drivers/nvme/host/fabrics.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * NVMe over Fabrics common host code.
   4  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
   5  */
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7 #include <linux/init.h>
   8 #include <linux/miscdevice.h>
   9 #include <linux/module.h>
  10 #include <linux/mutex.h>
  11 #include <linux/parser.h>
  12 #include <linux/seq_file.h>
  13 #include "nvme.h"
  14 #include "fabrics.h"
  15
  16 static LIST_HEAD(nvmf_transports);
  17 static DECLARE_RWSEM(nvmf_transports_rwsem);
  18
  19 static LIST_HEAD(nvmf_hosts);
  20 static DEFINE_MUTEX(nvmf_hosts_mutex);
  21
  22 static struct nvmf_host *nvmf_default_host;
  23
  24 static struct nvmf_host *__nvmf_host_find(const char *hostnqn)
  25 {
  26         struct nvmf_host *host;
  27
  28         list_for_each_entry(host, &nvmf_hosts, list) {
  29                 if (!strcmp(host->nqn, hostnqn))
  30                         return host;
  31         }
  32
  33         return NULL;
  34 }
  35
  36 static struct nvmf_host *nvmf_host_add(const char *hostnqn)
  37 {
  38         struct nvmf_host *host;
  39
  40         mutex_lock(&nvmf_hosts_mutex);
  41         host = __nvmf_host_find(hostnqn);
  42         if (host) {
  43                 kref_get(&host->ref);
  44                 goto out_unlock;
  45         }
  46
  47         host = kmalloc(sizeof(*host), GFP_KERNEL);
  48         if (!host)
  49                 goto out_unlock;
  50
  51         kref_init(&host->ref);
  52         strlcpy(host->nqn, hostnqn, NVMF_NQN_SIZE);
  53
  54         list_add_tail(&host->list, &nvmf_hosts);
  55 out_unlock:
  56         mutex_unlock(&nvmf_hosts_mutex);
  57         return host;
  58 }
  59
  60 static struct nvmf_host *nvmf_host_default(void)
  61 {
  62         struct nvmf_host *host;
  63
  64         host = kmalloc(sizeof(*host), GFP_KERNEL);
  65         if (!host)
  66                 return NULL;
  67
  68         kref_init(&host->ref);
  69         uuid_gen(&host->id);
  70         snprintf(host->nqn, NVMF_NQN_SIZE,
  71                 "nqn.2014-08.org.nvmexpress:uuid:%pUb", &host->id);
  72
  73         mutex_lock(&nvmf_hosts_mutex);
  74         list_add_tail(&host->list, &nvmf_hosts);
  75         mutex_unlock(&nvmf_hosts_mutex);
  76
  77         return host;
  78 }
  79
  80 static void nvmf_host_destroy(struct kref *ref)
  81 {
  82         struct nvmf_host *host = container_of(ref, struct nvmf_host, ref);
  83
  84         mutex_lock(&nvmf_hosts_mutex);
  85         list_del(&host->list);
  86         mutex_unlock(&nvmf_hosts_mutex);
  87
  88         kfree(host);
  89 }
  90
  91 static void nvmf_host_put(struct nvmf_host *host)
  92 {
  93         if (host)
  94                 kref_put(&host->ref, nvmf_host_destroy);
  95 }
  96
  97 /**
  98  * nvmf_get_address() -  Get address/port
  99  * @ctrl:       Host NVMe controller instance which we got the address
 100  * @buf:        OUTPUT parameter that will contain the address/port
 101  * @size:       buffer size
 102  */
 103 int nvmf_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
 104 {
 105         int len = 0;
 106
 107         if (ctrl->opts->mask & NVMF_OPT_TRADDR)
 108                 len += scnprintf(buf, size, "traddr=%s", ctrl->opts->traddr);
 109         if (ctrl->opts->mask & NVMF_OPT_TRSVCID)
 110                 len += scnprintf(buf + len, size - len, "%strsvcid=%s",
 111                                 (len) ? "," : "", ctrl->opts->trsvcid);
 112         if (ctrl->opts->mask & NVMF_OPT_HOST_TRADDR)
 113                 len += scnprintf(buf + len, size - len, "%shost_traddr=%s",
 114                                 (len) ? "," : "", ctrl->opts->host_traddr);
 115         if (ctrl->opts->mask & NVMF_OPT_HOST_IFACE)
 116                 len += scnprintf(buf + len, size - len, "%shost_iface=%s",
 117                                 (len) ? "," : "", ctrl->opts->host_iface);
 118         len += scnprintf(buf + len, size - len, "\n");
 119
 120         return len;
 121 }
 122 EXPORT_SYMBOL_GPL(nvmf_get_address);
 123
 124 /**
 125  * nvmf_reg_read32() -  NVMe Fabrics "Property Get" API function.
 126  * @ctrl:       Host NVMe controller instance maintaining the admin
 127  *              queue used to submit the property read command to
 128  *              the allocated NVMe controller resource on the target system.
 129  * @off:        Starting offset value of the targeted property
 130  *              register (see the fabrics section of the NVMe standard).
 131  * @val:        OUTPUT parameter that will contain the value of
 132  *              the property after a successful read.
 133  *
 134  * Used by the host system to retrieve a 32-bit capsule property value
 135  * from an NVMe controller on the target system.
 136  *
 137  * ("Capsule property" is an "PCIe register concept" applied to the
 138  * NVMe fabrics space.)
