fc_local_port *localport,
 672                                 struct nvme_fc_port_info *pinfo,
 673                                 struct nvme_fc_remote_port **portptr)
 674 {
 675         struct nvme_fc_lport *lport = localport_to_lport(localport);
 676         struct nvme_fc_rport *newrec;
 677         unsigned long flags;
 678         int ret, idx;
 679
 680         if (!nvme_fc_lport_get(lport)) {
 681                 ret = -ESHUTDOWN;
 682                 goto out_reghost_failed;
 683         }
 684
 685         /*
 686          * look to see if there is already a remoteport that is waiting
 687          * for a reconnect (within dev_loss_tmo) with the same WWN's.
 688          * If so, transition to it and reconnect.
 689          */
 690         newrec = nvme_fc_attach_to_suspended_rport(lport, pinfo);
 691
 692         /* found an rport, but something about its state is bad */
 693         if (IS_ERR(newrec)) {
 694                 ret = PTR_ERR(newrec);
 695                 goto out_lport_put;
 696
 697         /* found existing rport, which was resumed */
 698         } else if (newrec) {
 699                 nvme_fc_lport_put(lport);
 700                 __nvme_fc_set_dev_loss_tmo(newrec, pinfo);
 701                 nvme_fc_signal_discovery_scan(lport, newrec);
 702                 *portptr = &newrec->remoteport;
 703                 return 0;
 704         }
 705
 706         /* nothing found - allocate a new remoteport struct */
 707
 708         newrec = kmalloc((sizeof(*newrec) + lport->ops->remote_priv_sz),
 709                          GFP_KERNEL);
 710         if (!newrec) {
 711                 ret = -ENOMEM;
 712                 goto out_lport_put;
 713         }
 714
 715         idx = ida_simple_get(&lport->endp_cnt, 0, 0, GFP_KERNEL);
 716         if (idx < 0) {
 717                 ret = -ENOSPC;
 718                 goto out_kfree_rport;
 719         }
 720
 721         INIT_LIST_HEAD(&newrec->endp_list);
 722         INIT_LIST_HEAD(&newrec->ctrl_list);
 723         INIT_LIST_HEAD(&newrec->ls_req_list);
 724         INIT_LIST_HEAD(&newrec->disc_list);
 725         kref_init(&newrec->ref);
 726         atomic_set(&newrec->act_ctrl_cnt, 0);
 727         spin_lock_init(&newrec->lock);
 728         newrec->remoteport.localport = &lport->localport;
 729         INIT_LIST_HEAD(&newrec->ls_rcv_list);
 730         newrec->dev = lport->dev;
 731         newrec->lport = lport;
 732         if (lport->ops->remote_priv_sz)
 733                 newrec->remoteport.private = &newrec[1];
 734         else
 735                 newrec->remoteport.private = NULL;
 736         newrec->remoteport.port_role = pinfo->port_role;
 737         newrec->remoteport.node_name = pinfo->node_name;
 738         newrec->remoteport.port_name = pinfo->port_name;
 739         newrec->remoteport.port_id = pinfo->port_id;
 740         newrec->remoteport.port_state = FC_OBJSTATE_ONLINE;
 741         newrec->remoteport.port_num = idx;
 742         __nvme_fc_set_dev_loss_tmo(newrec, pinfo);
 743         INIT_WORK(&newrec->lsrcv_work, nvme_fc_handle_ls_rqst_work);
 744
 745         spin_lock_irqsave(&nvme_fc_lock, flags);
 746         list_add_tail(&newrec->endp_list, &lport->endp_list);
 747         spin_unlock_irqrestore(&nvme_fc_lock, flags);
 748
 749         nvme_fc_signal_discovery_scan(lport, newrec);
 750
 751         *portptr = &newrec->remoteport;
 752         return 0;
 753
 754 out_kfree_rport:
 755         kfree(newrec);
 756 out_lport_put:
 757         nvme_fc_lport_put(lport);
 758 out_reghost_failed:
 759         *portptr = NULL;
 760         return ret;
 761 }
 762 EXPORT_SYMBOL_GPL(nvme_fc_register_remoteport);
 763
 764 static int
 765 nvme_fc_abort_lsops(struct nvme_fc_rport *rport)
 766 {
 767         struct nvmefc_ls_req_op *lsop;
 768         unsigned long flags;
 769
 770 restart:
 771         spin_lock_irqsave(&rport->lock, flags);
 772
 773         list_for_each_entry(lsop, &rport->ls_req_list, lsreq_list) {
 774                 if (!(lsop->flags & FCOP_FLAGS_TERMIO)) {
 775                         lsop->flags |= FCOP_FLAGS_TERMIO;
 776                         spin_unlock_irqrestore(&rport->lock, flags);
 777                         rport->lport->ops->ls_abort(&rport->lport->localport,
 778                                                 &rport->remoteport,
 779                                                 &lsop->ls_req);
 780                         goto restart;
 781                 }
 782         }
 783         spin_unlock_irqrestore(&rport->lock, flags);
 784
 785         return 0;
 786 }
 787
 788 static void
 789 nvme_fc_ctrl_connectivity_loss(struct nvme_fc_ctrl *ctrl)
 790 {
 791         dev_info(ctrl->ctrl.device,
 792                 "NVME-FC{%d}: controller connectivity lost. Awaiting "
 793                 "Reconnect", ctrl->cnum);
 794
 795         switch (ctrl->ctrl.state) {
 796         case NVME_CTRL_NEW:
 797         case NVME_CTRL_LIVE:
 798                 /*
 799                  * Schedule a controller reset. The reset will terminate the
 800                  * association and schedule the reconnect timer.  Reconnects
 801                  * will be attempted until either the ctlr_loss_tmo
 802                  * (max_retries * connect_delay) expires or the remoteport's
 803                  * dev_loss_tmo expires.
 804                  */
 805                 if (nvme_reset_ctrl(&ctrl->ctrl)) {
 806                         dev_warn(ctrl->ctrl.device,
 807                                 "NVME-FC{%d}: Couldn't schedule reset.\n",
 808                                 ctrl->cnum);
 809                         nvme_delete_ctrl(&ctrl->ctrl);
 810                 }
 811                 break;
 812
 813         case NVME_CTRL_CONNECTING:
 814                 /*
 815                  * The association has already been terminated and the
 816                  * controller is attempting reconnects.  No need to do anything
 817                  * futher.  Reconnects will be attempted until either the
 818                  * ctlr_loss_tmo (max_retries * connect_delay) expires or the
 819                  * remoteport's dev_loss_tmo expires.
 820                  */
 821                 break;
 822
 823         case NVME_CTRL_RESETTING:
 824                 /*
 825                  * Controller is already in the process of terminating the
 826                  * association.  No need to do anything further. The reconnect
 827                  * step will kick in naturally after the association is
 828                  * terminated.
 829                  */
 830                 break;
 831
 832         case NVME_CTRL_DELETING:
 833         case NVME_CTRL_DELETING_NOIO:
 834         default:
 835                 /* no action to take - let it delete */
 836                 break;
 837         }
 838 }
 839
 840 /**
 841  * nvme_fc_unregister_remoteport - transport entry point called by an
 842  *                              LLDD to deregister/remove a previously
 843  *                              registered a NVME subsystem FC port.
 844  * @portptr: pointer to the (registered) remote port that is to be
 845  *           deregistered.
 846  *
 847  * Returns:
 848  * a completion status. Must be 0 upon success; a negative errno
 849  * (ex: -ENXIO) upon failure.
 850  */
 851 int
 852 nvme_fc_unregister_remoteport(struct nvme_fc_remote_port *portptr)
 853 {
 854         struct nvme_fc_rport *rport = remoteport_to_rport(portptr);
 855         struct nvme_fc_ctrl *ctrl;
 856         unsigned long flags;
 857
 858         if (!portptr)
 859                 return -EINVAL;
 860
 861         spin_lock_irqsave(&rport->lock, flags);
 862
 863         if (portptr->port_state != FC_OBJSTATE_ONLINE) {
 864                 spin_unlock_irqrestore(&rport->lock, flags);
 865                 return -EINVAL;
 866         }
 867         portptr->port_state = FC_OBJSTATE_DELETED;
 868
 869         rport->dev_loss_end = jiffies + (portptr->dev_loss_tmo * HZ);
 870
 871         list_for_each_entry(ctrl, &rport->ctrl_list, ctrl_list) {
 872                 /* if dev_loss_tmo==0, dev loss is immediate */
 873                 if (!portptr->dev_loss_tmo) {
 874                         dev_warn(ctrl->ctrl.device,
 875                                 "NVME-FC{%d}: controller connectivity lost.\n",
 876                                 ctrl->cnum);
 877                         nvme_delete_ctrl(&ctrl->ctrl);
 878                 } else
 879                         nvme_fc_ctrl_connectivity_loss(ctrl);
 880         }
 881
 882         spin_unlock_irqrestore(&rport->lock, flags);
 883
 884         nvme_fc_abort_lsops(rport);
 885
 886         if (atomic_read(&rport->act_ctrl_cnt) == 0)
 887                 rport->lport->ops->remoteport_delete(portptr);
 888
 889         /*
 890          * release the reference, which will allow, if all controllers
 891          * go away, which should only occur after dev_loss_tmo occurs,
 892          * for the rport to be torn down.
 893          */
 894         nvme_fc_rport_put(rport);
 895
 896         return 0;
 897 }
 898 EXPORT_SYMBOL_GPL(nvme_fc_unregister_remoteport);
 899
 900 /**
 901  * nvme_fc_rescan_remoteport - transport entry point called by an
 902  *                              LLDD to request a nvme device rescan.
 903  * @remoteport: pointer to the (registered) remote port that is to be
 904  *              rescanned.
 905  *
 906  * Returns: N/A
 907  */
 908 void
 909 nvme_fc_rescan_remoteport(struct nvme_fc_remote_port *remoteport)
 910 {
 911         struct nvme_fc_rport *rport = remoteport_to_rport(remoteport);
 912
 913         nvme_fc_signal_discovery_scan(rport->lport, rport);
 914 }
 915 EXPORT_SYMBOL_GPL(nvme_fc_rescan_remoteport);
 916
 917 int
 918 nvme_fc_set_remoteport_devloss(struct nvme_fc_remote_port *portptr,
 919                         u32 dev_loss_tmo)
 920 {
 921         struct nvme_fc_rport *rport = remoteport_to_rport(portptr);
 922         unsigned long flags;
 923
 924         spin_lock_irqsave(&rport->lock, flags);
 925
 926         if (portptr->port_state != FC_OBJSTATE_ONLINE) {
 927                 spin_unlock_irqrestore(&rport->lock, flags);
 928                 return -EINVAL;
 929         }
 930
 931         /* a dev_loss_tmo of 0 (immediate) is allowed to be set */
 932         rport->remoteport.dev_loss_tmo = dev_loss_tmo;
 933
 934         spin_unlock_irqrestore(&rport->lock, flags);
 935
 936         return 0;
 937 }
 938 EXPORT_SYMBOL_GPL(nvme_fc_set_remoteport_devloss);
 939
 940
 941 /* *********************** FC-NVME DMA Handling **************************** */
 942
 943 /*
 944  * The fcloop device passes in a NULL device pointer. Real LLD's will
 945  * pass in a valid device pointer. If NULL is passed to the dma mapping
 946  * routines, depending on the platform, it may or may not succeed, and
 947  * may crash.
 948  *
 949  * As such:
 950  * Wrapper all the dma routines and check the dev pointer.
 951  *
 952  * If simple mappings (return just a dma address, we'll noop them,
 953  * returning a dma address of 0.
 954  *
 955  * On more complex mappings (dma_map_sg), a pseudo routine fills
 956  * in the scatter list, setting all dma addresses to 0.
 957  */
 958
 959 static inline dma_addr_t
 960 fc_dma_map_single(struct device *dev, void *ptr, size_t size,
 961                 enum dma_data_direction dir)
 962 {
 963         return dev ? dma_map_single(dev, ptr, size, dir) : (dma_addr_t)0L;
 964 }
 965
 966 static inline int
 967 fc_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
 968 {
 969         return dev ? dma_mapping_error(dev, dma_addr) : 0;
 970 }
 971
 972 static inline void
 973 fc_dma_unmap_single(struct device *dev, dma_addr_t addr, size_t size,
 974         enum dma_data_direction dir)
 975 {
 976         if (dev)
 977                 dma_unmap_single(dev, addr, size, dir);
 978 }
 979
 980 static inline void
 981 fc_dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr, size_t size,
 982                 enum dma_data_direction dir)
 983 {
 984         if (dev)
 985                 dma_sync_single_for_cpu(dev, addr, size, dir);
 986 }
 987
 988 static inline void
 989 fc_dma_sync_single_for_device(struct device *dev, dma_addr_t addr, size_t size,
 990                 enum dma_data_direction dir)
 991 {
 992         if (dev)
 993                 dma_sync_single_for_device(dev, addr, size, dir);
 994 }
 995
 996 /* pseudo dma_map_sg call */
 997 static int
 998 fc_map_sg(struct scatterlist *sg, int nents)
 999 {
1000         struct scatterlist *s;
1001         int i;
1002
1003         WARN_ON(nents == 0 || sg[0].length == 0);
1004
1005         for_each_sg(sg, s, nents, i) {
1006                 s->dma_address = 0L;
1007 #ifdef CONFIG_NEED_SG_DMA_LENGTH
1008                 s->dma_length = s->length;
1009 #endif
1010         }
1011         return nents;
1012 }
1013
1014 static inline int
1015 fc_dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
1016                 enum dma_data_direction dir)
1017 {
1018         return dev ? dma_map_sg(dev, sg, nents, dir) : fc_map_sg(sg, nents);
1019 }
1020
1021 static inline void
1022 fc_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
1023                 enum dma_data_direction dir)
1024 {
1025         if (dev)
1026                 dma_unmap_sg(dev, sg, nents, dir);
1027 }
1028
1029 /* *********************** FC-NVME LS Handling **************************** */
1030
1031 static void nvme_fc_ctrl_put(struct nvme_fc_ctrl *);
1032 static int nvme_fc_ctrl_get(struct nvme_fc_ctrl *);
1033
1034 static void nvme_fc_error_recovery(struct nvme_fc_ctrl *ctrl, char *errmsg);
1035
1036 static void
1037 __nvme_fc_finish_ls_req(struct nvmefc_ls_req_op *lsop)
1038 {
1039         struct nvme_fc_rport *rport = lsop->rport;
1040         struct nvmefc_ls_req *lsreq = &lsop->ls_req;
1041         unsigned long flags;
1042
1043         spin_lock_irqsave(&rport->lock, flags);
1044
1045         if (!lsop->req_queued) {
1046                 spin_unlock_irqrestore(&rport->lock, flags);
1047                 return;
1048         }
1049
1050         list_del(&lsop->lsreq_list);
1051
1052         lsop->req_queued = false;
1053
1054         spin_unlock_irqrestore(&rport->lock, flags);
1055
1056         fc_dma_unmap_single(rport->dev, lsreq->rqstdma,
1057                                   (lsreq->rqstlen + lsreq->rsplen),
1058                                   DMA_BIDIRECTIONAL);
1059
1060         nvme_fc_rport_put(rport);
1061 }
1062
1063 static int
1064 __nvme_fc_send_ls_req(struct nvme_fc_rport *rport,
1065                 struct nvmefc_ls_req_op *lsop,
1066                 void (*done)(struct nvmefc_ls_req *req, int status))
1067 {
1068         struct nvmefc_ls_req *lsreq = &lsop->ls_req;
1069         unsigned long flags;
1070         int ret = 0;
1071
1072         if (rport->remoteport.port_state != FC_OBJSTATE_ONLINE)
1073                 return -ECONNREFUSED;
