drivers/nvme/target/core.c

   1 /*
   2  * Common code for the NVMe target.
   3  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  */
  14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  15 #include <linux/module.h>
  16 #include <linux/random.h>
  17 #include "nvmet.h"
  18
  19 static struct nvmet_fabrics_ops *nvmet_transports[NVMF_TRTYPE_MAX];
  20
  21 /*
  22  * This read/write semaphore is used to synchronize access to configuration
  23  * information on a target system that will result in discovery log page
  24  * information change for at least one host.
  25  * The full list of resources to protected by this semaphore is:
  26  *
  27  *  - subsystems list
  28  *  - per-subsystem allowed hosts list
  29  *  - allow_any_host subsystem attribute
  30  *  - nvmet_genctr
  31  *  - the nvmet_transports array
  32  *
  33  * When updating any of those lists/structures write lock should be obtained,
  34  * while when reading (popolating discovery log page or checking host-subsystem
  35  * link) read lock is obtained to allow concurrent reads.
  36  */
  37 DECLARE_RWSEM(nvmet_config_sem);
  38
  39 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
  40                 const char *subsysnqn);
  41
  42 u16 nvmet_copy_to_sgl(struct nvmet_req *req, off_t off, const void *buf,
  43                 size_t len)
  44 {
  45         if (sg_pcopy_from_buffer(req->sg, req->sg_cnt, buf, len, off) != len)
  46                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
  47         return 0;
  48 }
  49
  50 u16 nvmet_copy_from_sgl(struct nvmet_req *req, off_t off, void *buf, size_t len)
  51 {
  52         if (sg_pcopy_to_buffer(req->sg, req->sg_cnt, buf, len, off) != len)
  53                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
  54         return 0;
  55 }
  56
  57 static u32 nvmet_async_event_result(struct nvmet_async_event *aen)
  58 {
  59         return aen->event_type | (aen->event_info << 8) | (aen->log_page << 16);
  60 }
  61
  62 static void nvmet_async_events_free(struct nvmet_ctrl *ctrl)
  63 {
  64         struct nvmet_req *req;
  65
  66         while (1) {
  67                 mutex_lock(&ctrl->lock);
  68                 if (!ctrl->nr_async_event_cmds) {
  69                         mutex_unlock(&ctrl->lock);
  70                         return;
  71                 }
  72
  73                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
  74                 mutex_unlock(&ctrl->lock);
  75                 nvmet_req_complete(req, NVME_SC_INTERNAL | NVME_SC_DNR);
  76         }
  77 }
  78
  79 static void nvmet_async_event_work(struct work_struct *work)
  80 {
  81         struct nvmet_ctrl *ctrl =
  82                 container_of(work, struct nvmet_ctrl, async_event_work);
  83         struct nvmet_async_event *aen;
  84         struct nvmet_req *req;
  85
  86         while (1) {
  87                 mutex_lock(&ctrl->lock);
  88                 aen = list_first_entry_or_null(&ctrl->async_events,
  89                                 struct nvmet_async_event, entry);
  90                 if (!aen || !ctrl->nr_async_event_cmds) {
  91                         mutex_unlock(&ctrl->lock);
  92                         return;
  93                 }
  94
  95                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
  96                 nvmet_set_result(req, nvmet_async_event_result(aen));
  97
  98                 list_del(&aen->entry);
  99                 kfree(aen);
 100
 101                 mutex_unlock(&ctrl->lock);
 102                 nvmet_req_complete(req, 0);
 103         }
 104 }
 105
 106 static void nvmet_add_async_event(struct nvmet_ctrl *ctrl, u8 event_type,
 107                 u8 event_info, u8 log_page)
 108 {
 109         struct nvmet_async_event *aen;
 110
 111         aen = kmalloc(sizeof(*aen), GFP_KERNEL);
 112         if (!aen)
 113                 return;
 114
 115         aen->event_type = event_type;
 116         aen->event_info = event_info;
 117         aen->log_page = log_page;
 118
 119         mutex_lock(&ctrl->lock);
 120         list_add_tail(&aen->entry, &ctrl->async_events);
 121         mutex_unlock(&ctrl->lock);
 122
 123         schedule_work(&ctrl->async_event_work);
 124 }
 125
 126 int nvmet_register_transport(struct nvmet_fabrics_ops *ops)
 127 {
 128         int ret = 0;
 129
 130         down_write(&nvmet_config_sem);
 131         if (nvmet_transports[ops->type])
 132                 ret = -EINVAL;
 133         else
 134                 nvmet_transports[ops->type] = ops;
 135         up_write(&nvmet_config_sem);
 136
 137         return ret;
 138 }
 139 EXPORT_SYMBOL_GPL(nvmet_register_transport);
 140
 141 void nvmet_unregister_transport(struct nvmet_fabrics_ops *ops)
 142 {
 143         down_write(&nvmet_config_sem);
 144         nvmet_transports[ops->type] = NULL;