 139  *
 140  * Return:
 141  *      0: successful read
 142  *      > 0: NVMe error status code
 143  *      < 0: Linux errno error code
 144  */
 145 int nvmf_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
 146 {
 147         struct nvme_command cmd;
 148         union nvme_result res;
 149         int ret;
 150
 151         memset(&cmd, 0, sizeof(cmd));
 152         cmd.prop_get.opcode = nvme_fabrics_command;
 153         cmd.prop_get.fctype = nvme_fabrics_type_property_get;
 154         cmd.prop_get.offset = cpu_to_le32(off);
 155
 156         ret = __nvme_submit_sync_cmd(ctrl->fabrics_q, &cmd, &res, NULL, 0, 0,
 157                         NVME_QID_ANY, 0, 0, false);
 158
 159         if (ret >= 0)
 160                 *val = le64_to_cpu(res.u64);
 161         if (unlikely(ret != 0))
 162                 dev_err(ctrl->device,
 163                         "Property Get error: %d, offset %#x\n",
 164                         ret > 0 ? ret & ~NVME_SC_DNR : ret, off);
 165
 166         return ret;
 167 }
 168 EXPORT_SYMBOL_GPL(nvmf_reg_read32);
 169
 170 /**
 171  * nvmf_reg_read64() -  NVMe Fabrics "Property Get" API function.
 172  * @ctrl:       Host NVMe controller instance maintaining the admin
 173  *              queue used to submit the property read command to
 174  *              the allocated controller resource on the target system.
 175  * @off:        Starting offset value of the targeted property
 176  *              register (see the fabrics section of the NVMe standard).
 177  * @val:        OUTPUT parameter that will contain the value of
 178  *              the property after a successful read.
 179  *
 180  * Used by the host system to retrieve a 64-bit capsule property value
 181  * from an NVMe controller on the target system.
 182  *
 183  * ("Capsule property" is an "PCIe register concept" applied to the
 184  * NVMe fabrics space.)
 185  *
 186  * Return:
 187  *      0: successful read
 188  *      > 0: NVMe error status code
 189  *      < 0: Linux errno error code
 190  */
 191 int nvmf_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
 192 {
 193         struct nvme_command cmd = { };
 194         union nvme_result res;
 195         int ret;
 196
 197         cmd.prop_get.opcode = nvme_fabrics_command;
 198         cmd.prop_get.fctype = nvme_fabrics_type_property_get;
 199         cmd.prop_get.attrib = 1;
 200         cmd.prop_get.offset = cpu_to_le32(off);
 201
 202         ret = __nvme_submit_sync_cmd(ctrl->fabrics_q, &cmd, &res, NULL, 0, 0,
 203                         NVME_QID_ANY, 0, 0, false);
 204
 205         if (ret >= 0)
 206                 *val = le64_to_cpu(res.u64);
 207         if (unlikely(ret != 0))
 208                 dev_err(ctrl->device,
 209                         "Property Get error: %d, offset %#x\n",
 210                         ret > 0 ? ret & ~NVME_SC_DNR : ret, off);
 211         return ret;
 212 }
 213 EXPORT_SYMBOL_GPL(nvmf_reg_read64);
 214
 215 /**
 216  * nvmf_reg_write32() -  NVMe Fabrics "Property Write" API function.
 217  * @ctrl:       Host NVMe controller instance maintaining the admin
 218  *              queue used to submit the property read command to
 219  *              the allocated NVMe controller resource on the target system.
 220  * @off:        Starting offset value of the targeted property
 221  *              register (see the fabrics section of the NVMe standard).
 222  * @val:        Input parameter that contains the value to be
 223  *              written to the property.
 224  *
 225  * Used by the NVMe host system to write a 32-bit capsule property value
 226  * to an NVMe controller on the target system.
 227  *
 228  * ("Capsule property" is an "PCIe register concept" applied to the
 229  * NVMe fabrics space.)
 230  *
 231  * Return:
 232  *      0: successful write
 233  *      > 0: NVMe error status code
 234  *      < 0: Linux errno error code
 235  */
 236 int nvmf_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
 237 {
 238         struct nvme_command cmd = { };
 239         int ret;
 240
 241         cmd.prop_set.opcode = nvme_fabrics_command;
 242         cmd.prop_set.fctype = nvme_fabrics_type_property_set;
 243         cmd.prop_set.attrib = 0;
 244         cmd.prop_set.offset = cpu_to_le32(off);
 245         cmd.prop_set.value = cpu_to_le64(val);
 246
 247         ret = __nvme_submit_sync_cmd(ctrl->fabrics_q, &cmd, NULL, NULL, 0, 0,
 248                         NVME_QID_ANY, 0, 0, false);
 249         if (unlikely(ret))
 250                 dev_err(ctrl->device,
 251                         "Property Set error: %d, offset %#x\n",
 252                         ret > 0 ? ret & ~NVME_SC_DNR : ret, off);
 253         return ret;
 254 }
 255 EXPORT_SYMBOL_GPL(nvmf_reg_write32);
 256
 257 /**
 258  * nvmf_log_connect_error() - Error-parsing-diagnostic print out function for
 259  *                              connect() errors.
 260  * @ctrl:       The specific /dev/nvmeX device that had the error.
 261  * @errval:     Error code to be decoded in a more human-friendly
 262  *              printout.
 263  * @offset:     For use with the NVMe error code
 264  *              NVME_SC_CONNECT_INVALID_PARAM.
 265  * @cmd:        This is the SQE portion of a submission capsule.
 266  * @data:       This is the "Data" portion of a submission capsule.