1074
1075         if (!nvme_fc_rport_get(rport))
1076                 return -ESHUTDOWN;
1077
1078         lsreq->done = done;
1079         lsop->rport = rport;
1080         lsop->req_queued = false;
1081         INIT_LIST_HEAD(&lsop->lsreq_list);
1082         init_completion(&lsop->ls_done);
1083
1084         lsreq->rqstdma = fc_dma_map_single(rport->dev, lsreq->rqstaddr,
1085                                   lsreq->rqstlen + lsreq->rsplen,
1086                                   DMA_BIDIRECTIONAL);
1087         if (fc_dma_mapping_error(rport->dev, lsreq->rqstdma)) {
1088                 ret = -EFAULT;
1089                 goto out_putrport;
1090         }
1091         lsreq->rspdma = lsreq->rqstdma + lsreq->rqstlen;
1092
1093         spin_lock_irqsave(&rport->lock, flags);
1094
1095         list_add_tail(&lsop->lsreq_list, &rport->ls_req_list);
1096
1097         lsop->req_queued = true;
1098
1099         spin_unlock_irqrestore(&rport->lock, flags);
1100
1101         ret = rport->lport->ops->ls_req(&rport->lport->localport,
1102                                         &rport->remoteport, lsreq);
1103         if (ret)
1104                 goto out_unlink;
1105
1106         return 0;
1107
1108 out_unlink:
1109         lsop->ls_error = ret;
1110         spin_lock_irqsave(&rport->lock, flags);
1111         lsop->req_queued = false;
1112         list_del(&lsop->lsreq_list);
1113         spin_unlock_irqrestore(&rport->lock, flags);
1114         fc_dma_unmap_single(rport->dev, lsreq->rqstdma,
1115                                   (lsreq->rqstlen + lsreq->rsplen),
1116                                   DMA_BIDIRECTIONAL);
1117 out_putrport:
1118         nvme_fc_rport_put(rport);
1119
1120         return ret;
1121 }
1122
1123 static void
1124 nvme_fc_send_ls_req_done(struct nvmefc_ls_req *lsreq, int status)
1125 {
1126         struct nvmefc_ls_req_op *lsop = ls_req_to_lsop(lsreq);
1127
1128         lsop->ls_error = status;
1129         complete(&lsop->ls_done);
1130 }
1131
1132 static int
1133 nvme_fc_send_ls_req(struct nvme_fc_rport *rport, struct nvmefc_ls_req_op *lsop)
1134 {
1135         struct nvmefc_ls_req *lsreq = &lsop->ls_req;
1136         struct fcnvme_ls_rjt *rjt = lsreq->rspaddr;
1137         int ret;
1138
1139         ret = __nvme_fc_send_ls_req(rport, lsop, nvme_fc_send_ls_req_done);
1140
1141         if (!ret) {
1142                 /*
1143                  * No timeout/not interruptible as we need the struct
1144                  * to exist until the lldd calls us back. Thus mandate
1145                  * wait until driver calls back. lldd responsible for
1146                  * the timeout action
1147                  */
1148                 wait_for_completion(&lsop->ls_done);
1149
1150                 __nvme_fc_finish_ls_req(lsop);
1151
1152                 ret = lsop->ls_error;
1153         }
1154
1155         if (ret)
1156                 return ret;
1157
1158         /* ACC or RJT payload ? */
1159         if (rjt->w0.ls_cmd == FCNVME_LS_RJT)
1160                 return -ENXIO;
1161
1162         return 0;
1163 }
1164
1165 static int
1166 nvme_fc_send_ls_req_async(struct nvme_fc_rport *rport,
1167                 struct nvmefc_ls_req_op *lsop,
1168                 void (*done)(struct nvmefc_ls_req *req, int status))
1169 {
1170         /* don't wait for completion */
1171
1172         return __nvme_fc_send_ls_req(rport, lsop, done);
1173 }
1174
1175 static int
1176 nvme_fc_connect_admin_queue(struct nvme_fc_ctrl *ctrl,
1177         struct nvme_fc_queue *queue, u16 qsize, u16 ersp_ratio)
1178 {
1179         struct nvmefc_ls_req_op *lsop;
1180         struct nvmefc_ls_req *lsreq;
1181         struct fcnvme_ls_cr_assoc_rqst *assoc_rqst;
1182         struct fcnvme_ls_cr_assoc_acc *assoc_acc;
1183         unsigned long flags;
1184         int ret, fcret = 0;
1185
1186         lsop = kzalloc((sizeof(*lsop) +
1187                          sizeof(*assoc_rqst) + sizeof(*assoc_acc) +
1188                          ctrl->lport->ops->lsrqst_priv_sz), GFP_KERNEL);
1189         if (!lsop) {
1190                 dev_info(ctrl->ctrl.device,
1191                         "NVME-FC{%d}: send Create Association failed: ENOMEM\n",
1192                         ctrl->cnum);
1193                 ret = -ENOMEM;
1194                 goto out_no_memory;
1195         }
1196
1197         assoc_rqst = (struct fcnvme_ls_cr_assoc_rqst *)&lsop[1];
1198         assoc_acc = (struct fcnvme_ls_cr_assoc_acc *)&assoc_rqst[1];
1199         lsreq = &lsop->ls_req;
1200         if (ctrl->lport->ops->lsrqst_priv_sz)
1201                 lsreq->private = &assoc_acc[1];
1202         else
1203                 lsreq->private = NULL;
1204
1205         assoc_rqst->w0.ls_cmd = FCNVME_LS_CREATE_ASSOCIATION;
1206         assoc_rqst->desc_list_len =
1207                         cpu_to_be32(sizeof(struct fcnvme_lsdesc_cr_assoc_cmd));
1208
1209         assoc_rqst->assoc_cmd.desc_tag =
1210                         cpu_to_be32(FCNVME_LSDESC_CREATE_ASSOC_CMD);
1211         assoc_rqst->assoc_cmd.desc_len =
1212                         fcnvme_lsdesc_len(
1213                                 sizeof(struct fcnvme_lsdesc_cr_assoc_cmd));
1214
1215         assoc_rqst->assoc_cmd.ersp_ratio = cpu_to_be16(ersp_ratio);
1216         assoc_rqst->assoc_cmd.sqsize = cpu_to_be16(qsize - 1);
1217         /* Linux supports only Dynamic controllers */
1218         assoc_rqst->assoc_cmd.cntlid = cpu_to_be16(0xffff);
1219         uuid_copy(&assoc_rqst->assoc_cmd.hostid, &ctrl->ctrl.opts->host->id);
1220         strncpy(assoc_rqst->assoc_cmd.hostnqn, ctrl->ctrl.opts->host->nqn,
1221                 min(FCNVME_ASSOC_HOSTNQN_LEN, NVMF_NQN_SIZE));
1222         strncpy(assoc_rqst->assoc_cmd.subnqn, ctrl->ctrl.opts->subsysnqn,
1223                 min(FCNVME_ASSOC_SUBNQN_LEN, NVMF_NQN_SIZE));
1224
1225         lsop->queue = queue;
1226         lsreq->rqstaddr = assoc_rqst;
1227         lsreq->rqstlen = sizeof(*assoc_rqst);
1228         lsreq->rspaddr = assoc_acc;
1229         lsreq->rsplen = sizeof(*assoc_acc);
1230         lsreq->timeout = NVME_FC_LS_TIMEOUT_SEC;
1231
1232         ret = nvme_fc_send_ls_req(ctrl->rport, lsop);
1233         if (ret)
1234                 goto out_free_buffer;
1235
1236         /* process connect LS completion */
1237
1238         /* validate the ACC response */
1239         if (assoc_acc->hdr.w0.ls_cmd != FCNVME_LS_ACC)
1240                 fcret = VERR_LSACC;
1241         else if (assoc_acc->hdr.desc_list_len !=
1242                         fcnvme_lsdesc_len(
1243                                 sizeof(struct fcnvme_ls_cr_assoc_acc)))
1244                 fcret = VERR_CR_ASSOC_ACC_LEN;
1245         else if (assoc_acc->hdr.rqst.desc_tag !=
1246                         cpu_to_be32(FCNVME_LSDESC_RQST))
1247                 fcret = VERR_LSDESC_RQST;
1248         else if (assoc_acc->hdr.rqst.desc_len !=
1249                         fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_rqst)))
1250                 fcret = VERR_LSDESC_RQST_LEN;
1251         else if (assoc_acc->hdr.rqst.w0.ls_cmd != FCNVME_LS_CREATE_ASSOCIATION)
1252                 fcret = VERR_CR_ASSOC;
1253         else if (assoc_acc->associd.desc_tag !=
1254                         cpu_to_be32(FCNVME_LSDESC_ASSOC_ID))
1255                 fcret = VERR_ASSOC_ID;
1256         else if (assoc_acc->associd.desc_len !=
1257                         fcnvme_lsdesc_len(
1258                                 sizeof(struct fcnvme_lsdesc_assoc_id)))
1259                 fcret = VERR_ASSOC_ID_LEN;
1260         else if (assoc_acc->connectid.desc_tag !=
1261                         cpu_to_be32(FCNVME_LSDESC_CONN_ID))
1262                 fcret = VERR_CONN_ID;
1263         else if (assoc_acc->connectid.desc_len !=
1264                         fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_conn_id)))
1265                 fcret = VERR_CONN_ID_LEN;
1266
1267         if (fcret) {
1268                 ret = -EBADF;
1269                 dev_err(ctrl->dev,
1270                         "q %d Create Association LS failed: %s\n",
1271                         queue->qnum, validation_errors[fcret]);
1272         } else {
1273                 spin_lock_irqsave(&ctrl->lock, flags);
1274                 ctrl->association_id =
1275                         be64_to_cpu(assoc_acc->associd.association_id);
1276                 queue->connection_id =
1277                         be64_to_cpu(assoc_acc->connectid.connection_id);
1278                 set_bit(NVME_FC_Q_CONNECTED, &queue->flags);
1279                 spin_unlock_irqrestore(&ctrl->lock, flags);
1280         }
1281
1282 out_free_buffer:
1283         kfree(lsop);
1284 out_no_memory:
1285         if (ret)
1286                 dev_err(ctrl->dev,
1287                         "queue %d connect admin queue failed (%d).\n",
1288                         queue->qnum, ret);
1289         return ret;
1290 }
1291
1292 static int
1293 nvme_fc_connect_queue(struct nvme_fc_ctrl *ctrl, struct nvme_fc_queue *queue,
1294                         u16 qsize, u16 ersp_ratio)
1295 {
1296         struct nvmefc_ls_req_op *lsop;
1297         struct nvmefc_ls_req *lsreq;
1298         struct fcnvme_ls_cr_conn_rqst *conn_rqst;
1299         struct fcnvme_ls_cr_conn_acc *conn_acc;
1300         int ret, fcret = 0;
1301
1302         lsop = kzalloc((sizeof(*lsop) +
1303                          sizeof(*conn_rqst) + sizeof(*conn_acc) +
1304                          ctrl->lport->ops->lsrqst_priv_sz), GFP_KERNEL);
1305         if (!lsop) {
1306                 dev_info(ctrl->ctrl.device,
1307                         "NVME-FC{%d}: send Create Connection failed: ENOMEM\n",
1308                         ctrl->cnum);
1309                 ret = -ENOMEM;
1310                 goto out_no_memory;
1311         }
1312
1313         conn_rqst = (struct fcnvme_ls_cr_conn_rqst *)&lsop[1];
1314         conn_acc = (struct fcnvme_ls_cr_conn_acc *)&conn_rqst[1];
1315         lsreq = &lsop->ls_req;
1316         if (ctrl->lport->ops->lsrqst_priv_sz)
1317                 lsreq->private = (void *)&conn_acc[1];
1318         else
1319                 lsreq->private = NULL;
1320
1321         conn_rqst->w0.ls_cmd = FCNVME_LS_CREATE_CONNECTION;
1322         conn_rqst->desc_list_len = cpu_to_be32(
1323                                 sizeof(struct fcnvme_lsdesc_assoc_id) +
1324                                 sizeof(struct fcnvme_lsdesc_cr_conn_cmd));
1325
1326         conn_rqst->associd.desc_tag = cpu_to_be32(FCNVME_LSDESC_ASSOC_ID);
1327         conn_rqst->associd.desc_len =
1328                         fcnvme_lsdesc_len(
1329                                 sizeof(struct fcnvme_lsdesc_assoc_id));
1330         conn_rqst->associd.association_id = cpu_to_be64(ctrl->association_id);
1331         conn_rqst->connect_cmd.desc_tag =
1332                         cpu_to_be32(FCNVME_LSDESC_CREATE_CONN_CMD);
1333         conn_rqst->connect_cmd.desc_len =
1334                         fcnvme_lsdesc_len(
1335                                 sizeof(struct fcnvme_lsdesc_cr_conn_cmd));
1336         conn_rqst->connect_cmd.ersp_ratio = cpu_to_be16(ersp_ratio);
1337         conn_rqst->connect_cmd.qid  = cpu_to_be16(queue->qnum);
1338         conn_rqst->connect_cmd.sqsize = cpu_to_be16(qsize - 1);
1339
1340         lsop->queue = queue;
1341         lsreq->rqstaddr = conn_rqst;
1342         lsreq->rqstlen = sizeof(*conn_rqst);
1343         lsreq->rspaddr = conn_acc;
1344         lsreq->rsplen = sizeof(*conn_acc);
1345         lsreq->timeout = NVME_FC_LS_TIMEOUT_SEC;
1346
1347         ret = nvme_fc_send_ls_req(ctrl->rport, lsop);
1348         if (ret)
1349                 goto out_free_buffer;
1350
1351         /* process connect LS completion */
1352
1353         /* validate the ACC response */
1354         if (conn_acc->hdr.w0.ls_cmd != FCNVME_LS_ACC)
1355                 fcret = VERR_LSACC;
1356         else if (conn_acc->hdr.desc_list_len !=
1357                         fcnvme_lsdesc_len(sizeof(struct fcnvme_ls_cr_conn_acc)))
1358                 fcret = VERR_CR_CONN_ACC_LEN;
1359         else if (conn_acc->hdr.rqst.desc_tag != cpu_to_be32(FCNVME_LSDESC_RQST))
1360                 fcret = VERR_LSDESC_RQST;
1361         else if (conn_acc->hdr.rqst.desc_len !=
1362                         fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_rqst)))
1363                 fcret = VERR_LSDESC_RQST_LEN;
1364         else if (conn_acc->hdr.rqst.w0.ls_cmd != FCNVME_LS_CREATE_CONNECTION)
1365                 fcret = VERR_CR_CONN;
1366         else if (conn_acc->connectid.desc_tag !=
1367                         cpu_to_be32(FCNVME_LSDESC_CONN_ID))
1368                 fcret = VERR_CONN_ID;
1369         else if (conn_acc->connectid.desc_len !=
1370                         fcnvme_lsdesc_len(sizeof(struct fcnvme_lsdesc_conn_id)))
1371                 fcret = VERR_CONN_ID_LEN;
1372
1373         if (fcret) {
1374                 ret = -EBADF;
1375                 dev_err(ctrl->dev,
1376                         "q %d Create I/O Connection LS failed: %s\n",
1377                         queue->qnum, validation_errors[fcret]);
1378         } else {
1379                 queue->connection_id =
1380                         be64_to_cpu(conn_acc->connectid.connection_id);
1381                 set_bit(NVME_FC_Q_CONNECTED, &queue->flags);
1382         }
1383
1384 out_free_buffer:
1385         kfree(lsop);
1386 out_no_memory:
1387         if (ret)
1388                 dev_err(ctrl->dev,
1389                         "queue %d connect I/O queue failed (%d).\n",
1390                         queue->qnum, ret);
1391         return ret;
1392 }
1393
1394 static void
1395 nvme_fc_disconnect_assoc_done(struct nvmefc_ls_req *lsreq, int status)
1396 {
1397         struct nvmefc_ls_req_op *lsop = ls_req_to_lsop(lsreq);
1398
1399         __nvme_fc_finish_ls_req(lsop);
1400
1401         /* fc-nvme initiator doesn't care about success or failure of cmd */
1402
1403         kfree(lsop);
1404 }
1405
1406 /*
1407  * This routine sends a FC-NVME LS to disconnect (aka terminate)
1408  * the FC-NVME Association.  Terminating the association also
1409  * terminates the FC-NVME connections (per queue, both admin and io
1410  * queues) that are part of the association. E.g. things are torn
1411  * down, and the related FC-NVME Association ID and Connection IDs
1412  * become invalid.
1413  *
1414  * The behavior of the fc-nvme initiator is such that it's
1415  * understanding of the association and connections will implicitly
1416  * be torn down. The action is implicit as it may be due to a loss of
1417  * connectivity with the fc-nvme target, so you may never get a
1418  * response even if you tried.  As such, the action of this routine
1419  * is to asynchronously send the LS, ignore any results of the LS, and
1420  * continue on with terminating the association. If the fc-nvme target
1421  * is present and receives the LS, it too can tear down.