 145         up_write(&nvmet_config_sem);
 146 }
 147 EXPORT_SYMBOL_GPL(nvmet_unregister_transport);
 148
 149 int nvmet_enable_port(struct nvmet_port *port)
 150 {
 151         struct nvmet_fabrics_ops *ops;
 152         int ret;
 153
 154         lockdep_assert_held(&nvmet_config_sem);
 155
 156         ops = nvmet_transports[port->disc_addr.trtype];
 157         if (!ops) {
 158                 up_write(&nvmet_config_sem);
 159                 request_module("nvmet-transport-%d", port->disc_addr.trtype);
 160                 down_write(&nvmet_config_sem);
 161                 ops = nvmet_transports[port->disc_addr.trtype];
 162                 if (!ops) {
 163                         pr_err("transport type %d not supported\n",
 164                                 port->disc_addr.trtype);
 165                         return -EINVAL;
 166                 }
 167         }
 168
 169         if (!try_module_get(ops->owner))
 170                 return -EINVAL;
 171
 172         ret = ops->add_port(port);
 173         if (ret) {
 174                 module_put(ops->owner);
 175                 return ret;
 176         }
 177
 178         port->enabled = true;
 179         return 0;
 180 }
 181
 182 void nvmet_disable_port(struct nvmet_port *port)
 183 {
 184         struct nvmet_fabrics_ops *ops;
 185
 186         lockdep_assert_held(&nvmet_config_sem);
 187
 188         port->enabled = false;
 189
 190         ops = nvmet_transports[port->disc_addr.trtype];
 191         ops->remove_port(port);
 192         module_put(ops->owner);
 193 }
 194
 195 static void nvmet_keep_alive_timer(struct work_struct *work)
 196 {
 197         struct nvmet_ctrl *ctrl = container_of(to_delayed_work(work),
 198                         struct nvmet_ctrl, ka_work);
 199
 200         pr_err("ctrl %d keep-alive timer (%d seconds) expired!\n",
 201                 ctrl->cntlid, ctrl->kato);
 202
 203         nvmet_ctrl_fatal_error(ctrl);
 204 }
 205
 206 static void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl)
 207 {
 208         pr_debug("ctrl %d start keep-alive timer for %d secs\n",
 209                 ctrl->cntlid, ctrl->kato);
 210
 211         INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
 212         schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
 213 }
 214
 215 static void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
 216 {
 217         pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
 218
 219         cancel_delayed_work_sync(&ctrl->ka_work);
 220 }
 221
 222 static struct nvmet_ns *__nvmet_find_namespace(struct nvmet_ctrl *ctrl,
 223                 __le32 nsid)
 224 {
 225         struct nvmet_ns *ns;
 226
 227         list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link) {
 228                 if (ns->nsid == le32_to_cpu(nsid))
 229                         return ns;
 230         }
 231
 232         return NULL;
 233 }
 234
 235 struct nvmet_ns *nvmet_find_namespace(struct nvmet_ctrl *ctrl, __le32 nsid)
 236 {
 237         struct nvmet_ns *ns;
 238
 239         rcu_read_lock();
 240         ns = __nvmet_find_namespace(ctrl, nsid);
 241         if (ns)
 242                 percpu_ref_get(&ns->ref);
 243         rcu_read_unlock();
 244
 245         return ns;
 246 }
 247
 248 static void nvmet_destroy_namespace(struct percpu_ref *ref)
 249 {
 250         struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
 251
 252         complete(&ns->disable_done);
 253 }
 254
 255 void nvmet_put_namespace(struct nvmet_ns *ns)
 256 {
 257         percpu_ref_put(&ns->ref);
 258 }
 259
 260 int nvmet_ns_enable(struct nvmet_ns *ns)
 261 {
 262         struct nvmet_subsys *subsys = ns->subsys;
 263         struct nvmet_ctrl *ctrl;
 264         int ret = 0;
 265
 266         mutex_lock(&subsys->lock);
 267         if (ns->enabled)
 268                 goto out_unlock;
 269
 270         ns->bdev = blkdev_get_by_path(ns->device_path, FMODE_READ | FMODE_WRITE,
 271                         NULL);
 272         if (IS_ERR(ns->bdev)) {
 273                 pr_err("nvmet: failed to open block device %s: (%ld)\n",
 274                         ns->device_path, PTR_ERR(ns->bdev));
 275                 ret = PTR_ERR(ns->bdev);
 276                 ns->bdev = NULL;
 277                 goto out_unlock;
 278         }
 279
 280         ns->size = i_size_read(ns->bdev->bd_inode);
 281         ns->blksize_shift = blksize_bits(bdev_logical_block_size(ns->bdev));
 282
 283         ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
 284                                 0, GFP_KERNEL);
 285         if (ret)
 286                 goto out_blkdev_put;
 287
 288         if (ns->nsid > subsys->max_nsid)
 289                 subsys->max_nsid = ns->nsid;
 290
 291         /*
 292          * The namespaces list needs to be sorted to simplify the implementation
 293          * of the Identify Namepace List subcommand.