 267  */
 268 static void nvmf_log_connect_error(struct nvme_ctrl *ctrl,
 269                 int errval, int offset, struct nvme_command *cmd,
 270                 struct nvmf_connect_data *data)
 271 {
 272         int err_sctype = errval & ~NVME_SC_DNR;
 273
 274         switch (err_sctype) {
 275         case (NVME_SC_CONNECT_INVALID_PARAM):
 276                 if (offset >> 16) {
 277                         char *inv_data = "Connect Invalid Data Parameter";
 278
 279                         switch (offset & 0xffff) {
 280                         case (offsetof(struct nvmf_connect_data, cntlid)):
 281                                 dev_err(ctrl->device,
 282                                         "%s, cntlid: %d\n",
 283                                         inv_data, data->cntlid);
 284                                 break;
 285                         case (offsetof(struct nvmf_connect_data, hostnqn)):
 286                                 dev_err(ctrl->device,
 287                                         "%s, hostnqn \"%s\"\n",
 288                                         inv_data, data->hostnqn);
 289                                 break;
 290                         case (offsetof(struct nvmf_connect_data, subsysnqn)):
 291                                 dev_err(ctrl->device,
 292                                         "%s, subsysnqn \"%s\"\n",
 293                                         inv_data, data->subsysnqn);
 294                                 break;
 295                         default:
 296                                 dev_err(ctrl->device,
 297                                         "%s, starting byte offset: %d\n",
 298                                        inv_data, offset & 0xffff);
 299                                 break;
 300                         }
 301                 } else {
 302                         char *inv_sqe = "Connect Invalid SQE Parameter";
 303
 304                         switch (offset) {
 305                         case (offsetof(struct nvmf_connect_command, qid)):
 306                                 dev_err(ctrl->device,
 307                                        "%s, qid %d\n",
 308                                         inv_sqe, cmd->connect.qid);
 309                                 break;
 310                         default:
 311                                 dev_err(ctrl->device,
 312                                         "%s, starting byte offset: %d\n",
 313                                         inv_sqe, offset);
 314                         }
 315                 }
 316                 break;
 317         case NVME_SC_CONNECT_INVALID_HOST:
 318                 dev_err(ctrl->device,
 319                         "Connect for subsystem %s is not allowed, hostnqn: %s\n",
 320                         data->subsysnqn, data->hostnqn);
 321                 break;
 322         case NVME_SC_CONNECT_CTRL_BUSY:
 323                 dev_err(ctrl->device,
 324                         "Connect command failed: controller is busy or not available\n");
 325                 break;
 326         case NVME_SC_CONNECT_FORMAT:
 327                 dev_err(ctrl->device,
 328                         "Connect incompatible format: %d",
 329                         cmd->connect.recfmt);
 330                 break;
 331         case NVME_SC_HOST_PATH_ERROR:
 332                 dev_err(ctrl->device,
 333                         "Connect command failed: host path error\n");
 334                 break;
 335         default:
 336                 dev_err(ctrl->device,
 337                         "Connect command failed, error wo/DNR bit: %d\n",
 338                         err_sctype);
 339                 break;
 340         }
 341 }
 342
 343 /**
 344  * nvmf_connect_admin_queue() - NVMe Fabrics Admin Queue "Connect"
 345  *                              API function.
 346  * @ctrl:       Host nvme controller instance used to request
 347  *              a new NVMe controller allocation on the target
 348  *              system and  establish an NVMe Admin connection to
 349  *              that controller.
 350  *
 351  * This function enables an NVMe host device to request a new allocation of
 352  * an NVMe controller resource on a target system as well establish a
 353  * fabrics-protocol connection of the NVMe Admin queue between the
 354  * host system device and the allocated NVMe controller on the
 355  * target system via a NVMe Fabrics "Connect" command.
 356  *
 357  * Return:
 358  *      0: success
 359  *      > 0: NVMe error status code
 360  *      < 0: Linux errno error code
 361  *
 362  */
 363 int nvmf_connect_admin_queue(struct nvme_ctrl *ctrl)
 364 {
 365         struct nvme_command cmd = { };
 366         union nvme_result res;
 367         struct nvmf_connect_data *data;
 368         int ret;
 369
 370         cmd.connect.opcode = nvme_fabrics_command;
 371         cmd.connect.fctype = nvme_fabrics_type_connect;
 372         cmd.connect.qid = 0;
 373         cmd.connect.sqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
 374
 375         /*
 376          * Set keep-alive timeout in seconds granularity (ms * 1000)
 377          */
 378         cmd.connect.kato = cpu_to_le32(ctrl->kato * 1000);
 379
 380         if (ctrl->opts->disable_sqflow)
 381                 cmd.connect.cattr |= NVME_CONNECT_DISABLE_SQFLOW;
 382
 383         data = kzalloc(sizeof(*data), GFP_KERNEL);
 384         if (!data)
 385                 return -ENOMEM;
 386
 387         uuid_copy(&data->hostid, &ctrl->opts->host->id);
 388         data->cntlid = cpu_to_le16(0xffff);
 389         strncpy(data->subsysnqn, ctrl->opts->subsysnqn, NVMF_NQN_SIZE);
 390         strncpy(data->hostnqn, ctrl->opts->host->nqn, NVMF_NQN_SIZE);
 391
 392         ret = __nvme_submit_sync_cmd(ctrl->fabrics_q, &cmd, &res,
 393                         data, sizeof(*data), 0, NVME_QID_ANY, 1,
 394                         BLK_MQ_REQ_RESERVED | BLK_MQ_REQ_NOWAIT, false);
 395         if (ret) {
 396                 nvmf_log_connect_error(ctrl, ret, le32_to_cpu(res.u32),
 397                                        &cmd, data);
 398                 goto out_free_data;
 399         }
 400
 401         ctrl->cntlid = le16_to_cpu(res.u16);
 402
 403 out_free_data:
 404         kfree(data);
 405         return ret;
 406 }
 407 EXPORT_SYMBOL_GPL(nvmf_connect_admin_queue);
 408
 409 /**
 410  * nvmf_connect_io_queue() - NVMe Fabrics I/O Queue "Connect"
 411  *                           API function.
 412  * @ctrl:       Host nvme controller instance used to establish an
 413  *              NVMe I/O queue connection to the already allocated NVMe
 414  *              controller on the target system.
 415  * @qid:        NVMe I/O queue number for the new I/O connection between
 416  *              host and target (note qid == 0 is illegal as this is
 417  *              the Admin queue, per NVMe standard).
 418  * @poll:       Whether or not to poll for the completion of the connect cmd.
 419  *
 420  * This function issues a fabrics-protocol connection
 421  * of a NVMe I/O queue (via NVMe Fabrics "Connect" command)
 422  * between the host system device and the allocated NVMe controller
 423  * on the target system.