1422  */
1423 static void
1424 nvme_fc_xmt_disconnect_assoc(struct nvme_fc_ctrl *ctrl)
1425 {
1426         struct fcnvme_ls_disconnect_assoc_rqst *discon_rqst;
1427         struct fcnvme_ls_disconnect_assoc_acc *discon_acc;
1428         struct nvmefc_ls_req_op *lsop;
1429         struct nvmefc_ls_req *lsreq;
1430         int ret;
1431
1432         lsop = kzalloc((sizeof(*lsop) +
1433                         sizeof(*discon_rqst) + sizeof(*discon_acc) +
1434                         ctrl->lport->ops->lsrqst_priv_sz), GFP_KERNEL);
1435         if (!lsop) {
1436                 dev_info(ctrl->ctrl.device,
1437                         "NVME-FC{%d}: send Disconnect Association "
1438                         "failed: ENOMEM\n",
1439                         ctrl->cnum);
1440                 return;
1441         }
1442
1443         discon_rqst = (struct fcnvme_ls_disconnect_assoc_rqst *)&lsop[1];
1444         discon_acc = (struct fcnvme_ls_disconnect_assoc_acc *)&discon_rqst[1];
1445         lsreq = &lsop->ls_req;
1446         if (ctrl->lport->ops->lsrqst_priv_sz)
1447                 lsreq->private = (void *)&discon_acc[1];
1448         else
1449                 lsreq->private = NULL;
1450
1451         nvmefc_fmt_lsreq_discon_assoc(lsreq, discon_rqst, discon_acc,
1452                                 ctrl->association_id);
1453
1454         ret = nvme_fc_send_ls_req_async(ctrl->rport, lsop,
1455                                 nvme_fc_disconnect_assoc_done);
1456         if (ret)
1457                 kfree(lsop);
1458 }
1459
1460 static void
1461 nvme_fc_xmt_ls_rsp_done(struct nvmefc_ls_rsp *lsrsp)
1462 {
1463         struct nvmefc_ls_rcv_op *lsop = lsrsp->nvme_fc_private;
1464         struct nvme_fc_rport *rport = lsop->rport;
1465         struct nvme_fc_lport *lport = rport->lport;
1466         unsigned long flags;
1467
1468         spin_lock_irqsave(&rport->lock, flags);
1469         list_del(&lsop->lsrcv_list);
1470         spin_unlock_irqrestore(&rport->lock, flags);
1471
1472         fc_dma_sync_single_for_cpu(lport->dev, lsop->rspdma,
1473                                 sizeof(*lsop->rspbuf), DMA_TO_DEVICE);
1474         fc_dma_unmap_single(lport->dev, lsop->rspdma,
1475                         sizeof(*lsop->rspbuf), DMA_TO_DEVICE);
1476
1477         kfree(lsop);
1478
1479         nvme_fc_rport_put(rport);
1480 }
1481
1482 static void
1483 nvme_fc_xmt_ls_rsp(struct nvmefc_ls_rcv_op *lsop)
1484 {
1485         struct nvme_fc_rport *rport = lsop->rport;
1486         struct nvme_fc_lport *lport = rport->lport;
1487         struct fcnvme_ls_rqst_w0 *w0 = &lsop->rqstbuf->w0;
1488         int ret;
1489
1490         fc_dma_sync_single_for_device(lport->dev, lsop->rspdma,
1491                                   sizeof(*lsop->rspbuf), DMA_TO_DEVICE);
1492
1493         ret = lport->ops->xmt_ls_rsp(&lport->localport, &rport->remoteport,
1494                                      lsop->lsrsp);
1495         if (ret) {
1496                 dev_warn(lport->dev,
1497                         "LLDD rejected LS RSP xmt: LS %d status %d\n",
1498                         w0->ls_cmd, ret);
1499                 nvme_fc_xmt_ls_rsp_done(lsop->lsrsp);
1500                 return;
1501         }
1502 }
1503
1504 static struct nvme_fc_ctrl *
1505 nvme_fc_match_disconn_ls(struct nvme_fc_rport *rport,
1506                       struct nvmefc_ls_rcv_op *lsop)
1507 {
1508         struct fcnvme_ls_disconnect_assoc_rqst *rqst =
1509                                         &lsop->rqstbuf->rq_dis_assoc;
1510         struct nvme_fc_ctrl *ctrl, *ret = NULL;
1511         struct nvmefc_ls_rcv_op *oldls = NULL;
1512         u64 association_id = be64_to_cpu(rqst->associd.association_id);
1513         unsigned long flags;
1514
1515         spin_lock_irqsave(&rport->lock, flags);
1516
1517         list_for_each_entry(ctrl, &rport->ctrl_list, ctrl_list) {
1518                 if (!nvme_fc_ctrl_get(ctrl))
1519                         continue;
1520                 spin_lock(&ctrl->lock);
1521                 if (association_id == ctrl->association_id) {
1522                         oldls = ctrl->rcv_disconn;
1523                         ctrl->rcv_disconn = lsop;
1524                         ret = ctrl;
1525                 }
1526                 spin_unlock(&ctrl->lock);
1527                 if (ret)
1528                         /* leave the ctrl get reference */
1529                         break;
1530                 nvme_fc_ctrl_put(ctrl);
1531         }
1532
1533         spin_unlock_irqrestore(&rport->lock, flags);
1534
1535         /* transmit a response for anything that was pending */
1536         if (oldls) {
1537                 dev_info(rport->lport->dev,
1538                         "NVME-FC{%d}: Multiple Disconnect Association "
1539                         "LS's received\n", ctrl->cnum);
1540                 /* overwrite good response with bogus failure */
1541                 oldls->lsrsp->rsplen = nvme_fc_format_rjt(oldls->rspbuf,
1542                                                 sizeof(*oldls->rspbuf),
1543                                                 rqst->w0.ls_cmd,
1544                                                 FCNVME_RJT_RC_UNAB,
1545                                                 FCNVME_RJT_EXP_NONE, 0);
1546                 nvme_fc_xmt_ls_rsp(oldls);
1547         }
1548
1549         return ret;
1550 }
1551
1552 /*
1553  * returns true to mean LS handled and ls_rsp can be sent
1554  * returns false to defer ls_rsp xmt (will be done as part of
1555  *     association termination)
1556  */
1557 static bool
1558 nvme_fc_ls_disconnect_assoc(struct nvmefc_ls_rcv_op *lsop)
1559 {
1560         struct nvme_fc_rport *rport = lsop->rport;
1561         struct fcnvme_ls_disconnect_assoc_rqst *rqst =
1562                                         &lsop->rqstbuf->rq_dis_assoc;
1563         struct fcnvme_ls_disconnect_assoc_acc *acc =
1564                                         &lsop->rspbuf->rsp_dis_assoc;
1565         struct nvme_fc_ctrl *ctrl = NULL;
1566         int ret = 0;
1567
1568         memset(acc, 0, sizeof(*acc));
1569
1570         ret = nvmefc_vldt_lsreq_discon_assoc(lsop->rqstdatalen, rqst);
1571         if (!ret) {
1572                 /* match an active association */
1573                 ctrl = nvme_fc_match_disconn_ls(rport, lsop);
1574                 if (!ctrl)
1575                         ret = VERR_NO_ASSOC;
1576         }
1577
1578         if (ret) {
1579                 dev_info(rport->lport->dev,
1580                         "Disconnect LS failed: %s\n",
1581                         validation_errors[ret]);
1582                 lsop->lsrsp->rsplen = nvme_fc_format_rjt(acc,
1583                                         sizeof(*acc), rqst->w0.ls_cmd,
1584                                         (ret == VERR_NO_ASSOC) ?
1585                                                 FCNVME_RJT_RC_INV_ASSOC :
1586                                                 FCNVME_RJT_RC_LOGIC,
1587                                         FCNVME_RJT_EXP_NONE, 0);
1588                 return true;
1589         }
1590
1591         /* format an ACCept response */
1592
1593         lsop->lsrsp->rsplen = sizeof(*acc);
1594
1595         nvme_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1596                         fcnvme_lsdesc_len(
1597                                 sizeof(struct fcnvme_ls_disconnect_assoc_acc)),
1598                         FCNVME_LS_DISCONNECT_ASSOC);
1599
1600         /*
1601          * the transmit of the response will occur after the exchanges
1602          * for the association have been ABTS'd by
1603          * nvme_fc_delete_association().
1604          */
1605
1606         /* fail the association */
1607         nvme_fc_error_recovery(ctrl, "Disconnect Association LS received");
1608
1609         /* release the reference taken by nvme_fc_match_disconn_ls() */
1610         nvme_fc_ctrl_put(ctrl);
1611
1612         return false;
1613 }
1614
1615 /*
1616  * Actual Processing routine for received FC-NVME LS Requests from the LLD
1617  * returns true if a response should be sent afterward, false if rsp will
1618  * be sent asynchronously.
1619  */
1620 static bool
1621 nvme_fc_handle_ls_rqst(struct nvmefc_ls_rcv_op *lsop)
1622 {
1623         struct fcnvme_ls_rqst_w0 *w0 = &lsop->rqstbuf->w0;
1624         bool ret = true;
1625
1626         lsop->lsrsp->nvme_fc_private = lsop;
1627         lsop->lsrsp->rspbuf = lsop->rspbuf;
1628         lsop->lsrsp->rspdma = lsop->rspdma;
1629         lsop->lsrsp->done = nvme_fc_xmt_ls_rsp_done;
1630         /* Be preventative. handlers will later set to valid length */
1631         lsop->lsrsp->rsplen = 0;
1632
1633         /*
1634          * handlers:
1635          *   parse request input, execute the request, and format the
1636          *   LS response
1637          */
1638         switch (w0->ls_cmd) {
1639         case FCNVME_LS_DISCONNECT_ASSOC:
1640                 ret = nvme_fc_ls_disconnect_assoc(lsop);
1641                 break;
1642         case FCNVME_LS_DISCONNECT_CONN:
1643                 lsop->lsrsp->rsplen = nvme_fc_format_rjt(lsop->rspbuf,
1644                                 sizeof(*lsop->rspbuf), w0->ls_cmd,
1645                                 FCNVME_RJT_RC_UNSUP, FCNVME_RJT_EXP_NONE, 0);
1646                 break;
1647         case FCNVME_LS_CREATE_ASSOCIATION:
1648         case FCNVME_LS_CREATE_CONNECTION:
1649                 lsop->lsrsp->rsplen = nvme_fc_format_rjt(lsop->rspbuf,
1650                                 sizeof(*lsop->rspbuf), w0->ls_cmd,
1651                                 FCNVME_RJT_RC_LOGIC, FCNVME_RJT_EXP_NONE, 0);
1652                 break;
1653         default:
1654                 lsop->lsrsp->rsplen = nvme_fc_format_rjt(lsop->rspbuf,
1655                                 sizeof(*lsop->rspbuf), w0->ls_cmd,
1656                                 FCNVME_RJT_RC_INVAL, FCNVME_RJT_EXP_NONE, 0);
1657                 break;
1658         }
1659
1660         return(ret);
1661 }
1662
1663 static void
1664 nvme_fc_handle_ls_rqst_work(struct work_struct *work)
1665 {
1666         struct nvme_fc_rport *rport =
1667                 container_of(work, struct nvme_fc_rport, lsrcv_work);
1668         struct fcnvme_ls_rqst_w0 *w0;
1669         struct nvmefc_ls_rcv_op *lsop;
1670         unsigned long flags;
1671         bool sendrsp;
1672
1673 restart:
1674         sendrsp = true;
1675         spin_lock_irqsave(&rport->lock, flags);
1676         list_for_each_entry(lsop, &rport->ls_rcv_list, lsrcv_list) {
1677                 if (lsop->handled)
1678                         continue;
1679
1680                 lsop->handled = true;
1681                 if (rport->remoteport.port_state == FC_OBJSTATE_ONLINE) {
1682                         spin_unlock_irqrestore(&rport->lock, flags);
1683                         sendrsp = nvme_fc_handle_ls_rqst(lsop);
1684                 } else {
1685                         spin_unlock_irqrestore(&rport->lock, flags);
1686                         w0 = &lsop->rqstbuf->w0;
1687                         lsop->lsrsp->rsplen = nvme_fc_format_rjt(
1688                                                 lsop->rspbuf,
1689                                                 sizeof(*lsop->rspbuf),
1690                                                 w0->ls_cmd,
1691                                                 FCNVME_RJT_RC_UNAB,
1692                                                 FCNVME_RJT_EXP_NONE, 0);
1693                 }
1694                 if (sendrsp)
1695                         nvme_fc_xmt_ls_rsp(lsop);
1696                 goto restart;
1697         }
1698         spin_unlock_irqrestore(&rport->lock, flags);
1699 }
1700
1701 /**
1702  * nvme_fc_rcv_ls_req - transport entry point called by an LLDD
1703  *                       upon the reception of a NVME LS request.
1704  *
1705  * The nvme-fc layer will copy payload to an internal structure for
1706  * processing.  As such, upon completion of the routine, the LLDD may
1707  * immediately free/reuse the LS request buffer passed in the call.
1708  *
1709  * If this routine returns error, the LLDD should abort the exchange.
1710  *
1711  * @remoteport: pointer to the (registered) remote port that the LS
1712  *              was received from. The remoteport is associated with
1713  *              a specific localport.
1714  * @lsrsp:      pointer to a nvmefc_ls_rsp response structure to be
1715  *              used to reference the exchange corresponding to the LS
1716  *              when issuing an ls response.
1717  * @lsreqbuf:   pointer to the buffer containing the LS Request
1718  * @lsreqbuf_len: length, in bytes, of the received LS request
1719  */
1720 int
1721 nvme_fc_rcv_ls_req(struct nvme_fc_remote_port *portptr,
1722                         struct nvmefc_ls_rsp *lsrsp,
1723                         void *lsreqbuf, u32 lsreqbuf_len)
1724 {
1725         struct nvme_fc_rport *rport = remoteport_to_rport(portptr);
1726         struct nvme_fc_lport *lport = rport->lport;
1727         struct fcnvme_ls_rqst_w0 *w0 = (struct fcnvme_ls_rqst_w0 *)lsreqbuf;
1728         struct nvmefc_ls_rcv_op *lsop;
1729         unsigned long flags;
1730         int ret;
1731
1732         nvme_fc_rport_get(rport);
1733
1734         /* validate there's a routine to transmit a response */
1735         if (!lport->ops->xmt_ls_rsp) {
1736                 dev_info(lport->dev,
1737                         "RCV %s LS failed: no LLDD xmt_ls_rsp\n",
1738                         (w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
1739                                 nvmefc_ls_names[w0->ls_cmd] : "");
1740                 ret = -EINVAL;
1741                 goto out_put;
1742         }
1743
1744         if (lsreqbuf_len > sizeof(union nvmefc_ls_requests)) {
1745                 dev_info(lport->dev,
1746                         "RCV %s LS failed: payload too large\n",
1747                         (w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
1748                                 nvmefc_ls_names[w0->ls_cmd] : "");
1749                 ret = -E2BIG;
1750                 goto out_put;
1751         }
1752
1753         lsop = kzalloc(sizeof(*lsop) +
1754                         sizeof(union nvmefc_ls_requests) +
1755                         sizeof(union nvmefc_ls_responses),
1756                         GFP_KERNEL);
1757         if (!lsop) {
1758                 dev_info(lport->dev,
1759                         "RCV %s LS failed: No memory\n",
1760                         (w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
1761                                 nvmefc_ls_names[w0->ls_cmd] : "");
1762                 ret = -ENOMEM;
1763                 goto out_put;
1764         }
1765         lsop->rqstbuf = (union nvmefc_ls_requests *)&lsop[1];
1766         lsop->rspbuf = (union nvmefc_ls_responses *)&lsop->rqstbuf[1];
1767
1768         lsop->rspdma = fc_dma_map_single(lport->dev, lsop->rspbuf,
1769                                         sizeof(*lsop->rspbuf),
1770                                         DMA_TO_DEVICE);
1771         if (fc_dma_mapping_error(lport->dev, lsop->rspdma)) {
1772                 dev_info(lport->dev,
1773                         "RCV %s LS failed: DMA mapping failure\n",
1774                         (w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
1775                                 nvmefc_ls_names[w0->ls_cmd] : "");
1776                 ret = -EFAULT;
1777                 goto out_free;
1778         }
1779
1780         lsop->rport = rport;
1781         lsop->lsrsp = lsrsp;
1782
1783         memcpy(lsop->rqstbuf, lsreqbuf, lsreqbuf_len);
1784         lsop->rqstdatalen = lsreqbuf_len;
1785
1786         spin_lock_irqsave(&rport->lock, flags);
1787         if (rport->remoteport.port_state != FC_OBJSTATE_ONLINE) {
1788                 spin_unlock_irqrestore(&rport->lock, flags);
1789                 ret = -ENOTCONN;
1790                 goto out_unmap;
1791         }
1792         list_add_tail(&lsop->lsrcv_list, &rport->ls_rcv_list);
1793         spin_unlock_irqrestore(&rport->lock, flags);
1794
1795         schedule_work(&rport->lsrcv_work);
1796
1797         return 0;
1798
1799 out_unmap:
1800         fc_dma_unmap_single(lport->dev, lsop->rspdma,
1801                         sizeof(*lsop->rspbuf), DMA_TO_DEVICE);
1802 out_free:
1803         kfree(lsop);
1804 out_put:
1805         nvme_fc_rport_put(rport);
1806         return ret;
1807 }
1808 EXPORT_SYMBOL_GPL(nvme_fc_rcv_ls_req);
1809
1810
1811 /* *********************** NVME Ctrl Routines **************************** */
1812
1813 static void
1814 __nvme_fc_exit_request(struct nvme_fc_ctrl *ctrl,
1815                 struct nvme_fc_fcp_op *op)
1816 {
1817         fc_dma_unmap_single(ctrl->lport->dev, op->fcp_req.rspdma,
1818                                 sizeof(op->rsp_iu), DMA_FROM_DEVICE);
1819         fc_dma_unmap_single(ctrl->lport->dev, op->fcp_req.cmddma,
1820                                 sizeof(op->cmd_iu), DMA_TO_DEVICE);
1821
1822         atomic_set(&op->state, FCPOP_STATE_UNINIT);
1823 }
1824
1825 static void
1826 nvme_fc_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1827                 unsigned int hctx_idx)
1828 {
1829         struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
1830
1831         return __nvme_fc_exit_request(set->driver_data, op);
1832 }
1833
1834 static int
1835 __nvme_fc_abort_op(struct nvme_fc_ctrl *ctrl, struct nvme_fc_fcp_op *op)
1836 {
1837         unsigned long flags;
1838         int opstate;
1839
1840         spin_lock_irqsave(&ctrl->lock, flags);
1841         opstate = atomic_xchg(&op->state, FCPOP_STATE_ABORTED);
1842         if (opstate != FCPOP_STATE_ACTIVE)
1843                 atomic_set(&op->state, opstate);
1844         else if (test_bit(FCCTRL_TERMIO, &ctrl->flags)) {
1845                 op->flags |= FCOP_FLAGS_TERMIO;
1846                 ctrl->iocnt++;
1847         }
1848         spin_unlock_irqrestore(&ctrl->lock, flags);
1849
1850         if (opstate != FCPOP_STATE_ACTIVE)
1851                 return -ECANCELED;
1852
1853         ctrl->lport->ops->fcp_abort(&ctrl->lport->localport,
1854                                         &ctrl->rport->remoteport,
1855                                         op->queue->lldd_handle,
1856                                         &op->fcp_req);
1857
1858         return 0;
1859 }
1860
1861 static void
1862 nvme_fc_abort_aen_ops(struct nvme_fc_ctrl *ctrl)
1863 {
1864         struct nvme_fc_fcp_op *aen_op = ctrl->aen_ops;
1865         int i;
1866
1867         /* ensure we've initialized the ops once */
1868         if (!(aen_op->flags & FCOP_FLAGS_AEN))
1869                 return;
1870
1871         for (i = 0; i < NVME_NR_AEN_COMMANDS; i++, aen_op++)
1872                 __nvme_fc_abort_op(ctrl, aen_op);
1873 }
1874
1875 static inline void
1876 __nvme_fc_fcpop_chk_teardowns(struct nvme_fc_ctrl *ctrl,
1877                 struct nvme_fc_fcp_op *op, int opstate)
1878 {
1879         unsigned long flags;
1880
1881         if (opstate == FCPOP_STATE_ABORTED) {
1882                 spin_lock_irqsave(&ctrl->lock, flags);
1883                 if (test_bit(FCCTRL_TERMIO, &ctrl->flags) &&
1884                     op->flags & FCOP_FLAGS_TERMIO) {
1885                         if (!--ctrl->iocnt)
1886                                 wake_up(&ctrl->ioabort_wait);
1887                 }
1888                 spin_unlock_irqrestore(&ctrl->lock, flags);
1889         }
1890 }
1891
1892 static void
1893 nvme_fc_ctrl_ioerr_work(struct work_struct *work)
1894 {
1895         struct nvme_fc_ctrl *ctrl =
1896                         container_of(work, struct nvme_fc_ctrl, ioerr_work);
1897
1898         nvme_fc_error_recovery(ctrl, "transport detected io error");
1899 }
1900
1901 static void
1902 nvme_fc_fcpio_done(struct nvmefc_fcp_req *req)
1903 {
1904         struct nvme_fc_fcp_op *op = fcp_req_to_fcp_op(req);
1905         struct request *rq = op->rq;
1906         struct nvmefc_fcp_req *freq = &op->fcp_req;
1907         struct nvme_fc_ctrl *ctrl = op->ctrl;
1908         struct nvme_fc_queue *queue = op->queue;
1909         struct nvme_completion *cqe = &op->rsp_iu.cqe;
1910         struct nvme_command *sqe = &op->cmd_iu.sqe;
1911         __le16 status = cpu_to_le16(NVME_SC_SUCCESS << 1);
1912         union nvme_result result;
1913         bool terminate_assoc = true;
1914         int opstate;
1915
1916         /*
1917          * WARNING:
1918          * The current linux implementation of a nvme controller
1919          * allocates a single tag set for all io queues and sizes
1920          * the io queues to fully hold all possible tags. Thus, the
1921          * implementation does not reference or care about the sqhd
1922          * value as it never needs to use the sqhd/sqtail pointers
1923          * for submission pacing.