 294          */
 295         if (list_empty(&subsys->namespaces)) {
 296                 list_add_tail_rcu(&ns->dev_link, &subsys->namespaces);
 297         } else {
 298                 struct nvmet_ns *old;
 299
 300                 list_for_each_entry_rcu(old, &subsys->namespaces, dev_link) {
 301                         BUG_ON(ns->nsid == old->nsid);
 302                         if (ns->nsid < old->nsid)
 303                                 break;
 304                 }
 305
 306                 list_add_tail_rcu(&ns->dev_link, &old->dev_link);
 307         }
 308
 309         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 310                 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE, 0, 0);
 311
 312         ns->enabled = true;
 313         ret = 0;
 314 out_unlock:
 315         mutex_unlock(&subsys->lock);
 316         return ret;
 317 out_blkdev_put:
 318         blkdev_put(ns->bdev, FMODE_WRITE|FMODE_READ);
 319         ns->bdev = NULL;
 320         goto out_unlock;
 321 }
 322
 323 void nvmet_ns_disable(struct nvmet_ns *ns)
 324 {
 325         struct nvmet_subsys *subsys = ns->subsys;
 326         struct nvmet_ctrl *ctrl;
 327
 328         mutex_lock(&subsys->lock);
 329         if (!ns->enabled)
 330                 goto out_unlock;
 331
 332         ns->enabled = false;
 333         list_del_rcu(&ns->dev_link);
 334         mutex_unlock(&subsys->lock);
 335
 336         /*
 337          * Now that we removed the namespaces from the lookup list, we
 338          * can kill the per_cpu ref and wait for any remaining references
 339          * to be dropped, as well as a RCU grace period for anyone only
 340          * using the namepace under rcu_read_lock().  Note that we can't
 341          * use call_rcu here as we need to ensure the namespaces have
 342          * been fully destroyed before unloading the module.