 424  *
 425  * Return:
 426  *      0: success
 427  *      > 0: NVMe error status code
 428  *      < 0: Linux errno error code
 429  */
 430 int nvmf_connect_io_queue(struct nvme_ctrl *ctrl, u16 qid, bool poll)
 431 {
 432         struct nvme_command cmd;
 433         struct nvmf_connect_data *data;
 434         union nvme_result res;
 435         int ret;
 436
 437         memset(&cmd, 0, sizeof(cmd));
 438         cmd.connect.opcode = nvme_fabrics_command;
 439         cmd.connect.fctype = nvme_fabrics_type_connect;
 440         cmd.connect.qid = cpu_to_le16(qid);
 441         cmd.connect.sqsize = cpu_to_le16(ctrl->sqsize);
 442
 443         if (ctrl->opts->disable_sqflow)
 444                 cmd.connect.cattr |= NVME_CONNECT_DISABLE_SQFLOW;
 445
 446         data = kzalloc(sizeof(*data), GFP_KERNEL);
 447         if (!data)
 448                 return -ENOMEM;
 449
 450         uuid_copy(&data->hostid, &ctrl->opts->host->id);
 451         data->cntlid = cpu_to_le16(ctrl->cntlid);
 452         strncpy(data->subsysnqn, ctrl->opts->subsysnqn, NVMF_NQN_SIZE);
 453         strncpy(data->hostnqn, ctrl->opts->host->nqn, NVMF_NQN_SIZE);
 454
 455         ret = __nvme_submit_sync_cmd(ctrl->connect_q, &cmd, &res,
 456                         data, sizeof(*data), 0, qid, 1,
 457                         BLK_MQ_REQ_RESERVED | BLK_MQ_REQ_NOWAIT, poll);
 458         if (ret) {
 459                 nvmf_log_connect_error(ctrl, ret, le32_to_cpu(res.u32),
 460                                        &cmd, data);
 461         }
 462         kfree(data);
 463         return ret;
 464 }
 465 EXPORT_SYMBOL_GPL(nvmf_connect_io_queue);
 466
 467 bool nvmf_should_reconnect(struct nvme_ctrl *ctrl)
 468 {
 469         if (ctrl->opts->max_reconnects == -1 ||
 470             ctrl->nr_reconnects < ctrl->opts->max_reconnects)
 471                 return true;
 472
 473         return false;
 474 }
 475 EXPORT_SYMBOL_GPL(nvmf_should_reconnect);
 476
 477 /**
 478  * nvmf_register_transport() - NVMe Fabrics Library registration function.
 479  * @ops:        Transport ops instance to be registered to the
 480  *              common fabrics library.
 481  *
 482  * API function that registers the type of specific transport fabric
 483  * being implemented to the common NVMe fabrics library. Part of
 484  * the overall init sequence of starting up a fabrics driver.
 485  */
 486 int nvmf_register_transport(struct nvmf_transport_ops *ops)
 487 {
 488         if (!ops->create_ctrl)
 489                 return -EINVAL;
 490
 491         down_write(&nvmf_transports_rwsem);
 492         list_add_tail(&ops->entry, &nvmf_transports);
 493         up_write(&nvmf_transports_rwsem);
 494
 495         return 0;
 496 }
 497 EXPORT_SYMBOL_GPL(nvmf_register_transport);
 498
 499 /**
 500  * nvmf_unregister_transport() - NVMe Fabrics Library unregistration function.
 501  * @ops:        Transport ops instance to be unregistered from the
 502  *              common fabrics library.
 503  *
 504  * Fabrics API function that unregisters the type of specific transport
 505  * fabric being implemented from the common NVMe fabrics library.
 506  * Part of the overall exit sequence of unloading the implemented driver.
 507  */
 508 void nvmf_unregister_transport(struct nvmf_transport_ops *ops)
 509 {
 510         down_write(&nvmf_transports_rwsem);
 511         list_del(&ops->entry);
 512         up_write(&nvmf_transports_rwsem);
 513 }
 514 EXPORT_SYMBOL_GPL(nvmf_unregister_transport);
 515
 516 static struct nvmf_transport_ops *nvmf_lookup_transport(
 517                 struct nvmf_ctrl_options *opts)
 518 {
 519         struct nvmf_transport_ops *ops;
 520
 521         lockdep_assert_held(&nvmf_transports_rwsem);
 522
 523         list_for_each_entry(ops, &nvmf_transports, entry) {
 524                 if (strcmp(ops->name, opts->transport) == 0)
 525                         return ops;
 526         }
 527
 528         return NULL;
 529 }
 530
 531 static const match_table_t opt_tokens = {
 532         { NVMF_OPT_TRANSPORT,           "transport=%s"          },
 533         { NVMF_OPT_TRADDR,              "traddr=%s"             },
 534         { NVMF_OPT_TRSVCID,             "trsvcid=%s"            },
 535         { NVMF_OPT_NQN,                 "nqn=%s"                },
 536         { NVMF_OPT_QUEUE_SIZE,          "queue_size=%d"         },
 537         { NVMF_OPT_NR_IO_QUEUES,        "nr_io_queues=%d"       },
 538         { NVMF_OPT_RECONNECT_DELAY,     "reconnect_delay=%d"    },
 539         { NVMF_OPT_CTRL_LOSS_TMO,       "ctrl_loss_tmo=%d"      },
 540         { NVMF_OPT_KATO,                "keep_alive_tmo=%d"     },
 541         { NVMF_OPT_HOSTNQN,             "hostnqn=%s"            },
 542         { NVMF_OPT_HOST_TRADDR,         "host_traddr=%s"        },
 543         { NVMF_OPT_HOST_IFACE,          "host_iface=%s"         },
 544         { NVMF_OPT_HOST_ID,             "hostid=%s"             },
 545         { NVMF_OPT_DUP_CONNECT,         "duplicate_connect"     },
 546         { NVMF_OPT_DISABLE_SQFLOW,      "disable_sqflow"        },
 547         { NVMF_OPT_HDR_DIGEST,          "hdr_digest"            },
 548         { NVMF_OPT_DATA_DIGEST,         "data_digest"           },
 549         { NVMF_OPT_NR_WRITE_QUEUES,     "nr_write_queues=%d"    },
 550         { NVMF_OPT_NR_POLL_QUEUES,      "nr_poll_queues=%d"     },
 551         { NVMF_OPT_TOS,                 "tos=%d"                },
 552         { NVMF_OPT_FAIL_FAST_TMO,       "fast_io_fail_tmo=%d"   },
 553         { NVMF_OPT_ERR,                 NULL                    }
 554 };
 555
 556 static int nvmf_parse_options(struct nvmf_ctrl_options *opts,
 557                 const char *buf)
 558 {
 559         substring_t args[MAX_OPT_ARGS];
 560         char *options, *o, *p;
 561         int token, ret = 0;
 562         size_t nqnlen  = 0;
 563         int ctrl_loss_tmo = NVMF_DEF_CTRL_LOSS_TMO;
 564         uuid_t hostid;
 565
 566         /* Set defaults */
 567         opts->queue_size = NVMF_DEF_QUEUE_SIZE;
 568         opts->nr_io_queues = num_online_cpus();
 569         opts->reconnect_delay = NVMF_DEF_RECONNECT_DELAY;
 570         opts->kato = 0;
 571         opts->duplicate_connect = false;
 572         opts->fast_io_fail_tmo = NVMF_DEF_FAIL_FAST_TMO;
 573         opts->hdr_digest = false;
 574         opts->data_digest = false;
 575         opts->tos = -1; /* < 0 == use transport default */
 576
 577         options = o = kstrdup(buf, GFP_KERNEL);
 578         if (!options)
 579                 return -ENOMEM;
 580
 581         uuid_gen(&hostid);
 582
 583         while ((p = strsep(&o, ",\n")) != NULL) {
 584                 if (!*p)
 585                         continue;
 586
 587                 token = match_token(p, opt_tokens, args);
 588                 opts->mask |= token;
 589                 switch (token) {
 590                 case NVMF_OPT_TRANSPORT:
 591                         p = match_strdup(args);
 592                         if (!p) {
 593                                 ret = -ENOMEM;