1924          *
1925          * This affects the FC-NVME implementation in two ways:
1926          * 1) As the value doesn't matter, we don't need to waste
1927          *    cycles extracting it from ERSPs and stamping it in the
1928          *    cases where the transport fabricates CQEs on successful
1929          *    completions.
1930          * 2) The FC-NVME implementation requires that delivery of
1931          *    ERSP completions are to go back to the nvme layer in order
1932          *    relative to the rsn, such that the sqhd value will always
1933          *    be "in order" for the nvme layer. As the nvme layer in
1934          *    linux doesn't care about sqhd, there's no need to return
1935          *    them in order.
1936          *
1937          * Additionally:
1938          * As the core nvme layer in linux currently does not look at
1939          * every field in the cqe - in cases where the FC transport must
1940          * fabricate a CQE, the following fields will not be set as they
1941          * are not referenced:
1942          *      cqe.sqid,  cqe.sqhd,  cqe.command_id
1943          *
1944          * Failure or error of an individual i/o, in a transport
1945          * detected fashion unrelated to the nvme completion status,
1946          * potentially cause the initiator and target sides to get out
1947          * of sync on SQ head/tail (aka outstanding io count allowed).
1948          * Per FC-NVME spec, failure of an individual command requires
1949          * the connection to be terminated, which in turn requires the
1950          * association to be terminated.
1951          */
1952
1953         opstate = atomic_xchg(&op->state, FCPOP_STATE_COMPLETE);
1954
1955         fc_dma_sync_single_for_cpu(ctrl->lport->dev, op->fcp_req.rspdma,
1956                                 sizeof(op->rsp_iu), DMA_FROM_DEVICE);
1957
1958         if (opstate == FCPOP_STATE_ABORTED)
1959                 status = cpu_to_le16(NVME_SC_HOST_ABORTED_CMD << 1);
1960         else if (freq->status) {
1961                 status = cpu_to_le16(NVME_SC_HOST_PATH_ERROR << 1);
1962                 dev_info(ctrl->ctrl.device,
1963                         "NVME-FC{%d}: io failed due to lldd error %d\n",
1964                         ctrl->cnum, freq->status);
1965         }
1966
1967         /*
1968          * For the linux implementation, if we have an unsuccesful
1969          * status, they blk-mq layer can typically be called with the
1970          * non-zero status and the content of the cqe isn't important.
1971          */
1972         if (status)
1973                 goto done;
1974
1975         /*
1976          * command completed successfully relative to the wire
1977          * protocol. However, validate anything received and
1978          * extract the status and result from the cqe (create it
1979          * where necessary).
1980          */
1981
1982         switch (freq->rcv_rsplen) {
1983
1984         case 0:
1985         case NVME_FC_SIZEOF_ZEROS_RSP:
1986                 /*
1987                  * No response payload or 12 bytes of payload (which
1988                  * should all be zeros) are considered successful and
1989                  * no payload in the CQE by the transport.
1990                  */
1991                 if (freq->transferred_length !=
1992                     be32_to_cpu(op->cmd_iu.data_len)) {
1993                         status = cpu_to_le16(NVME_SC_HOST_PATH_ERROR << 1);
1994                         dev_info(ctrl->ctrl.device,
1995                                 "NVME-FC{%d}: io failed due to bad transfer "
1996                                 "length: %d vs expected %d\n",
1997                                 ctrl->cnum, freq->transferred_length,
1998                                 be32_to_cpu(op->cmd_iu.data_len));
1999                         goto done;
2000                 }
2001                 result.u64 = 0;
2002                 break;
2003
2004         case sizeof(struct nvme_fc_ersp_iu):
2005                 /*
2006                  * The ERSP IU contains a full completion with CQE.
2007                  * Validate ERSP IU and look at cqe.
2008                  */
2009                 if (unlikely(be16_to_cpu(op->rsp_iu.iu_len) !=
2010                                         (freq->rcv_rsplen / 4) ||
2011                              be32_to_cpu(op->rsp_iu.xfrd_len) !=
2012                                         freq->transferred_length ||
2013                              op->rsp_iu.ersp_result ||
2014                              sqe->common.command_id != cqe->command_id)) {
2015                         status = cpu_to_le16(NVME_SC_HOST_PATH_ERROR << 1);
2016                         dev_info(ctrl->ctrl.device,
2017                                 "NVME-FC{%d}: io failed due to bad NVMe_ERSP: "
2018                                 "iu len %d, xfr len %d vs %d, status code "
2019                                 "%d, cmdid %d vs %d\n",
2020                                 ctrl->cnum, be16_to_cpu(op->rsp_iu.iu_len),
2021                                 be32_to_cpu(op->rsp_iu.xfrd_len),
2022                                 freq->transferred_length,
2023                                 op->rsp_iu.ersp_result,
2024                                 sqe->common.command_id,
2025                                 cqe->command_id);
2026                         goto done;
2027                 }
2028                 result = cqe->result;
2029                 status = cqe->status;
2030                 break;
2031
2032         default:
2033                 status = cpu_to_le16(NVME_SC_HOST_PATH_ERROR << 1);
2034                 dev_info(ctrl->ctrl.device,
2035                         "NVME-FC{%d}: io failed due to odd NVMe_xRSP iu "
2036                         "len %d\n",
2037                         ctrl->cnum, freq->rcv_rsplen);
2038                 goto done;
2039         }
2040
2041         terminate_assoc = false;
2042
2043 done:
2044         if (op->flags & FCOP_FLAGS_AEN) {
2045                 nvme_complete_async_event(&queue->ctrl->ctrl, status, &result);
2046                 __nvme_fc_fcpop_chk_teardowns(ctrl, op, opstate);
2047                 atomic_set(&op->state, FCPOP_STATE_IDLE);
2048                 op->flags = FCOP_FLAGS_AEN;     /* clear other flags */
2049                 nvme_fc_ctrl_put(ctrl);
2050                 goto check_error;
2051         }
2052
2053         __nvme_fc_fcpop_chk_teardowns(ctrl, op, opstate);
2054         if (!nvme_try_complete_req(rq, status, result))
2055                 nvme_fc_complete_rq(rq);
2056
2057 check_error:
2058         if (terminate_assoc && ctrl->ctrl.state != NVME_CTRL_RESETTING)
2059                 queue_work(nvme_reset_wq, &ctrl->ioerr_work);
2060 }
2061
2062 static int
2063 __nvme_fc_init_request(struct nvme_fc_ctrl *ctrl,
2064                 struct nvme_fc_queue *queue, struct nvme_fc_fcp_op *op,
2065                 struct request *rq, u32 rqno)
2066 {
2067         struct nvme_fcp_op_w_sgl *op_w_sgl =
2068                 container_of(op, typeof(*op_w_sgl), op);
2069         struct nvme_fc_cmd_iu *cmdiu = &op->cmd_iu;
2070         int ret = 0;
2071
2072         memset(op, 0, sizeof(*op));
2073         op->fcp_req.cmdaddr = &op->cmd_iu;
2074         op->fcp_req.cmdlen = sizeof(op->cmd_iu);
2075         op->fcp_req.rspaddr = &op->rsp_iu;
2076         op->fcp_req.rsplen = sizeof(op->rsp_iu);
2077         op->fcp_req.done = nvme_fc_fcpio_done;
2078         op->ctrl = ctrl;
2079         op->queue = queue;
2080         op->rq = rq;
2081         op->rqno = rqno;
2082
2083         cmdiu->format_id = NVME_CMD_FORMAT_ID;
2084         cmdiu->fc_id = NVME_CMD_FC_ID;
2085         cmdiu->iu_len = cpu_to_be16(sizeof(*cmdiu) / sizeof(u32));
2086         if (queue->qnum)
2087                 cmdiu->rsv_cat = fccmnd_set_cat_css(0,
2088                                         (NVME_CC_CSS_NVM >> NVME_CC_CSS_SHIFT));
2089         else
2090                 cmdiu->rsv_cat = fccmnd_set_cat_admin(0);
2091
2092         op->fcp_req.cmddma = fc_dma_map_single(ctrl->lport->dev,
2093                                 &op->cmd_iu, sizeof(op->cmd_iu), DMA_TO_DEVICE);
2094         if (fc_dma_mapping_error(ctrl->lport->dev, op->fcp_req.cmddma)) {
2095                 dev_err(ctrl->dev,
2096                         "FCP Op failed - cmdiu dma mapping failed.\n");
2097                 ret = -EFAULT;
2098                 goto out_on_error;
2099         }
2100
2101         op->fcp_req.rspdma = fc_dma_map_single(ctrl->lport->dev,
2102                                 &op->rsp_iu, sizeof(op->rsp_iu),
2103                                 DMA_FROM_DEVICE);
2104         if (fc_dma_mapping_error(ctrl->lport->dev, op->fcp_req.rspdma)) {
2105                 dev_err(ctrl->dev,
2106                         "FCP Op failed - rspiu dma mapping failed.\n");
2107                 ret = -EFAULT;
2108         }
2109
2110         atomic_set(&op->state, FCPOP_STATE_IDLE);
2111 out_on_error:
2112         return ret;
2113 }
2114
2115 static int
2116 nvme_fc_init_request(struct blk_mq_tag_set *set, struct request *rq,
2117                 unsigned int hctx_idx, unsigned int numa_node)
2118 {
2119         struct nvme_fc_ctrl *ctrl = set->driver_data;
2120         struct nvme_fcp_op_w_sgl *op = blk_mq_rq_to_pdu(rq);
2121         int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
2122         struct nvme_fc_queue *queue = &ctrl->queues[queue_idx];
2123         int res;
2124
2125         res = __nvme_fc_init_request(ctrl, queue, &op->op, rq, queue->rqcnt++);
2126         if (res)
2127                 return res;
2128         op->op.fcp_req.first_sgl = op->sgl;
2129         op->op.fcp_req.private = &op->priv[0];
2130         nvme_req(rq)->ctrl = &ctrl->ctrl;
2131         return res;
2132 }
2133
2134 static int
2135 nvme_fc_init_aen_ops(struct nvme_fc_ctrl *ctrl)
2136 {
2137         struct nvme_fc_fcp_op *aen_op;
2138         struct nvme_fc_cmd_iu *cmdiu;
2139         struct nvme_command *sqe;
2140         void *private = NULL;
2141         int i, ret;
2142
2143         aen_op = ctrl->aen_ops;
2144         for (i = 0; i < NVME_NR_AEN_COMMANDS; i++, aen_op++) {
2145                 if (ctrl->lport->ops->fcprqst_priv_sz) {
2146                         private = kzalloc(ctrl->lport->ops->fcprqst_priv_sz,
2147                                                 GFP_KERNEL);
2148                         if (!private)
2149                                 return -ENOMEM;
2150                 }
2151
2152                 cmdiu = &aen_op->cmd_iu;
2153                 sqe = &cmdiu->sqe;
2154                 ret = __nvme_fc_init_request(ctrl, &ctrl->queues[0],
2155                                 aen_op, (struct request *)NULL,
2156                                 (NVME_AQ_BLK_MQ_DEPTH + i));
2157                 if (ret) {
2158                         kfree(private);
2159                         return ret;
2160                 }
2161
2162                 aen_op->flags = FCOP_FLAGS_AEN;
2163                 aen_op->fcp_req.private = private;
2164
2165                 memset(sqe, 0, sizeof(*sqe));
2166                 sqe->common.opcode = nvme_admin_async_event;
2167                 /* Note: core layer may overwrite the sqe.command_id value */
2168                 sqe->common.command_id = NVME_AQ_BLK_MQ_DEPTH + i;
2169         }
2170         return 0;
2171 }
2172
2173 static void
2174 nvme_fc_term_aen_ops(struct nvme_fc_ctrl *ctrl)
2175 {
2176         struct nvme_fc_fcp_op *aen_op;
2177         int i;
2178
2179         cancel_work_sync(&ctrl->ctrl.async_event_work);
2180         aen_op = ctrl->aen_ops;
2181         for (i = 0; i < NVME_NR_AEN_COMMANDS; i++, aen_op++) {
2182                 __nvme_fc_exit_request(ctrl, aen_op);
2183
2184                 kfree(aen_op->fcp_req.private);
2185                 aen_op->fcp_req.private = NULL;
2186         }
2187 }
2188
2189 static inline void
2190 __nvme_fc_init_hctx(struct blk_mq_hw_ctx *hctx, struct nvme_fc_ctrl *ctrl,
2191                 unsigned int qidx)
2192 {
2193         struct nvme_fc_queue *queue = &ctrl->queues[qidx];
2194
2195         hctx->driver_data = queue;
2196         queue->hctx = hctx;
2197 }
2198
2199 static int
2200 nvme_fc_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
2201                 unsigned int hctx_idx)
2202 {
2203         struct nvme_fc_ctrl *ctrl = data;
2204
2205         __nvme_fc_init_hctx(hctx, ctrl, hctx_idx + 1);
2206
2207         return 0;
2208 }
2209
2210 static int
2211 nvme_fc_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
2212                 unsigned int hctx_idx)
2213 {
2214         struct nvme_fc_ctrl *ctrl = data;
2215
2216         __nvme_fc_init_hctx(hctx, ctrl, hctx_idx);
2217
2218         return 0;
2219 }
2220
2221 static void
2222 nvme_fc_init_queue(struct nvme_fc_ctrl *ctrl, int idx)
2223 {
2224         struct nvme_fc_queue *queue;
2225
2226         queue = &ctrl->queues[idx];
2227         memset(queue, 0, sizeof(*queue));
2228         queue->ctrl = ctrl;
2229         queue->qnum = idx;
2230         atomic_set(&queue->csn, 0);
2231         queue->dev = ctrl->dev;
2232
2233         if (idx > 0)
2234                 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
2235         else
2236                 queue->cmnd_capsule_len = sizeof(struct nvme_command);
2237
2238         /*
2239          * Considered whether we should allocate buffers for all SQEs
2240          * and CQEs and dma map them - mapping their respective entries
2241          * into the request structures (kernel vm addr and dma address)
2242          * thus the driver could use the buffers/mappings directly.
2243          * It only makes sense if the LLDD would use them for its
2244          * messaging api. It's very unlikely most adapter api's would use
2245          * a native NVME sqe/cqe. More reasonable if FC-NVME IU payload
2246          * structures were used instead.
2247          */
2248 }
2249
2250 /*
2251  * This routine terminates a queue at the transport level.
2252  * The transport has already ensured that all outstanding ios on
2253  * the queue have been terminated.
2254  * The transport will send a Disconnect LS request to terminate
2255  * the queue's connection. Termination of the admin queue will also
2256  * terminate the association at the target.
2257  */
2258 static void
2259 nvme_fc_free_queue(struct nvme_fc_queue *queue)
2260 {
2261         if (!test_and_clear_bit(NVME_FC_Q_CONNECTED, &queue->flags))
2262                 return;
2263
2264         clear_bit(NVME_FC_Q_LIVE, &queue->flags);
2265         /*
2266          * Current implementation never disconnects a single queue.