 343          */
 344         percpu_ref_kill(&ns->ref);
 345         synchronize_rcu();
 346         wait_for_completion(&ns->disable_done);
 347         percpu_ref_exit(&ns->ref);
 348
 349         mutex_lock(&subsys->lock);
 350         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 351                 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE, 0, 0);
 352
 353         if (ns->bdev)
 354                 blkdev_put(ns->bdev, FMODE_WRITE|FMODE_READ);
 355 out_unlock:
 356         mutex_unlock(&subsys->lock);
 357 }
 358
 359 void nvmet_ns_free(struct nvmet_ns *ns)
 360 {
 361         nvmet_ns_disable(ns);
 362
 363         kfree(ns->device_path);
 364         kfree(ns);
 365 }
 366
 367 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
 368 {
 369         struct nvmet_ns *ns;
 370
 371         ns = kzalloc(sizeof(*ns), GFP_KERNEL);
 372         if (!ns)
 373                 return NULL;
 374
 375         INIT_LIST_HEAD(&ns->dev_link);
 376         init_completion(&ns->disable_done);
 377
 378         ns->nsid = nsid;
 379         ns->subsys = subsys;
 380
 381         return ns;
 382 }
 383
 384 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
 385 {
 386         if (status)
 387                 nvmet_set_status(req, status);
 388
 389         /* XXX: need to fill in something useful for sq_head */
 390         req->rsp->sq_head = 0;
 391         if (likely(req->sq)) /* may happen during early failure */
 392                 req->rsp->sq_id = cpu_to_le16(req->sq->qid);
 393         req->rsp->command_id = req->cmd->common.command_id;
 394
 395         if (req->ns)
 396                 nvmet_put_namespace(req->ns);
 397         req->ops->queue_response(req);
 398 }
 399
 400 void nvmet_req_complete(struct nvmet_req *req, u16 status)
 401 {
 402         __nvmet_req_complete(req, status);
 403         percpu_ref_put(&req->sq->ref);
 404 }
 405 EXPORT_SYMBOL_GPL(nvmet_req_complete);
 406
 407 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
 408                 u16 qid, u16 size)
 409 {
 410         cq->qid = qid;
 411         cq->size = size;
 412
 413         ctrl->cqs[qid] = cq;
 414 }
 415
 416 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
 417                 u16 qid, u16 size)
 418 {
 419         sq->qid = qid;
 420         sq->size = size;
 421
 422         ctrl->sqs[qid] = sq;
 423 }
 424
 425 void nvmet_sq_destroy(struct nvmet_sq *sq)
 426 {
 427         /*
 428          * If this is the admin queue, complete all AERs so that our
 429          * queue doesn't have outstanding requests on it.
 430          */
 431         if (sq->ctrl && sq->ctrl->sqs && sq->ctrl->sqs[0] == sq)
 432                 nvmet_async_events_free(sq->ctrl);
 433         percpu_ref_kill(&sq->ref);
 434         wait_for_completion(&sq->free_done);
 435         percpu_ref_exit(&sq->ref);
 436
 437         if (sq->ctrl) {
 438                 nvmet_ctrl_put(sq->ctrl);
 439                 sq->ctrl = NULL; /* allows reusing the queue later */
 440         }
 441 }
 442 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
 443
 444 static void nvmet_sq_free(struct percpu_ref *ref)
 445 {
 446         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 447
 448         complete(&sq->free_done);
 449 }
 450
 451 int nvmet_sq_init(struct nvmet_sq *sq)
 452 {
 453         int ret;
 454
 455         ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
 456         if (ret) {
 457                 pr_err("percpu_ref init failed!\n");
 458                 return ret;
 459         }
 460         init_completion(&sq->free_done);
 461
 462         return 0;
 463 }
 464 EXPORT_SYMBOL_GPL(nvmet_sq_init);
 465
 466 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
 467                 struct nvmet_sq *sq, struct nvmet_fabrics_ops *ops)
 468 {
 469         u8 flags = req->cmd->common.flags;
 470         u16 status;
 471
 472         req->cq = cq;
 473         req->sq = sq;
 474         req->ops = ops;
 475         req->sg = NULL;
 476         req->sg_cnt = 0;
 477         req->rsp->status = 0;
 478
 479         /* no support for fused commands yet */
 480         if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
 481                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 482                 goto fail;
 483         }
 484
 485         /* either variant of SGLs is fine, as we don't support metadata */
 486         if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF &&
 487                      (flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METASEG)) {
 488                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 489                 goto fail;
 490         }
 491
 492         if (unlikely(!req->sq->ctrl))
 493                 /* will return an error for any Non-connect command: */
 494                 status = nvmet_parse_connect_cmd(req);
 495         else if (likely(req->sq->qid != 0))
 496                 status = nvmet_parse_io_cmd(req);
 497         else if (req->cmd->common.opcode == nvme_fabrics_command)
 498                 status = nvmet_parse_fabrics_cmd(req);