 594                                 goto out;
 595                         }
 596                         kfree(opts->transport);
 597                         opts->transport = p;
 598                         break;
 599                 case NVMF_OPT_NQN:
 600                         p = match_strdup(args);
 601                         if (!p) {
 602                                 ret = -ENOMEM;
 603                                 goto out;
 604                         }
 605                         kfree(opts->subsysnqn);
 606                         opts->subsysnqn = p;
 607                         nqnlen = strlen(opts->subsysnqn);
 608                         if (nqnlen >= NVMF_NQN_SIZE) {
 609                                 pr_err("%s needs to be < %d bytes\n",
 610                                         opts->subsysnqn, NVMF_NQN_SIZE);
 611                                 ret = -EINVAL;
 612                                 goto out;
 613                         }
 614                         opts->discovery_nqn =
 615                                 !(strcmp(opts->subsysnqn,
 616                                          NVME_DISC_SUBSYS_NAME));
 617                         break;
 618                 case NVMF_OPT_TRADDR:
 619                         p = match_strdup(args);
 620                         if (!p) {
 621                                 ret = -ENOMEM;
 622                                 goto out;
 623                         }
 624                         kfree(opts->traddr);
 625                         opts->traddr = p;
 626                         break;
 627                 case NVMF_OPT_TRSVCID:
 628                         p = match_strdup(args);
 629                         if (!p) {
 630                                 ret = -ENOMEM;
 631                                 goto out;
 632                         }
 633                         kfree(opts->trsvcid);
 634                         opts->trsvcid = p;
 635                         break;
 636                 case NVMF_OPT_QUEUE_SIZE:
 637                         if (match_int(args, &token)) {
 638                                 ret = -EINVAL;
 639                                 goto out;
 640                         }
 641                         if (token < NVMF_MIN_QUEUE_SIZE ||
 642                             token > NVMF_MAX_QUEUE_SIZE) {
 643                                 pr_err("Invalid queue_size %d\n", token);
 644                                 ret = -EINVAL;
 645                                 goto out;
 646                         }
 647                         opts->queue_size = token;
 648                         break;
 649                 case NVMF_OPT_NR_IO_QUEUES:
 650                         if (match_int(args, &token)) {
 651                                 ret = -EINVAL;
 652                                 goto out;
 653                         }
 654                         if (token <= 0) {
 655                                 pr_err("Invalid number of IOQs %d\n", token);
 656                                 ret = -EINVAL;
 657                                 goto out;
 658                         }
 659                         if (opts->discovery_nqn) {
 660                                 pr_debug("Ignoring nr_io_queues value for discovery controller\n");
 661                                 break;
 662                         }
 663
 664                         opts->nr_io_queues = min_t(unsigned int,
 665                                         num_online_cpus(), token);
 666                         break;
 667                 case NVMF_OPT_KATO:
 668                         if (match_int(args, &token)) {
 669                                 ret = -EINVAL;
 670                                 goto out;
 671                         }
 672
 673                         if (token < 0) {
 674                                 pr_err("Invalid keep_alive_tmo %d\n", token);
 675                                 ret = -EINVAL;
 676                                 goto out;
 677                         } else if (token == 0 && !opts->discovery_nqn) {
 678                                 /* Allowed for debug */
 679                                 pr_warn("keep_alive_tmo 0 won't execute keep alives!!!\n");
 680                         }
 681                         opts->kato = token;
 682                         break;
 683                 case NVMF_OPT_CTRL_LOSS_TMO:
 684                         if (match_int(args, &token)) {
 685                                 ret = -EINVAL;
 686                                 goto out;
 687                         }
 688
 689                         if (token < 0)
 690                                 pr_warn("ctrl_loss_tmo < 0 will reconnect forever\n");
 691                         ctrl_loss_tmo = token;
 692                         break;
 693                 case NVMF_OPT_FAIL_FAST_TMO:
 694                         if (match_int(args, &token)) {
 695                                 ret = -EINVAL;
 696                                 goto out;
 697                         }
 698
 699                         if (token >= 0)
 700                                 pr_warn("I/O fail on reconnect controller after %d sec\n",
 701                                         token);
 702                         opts->fast_io_fail_tmo = token;
 703                         break;
 704                 case NVMF_OPT_HOSTNQN:
 705                         if (opts->host) {
 706                                 pr_err("hostnqn already user-assigned: %s\n",
 707                                        opts->host->nqn);
 708                                 ret = -EADDRINUSE;
 709                                 goto out;
 710                         }
 711                         p = match_strdup(args);
 712                         if (!p) {
 713                                 ret = -ENOMEM;
 714                                 goto out;
 715                         }
 716                         nqnlen = strlen(p);
 717                         if (nqnlen >= NVMF_NQN_SIZE) {
 718                                 pr_err("%s needs to be < %d bytes\n",
 719                                         p, NVMF_NQN_SIZE);
 720                                 kfree(p);
 721                                 ret = -EINVAL;
 722                                 goto out;
 723                         }
 724                         nvmf_host_put(opts->host);
 725                         opts->host = nvmf_host_add(p);
 726                         kfree(p);
 727                         if (!opts->host) {
 728                                 ret = -ENOMEM;
 729                                 goto out;
 730                         }
 731                         break;
 732                 case NVMF_OPT_RECONNECT_DELAY:
 733                         if (match_int(args, &token)) {
 734                                 ret = -EINVAL;
 735                                 goto out;
 736                         }
 737                         if (token <= 0) {