2267          * It always terminates a whole association. So there is never
2268          * a disconnect(queue) LS sent to the target.
2269          */
2270
2271         queue->connection_id = 0;
2272         atomic_set(&queue->csn, 0);
2273 }
2274
2275 static void
2276 __nvme_fc_delete_hw_queue(struct nvme_fc_ctrl *ctrl,
2277         struct nvme_fc_queue *queue, unsigned int qidx)
2278 {
2279         if (ctrl->lport->ops->delete_queue)
2280                 ctrl->lport->ops->delete_queue(&ctrl->lport->localport, qidx,
2281                                 queue->lldd_handle);
2282         queue->lldd_handle = NULL;
2283 }
2284
2285 static void
2286 nvme_fc_free_io_queues(struct nvme_fc_ctrl *ctrl)
2287 {
2288         int i;
2289
2290         for (i = 1; i < ctrl->ctrl.queue_count; i++)
2291                 nvme_fc_free_queue(&ctrl->queues[i]);
2292 }
2293
2294 static int
2295 __nvme_fc_create_hw_queue(struct nvme_fc_ctrl *ctrl,
2296         struct nvme_fc_queue *queue, unsigned int qidx, u16 qsize)
2297 {
2298         int ret = 0;
2299
2300         queue->lldd_handle = NULL;
2301         if (ctrl->lport->ops->create_queue)
2302                 ret = ctrl->lport->ops->create_queue(&ctrl->lport->localport,
2303                                 qidx, qsize, &queue->lldd_handle);
2304
2305         return ret;
2306 }
2307
2308 static void
2309 nvme_fc_delete_hw_io_queues(struct nvme_fc_ctrl *ctrl)
2310 {
2311         struct nvme_fc_queue *queue = &ctrl->queues[ctrl->ctrl.queue_count - 1];
2312         int i;
2313
2314         for (i = ctrl->ctrl.queue_count - 1; i >= 1; i--, queue--)
2315                 __nvme_fc_delete_hw_queue(ctrl, queue, i);
2316 }
2317
2318 static int
2319 nvme_fc_create_hw_io_queues(struct nvme_fc_ctrl *ctrl, u16 qsize)
2320 {
2321         struct nvme_fc_queue *queue = &ctrl->queues[1];
2322         int i, ret;
2323
2324         for (i = 1; i < ctrl->ctrl.queue_count; i++, queue++) {
2325                 ret = __nvme_fc_create_hw_queue(ctrl, queue, i, qsize);
2326                 if (ret)
2327                         goto delete_queues;
2328         }
2329
2330         return 0;
2331
2332 delete_queues:
2333         for (; i > 0; i--)
2334                 __nvme_fc_delete_hw_queue(ctrl, &ctrl->queues[i], i);
2335         return ret;
2336 }
2337
2338 static int
2339 nvme_fc_connect_io_queues(struct nvme_fc_ctrl *ctrl, u16 qsize)
2340 {
2341         int i, ret = 0;
2342
2343         for (i = 1; i < ctrl->ctrl.queue_count; i++) {
2344                 ret = nvme_fc_connect_queue(ctrl, &ctrl->queues[i], qsize,
2345                                         (qsize / 5));
2346                 if (ret)
2347                         break;
2348                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i, false);
2349                 if (ret)
2350                         break;
2351
2352                 set_bit(NVME_FC_Q_LIVE, &ctrl->queues[i].flags);
2353         }
2354
2355         return ret;
2356 }
2357
2358 static void
2359 nvme_fc_init_io_queues(struct nvme_fc_ctrl *ctrl)
2360 {
2361         int i;
2362
2363         for (i = 1; i < ctrl->ctrl.queue_count; i++)
2364                 nvme_fc_init_queue(ctrl, i);
2365 }
2366
2367 static void
2368 nvme_fc_ctrl_free(struct kref *ref)
2369 {
2370         struct nvme_fc_ctrl *ctrl =
2371                 container_of(ref, struct nvme_fc_ctrl, ref);
2372         unsigned long flags;
2373
2374         if (ctrl->ctrl.tagset) {
2375                 blk_cleanup_queue(ctrl->ctrl.connect_q);
2376                 blk_mq_free_tag_set(&ctrl->tag_set);
2377         }
2378
2379         /* remove from rport list */
2380         spin_lock_irqsave(&ctrl->rport->lock, flags);
2381         list_del(&ctrl->ctrl_list);
2382         spin_unlock_irqrestore(&ctrl->rport->lock, flags);
2383
2384         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
2385         blk_cleanup_queue(ctrl->ctrl.admin_q);
2386         blk_cleanup_queue(ctrl->ctrl.fabrics_q);
2387         blk_mq_free_tag_set(&ctrl->admin_tag_set);
2388
2389         kfree(ctrl->queues);
2390
2391         put_device(ctrl->dev);
2392         nvme_fc_rport_put(ctrl->rport);
2393
2394         ida_simple_remove(&nvme_fc_ctrl_cnt, ctrl->cnum);
2395         if (ctrl->ctrl.opts)
2396                 nvmf_free_options(ctrl->ctrl.opts);
2397         kfree(ctrl);
2398 }
2399
2400 static void
2401 nvme_fc_ctrl_put(struct nvme_fc_ctrl *ctrl)
2402 {
2403         kref_put(&ctrl->ref, nvme_fc_ctrl_free);
2404 }
2405
2406 static int
2407 nvme_fc_ctrl_get(struct nvme_fc_ctrl *ctrl)
2408 {
2409         return kref_get_unless_zero(&ctrl->ref);
2410 }
2411
2412 /*
2413  * All accesses from nvme core layer done - can now free the
2414  * controller. Called after last nvme_put_ctrl() call
2415  */
2416 static void
2417 nvme_fc_nvme_ctrl_freed(struct nvme_ctrl *nctrl)
2418 {
2419         struct nvme_fc_ctrl *ctrl = to_fc_ctrl(nctrl);
2420
2421         WARN_ON(nctrl != &ctrl->ctrl);
2422
2423         nvme_fc_ctrl_put(ctrl);
2424 }
2425
2426 /*
2427  * This routine is used by the transport when it needs to find active
2428  * io on a queue that is to be terminated. The transport uses
2429  * blk_mq_tagset_busy_itr() to find the busy requests, which then invoke
2430  * this routine to kill them on a 1 by 1 basis.
2431  *
2432  * As FC allocates FC exchange for each io, the transport must contact
2433  * the LLDD to terminate the exchange, thus releasing the FC exchange.
2434  * After terminating the exchange the LLDD will call the transport's
2435  * normal io done path for the request, but it will have an aborted
2436  * status. The done path will return the io request back to the block
2437  * layer with an error status.
2438  */
2439 static bool
2440 nvme_fc_terminate_exchange(struct request *req, void *data, bool reserved)
2441 {
2442         struct nvme_ctrl *nctrl = data;
2443         struct nvme_fc_ctrl *ctrl = to_fc_ctrl(nctrl);
2444         struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(req);
2445
2446         op->nreq.flags |= NVME_REQ_CANCELLED;
2447         __nvme_fc_abort_op(ctrl, op);
2448         return true;
2449 }
2450
2451 /*
2452  * This routine runs through all outstanding commands on the association
2453  * and aborts them.  This routine is typically be called by the
2454  * delete_association routine. It is also called due to an error during
2455  * reconnect. In that scenario, it is most likely a command that initializes
2456  * the controller, including fabric Connect commands on io queues, that
2457  * may have timed out or failed thus the io must be killed for the connect
2458  * thread to see the error.
2459  */
2460 static void
2461 __nvme_fc_abort_outstanding_ios(struct nvme_fc_ctrl *ctrl, bool start_queues)
2462 {
2463         /*
2464          * If io queues are present, stop them and terminate all outstanding
2465          * ios on them. As FC allocates FC exchange for each io, the
2466          * transport must contact the LLDD to terminate the exchange,
2467          * thus releasing the FC exchange. We use blk_mq_tagset_busy_itr()
2468          * to tell us what io's are busy and invoke a transport routine
2469          * to kill them with the LLDD.  After terminating the exchange
2470          * the LLDD will call the transport's normal io done path, but it
2471          * will have an aborted status. The done path will return the
2472          * io requests back to the block layer as part of normal completions
2473          * (but with error status).
2474          */
2475         if (ctrl->ctrl.queue_count > 1) {
2476                 nvme_stop_queues(&ctrl->ctrl);
2477                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
2478                                 nvme_fc_terminate_exchange, &ctrl->ctrl);
2479                 blk_mq_tagset_wait_completed_request(&ctrl->tag_set);
2480                 if (start_queues)
2481                         nvme_start_queues(&ctrl->ctrl);
2482         }
2483
2484         /*
2485          * Other transports, which don't have link-level contexts bound
2486          * to sqe's, would try to gracefully shutdown the controller by
2487          * writing the registers for shutdown and polling (call
2488          * nvme_shutdown_ctrl()). Given a bunch of i/o was potentially
2489          * just aborted and we will wait on those contexts, and given
2490          * there was no indication of how live the controlelr is on the
2491          * link, don't send more io to create more contexts for the
2492          * shutdown. Let the controller fail via keepalive failure if
2493          * its still present.
2494          */
2495
2496         /*
2497          * clean up the admin queue. Same thing as above.
2498          */
2499         blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
2500         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
2501                                 nvme_fc_terminate_exchange, &ctrl->ctrl);
2502         blk_mq_tagset_wait_completed_request(&ctrl->admin_tag_set);
2503 }
2504
2505 static void
2506 nvme_fc_error_recovery(struct nvme_fc_ctrl *ctrl, char *errmsg)
2507 {
2508         /*
2509          * if an error (io timeout, etc) while (re)connecting, the remote
2510          * port requested terminating of the association (disconnect_ls)
2511          * or an error (timeout or abort) occurred on an io while creating
2512          * the controller.  Abort any ios on the association and let the
2513          * create_association error path resolve things.
2514          */
2515         if (ctrl->ctrl.state == NVME_CTRL_CONNECTING) {
2516                 __nvme_fc_abort_outstanding_ios(ctrl, true);
2517                 set_bit(ASSOC_FAILED, &ctrl->flags);
2518                 return;
2519         }
2520
2521         /* Otherwise, only proceed if in LIVE state - e.g. on first error */
2522         if (ctrl->ctrl.state != NVME_CTRL_LIVE)
2523                 return;
2524
2525         dev_warn(ctrl->ctrl.device,
2526                 "NVME-FC{%d}: transport association event: %s\n",
2527                 ctrl->cnum, errmsg);
2528         dev_warn(ctrl->ctrl.device,
2529                 "NVME-FC{%d}: resetting controller\n", ctrl->cnum);
2530
2531         nvme_reset_ctrl(&ctrl->ctrl);
2532 }
2533
2534 static enum blk_eh_timer_return
2535 nvme_fc_timeout(struct request *rq, bool reserved)
2536 {
2537         struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
2538         struct nvme_fc_ctrl *ctrl = op->ctrl;
2539         struct nvme_fc_cmd_iu *cmdiu = &op->cmd_iu;
2540         struct nvme_command *sqe = &cmdiu->sqe;
2541
2542         /*
2543          * Attempt to abort the offending command. Command completion
2544          * will detect the aborted io and will fail the connection.
2545          */
2546         dev_info(ctrl->ctrl.device,
2547                 "NVME-FC{%d.%d}: io timeout: opcode %d fctype %d w10/11: "
2548                 "x%08x/x%08x\n",
2549                 ctrl->cnum, op->queue->qnum, sqe->common.opcode,
2550                 sqe->connect.fctype, sqe->common.cdw10, sqe->common.cdw11);
2551         if (__nvme_fc_abort_op(ctrl, op))
2552                 nvme_fc_error_recovery(ctrl, "io timeout abort failed");
2553
2554         /*
2555          * the io abort has been initiated. Have the reset timer
2556          * restarted and the abort completion will complete the io
2557          * shortly. Avoids a synchronous wait while the abort finishes.
2558          */
2559         return BLK_EH_RESET_TIMER;
2560 }
2561
2562 static int
2563 nvme_fc_map_data(struct nvme_fc_ctrl *ctrl, struct request *rq,
2564                 struct nvme_fc_fcp_op *op)
2565 {
2566         struct nvmefc_fcp_req *freq = &op->fcp_req;
2567         int ret;
2568
2569         freq->sg_cnt = 0;
2570
2571         if (!blk_rq_nr_phys_segments(rq))
2572                 return 0;
2573
2574         freq->sg_table.sgl = freq->first_sgl;
2575         ret = sg_alloc_table_chained(&freq->sg_table,
2576                         blk_rq_nr_phys_segments(rq), freq->sg_table.sgl,
2577                         NVME_INLINE_SG_CNT);
2578         if (ret)
2579                 return -ENOMEM;
2580
2581         op->nents = blk_rq_map_sg(rq->q, rq, freq->sg_table.sgl);
2582         WARN_ON(op->nents > blk_rq_nr_phys_segments(rq));
2583         freq->sg_cnt = fc_dma_map_sg(ctrl->lport->dev, freq->sg_table.sgl,
2584                                 op->nents, rq_dma_dir(rq));
2585         if (unlikely(freq->sg_cnt <= 0)) {
2586                 sg_free_table_chained(&freq->sg_table, NVME_INLINE_SG_CNT);
2587                 freq->sg_cnt = 0;
2588                 return -EFAULT;
2589         }
2590
2591         /*
2592          * TODO: blk_integrity_rq(rq)  for DIF
2593          */
2594         return 0;
2595 }
2596
2597 static void
2598 nvme_fc_unmap_data(struct nvme_fc_ctrl *ctrl, struct request *rq,
2599                 struct nvme_fc_fcp_op *op)
2600 {
2601         struct nvmefc_fcp_req *freq = &op->fcp_req;
2602
2603         if (!freq->sg_cnt)
2604                 return;
2605
2606         fc_dma_unmap_sg(ctrl->lport->dev, freq->sg_table.sgl, op->nents,
2607                         rq_dma_dir(rq));
2608
2609         sg_free_table_chained(&freq->sg_table, NVME_INLINE_SG_CNT);
2610
2611         freq->sg_cnt = 0;
2612 }
2613
2614 /*
2615  * In FC, the queue is a logical thing. At transport connect, the target
2616  * creates its "queue" and returns a handle that is to be given to the
2617  * target whenever it posts something to the corresponding SQ.  When an
2618  * SQE is sent on a SQ, FC effectively considers the SQE, or rather the
2619  * command contained within the SQE, an io, and assigns a FC exchange
2620  * to it. The SQE and the associated SQ handle are sent in the initial
2621  * CMD IU sents on the exchange. All transfers relative to the io occur
2622  * as part of the exchange.  The CQE is the last thing for the io,
2623  * which is transferred (explicitly or implicitly) with the RSP IU
2624  * sent on the exchange. After the CQE is received, the FC exchange is
2625  * terminaed and the Exchange may be used on a different io.
2626  *
2627  * The transport to LLDD api has the transport making a request for a
2628  * new fcp io request to the LLDD. The LLDD then allocates a FC exchange
2629  * resource and transfers the command. The LLDD will then process all
2630  * steps to complete the io. Upon completion, the transport done routine
2631  * is called.
2632  *
2633  * So - while the operation is outstanding to the LLDD, there is a link
2634  * level FC exchange resource that is also outstanding. This must be
2635  * considered in all cleanup operations.
2636  */
2637 static blk_status_t
2638 nvme_fc_start_fcp_op(struct nvme_fc_ctrl *ctrl, struct nvme_fc_queue *queue,
2639         struct nvme_fc_fcp_op *op, u32 data_len,
2640         enum nvmefc_fcp_datadir io_dir)
2641 {
2642         struct nvme_fc_cmd_iu *cmdiu = &op->cmd_iu;
2643         struct nvme_command *sqe = &cmdiu->sqe;
2644         int ret, opstate;
2645
2646         /*
2647          * before attempting to send the io, check to see if we believe
2648          * the target device is present
2649          */
2650         if (ctrl->rport->remoteport.port_state != FC_OBJSTATE_ONLINE)
2651                 return BLK_STS_RESOURCE;
2652
2653         if (!nvme_fc_ctrl_get(ctrl))
2654                 return BLK_STS_IOERR;
2655
2656         /* format the FC-NVME CMD IU and fcp_req */
2657         cmdiu->connection_id = cpu_to_be64(queue->connection_id);
2658         cmdiu->data_len = cpu_to_be32(data_len);
2659         switch (io_dir) {
2660         case NVMEFC_FCP_WRITE:
2661                 cmdiu->flags = FCNVME_CMD_FLAGS_WRITE;
2662                 break;
2663         case NVMEFC_FCP_READ:
2664                 cmdiu->flags = FCNVME_CMD_FLAGS_READ;
2665                 break;
2666         case NVMEFC_FCP_NODATA:
2667                 cmdiu->flags = 0;
2668                 break;
2669         }
2670         op->fcp_req.payload_length = data_len;
2671         op->fcp_req.io_dir = io_dir;
2672         op->fcp_req.transferred_length = 0;
2673         op->fcp_req.rcv_rsplen = 0;
2674         op->fcp_req.status = NVME_SC_SUCCESS;
2675         op->fcp_req.sqid = cpu_to_le16(queue->qnum);
2676
2677         /*
2678          * validate per fabric rules, set fields mandated by fabric spec
2679          * as well as those by FC-NVME spec.