 499         else if (req->sq->ctrl->subsys->type == NVME_NQN_DISC)
 500                 status = nvmet_parse_discovery_cmd(req);
 501         else
 502                 status = nvmet_parse_admin_cmd(req);
 503
 504         if (status)
 505                 goto fail;
 506
 507         if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
 508                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 509                 goto fail;
 510         }
 511
 512         return true;
 513
 514 fail:
 515         __nvmet_req_complete(req, status);
 516         return false;
 517 }
 518 EXPORT_SYMBOL_GPL(nvmet_req_init);
 519
 520 static inline bool nvmet_cc_en(u32 cc)
 521 {
 522         return cc & 0x1;
 523 }
 524
 525 static inline u8 nvmet_cc_css(u32 cc)
 526 {
 527         return (cc >> 4) & 0x7;
 528 }
 529
 530 static inline u8 nvmet_cc_mps(u32 cc)
 531 {
 532         return (cc >> 7) & 0xf;
 533 }
 534
 535 static inline u8 nvmet_cc_ams(u32 cc)
 536 {
 537         return (cc >> 11) & 0x7;
 538 }
 539
 540 static inline u8 nvmet_cc_shn(u32 cc)
 541 {
 542         return (cc >> 14) & 0x3;
 543 }
 544
 545 static inline u8 nvmet_cc_iosqes(u32 cc)
 546 {
 547         return (cc >> 16) & 0xf;
 548 }
 549
 550 static inline u8 nvmet_cc_iocqes(u32 cc)
 551 {
 552         return (cc >> 20) & 0xf;
 553 }
 554
 555 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
 556 {
 557         lockdep_assert_held(&ctrl->lock);
 558
 559         if (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
 560             nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES ||
 561             nvmet_cc_mps(ctrl->cc) != 0 ||
 562             nvmet_cc_ams(ctrl->cc) != 0 ||
 563             nvmet_cc_css(ctrl->cc) != 0) {
 564                 ctrl->csts = NVME_CSTS_CFS;
 565                 return;
 566         }
 567
 568         ctrl->csts = NVME_CSTS_RDY;
 569 }
 570
 571 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
 572 {
 573         lockdep_assert_held(&ctrl->lock);
 574
 575         /* XXX: tear down queues? */
 576         ctrl->csts &= ~NVME_CSTS_RDY;
 577         ctrl->cc = 0;
 578 }
 579
 580 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
 581 {
 582         u32 old;
 583
 584         mutex_lock(&ctrl->lock);
 585         old = ctrl->cc;
 586         ctrl->cc = new;
 587
 588         if (nvmet_cc_en(new) && !nvmet_cc_en(old))
 589                 nvmet_start_ctrl(ctrl);
 590         if (!nvmet_cc_en(new) && nvmet_cc_en(old))
 591                 nvmet_clear_ctrl(ctrl);
 592         if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
 593                 nvmet_clear_ctrl(ctrl);
 594                 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
 595         }
 596         if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
 597                 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
 598         mutex_unlock(&ctrl->lock);
 599 }
 600
 601 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
 602 {
 603         /* command sets supported: NVMe command set: */
 604         ctrl->cap = (1ULL << 37);
 605         /* CC.EN timeout in 500msec units: */
 606         ctrl->cap |= (15ULL << 24);
 607         /* maximum queue entries supported: */
 608         ctrl->cap |= NVMET_QUEUE_SIZE - 1;
 609 }
 610
 611 u16 nvmet_ctrl_find_get(const char *subsysnqn, const char *hostnqn, u16 cntlid,
 612                 struct nvmet_req *req, struct nvmet_ctrl **ret)
 613 {
 614         struct nvmet_subsys *subsys;
 615         struct nvmet_ctrl *ctrl;
 616         u16 status = 0;
 617
 618         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
 619         if (!subsys) {
 620                 pr_warn("connect request for invalid subsystem %s!\n",
 621                         subsysnqn);
 622                 req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
 623                 return NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
 624         }
 625
 626         mutex_lock(&subsys->lock);
 627         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 628                 if (ctrl->cntlid == cntlid) {
 629                         if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
 630                                 pr_warn("hostnqn mismatch.\n");
 631                                 continue;
 632                         }
 633                         if (!kref_get_unless_zero(&ctrl->ref))
 634                                 continue;
 635
 636                         *ret = ctrl;
 637                         goto out;
 638                 }
 639         }
 640
 641         pr_warn("could not find controller %d for subsys %s / host %s\n",
 642                 cntlid, subsysnqn, hostnqn);
 643         req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
 644         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
 645
 646 out:
 647         mutex_unlock(&subsys->lock);
 648         nvmet_subsys_put(subsys);
 649         return status;
 650 }
 651
 652 static bool __nvmet_host_allowed(struct nvmet_subsys *subsys,