 738                                 pr_err("Invalid reconnect_delay %d\n", token);
 739                                 ret = -EINVAL;
 740                                 goto out;
 741                         }
 742                         opts->reconnect_delay = token;
 743                         break;
 744                 case NVMF_OPT_HOST_TRADDR:
 745                         p = match_strdup(args);
 746                         if (!p) {
 747                                 ret = -ENOMEM;
 748                                 goto out;
 749                         }
 750                         kfree(opts->host_traddr);
 751                         opts->host_traddr = p;
 752                         break;
 753                 case NVMF_OPT_HOST_IFACE:
 754                         p = match_strdup(args);
 755                         if (!p) {
 756                                 ret = -ENOMEM;
 757                                 goto out;
 758                         }
 759                         kfree(opts->host_iface);
 760                         opts->host_iface = p;
 761                         break;
 762                 case NVMF_OPT_HOST_ID:
 763                         p = match_strdup(args);
 764                         if (!p) {
 765                                 ret = -ENOMEM;
 766                                 goto out;
 767                         }
 768                         ret = uuid_parse(p, &hostid);
 769                         if (ret) {
 770                                 pr_err("Invalid hostid %s\n", p);
 771                                 ret = -EINVAL;
 772                                 kfree(p);
 773                                 goto out;
 774                         }
 775                         kfree(p);
 776                         break;
 777                 case NVMF_OPT_DUP_CONNECT:
 778                         opts->duplicate_connect = true;
 779                         break;
 780                 case NVMF_OPT_DISABLE_SQFLOW:
 781                         opts->disable_sqflow = true;
 782                         break;
 783                 case NVMF_OPT_HDR_DIGEST:
 784                         opts->hdr_digest = true;
 785                         break;
 786                 case NVMF_OPT_DATA_DIGEST:
 787                         opts->data_digest = true;
 788                         break;
 789                 case NVMF_OPT_NR_WRITE_QUEUES:
 790                         if (match_int(args, &token)) {
 791                                 ret = -EINVAL;
 792                                 goto out;
 793                         }
 794                         if (token <= 0) {
 795                                 pr_err("Invalid nr_write_queues %d\n", token);
 796                                 ret = -EINVAL;
 797                                 goto out;
 798                         }
 799                         opts->nr_write_queues = token;
 800                         break;
 801                 case NVMF_OPT_NR_POLL_QUEUES:
 802                         if (match_int(args, &token)) {
 803                                 ret = -EINVAL;
 804                                 goto out;
 805                         }
 806                         if (token <= 0) {
 807                                 pr_err("Invalid nr_poll_queues %d\n", token);
 808                                 ret = -EINVAL;
 809                                 goto out;
 810                         }
 811                         opts->nr_poll_queues = token;
 812                         break;
 813                 case NVMF_OPT_TOS:
 814                         if (match_int(args, &token)) {
 815                                 ret = -EINVAL;
 816                                 goto out;
 817                         }
 818                         if (token < 0) {
 819                                 pr_err("Invalid type of service %d\n", token);
 820                                 ret = -EINVAL;
 821                                 goto out;
 822                         }
 823                         if (token > 255) {
 824                                 pr_warn("Clamping type of service to 255\n");
 825                                 token = 255;
 826                         }
 827                         opts->tos = token;
 828                         break;
 829                 default:
 830                         pr_warn("unknown parameter or missing value '%s' in ctrl creation request\n",
 831                                 p);
 832                         ret = -EINVAL;
 833                         goto out;
 834                 }
 835         }
 836
 837         if (opts->discovery_nqn) {
 838                 opts->nr_io_queues = 0;
 839                 opts->nr_write_queues = 0;
 840                 opts->nr_poll_queues = 0;
 841                 opts->duplicate_connect = true;
 842         } else {
 843                 if (!opts->kato)
 844                         opts->kato = NVME_DEFAULT_KATO;
 845         }
 846         if (ctrl_loss_tmo < 0) {
 847                 opts->max_reconnects = -1;
 848         } else {
 849                 opts->max_reconnects = DIV_ROUND_UP(ctrl_loss_tmo,
 850                                                 opts->reconnect_delay);
 851                 if (ctrl_loss_tmo < opts->fast_io_fail_tmo)
 852                         pr_warn("failfast tmo (%d) larger than controller loss tmo (%d)\n",
 853                                 opts->fast_io_fail_tmo, ctrl_loss_tmo);
 854         }
 855
 856         if (!opts->host) {
 857                 kref_get(&nvmf_default_host->ref);
 858                 opts->host = nvmf_default_host;
 859         }
 860
 861         uuid_copy(&opts->host->id, &hostid);
 862
 863 out:
 864         kfree(options);
 865         return ret;
 866 }
 867
 868 static int nvmf_check_required_opts(struct nvmf_ctrl_options *opts,
 869                 unsigned int required_opts)
 870 {
 871         if ((opts->mask & required_opts) != required_opts) {
 872                 int i;
 873
 874                 for (i = 0; i < ARRAY_SIZE(opt_tokens); i++) {
 875                         if ((opt_tokens[i].token & required_opts) &&
 876                             !(opt_tokens[i].token & opts->mask)) {
 877                                 pr_warn("missing parameter '%s'\n",
 878                                         opt_tokens[i].pattern);
 879                         }
 880                 }
 881
 882                 return -EINVAL;
 883         }
 884
 885         return 0;
 886 }
 887
 888 bool nvmf_ip_options_match(struct nvme_ctrl *ctrl,
 889                 struct nvmf_ctrl_options *opts)
 890 {
 891         if (!nvmf_ctlr_matches_baseopts(ctrl, opts) ||
 892             strcmp(opts->traddr, ctrl->opts->traddr) ||
 893             strcmp(opts->trsvcid, ctrl->opts->trsvcid))
 894                 return false;
 895
 896         /*
 897          * Checking the local address is rough. In most cases, none is specified
 898          * and the host port is selected by the stack.