2680          */
2681         WARN_ON_ONCE(sqe->common.metadata);
2682         sqe->common.flags |= NVME_CMD_SGL_METABUF;
2683
2684         /*
2685          * format SQE DPTR field per FC-NVME rules:
2686          *    type=0x5     Transport SGL Data Block Descriptor
2687          *    subtype=0xA  Transport-specific value
2688          *    address=0
2689          *    length=length of the data series
2690          */
2691         sqe->rw.dptr.sgl.type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2692                                         NVME_SGL_FMT_TRANSPORT_A;
2693         sqe->rw.dptr.sgl.length = cpu_to_le32(data_len);
2694         sqe->rw.dptr.sgl.addr = 0;
2695
2696         if (!(op->flags & FCOP_FLAGS_AEN)) {
2697                 ret = nvme_fc_map_data(ctrl, op->rq, op);
2698                 if (ret < 0) {
2699                         nvme_cleanup_cmd(op->rq);
2700                         nvme_fc_ctrl_put(ctrl);
2701                         if (ret == -ENOMEM || ret == -EAGAIN)
2702                                 return BLK_STS_RESOURCE;
2703                         return BLK_STS_IOERR;
2704                 }
2705         }
2706
2707         fc_dma_sync_single_for_device(ctrl->lport->dev, op->fcp_req.cmddma,
2708                                   sizeof(op->cmd_iu), DMA_TO_DEVICE);
2709
2710         atomic_set(&op->state, FCPOP_STATE_ACTIVE);
2711
2712         if (!(op->flags & FCOP_FLAGS_AEN))
2713                 blk_mq_start_request(op->rq);
2714
2715         cmdiu->csn = cpu_to_be32(atomic_inc_return(&queue->csn));
2716         ret = ctrl->lport->ops->fcp_io(&ctrl->lport->localport,
2717                                         &ctrl->rport->remoteport,
2718                                         queue->lldd_handle, &op->fcp_req);
2719
2720         if (ret) {
2721                 /*
2722                  * If the lld fails to send the command is there an issue with
2723                  * the csn value?  If the command that fails is the Connect,
2724                  * no - as the connection won't be live.  If it is a command
2725                  * post-connect, it's possible a gap in csn may be created.
2726                  * Does this matter?  As Linux initiators don't send fused
2727                  * commands, no.  The gap would exist, but as there's nothing
2728                  * that depends on csn order to be delivered on the target
2729                  * side, it shouldn't hurt.  It would be difficult for a
2730                  * target to even detect the csn gap as it has no idea when the
2731                  * cmd with the csn was supposed to arrive.
2732                  */
2733                 opstate = atomic_xchg(&op->state, FCPOP_STATE_COMPLETE);
2734                 __nvme_fc_fcpop_chk_teardowns(ctrl, op, opstate);
2735
2736                 if (!(op->flags & FCOP_FLAGS_AEN)) {
2737                         nvme_fc_unmap_data(ctrl, op->rq, op);
2738                         nvme_cleanup_cmd(op->rq);
2739                 }
2740
2741                 nvme_fc_ctrl_put(ctrl);
2742
2743                 if (ctrl->rport->remoteport.port_state == FC_OBJSTATE_ONLINE &&
2744                                 ret != -EBUSY)
2745                         return BLK_STS_IOERR;
2746
2747                 return BLK_STS_RESOURCE;
2748         }
2749
2750         return BLK_STS_OK;
2751 }
2752
2753 static blk_status_t
2754 nvme_fc_queue_rq(struct blk_mq_hw_ctx *hctx,
2755                         const struct blk_mq_queue_data *bd)
2756 {
2757         struct nvme_ns *ns = hctx->queue->queuedata;
2758         struct nvme_fc_queue *queue = hctx->driver_data;
2759         struct nvme_fc_ctrl *ctrl = queue->ctrl;
2760         struct request *rq = bd->rq;
2761         struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
2762         struct nvme_fc_cmd_iu *cmdiu = &op->cmd_iu;
2763         struct nvme_command *sqe = &cmdiu->sqe;
2764         enum nvmefc_fcp_datadir io_dir;
2765         bool queue_ready = test_bit(NVME_FC_Q_LIVE, &queue->flags);
2766         u32 data_len;
2767         blk_status_t ret;
2768
2769         if (ctrl->rport->remoteport.port_state != FC_OBJSTATE_ONLINE ||
2770             !nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2771                 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
2772
2773         ret = nvme_setup_cmd(ns, rq, sqe);
2774         if (ret)
2775                 return ret;
2776
2777         /*
2778          * nvme core doesn't quite treat the rq opaquely. Commands such
2779          * as WRITE ZEROES will return a non-zero rq payload_bytes yet
2780          * there is no actual payload to be transferred.
2781          * To get it right, key data transmission on there being 1 or
2782          * more physical segments in the sg list. If there is no
2783          * physical segments, there is no payload.
2784          */
2785         if (blk_rq_nr_phys_segments(rq)) {
2786                 data_len = blk_rq_payload_bytes(rq);
2787                 io_dir = ((rq_data_dir(rq) == WRITE) ?
2788                                         NVMEFC_FCP_WRITE : NVMEFC_FCP_READ);
2789         } else {
2790                 data_len = 0;
2791                 io_dir = NVMEFC_FCP_NODATA;
2792         }
2793
2794
2795         return nvme_fc_start_fcp_op(ctrl, queue, op, data_len, io_dir);
2796 }
2797
2798 static void
2799 nvme_fc_submit_async_event(struct nvme_ctrl *arg)
2800 {
2801         struct nvme_fc_ctrl *ctrl = to_fc_ctrl(arg);
2802         struct nvme_fc_fcp_op *aen_op;
2803         blk_status_t ret;
2804
2805         if (test_bit(FCCTRL_TERMIO, &ctrl->flags))
2806                 return;
2807
2808         aen_op = &ctrl->aen_ops[0];
2809
2810         ret = nvme_fc_start_fcp_op(ctrl, aen_op->queue, aen_op, 0,
2811                                         NVMEFC_FCP_NODATA);
2812         if (ret)
2813                 dev_err(ctrl->ctrl.device,
2814                         "failed async event work\n");
2815 }
2816
2817 static void
2818 nvme_fc_complete_rq(struct request *rq)
2819 {
2820         struct nvme_fc_fcp_op *op = blk_mq_rq_to_pdu(rq);
2821         struct nvme_fc_ctrl *ctrl = op->ctrl;
2822
2823         atomic_set(&op->state, FCPOP_STATE_IDLE);
2824         op->flags &= ~FCOP_FLAGS_TERMIO;
2825
2826         nvme_fc_unmap_data(ctrl, rq, op);
2827         nvme_complete_rq(rq);
2828         nvme_fc_ctrl_put(ctrl);
2829 }
2830
2831
2832 static const struct blk_mq_ops nvme_fc_mq_ops = {
2833         .queue_rq       = nvme_fc_queue_rq,
2834         .complete       = nvme_fc_complete_rq,
2835         .init_request   = nvme_fc_init_request,
2836         .exit_request   = nvme_fc_exit_request,
2837         .init_hctx      = nvme_fc_init_hctx,
2838         .timeout        = nvme_fc_timeout,
2839 };
2840
2841 static int
2842 nvme_fc_create_io_queues(struct nvme_fc_ctrl *ctrl)
2843 {
2844         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2845         unsigned int nr_io_queues;
2846         int ret;
2847
2848         nr_io_queues = min(min(opts->nr_io_queues, num_online_cpus()),
2849                                 ctrl->lport->ops->max_hw_queues);
2850         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
2851         if (ret) {
2852                 dev_info(ctrl->ctrl.device,
2853                         "set_queue_count failed: %d\n", ret);
2854                 return ret;
2855         }
2856
2857         ctrl->ctrl.queue_count = nr_io_queues + 1;
2858         if (!nr_io_queues)
2859                 return 0;
2860
2861         nvme_fc_init_io_queues(ctrl);
2862
2863         memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
2864         ctrl->tag_set.ops = &nvme_fc_mq_ops;
2865         ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
2866         ctrl->tag_set.reserved_tags = NVMF_RESERVED_TAGS;
2867         ctrl->tag_set.numa_node = ctrl->ctrl.numa_node;
2868         ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
2869         ctrl->tag_set.cmd_size =
2870                 struct_size((struct nvme_fcp_op_w_sgl *)NULL, priv,
2871                             ctrl->lport->ops->fcprqst_priv_sz);
2872         ctrl->tag_set.driver_data = ctrl;
2873         ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
2874         ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
2875
2876         ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
2877         if (ret)
2878                 return ret;
2879
2880         ctrl->ctrl.tagset = &ctrl->tag_set;
2881
2882         ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
2883         if (IS_ERR(ctrl->ctrl.connect_q)) {
2884                 ret = PTR_ERR(ctrl->ctrl.connect_q);
2885                 goto out_free_tag_set;
2886         }
2887
2888         ret = nvme_fc_create_hw_io_queues(ctrl, ctrl->ctrl.sqsize + 1);
2889         if (ret)
2890                 goto out_cleanup_blk_queue;
2891
2892         ret = nvme_fc_connect_io_queues(ctrl, ctrl->ctrl.sqsize + 1);
2893         if (ret)
2894                 goto out_delete_hw_queues;
2895
2896         ctrl->ioq_live = true;
2897
2898         return 0;
2899
2900 out_delete_hw_queues:
2901         nvme_fc_delete_hw_io_queues(ctrl);
2902 out_cleanup_blk_queue:
2903         blk_cleanup_queue(ctrl->ctrl.connect_q);
2904 out_free_tag_set:
2905         blk_mq_free_tag_set(&ctrl->tag_set);
2906         nvme_fc_free_io_queues(ctrl);
2907
2908         /* force put free routine to ignore io queues */
2909         ctrl->ctrl.tagset = NULL;
2910
2911         return ret;
2912 }
2913
2914 static int
2915 nvme_fc_recreate_io_queues(struct nvme_fc_ctrl *ctrl)
2916 {
2917         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2918         u32 prior_ioq_cnt = ctrl->ctrl.queue_count - 1;
2919         unsigned int nr_io_queues;
2920         int ret;
2921
2922         nr_io_queues = min(min(opts->nr_io_queues, num_online_cpus()),
2923                                 ctrl->lport->ops->max_hw_queues);
2924         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
2925         if (ret) {
2926                 dev_info(ctrl->ctrl.device,
2927                         "set_queue_count failed: %d\n", ret);
2928                 return ret;
2929         }
2930
2931         if (!nr_io_queues && prior_ioq_cnt) {
2932                 dev_info(ctrl->ctrl.device,
2933                         "Fail Reconnect: At least 1 io queue "
2934                         "required (was %d)\n", prior_ioq_cnt);
2935                 return -ENOSPC;
2936         }
2937
2938         ctrl->ctrl.queue_count = nr_io_queues + 1;
2939         /* check for io queues existing */
2940         if (ctrl->ctrl.queue_count == 1)
2941                 return 0;
2942
2943         ret = nvme_fc_create_hw_io_queues(ctrl, ctrl->ctrl.sqsize + 1);
2944         if (ret)
2945                 goto out_free_io_queues;
2946
2947         ret = nvme_fc_connect_io_queues(ctrl, ctrl->ctrl.sqsize + 1);
2948         if (ret)
2949                 goto out_delete_hw_queues;
2950
2951         if (prior_ioq_cnt != nr_io_queues) {
2952                 dev_info(ctrl->ctrl.device,
2953                         "reconnect: revising io queue count from %d to %d\n",
2954                         prior_ioq_cnt, nr_io_queues);
2955                 nvme_wait_freeze(&ctrl->ctrl);
2956                 blk_mq_update_nr_hw_queues(&ctrl->tag_set, nr_io_queues);
2957                 nvme_unfreeze(&ctrl->ctrl);
2958         }
2959
2960         return 0;
2961
2962 out_delete_hw_queues:
2963         nvme_fc_delete_hw_io_queues(ctrl);
2964 out_free_io_queues:
2965         nvme_fc_free_io_queues(ctrl);
2966         return ret;
2967 }
2968
2969 static void
2970 nvme_fc_rport_active_on_lport(struct nvme_fc_rport *rport)
2971 {
2972         struct nvme_fc_lport *lport = rport->lport;
2973
2974         atomic_inc(&lport->act_rport_cnt);
2975 }
2976
2977 static void
2978 nvme_fc_rport_inactive_on_lport(struct nvme_fc_rport *rport)
2979 {
2980         struct nvme_fc_lport *lport = rport->lport;
2981         u32 cnt;
2982
2983         cnt = atomic_dec_return(&lport->act_rport_cnt);
2984         if (cnt == 0 && lport->localport.port_state == FC_OBJSTATE_DELETED)
2985                 lport->ops->localport_delete(&lport->localport);
2986 }
2987
2988 static int
2989 nvme_fc_ctlr_active_on_rport(struct nvme_fc_ctrl *ctrl)
2990 {
2991         struct nvme_fc_rport *rport = ctrl->rport;
2992         u32 cnt;
2993
2994         if (test_and_set_bit(ASSOC_ACTIVE, &ctrl->flags))
2995                 return 1;
2996
2997         cnt = atomic_inc_return(&rport->act_ctrl_cnt);
2998         if (cnt == 1)
2999                 nvme_fc_rport_active_on_lport(rport);
3000
3001         return 0;
3002 }
3003
3004 static int
3005 nvme_fc_ctlr_inactive_on_rport(struct nvme_fc_ctrl *ctrl)
3006 {
3007         struct nvme_fc_rport *rport = ctrl->rport;
3008         struct nvme_fc_lport *lport = rport->lport;
3009         u32 cnt;
3010
3011         /* clearing of ctrl->flags ASSOC_ACTIVE bit is in association delete */
3012
3013         cnt = atomic_dec_return(&rport->act_ctrl_cnt);
3014         if (cnt == 0) {
3015                 if (rport->remoteport.port_state == FC_OBJSTATE_DELETED)
3016                         lport->ops->remoteport_delete(&rport->remoteport);
3017                 nvme_fc_rport_inactive_on_lport(rport);
3018         }
3019
3020         return 0;
3021 }
3022
3023 /*
3024  * This routine restarts the controller on the host side, and
3025  * on the link side, recreates the controller association.
3026  */
3027 static int
3028 nvme_fc_create_association(struct nvme_fc_ctrl *ctrl)
3029 {
3030         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
3031         struct nvmefc_ls_rcv_op *disls = NULL;
3032         unsigned long flags;
3033         int ret;
3034         bool changed;
3035
3036         ++ctrl->ctrl.nr_reconnects;
3037
3038         if (ctrl->rport->remoteport.port_state != FC_OBJSTATE_ONLINE)
3039                 return -ENODEV;
3040
3041         if (nvme_fc_ctlr_active_on_rport(ctrl))
3042                 return -ENOTUNIQ;
3043
3044         dev_info(ctrl->ctrl.device,
3045                 "NVME-FC{%d}: create association : host wwpn 0x%016llx "
3046                 " rport wwpn 0x%016llx: NQN \"%s\"\n",
3047                 ctrl->cnum, ctrl->lport->localport.port_name,
3048                 ctrl->rport->remoteport.port_name, ctrl->ctrl.opts->subsysnqn);
3049
3050         clear_bit(ASSOC_FAILED, &ctrl->flags);
3051
3052         /*
3053          * Create the admin queue
3054          */
3055
3056         ret = __nvme_fc_create_hw_queue(ctrl, &ctrl->queues[0], 0,
3057                                 NVME_AQ_DEPTH);
3058         if (ret)
3059                 goto out_free_queue;
3060
3061         ret = nvme_fc_connect_admin_queue(ctrl, &ctrl->queues[0],
3062                                 NVME_AQ_DEPTH, (NVME_AQ_DEPTH / 4));
3063         if (ret)
3064                 goto out_delete_hw_queue;
3065
3066         ret = nvmf_connect_admin_queue(&ctrl->ctrl);
3067         if (ret)
3068                 goto out_disconnect_admin_queue;
3069
3070         set_bit(NVME_FC_Q_LIVE, &ctrl->queues[0].flags);
3071
3072         /*
3073          * Check controller capabilities
3074          *
3075          * todo:- add code to check if ctrl attributes changed from
3076          * prior connection values
3077          */
3078
3079         ret = nvme_enable_ctrl(&ctrl->ctrl);
3080         if (ret || test_bit(ASSOC_FAILED, &ctrl->flags))
3081                 goto out_disconnect_admin_queue;
3082
3083         ctrl->ctrl.max_segments = ctrl->lport->ops->max_sgl_segments;
3084         ctrl->ctrl.max_hw_sectors = ctrl->ctrl.max_segments <<
3085                                                 (ilog2(SZ_4K) - 9);
3086
3087         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
3088
3089         ret = nvme_init_identify(&ctrl->ctrl);
3090         if (ret || test_bit(ASSOC_FAILED, &ctrl->flags))
3091                 goto out_disconnect_admin_queue;
3092
3093         /* sanity checks */
3094
3095         /* FC-NVME does not have other data in the capsule */
3096         if (ctrl->ctrl.icdoff) {
3097                 dev_err(ctrl->ctrl.device, "icdoff %d is not supported!\n",
3098                                 ctrl->ctrl.icdoff);
3099                 goto out_disconnect_admin_queue;
3100         }
3101
3102         /* FC-NVME supports normal SGL Data Block Descriptors */
3103
3104         if (opts->queue_size > ctrl->ctrl.maxcmd) {
3105                 /* warn if maxcmd is lower than queue_size */
3106                 dev_warn(ctrl->ctrl.device,
3107                         "queue_size %zu > ctrl maxcmd %u, reducing "
3108                         "to maxcmd\n",
3109                         opts->queue_size, ctrl->ctrl.maxcmd);
3110                 opts->queue_size = ctrl->ctrl.maxcmd;
3111         }
3112
3113         if (opts->queue_size > ctrl->ctrl.sqsize + 1) {
3114                 /* warn if sqsize is lower than queue_size */
3115                 dev_warn(ctrl->ctrl.device,
3116                         "queue_size %zu > ctrl sqsize %u, reducing "
3117                         "to sqsize\n",
3118                         opts->queue_size, ctrl->ctrl.sqsize + 1);
3119                 opts->queue_size = ctrl->ctrl.sqsize + 1;
3120         }
3121
3122         ret = nvme_fc_init_aen_ops(ctrl);
3123         if (ret)
3124                 goto out_term_aen_ops;
3125
3126         /*
3127          * Create the io queues
3128          */
3129
3130         if (ctrl->ctrl.queue_count > 1) {
3131                 if (!ctrl->ioq_live)
3132                         ret = nvme_fc_create_io_queues(ctrl);
3133                 else
3134                         ret = nvme_fc_recreate_io_queues(ctrl);
3135         }
3136         if (ret || test_bit(ASSOC_FAILED, &ctrl->flags))
3137                 goto out_term_aen_ops;
3138
3139         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
3140
3141         ctrl->ctrl.nr_reconnects = 0;
3142
3143         if (changed)
3144                 nvme_start_ctrl(&ctrl->ctrl);
3145
3146         return 0;       /* Success */
3147
3148 out_term_aen_ops:
3149         nvme_fc_term_aen_ops(ctrl);
3150 out_disconnect_admin_queue:
3151         /* send a Disconnect(association) LS to fc-nvme target */
3152         nvme_fc_xmt_disconnect_assoc(ctrl);
3153         spin_lock_irqsave(&ctrl->lock, flags);
3154         ctrl->association_id = 0;
3155         disls = ctrl->rcv_disconn;
3156         ctrl->rcv_disconn = NULL;
3157         spin_unlock_irqrestore(&ctrl->lock, flags);
3158         if (disls)
3159                 nvme_fc_xmt_ls_rsp(disls);
3160 out_delete_hw_queue:
3161         __nvme_fc_delete_hw_queue(ctrl, &ctrl->queues[0], 0);
3162 out_free_queue:
3163         nvme_fc_free_queue(&ctrl->queues[0]);
3164         clear_bit(ASSOC_ACTIVE, &ctrl->flags);
3165         nvme_fc_ctlr_inactive_on_rport(ctrl);
3166
3167         return ret;
3168 }
3169
3170
3171 /*
3172  * This routine stops operation of the controller on the host side.