 653                 const char *hostnqn)
 654 {
 655         struct nvmet_host_link *p;
 656
 657         if (subsys->allow_any_host)
 658                 return true;
 659
 660         list_for_each_entry(p, &subsys->hosts, entry) {
 661                 if (!strcmp(nvmet_host_name(p->host), hostnqn))
 662                         return true;
 663         }
 664
 665         return false;
 666 }
 667
 668 static bool nvmet_host_discovery_allowed(struct nvmet_req *req,
 669                 const char *hostnqn)
 670 {
 671         struct nvmet_subsys_link *s;
 672
 673         list_for_each_entry(s, &req->port->subsystems, entry) {
 674                 if (__nvmet_host_allowed(s->subsys, hostnqn))
 675                         return true;
 676         }
 677
 678         return false;
 679 }
 680
 681 bool nvmet_host_allowed(struct nvmet_req *req, struct nvmet_subsys *subsys,
 682                 const char *hostnqn)
 683 {
 684         lockdep_assert_held(&nvmet_config_sem);
 685
 686         if (subsys->type == NVME_NQN_DISC)
 687                 return nvmet_host_discovery_allowed(req, hostnqn);
 688         else
 689                 return __nvmet_host_allowed(subsys, hostnqn);
 690 }
 691
 692 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
 693                 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
 694 {
 695         struct nvmet_subsys *subsys;
 696         struct nvmet_ctrl *ctrl;
 697         int ret;
 698         u16 status;
 699
 700         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
 701         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
 702         if (!subsys) {
 703                 pr_warn("connect request for invalid subsystem %s!\n",
 704                         subsysnqn);
 705                 req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
 706                 goto out;
 707         }
 708
 709         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
 710         down_read(&nvmet_config_sem);
 711         if (!nvmet_host_allowed(req, subsys, hostnqn)) {
 712                 pr_info("connect by host %s for subsystem %s not allowed\n",
 713                         hostnqn, subsysnqn);
 714                 req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
 715                 up_read(&nvmet_config_sem);
 716                 goto out_put_subsystem;
 717         }
 718         up_read(&nvmet_config_sem);
 719
 720         status = NVME_SC_INTERNAL;
 721         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
 722         if (!ctrl)
 723                 goto out_put_subsystem;
 724         mutex_init(&ctrl->lock);
 725
 726         nvmet_init_cap(ctrl);
 727
 728         INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
 729         INIT_LIST_HEAD(&ctrl->async_events);
 730
 731         memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
 732         memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
 733
 734         /* generate a random serial number as our controllers are ephemeral: */
 735         get_random_bytes(&ctrl->serial, sizeof(ctrl->serial));
 736
 737         kref_init(&ctrl->ref);
 738         ctrl->subsys = subsys;
 739
 740         ctrl->cqs = kcalloc(subsys->max_qid + 1,
 741                         sizeof(struct nvmet_cq *),
 742                         GFP_KERNEL);
 743         if (!ctrl->cqs)
 744                 goto out_free_ctrl;
 745
 746         ctrl->sqs = kcalloc(subsys->max_qid + 1,
 747                         sizeof(struct nvmet_sq *),
 748                         GFP_KERNEL);
 749         if (!ctrl->sqs)
 750                 goto out_free_cqs;
 751
 752         ret = ida_simple_get(&subsys->cntlid_ida,
 753                              NVME_CNTLID_MIN, NVME_CNTLID_MAX,
 754                              GFP_KERNEL);
 755         if (ret < 0) {
 756                 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
 757                 goto out_free_sqs;
 758         }
 759         ctrl->cntlid = ret;
 760
 761         ctrl->ops = req->ops;
 762         if (ctrl->subsys->type == NVME_NQN_DISC) {
 763                 /* Don't accept keep-alive timeout for discovery controllers */
 764                 if (kato) {
 765                         status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 766                         goto out_free_sqs;
 767                 }
 768
 769                 /*
 770                  * Discovery controllers use some arbitrary high value in order
 771                  * to cleanup stale discovery sessions
 772                  *
 773                  * From the latest base diff RC:
 774                  * "The Keep Alive command is not supported by
 775                  * Discovery controllers. A transport may specify a
 776                  * fixed Discovery controller activity timeout value
 777                  * (e.g., 2 minutes).  If no commands are received
 778                  * by a Discovery controller within that time
 779                  * period, the controller may perform the
 780                  * actions for Keep Alive Timer expiration".