 899          *
 900          * Assume no match if:
 901          * -  local address is specified and address is not the same
 902          * -  local address is not specified but remote is, or vice versa
 903          *    (admin using specific host_traddr when it matters).
 904          */
 905         if ((opts->mask & NVMF_OPT_HOST_TRADDR) &&
 906             (ctrl->opts->mask & NVMF_OPT_HOST_TRADDR)) {
 907                 if (strcmp(opts->host_traddr, ctrl->opts->host_traddr))
 908                         return false;
 909         } else if ((opts->mask & NVMF_OPT_HOST_TRADDR) ||
 910                    (ctrl->opts->mask & NVMF_OPT_HOST_TRADDR)) {
 911                 return false;
 912         }
 913
 914         return true;
 915 }
 916 EXPORT_SYMBOL_GPL(nvmf_ip_options_match);
 917
 918 static int nvmf_check_allowed_opts(struct nvmf_ctrl_options *opts,
 919                 unsigned int allowed_opts)
 920 {
 921         if (opts->mask & ~allowed_opts) {
 922                 int i;
 923
 924                 for (i = 0; i < ARRAY_SIZE(opt_tokens); i++) {
 925                         if ((opt_tokens[i].token & opts->mask) &&
 926                             (opt_tokens[i].token & ~allowed_opts)) {
 927                                 pr_warn("invalid parameter '%s'\n",
 928                                         opt_tokens[i].pattern);
 929                         }
 930                 }
 931
 932                 return -EINVAL;
 933         }
 934
 935         return 0;
 936 }
 937
 938 void nvmf_free_options(struct nvmf_ctrl_options *opts)
 939 {
 940         nvmf_host_put(opts->host);
 941         kfree(opts->transport);
 942         kfree(opts->traddr);
 943         kfree(opts->trsvcid);
 944         kfree(opts->subsysnqn);
 945         kfree(opts->host_traddr);
 946         kfree(opts->host_iface);
 947         kfree(opts);
 948 }
 949 EXPORT_SYMBOL_GPL(nvmf_free_options);
 950
 951 #define NVMF_REQUIRED_OPTS      (NVMF_OPT_TRANSPORT | NVMF_OPT_NQN)
 952 #define NVMF_ALLOWED_OPTS       (NVMF_OPT_QUEUE_SIZE | NVMF_OPT_NR_IO_QUEUES | \
 953                                  NVMF_OPT_KATO | NVMF_OPT_HOSTNQN | \
 954                                  NVMF_OPT_HOST_ID | NVMF_OPT_DUP_CONNECT |\
 955                                  NVMF_OPT_DISABLE_SQFLOW |\
 956                                  NVMF_OPT_FAIL_FAST_TMO)
 957
 958 static struct nvme_ctrl *
 959 nvmf_create_ctrl(struct device *dev, const char *buf)
 960 {
 961         struct nvmf_ctrl_options *opts;
 962         struct nvmf_transport_ops *ops;
 963         struct nvme_ctrl *ctrl;
 964         int ret;
 965
 966         opts = kzalloc(sizeof(*opts), GFP_KERNEL);
 967         if (!opts)
 968                 return ERR_PTR(-ENOMEM);
 969
 970         ret = nvmf_parse_options(opts, buf);
 971         if (ret)
 972                 goto out_free_opts;
 973
 974
 975         request_module("nvme-%s", opts->transport);
 976
 977         /*
 978          * Check the generic options first as we need a valid transport for
 979          * the lookup below.  Then clear the generic flags so that transport
 980          * drivers don't have to care about them.