3173  * On the host os stack side: Admin and IO queues are stopped,
3174  *   outstanding ios on them terminated via FC ABTS.
3175  * On the link side: the association is terminated.
3176  */
3177 static void
3178 nvme_fc_delete_association(struct nvme_fc_ctrl *ctrl)
3179 {
3180         struct nvmefc_ls_rcv_op *disls = NULL;
3181         unsigned long flags;
3182
3183         if (!test_and_clear_bit(ASSOC_ACTIVE, &ctrl->flags))
3184                 return;
3185
3186         spin_lock_irqsave(&ctrl->lock, flags);
3187         set_bit(FCCTRL_TERMIO, &ctrl->flags);
3188         ctrl->iocnt = 0;
3189         spin_unlock_irqrestore(&ctrl->lock, flags);
3190
3191         __nvme_fc_abort_outstanding_ios(ctrl, false);
3192
3193         /* kill the aens as they are a separate path */
3194         nvme_fc_abort_aen_ops(ctrl);
3195
3196         /* wait for all io that had to be aborted */
3197         spin_lock_irq(&ctrl->lock);
3198         wait_event_lock_irq(ctrl->ioabort_wait, ctrl->iocnt == 0, ctrl->lock);
3199         clear_bit(FCCTRL_TERMIO, &ctrl->flags);
3200         spin_unlock_irq(&ctrl->lock);
3201
3202         nvme_fc_term_aen_ops(ctrl);
3203
3204         /*
3205          * send a Disconnect(association) LS to fc-nvme target
3206          * Note: could have been sent at top of process, but
3207          * cleaner on link traffic if after the aborts complete.
3208          * Note: if association doesn't exist, association_id will be 0
3209          */
3210         if (ctrl->association_id)
3211                 nvme_fc_xmt_disconnect_assoc(ctrl);
3212
3213         spin_lock_irqsave(&ctrl->lock, flags);
3214         ctrl->association_id = 0;
3215         disls = ctrl->rcv_disconn;
3216         ctrl->rcv_disconn = NULL;
3217         spin_unlock_irqrestore(&ctrl->lock, flags);
3218         if (disls)
3219                 /*
3220                  * if a Disconnect Request was waiting for a response, send
3221                  * now that all ABTS's have been issued (and are complete).
3222                  */
3223                 nvme_fc_xmt_ls_rsp(disls);
3224
3225         if (ctrl->ctrl.tagset) {
3226                 nvme_fc_delete_hw_io_queues(ctrl);
3227                 nvme_fc_free_io_queues(ctrl);
3228         }
3229
3230         __nvme_fc_delete_hw_queue(ctrl, &ctrl->queues[0], 0);
3231         nvme_fc_free_queue(&ctrl->queues[0]);
3232
3233         /* re-enable the admin_q so anything new can fast fail */
3234         blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
3235
3236         /* resume the io queues so that things will fast fail */
3237         nvme_start_queues(&ctrl->ctrl);
3238
3239         nvme_fc_ctlr_inactive_on_rport(ctrl);
3240 }
3241
3242 static void
3243 nvme_fc_delete_ctrl(struct nvme_ctrl *nctrl)
3244 {
3245         struct nvme_fc_ctrl *ctrl = to_fc_ctrl(nctrl);
3246
3247         cancel_work_sync(&ctrl->ioerr_work);
3248         cancel_delayed_work_sync(&ctrl->connect_work);
3249         /*
3250          * kill the association on the link side.  this will block
3251          * waiting for io to terminate
3252          */
3253         nvme_fc_delete_association(ctrl);
3254 }
3255
3256 static void
3257 nvme_fc_reconnect_or_delete(struct nvme_fc_ctrl *ctrl, int status)
3258 {
3259         struct nvme_fc_rport *rport = ctrl->rport;
3260         struct nvme_fc_remote_port *portptr = &rport->remoteport;
3261         unsigned long recon_delay = ctrl->ctrl.opts->reconnect_delay * HZ;
3262         bool recon = true;
3263
3264         if (ctrl->ctrl.state != NVME_CTRL_CONNECTING)
3265                 return;
3266
3267         if (portptr->port_state == FC_OBJSTATE_ONLINE)
3268                 dev_info(ctrl->ctrl.device,
3269                         "NVME-FC{%d}: reset: Reconnect attempt failed (%d)\n",
3270                         ctrl->cnum, status);
3271         else if (time_after_eq(jiffies, rport->dev_loss_end))
3272                 recon = false;
3273
3274         if (recon && nvmf_should_reconnect(&ctrl->ctrl)) {
3275                 if (portptr->port_state == FC_OBJSTATE_ONLINE)
3276                         dev_info(ctrl->ctrl.device,
3277                                 "NVME-FC{%d}: Reconnect attempt in %ld "
3278                                 "seconds\n",
3279                                 ctrl->cnum, recon_delay / HZ);
3280                 else if (time_after(jiffies + recon_delay, rport->dev_loss_end))
3281                         recon_delay = rport->dev_loss_end - jiffies;
3282
3283                 queue_delayed_work(nvme_wq, &ctrl->connect_work, recon_delay);
3284         } else {
3285                 if (portptr->port_state == FC_OBJSTATE_ONLINE)
3286                         dev_warn(ctrl->ctrl.device,
3287                                 "NVME-FC{%d}: Max reconnect attempts (%d) "
3288                                 "reached.\n",
3289                                 ctrl->cnum, ctrl->ctrl.nr_reconnects);
3290                 else
3291                         dev_warn(ctrl->ctrl.device,
3292                                 "NVME-FC{%d}: dev_loss_tmo (%d) expired "
3293                                 "while waiting for remoteport connectivity.\n",
3294                                 ctrl->cnum, min_t(int, portptr->dev_loss_tmo,
3295                                         (ctrl->ctrl.opts->max_reconnects *
3296                                          ctrl->ctrl.opts->reconnect_delay)));
3297                 WARN_ON(nvme_delete_ctrl(&ctrl->ctrl));
3298         }
3299 }
3300
3301 static void
3302 nvme_fc_reset_ctrl_work(struct work_struct *work)
3303 {
3304         struct nvme_fc_ctrl *ctrl =
3305                 container_of(work, struct nvme_fc_ctrl, ctrl.reset_work);
3306
3307         nvme_stop_ctrl(&ctrl->ctrl);
3308
3309         /* will block will waiting for io to terminate */
3310         nvme_fc_delete_association(ctrl);
3311
3312         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING))
3313                 dev_err(ctrl->ctrl.device,
3314                         "NVME-FC{%d}: error_recovery: Couldn't change state "
3315                         "to CONNECTING\n", ctrl->cnum);
3316
3317         if (ctrl->rport->remoteport.port_state == FC_OBJSTATE_ONLINE) {
3318                 if (!queue_delayed_work(nvme_wq, &ctrl->connect_work, 0)) {
3319                         dev_err(ctrl->ctrl.device,
3320                                 "NVME-FC{%d}: failed to schedule connect "
3321                                 "after reset\n", ctrl->cnum);
3322                 } else {
3323                         flush_delayed_work(&ctrl->connect_work);
3324                 }
3325         } else {
3326                 nvme_fc_reconnect_or_delete(ctrl, -ENOTCONN);
3327         }
3328 }
3329
3330
3331 static const struct nvme_ctrl_ops nvme_fc_ctrl_ops = {
3332         .name                   = "fc",
3333         .module                 = THIS_MODULE,
3334         .flags                  = NVME_F_FABRICS,
3335         .reg_read32             = nvmf_reg_read32,
3336         .reg_read64             = nvmf_reg_read64,
3337         .reg_write32            = nvmf_reg_write32,
3338         .free_ctrl              = nvme_fc_nvme_ctrl_freed,
3339         .submit_async_event     = nvme_fc_submit_async_event,
3340         .delete_ctrl            = nvme_fc_delete_ctrl,
3341         .get_address            = nvmf_get_address,
3342 };
3343
3344 static void
3345 nvme_fc_connect_ctrl_work(struct work_struct *work)
3346 {
3347         int ret;
3348
3349         struct nvme_fc_ctrl *ctrl =
3350                         container_of(to_delayed_work(work),
3351                                 struct nvme_fc_ctrl, connect_work);
3352
3353         ret = nvme_fc_create_association(ctrl);
3354         if (ret)
3355                 nvme_fc_reconnect_or_delete(ctrl, ret);
3356         else
3357                 dev_info(ctrl->ctrl.device,
3358                         "NVME-FC{%d}: controller connect complete\n",
3359                         ctrl->cnum);
3360 }
3361
3362
3363 static const struct blk_mq_ops nvme_fc_admin_mq_ops = {
3364         .queue_rq       = nvme_fc_queue_rq,
3365         .complete       = nvme_fc_complete_rq,
3366         .init_request   = nvme_fc_init_request,
3367         .exit_request   = nvme_fc_exit_request,
3368         .init_hctx      = nvme_fc_init_admin_hctx,
3369         .timeout        = nvme_fc_timeout,
3370 };
3371
3372
3373 /*
3374  * Fails a controller request if it matches an existing controller
3375  * (association) with the same tuple:
3376  * <Host NQN, Host ID, local FC port, remote FC port, SUBSYS NQN>
3377  *
3378  * The ports don't need to be compared as they are intrinsically
3379  * already matched by the port pointers supplied.
3380  */
3381 static bool
3382 nvme_fc_existing_controller(struct nvme_fc_rport *rport,
3383                 struct nvmf_ctrl_options *opts)
3384 {
3385         struct nvme_fc_ctrl *ctrl;
3386         unsigned long flags;
3387         bool found = false;
3388
3389         spin_lock_irqsave(&rport->lock, flags);
3390         list_for_each_entry(ctrl, &rport->ctrl_list, ctrl_list) {
3391                 found = nvmf_ctlr_matches_baseopts(&ctrl->ctrl, opts);
3392                 if (found)
3393                         break;
3394         }
3395         spin_unlock_irqrestore(&rport->lock, flags);
3396
3397         return found;
3398 }
3399
3400 static struct nvme_ctrl *
3401 nvme_fc_init_ctrl(struct device *dev, struct nvmf_ctrl_options *opts,
3402         struct nvme_fc_lport *lport, struct nvme_fc_rport *rport)
3403 {
3404         struct nvme_fc_ctrl *ctrl;
3405         unsigned long flags;
3406         int ret, idx, ctrl_loss_tmo;
3407
3408         if (!(rport->remoteport.port_role &
3409             (FC_PORT_ROLE_NVME_DISCOVERY | FC_PORT_ROLE_NVME_TARGET))) {
3410                 ret = -EBADR;
3411                 goto out_fail;
3412         }
3413
3414         if (!opts->duplicate_connect &&
3415             nvme_fc_existing_controller(rport, opts)) {
3416                 ret = -EALREADY;
3417                 goto out_fail;
3418         }
3419
3420         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
3421         if (!ctrl) {
3422                 ret = -ENOMEM;
3423                 goto out_fail;
3424         }
3425
3426         idx = ida_simple_get(&nvme_fc_ctrl_cnt, 0, 0, GFP_KERNEL);
3427         if (idx < 0) {
3428                 ret = -ENOSPC;
3429                 goto out_free_ctrl;
3430         }
3431
3432         /*
3433          * if ctrl_loss_tmo is being enforced and the default reconnect delay
3434          * is being used, change to a shorter reconnect delay for FC.
3435          */
3436         if (opts->max_reconnects != -1 &&
3437             opts->reconnect_delay == NVMF_DEF_RECONNECT_DELAY &&
3438             opts->reconnect_delay > NVME_FC_DEFAULT_RECONNECT_TMO) {
3439                 ctrl_loss_tmo = opts->max_reconnects * opts->reconnect_delay;
3440                 opts->reconnect_delay = NVME_FC_DEFAULT_RECONNECT_TMO;
3441                 opts->max_reconnects = DIV_ROUND_UP(ctrl_loss_tmo,
3442                                                 opts->reconnect_delay);
3443         }
3444
3445         ctrl->ctrl.opts = opts;
3446         ctrl->ctrl.nr_reconnects = 0;
3447         if (lport->dev)
3448                 ctrl->ctrl.numa_node = dev_to_node(lport->dev);
3449         else
3450                 ctrl->ctrl.numa_node = NUMA_NO_NODE;
3451         INIT_LIST_HEAD(&ctrl->ctrl_list);
3452         ctrl->lport = lport;
3453         ctrl->rport = rport;
3454         ctrl->dev = lport->dev;
3455         ctrl->cnum = idx;
3456         ctrl->ioq_live = false;
3457         init_waitqueue_head(&ctrl->ioabort_wait);
3458
3459         get_device(ctrl->dev);
3460         kref_init(&ctrl->ref);
3461
3462         INIT_WORK(&ctrl->ctrl.reset_work, nvme_fc_reset_ctrl_work);
3463         INIT_DELAYED_WORK(&ctrl->connect_work, nvme_fc_connect_ctrl_work);
3464         INIT_WORK(&ctrl->ioerr_work, nvme_fc_ctrl_ioerr_work);
3465         spin_lock_init(&ctrl->lock);
3466
3467         /* io queue count */
3468         ctrl->ctrl.queue_count = min_t(unsigned int,
3469                                 opts->nr_io_queues,
3470                                 lport->ops->max_hw_queues);
3471         ctrl->ctrl.queue_count++;       /* +1 for admin queue */
3472
3473         ctrl->ctrl.sqsize = opts->queue_size - 1;
3474         ctrl->ctrl.kato = opts->kato;
3475         ctrl->ctrl.cntlid = 0xffff;
3476
3477         ret = -ENOMEM;
3478         ctrl->queues = kcalloc(ctrl->ctrl.queue_count,
3479                                 sizeof(struct nvme_fc_queue), GFP_KERNEL);
3480         if (!ctrl->queues)
3481                 goto out_free_ida;
3482
3483         nvme_fc_init_queue(ctrl, 0);
3484
3485         memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
3486         ctrl->admin_tag_set.ops = &nvme_fc_admin_mq_ops;
3487         ctrl->admin_tag_set.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
3488         ctrl->admin_tag_set.reserved_tags = NVMF_RESERVED_TAGS;
3489         ctrl->admin_tag_set.numa_node = ctrl->ctrl.numa_node;
3490         ctrl->admin_tag_set.cmd_size =
3491                 struct_size((struct nvme_fcp_op_w_sgl *)NULL, priv,
3492                             ctrl->lport->ops->fcprqst_priv_sz);
3493         ctrl->admin_tag_set.driver_data = ctrl;
3494         ctrl->admin_tag_set.nr_hw_queues = 1;
3495         ctrl->admin_tag_set.timeout = NVME_ADMIN_TIMEOUT;
3496         ctrl->admin_tag_set.flags = BLK_MQ_F_NO_SCHED;
3497
3498         ret = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
3499         if (ret)
3500                 goto out_free_queues;
3501         ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
3502
3503         ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
3504         if (IS_ERR(ctrl->ctrl.fabrics_q)) {
3505                 ret = PTR_ERR(ctrl->ctrl.fabrics_q);
3506                 goto out_free_admin_tag_set;
3507         }
3508
3509         ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
3510         if (IS_ERR(ctrl->ctrl.admin_q)) {
3511                 ret = PTR_ERR(ctrl->ctrl.admin_q);
3512                 goto out_cleanup_fabrics_q;
3513         }
3514
3515         /*
3516          * Would have been nice to init io queues tag set as well.