 981          */
 982         ret = nvmf_check_required_opts(opts, NVMF_REQUIRED_OPTS);
 983         if (ret)
 984                 goto out_free_opts;
 985         opts->mask &= ~NVMF_REQUIRED_OPTS;
 986
 987         down_read(&nvmf_transports_rwsem);
 988         ops = nvmf_lookup_transport(opts);
 989         if (!ops) {
 990                 pr_info("no handler found for transport %s.\n",
 991                         opts->transport);
 992                 ret = -EINVAL;
 993                 goto out_unlock;
 994         }
 995
 996         if (!try_module_get(ops->module)) {
 997                 ret = -EBUSY;
 998                 goto out_unlock;
 999         }
1000         up_read(&nvmf_transports_rwsem);
1001
1002         ret = nvmf_check_required_opts(opts, ops->required_opts);
1003         if (ret)
1004                 goto out_module_put;
1005         ret = nvmf_check_allowed_opts(opts, NVMF_ALLOWED_OPTS |
1006                                 ops->allowed_opts | ops->required_opts);
1007         if (ret)
1008                 goto out_module_put;
1009
1010         ctrl = ops->create_ctrl(dev, opts);
1011         if (IS_ERR(ctrl)) {
1012                 ret = PTR_ERR(ctrl);
1013                 goto out_module_put;
1014         }
1015
1016         module_put(ops->module);
1017         return ctrl;
1018
1019 out_module_put:
1020         module_put(ops->module);
1021         goto out_free_opts;
1022 out_unlock:
1023         up_read(&nvmf_transports_rwsem);
1024 out_free_opts:
1025         nvmf_free_options(opts);
1026         return ERR_PTR(ret);
1027 }
1028
1029 static struct class *nvmf_class;
1030 static struct device *nvmf_device;
1031 static DEFINE_MUTEX(nvmf_dev_mutex);
1032
1033 static ssize_t nvmf_dev_write(struct file *file, const char __user *ubuf,
1034                 size_t count, loff_t *pos)
1035 {
1036         struct seq_file *seq_file = file->private_data;
1037         struct nvme_ctrl *ctrl;
1038         const char *buf;
1039         int ret = 0;
1040
1041         if (count > PAGE_SIZE)
1042                 return -ENOMEM;
1043
1044         buf = memdup_user_nul(ubuf, count);
1045         if (IS_ERR(buf))
1046                 return PTR_ERR(buf);
1047
1048         mutex_lock(&nvmf_dev_mutex);
1049         if (seq_file->private) {
1050                 ret = -EINVAL;
1051                 goto out_unlock;
1052         }
1053
1054         ctrl = nvmf_create_ctrl(nvmf_device, buf);
1055         if (IS_ERR(ctrl)) {
1056                 ret = PTR_ERR(ctrl);
1057                 goto out_unlock;
1058         }
1059
1060         seq_file->private = ctrl;
1061
1062 out_unlock:
1063         mutex_unlock(&nvmf_dev_mutex);
1064         kfree(buf);
1065         return ret ? ret : count;
1066 }
1067
1068 static int nvmf_dev_show(struct seq_file *seq_file, void *private)
1069 {
1070         struct nvme_ctrl *ctrl;
1071         int ret = 0;
1072
1073         mutex_lock(&nvmf_dev_mutex);
1074         ctrl = seq_file->private;
1075         if (!ctrl) {
1076                 ret = -EINVAL;
1077                 goto out_unlock;
1078         }
1079
1080         seq_printf(seq_file, "instance=%d,cntlid=%d\n",
1081                         ctrl->instance, ctrl->cntlid);
1082
1083 out_unlock:
1084         mutex_unlock(&nvmf_dev_mutex);
1085         return ret;
1086 }
1087
1088 static int nvmf_dev_open(struct inode *inode, struct file *file)
1089 {
1090         /*
1091          * The miscdevice code initializes file->private_data, but doesn't
1092          * make use of it later.
1093          */
1094         file->private_data = NULL;
1095         return single_open(file, nvmf_dev_show, NULL);
1096 }
1097
1098 static int nvmf_dev_release(struct inode *inode, struct file *file)
1099 {
1100         struct seq_file *seq_file = file->private_data;
1101         struct nvme_ctrl *ctrl = seq_file->private;
1102
1103         if (ctrl)
1104                 nvme_put_ctrl(ctrl);
1105         return single_release(inode, file);
1106 }
1107
1108 static const struct file_operations nvmf_dev_fops = {
1109         .owner          = THIS_MODULE,
1110         .write          = nvmf_dev_write,
1111         .read           = seq_read,
1112         .open           = nvmf_dev_open,
1113         .release        = nvmf_dev_release,
1114 };
1115
1116 static struct miscdevice nvmf_misc = {
1117         .minor          = MISC_DYNAMIC_MINOR,
1118         .name           = "nvme-fabrics",
1119         .fops           = &nvmf_dev_fops,
1120 };
1121
1122 static int __init nvmf_init(void)
1123 {
1124         int ret;
1125
1126         nvmf_default_host = nvmf_host_default();
1127         if (!nvmf_default_host)
1128                 return -ENOMEM;
1129
1130         nvmf_class = class_create(THIS_MODULE, "nvme-fabrics");
1131         if (IS_ERR(nvmf_class)) {
1132                 pr_err("couldn't register class nvme-fabrics\n");
1133                 ret = PTR_ERR(nvmf_class);
1134                 goto out_free_host;
1135         }
1136
1137         nvmf_device =
1138                 device_create(nvmf_class, NULL, MKDEV(0, 0), NULL, "ctl");
1139         if (IS_ERR(nvmf_device)) {
1140                 pr_err("couldn't create nvme-fabris device!\n");
1141                 ret = PTR_ERR(nvmf_device);
1142                 goto out_destroy_class;
1143         }
1144
1145         ret = misc_register(&nvmf_misc);
1146         if (ret) {
1147                 pr_err("couldn't register misc device: %d\n", ret);
1148                 goto out_destroy_device;
1149         }
1150
1151         return 0;
1152
1153 out_destroy_device:
1154         device_destroy(nvmf_class, MKDEV(0, 0));
1155 out_destroy_class:
1156         class_destroy(nvmf_class);
1157 out_free_host:
1158         nvmf_host_put(nvmf_default_host);
1159         return ret;
1160 }
1161
1162 static void __exit nvmf_exit(void)
1163 {
1164         misc_deregister(&nvmf_misc);
1165         device_destroy(nvmf_class, MKDEV(0, 0));
1166         class_destroy(nvmf_class);
1167         nvmf_host_put(nvmf_default_host);
1168
1169         BUILD_BUG_ON(sizeof(struct nvmf_common_command) != 64);
1170         BUILD_BUG_ON(sizeof(struct nvmf_connect_command) != 64);
1171         BUILD_BUG_ON(sizeof(struct nvmf_property_get_command) != 64);
1172         BUILD_BUG_ON(sizeof(struct nvmf_property_set_command) != 64);
1173         BUILD_BUG_ON(sizeof(struct nvmf_connect_data) != 1024);
1174 }
1175
1176 MODULE_LICENSE("GPL v2");
1177
1178 module_init(nvmf_init);
1179 module_exit(nvmf_exit);