3517          * However, we require interaction from the controller
3518          * for max io queue count before we can do so.
3519          * Defer this to the connect path.
3520          */
3521
3522         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_fc_ctrl_ops, 0);
3523         if (ret)
3524                 goto out_cleanup_admin_q;
3525
3526         /* at this point, teardown path changes to ref counting on nvme ctrl */
3527
3528         spin_lock_irqsave(&rport->lock, flags);
3529         list_add_tail(&ctrl->ctrl_list, &rport->ctrl_list);
3530         spin_unlock_irqrestore(&rport->lock, flags);
3531
3532         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING) ||
3533             !nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
3534                 dev_err(ctrl->ctrl.device,
3535                         "NVME-FC{%d}: failed to init ctrl state\n", ctrl->cnum);
3536                 goto fail_ctrl;
3537         }
3538
3539         if (!queue_delayed_work(nvme_wq, &ctrl->connect_work, 0)) {
3540                 dev_err(ctrl->ctrl.device,
3541                         "NVME-FC{%d}: failed to schedule initial connect\n",
3542                         ctrl->cnum);
3543                 goto fail_ctrl;
3544         }
3545
3546         flush_delayed_work(&ctrl->connect_work);
3547
3548         dev_info(ctrl->ctrl.device,
3549                 "NVME-FC{%d}: new ctrl: NQN \"%s\"\n",
3550                 ctrl->cnum, ctrl->ctrl.opts->subsysnqn);
3551
3552         return &ctrl->ctrl;
3553
3554 fail_ctrl:
3555         nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING);
3556         cancel_work_sync(&ctrl->ioerr_work);
3557         cancel_work_sync(&ctrl->ctrl.reset_work);
3558         cancel_delayed_work_sync(&ctrl->connect_work);
3559
3560         ctrl->ctrl.opts = NULL;
3561
3562         /* initiate nvme ctrl ref counting teardown */
3563         nvme_uninit_ctrl(&ctrl->ctrl);
3564
3565         /* Remove core ctrl ref. */
3566         nvme_put_ctrl(&ctrl->ctrl);
3567
3568         /* as we're past the point where we transition to the ref
3569          * counting teardown path, if we return a bad pointer here,
3570          * the calling routine, thinking it's prior to the
3571          * transition, will do an rport put. Since the teardown
3572          * path also does a rport put, we do an extra get here to
3573          * so proper order/teardown happens.
3574          */
3575         nvme_fc_rport_get(rport);
3576
3577         return ERR_PTR(-EIO);
3578
3579 out_cleanup_admin_q:
3580         blk_cleanup_queue(ctrl->ctrl.admin_q);
3581 out_cleanup_fabrics_q:
3582         blk_cleanup_queue(ctrl->ctrl.fabrics_q);
3583 out_free_admin_tag_set:
3584         blk_mq_free_tag_set(&ctrl->admin_tag_set);
3585 out_free_queues:
3586         kfree(ctrl->queues);
3587 out_free_ida:
3588         put_device(ctrl->dev);
3589         ida_simple_remove(&nvme_fc_ctrl_cnt, ctrl->cnum);
3590 out_free_ctrl:
3591         kfree(ctrl);
3592 out_fail:
3593         /* exit via here doesn't follow ctlr ref points */
3594         return ERR_PTR(ret);
3595 }
3596
3597
3598 struct nvmet_fc_traddr {
3599         u64     nn;
3600         u64     pn;
3601 };
3602
3603 static int
3604 __nvme_fc_parse_u64(substring_t *sstr, u64 *val)
3605 {
3606         u64 token64;
3607
3608         if (match_u64(sstr, &token64))
3609                 return -EINVAL;
3610         *val = token64;
3611
3612         return 0;
3613 }
3614
3615 /*
3616  * This routine validates and extracts the WWN's from the TRADDR string.
3617  * As kernel parsers need the 0x to determine number base, universally
3618  * build string to parse with 0x prefix before parsing name strings.
3619  */
3620 static int
3621 nvme_fc_parse_traddr(struct nvmet_fc_traddr *traddr, char *buf, size_t blen)
3622 {
3623         char name[2 + NVME_FC_TRADDR_HEXNAMELEN + 1];
3624         substring_t wwn = { name, &name[sizeof(name)-1] };
3625         int nnoffset, pnoffset;
3626
3627         /* validate if string is one of the 2 allowed formats */
3628         if (strnlen(buf, blen) == NVME_FC_TRADDR_MAXLENGTH &&
3629                         !strncmp(buf, "nn-0x", NVME_FC_TRADDR_OXNNLEN) &&
3630                         !strncmp(&buf[NVME_FC_TRADDR_MAX_PN_OFFSET],
3631                                 "pn-0x", NVME_FC_TRADDR_OXNNLEN)) {
3632                 nnoffset = NVME_FC_TRADDR_OXNNLEN;
3633                 pnoffset = NVME_FC_TRADDR_MAX_PN_OFFSET +
3634                                                 NVME_FC_TRADDR_OXNNLEN;
3635         } else if ((strnlen(buf, blen) == NVME_FC_TRADDR_MINLENGTH &&
3636                         !strncmp(buf, "nn-", NVME_FC_TRADDR_NNLEN) &&
3637                         !strncmp(&buf[NVME_FC_TRADDR_MIN_PN_OFFSET],
3638                                 "pn-", NVME_FC_TRADDR_NNLEN))) {
3639                 nnoffset = NVME_FC_TRADDR_NNLEN;
3640                 pnoffset = NVME_FC_TRADDR_MIN_PN_OFFSET + NVME_FC_TRADDR_NNLEN;
3641         } else
3642                 goto out_einval;
3643
3644         name[0] = '0';
3645         name[1] = 'x';
3646         name[2 + NVME_FC_TRADDR_HEXNAMELEN] = 0;
3647
3648         memcpy(&name[2], &buf[nnoffset], NVME_FC_TRADDR_HEXNAMELEN);
3649         if (__nvme_fc_parse_u64(&wwn, &traddr->nn))
3650                 goto out_einval;
3651
3652         memcpy(&name[2], &buf[pnoffset], NVME_FC_TRADDR_HEXNAMELEN);
3653         if (__nvme_fc_parse_u64(&wwn, &traddr->pn))
3654                 goto out_einval;
3655
3656         return 0;
3657
3658 out_einval:
3659         pr_warn("%s: bad traddr string\n", __func__);
3660         return -EINVAL;
3661 }
3662
3663 static struct nvme_ctrl *
3664 nvme_fc_create_ctrl(struct device *dev, struct nvmf_ctrl_options *opts)
3665 {
3666         struct nvme_fc_lport *lport;
3667         struct nvme_fc_rport *rport;
3668         struct nvme_ctrl *ctrl;
3669         struct nvmet_fc_traddr laddr = { 0L, 0L };
3670         struct nvmet_fc_traddr raddr = { 0L, 0L };
3671         unsigned long flags;
3672         int ret;
3673
3674         ret = nvme_fc_parse_traddr(&raddr, opts->traddr, NVMF_TRADDR_SIZE);
3675         if (ret || !raddr.nn || !raddr.pn)
3676                 return ERR_PTR(-EINVAL);
3677
3678         ret = nvme_fc_parse_traddr(&laddr, opts->host_traddr, NVMF_TRADDR_SIZE);
3679         if (ret || !laddr.nn || !laddr.pn)
3680                 return ERR_PTR(-EINVAL);
3681
3682         /* find the host and remote ports to connect together */
3683         spin_lock_irqsave(&nvme_fc_lock, flags);
3684         list_for_each_entry(lport, &nvme_fc_lport_list, port_list) {
3685                 if (lport->localport.node_name != laddr.nn ||
3686                     lport->localport.port_name != laddr.pn ||
3687                     lport->localport.port_state != FC_OBJSTATE_ONLINE)
3688                         continue;
3689
3690                 list_for_each_entry(rport, &lport->endp_list, endp_list) {
3691                         if (rport->remoteport.node_name != raddr.nn ||
3692                             rport->remoteport.port_name != raddr.pn ||
3693                             rport->remoteport.port_state != FC_OBJSTATE_ONLINE)
3694                                 continue;
3695
3696                         /* if fail to get reference fall through. Will error */
3697                         if (!nvme_fc_rport_get(rport))
3698                                 break;
3699
3700                         spin_unlock_irqrestore(&nvme_fc_lock, flags);
3701
3702                         ctrl = nvme_fc_init_ctrl(dev, opts, lport, rport);
3703                         if (IS_ERR(ctrl))
3704                                 nvme_fc_rport_put(rport);
3705                         return ctrl;
3706                 }
3707         }
3708         spin_unlock_irqrestore(&nvme_fc_lock, flags);
3709
3710         pr_warn("%s: %s - %s combination not found\n",
3711                 __func__, opts->traddr, opts->host_traddr);
3712         return ERR_PTR(-ENOENT);
3713 }
3714
3715
3716 static struct nvmf_transport_ops nvme_fc_transport = {
3717         .name           = "fc",
3718         .module         = THIS_MODULE,
3719         .required_opts  = NVMF_OPT_TRADDR | NVMF_OPT_HOST_TRADDR,
3720         .allowed_opts   = NVMF_OPT_RECONNECT_DELAY | NVMF_OPT_CTRL_LOSS_TMO,
3721         .create_ctrl    = nvme_fc_create_ctrl,
3722 };
3723
3724 /* Arbitrary successive failures max. With lots of subsystems could be high */
3725 #define DISCOVERY_MAX_FAIL      20
3726
3727 static ssize_t nvme_fc_nvme_discovery_store(struct device *dev,
3728                 struct device_attribute *attr, const char *buf, size_t count)
3729 {
3730         unsigned long flags;
3731         LIST_HEAD(local_disc_list);
3732         struct nvme_fc_lport *lport;
3733         struct nvme_fc_rport *rport;
3734         int failcnt = 0;
3735
3736         spin_lock_irqsave(&nvme_fc_lock, flags);
3737 restart:
3738         list_for_each_entry(lport, &nvme_fc_lport_list, port_list) {
3739                 list_for_each_entry(rport, &lport->endp_list, endp_list) {
3740                         if (!nvme_fc_lport_get(lport))
3741                                 continue;
3742                         if (!nvme_fc_rport_get(rport)) {
3743                                 /*
3744                                  * This is a temporary condition. Upon restart
3745                                  * this rport will be gone from the list.
3746                                  *
3747                                  * Revert the lport put and retry.  Anything
3748                                  * added to the list already will be skipped (as
3749                                  * they are no longer list_empty).  Loops should
3750                                  * resume at rports that were not yet seen.
3751                                  */
3752                                 nvme_fc_lport_put(lport);
3753
3754                                 if (failcnt++ < DISCOVERY_MAX_FAIL)
3755                                         goto restart;
3756
3757                                 pr_err("nvme_discovery: too many reference "
3758                                        "failures\n");
3759                                 goto process_local_list;
3760                         }
3761                         if (list_empty(&rport->disc_list))
3762                                 list_add_tail(&rport->disc_list,
3763                                               &local_disc_list);
3764                 }
3765         }
3766
3767 process_local_list:
3768         while (!list_empty(&local_disc_list)) {
3769                 rport = list_first_entry(&local_disc_list,
3770                                          struct nvme_fc_rport, disc_list);
3771                 list_del_init(&rport->disc_list);
3772                 spin_unlock_irqrestore(&nvme_fc_lock, flags);
3773
3774                 lport = rport->lport;
3775                 /* signal discovery. Won't hurt if it repeats */
3776                 nvme_fc_signal_discovery_scan(lport, rport);
3777                 nvme_fc_rport_put(rport);
3778                 nvme_fc_lport_put(lport);
3779
3780                 spin_lock_irqsave(&nvme_fc_lock, flags);
3781         }
3782         spin_unlock_irqrestore(&nvme_fc_lock, flags);
3783
3784         return count;
3785 }
3786 static DEVICE_ATTR(nvme_discovery, 0200, NULL, nvme_fc_nvme_discovery_store);
3787
3788 static struct attribute *nvme_fc_attrs[] = {
3789         &dev_attr_nvme_discovery.attr,
3790         NULL
3791 };
3792
3793 static const struct attribute_group nvme_fc_attr_group = {
3794         .attrs = nvme_fc_attrs,
3795 };
3796
3797 static const struct attribute_group *nvme_fc_attr_groups[] = {
3798         &nvme_fc_attr_group,
3799         NULL
3800 };
3801
3802 static struct class fc_class = {
3803         .name = "fc",
3804         .dev_groups = nvme_fc_attr_groups,
3805         .owner = THIS_MODULE,
3806 };
3807
3808 static int __init nvme_fc_init_module(void)
3809 {
3810         int ret;
3811
3812         nvme_fc_wq = alloc_workqueue("nvme_fc_wq", WQ_MEM_RECLAIM, 0);
3813         if (!nvme_fc_wq)
3814                 return -ENOMEM;
3815
3816         /*
3817          * NOTE:
3818          * It is expected that in the future the kernel will combine
3819          * the FC-isms that are currently under scsi and now being
3820          * added to by NVME into a new standalone FC class. The SCSI
3821          * and NVME protocols and their devices would be under this
3822          * new FC class.
3823          *
3824          * As we need something to post FC-specific udev events to,
3825          * specifically for nvme probe events, start by creating the
3826          * new device class.  When the new standalone FC class is
3827          * put in place, this code will move to a more generic
3828          * location for the class.
3829          */
3830         ret = class_register(&fc_class);
3831         if (ret) {
3832                 pr_err("couldn't register class fc\n");
3833                 goto out_destroy_wq;
3834         }
3835
3836         /*
3837          * Create a device for the FC-centric udev events
3838          */
3839         fc_udev_device = device_create(&fc_class, NULL, MKDEV(0, 0), NULL,
3840                                 "fc_udev_device");
3841         if (IS_ERR(fc_udev_device)) {
3842                 pr_err("couldn't create fc_udev device!\n");
3843                 ret = PTR_ERR(fc_udev_device);
3844                 goto out_destroy_class;
3845         }
3846
3847         ret = nvmf_register_transport(&nvme_fc_transport);
3848         if (ret)
3849                 goto out_destroy_device;
3850
3851         return 0;
3852
3853 out_destroy_device:
3854         device_destroy(&fc_class, MKDEV(0, 0));
3855 out_destroy_class:
3856         class_unregister(&fc_class);
3857 out_destroy_wq:
3858         destroy_workqueue(nvme_fc_wq);
3859
3860         return ret;
3861 }
3862
3863 static void
3864 nvme_fc_delete_controllers(struct nvme_fc_rport *rport)
3865 {
3866         struct nvme_fc_ctrl *ctrl;
3867
3868         spin_lock(&rport->lock);
3869         list_for_each_entry(ctrl, &rport->ctrl_list, ctrl_list) {
3870                 dev_warn(ctrl->ctrl.device,
3871                         "NVME-FC{%d}: transport unloading: deleting ctrl\n",
3872                         ctrl->cnum);
3873                 nvme_delete_ctrl(&ctrl->ctrl);
3874         }
3875         spin_unlock(&rport->lock);
3876 }
3877
3878 static void
3879 nvme_fc_cleanup_for_unload(void)
3880 {
3881         struct nvme_fc_lport *lport;
3882         struct nvme_fc_rport *rport;
3883
3884         list_for_each_entry(lport, &nvme_fc_lport_list, port_list) {
3885                 list_for_each_entry(rport, &lport->endp_list, endp_list) {
3886                         nvme_fc_delete_controllers(rport);
3887                 }
3888         }
3889 }
3890
3891 static void __exit nvme_fc_exit_module(void)
3892 {
3893         unsigned long flags;
3894         bool need_cleanup = false;
3895
3896         spin_lock_irqsave(&nvme_fc_lock, flags);
3897         nvme_fc_waiting_to_unload = true;
3898         if (!list_empty(&nvme_fc_lport_list)) {
3899                 need_cleanup = true;
3900                 nvme_fc_cleanup_for_unload();
3901         }
3902         spin_unlock_irqrestore(&nvme_fc_lock, flags);
3903         if (need_cleanup) {
3904                 pr_info("%s: waiting for ctlr deletes\n", __func__);
3905                 wait_for_completion(&nvme_fc_unload_proceed);
3906                 pr_info("%s: ctrl deletes complete\n", __func__);
3907         }
3908
3909         nvmf_unregister_transport(&nvme_fc_transport);
3910
3911         ida_destroy(&nvme_fc_local_port_cnt);
3912         ida_destroy(&nvme_fc_ctrl_cnt);
3913
3914         device_destroy(&fc_class, MKDEV(0, 0));
3915         class_unregister(&fc_class);
3916         destroy_workqueue(nvme_fc_wq);
3917 }
3918
3919 module_init(nvme_fc_init_module);
3920 module_exit(nvme_fc_exit_module);
3921
3922 MODULE_LICENSE("GPL v2");