 781                  */
 782                 ctrl->kato = NVMET_DISC_KATO;
 783         } else {
 784                 /* keep-alive timeout in seconds */
 785                 ctrl->kato = DIV_ROUND_UP(kato, 1000);
 786         }
 787         nvmet_start_keep_alive_timer(ctrl);
 788
 789         mutex_lock(&subsys->lock);
 790         list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
 791         mutex_unlock(&subsys->lock);
 792
 793         *ctrlp = ctrl;
 794         return 0;
 795
 796 out_free_sqs:
 797         kfree(ctrl->sqs);
 798 out_free_cqs:
 799         kfree(ctrl->cqs);
 800 out_free_ctrl:
 801         kfree(ctrl);
 802 out_put_subsystem:
 803         nvmet_subsys_put(subsys);
 804 out:
 805         return status;
 806 }
 807
 808 static void nvmet_ctrl_free(struct kref *ref)
 809 {
 810         struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
 811         struct nvmet_subsys *subsys = ctrl->subsys;
 812
 813         nvmet_stop_keep_alive_timer(ctrl);
 814
 815         mutex_lock(&subsys->lock);
 816         list_del(&ctrl->subsys_entry);
 817         mutex_unlock(&subsys->lock);
 818
 819         flush_work(&ctrl->async_event_work);
 820         cancel_work_sync(&ctrl->fatal_err_work);
 821
 822         ida_simple_remove(&subsys->cntlid_ida, ctrl->cntlid);
 823         nvmet_subsys_put(subsys);
 824
 825         kfree(ctrl->sqs);
 826         kfree(ctrl->cqs);
 827         kfree(ctrl);
 828 }
 829
 830 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
 831 {
 832         kref_put(&ctrl->ref, nvmet_ctrl_free);
 833 }
 834
 835 static void nvmet_fatal_error_handler(struct work_struct *work)
 836 {
 837         struct nvmet_ctrl *ctrl =
 838                         container_of(work, struct nvmet_ctrl, fatal_err_work);
 839
 840         pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
 841         ctrl->ops->delete_ctrl(ctrl);
 842 }
 843
 844 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
 845 {
 846         mutex_lock(&ctrl->lock);
 847         if (!(ctrl->csts & NVME_CSTS_CFS)) {
 848                 ctrl->csts |= NVME_CSTS_CFS;
 849                 INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
 850                 schedule_work(&ctrl->fatal_err_work);
 851         }
 852         mutex_unlock(&ctrl->lock);
 853 }
 854 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
 855
 856 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
 857                 const char *subsysnqn)
 858 {
 859         struct nvmet_subsys_link *p;
 860
 861         if (!port)
 862                 return NULL;
 863
 864         if (!strncmp(NVME_DISC_SUBSYS_NAME, subsysnqn,
 865                         NVMF_NQN_SIZE)) {
 866                 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
 867                         return NULL;
 868                 return nvmet_disc_subsys;
 869         }
 870
 871         down_read(&nvmet_config_sem);
 872         list_for_each_entry(p, &port->subsystems, entry) {
 873                 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
 874                                 NVMF_NQN_SIZE)) {
 875                         if (!kref_get_unless_zero(&p->subsys->ref))
 876                                 break;
 877                         up_read(&nvmet_config_sem);
 878                         return p->subsys;
 879                 }
 880         }
 881         up_read(&nvmet_config_sem);
 882         return NULL;
 883 }
 884
 885 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
 886                 enum nvme_subsys_type type)
 887 {
 888         struct nvmet_subsys *subsys;
 889
 890         subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
 891         if (!subsys)
 892                 return NULL;
 893
 894         subsys->ver = NVME_VS(1, 2, 1); /* NVMe 1.2.1 */
 895
 896         switch (type) {
 897         case NVME_NQN_NVME:
 898                 subsys->max_qid = NVMET_NR_QUEUES;
 899                 break;
 900         case NVME_NQN_DISC:
 901                 subsys->max_qid = 0;
 902                 break;
 903         default:
 904                 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
 905                 kfree(subsys);
 906                 return NULL;
 907         }
 908         subsys->type = type;
 909         subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
 910                         GFP_KERNEL);
 911         if (!subsys->subsysnqn) {
 912                 kfree(subsys);
 913                 return NULL;
 914         }
 915
 916         kref_init(&subsys->ref);
 917
 918         mutex_init(&subsys->lock);
 919         INIT_LIST_HEAD(&subsys->namespaces);
 920         INIT_LIST_HEAD(&subsys->ctrls);
 921
 922         ida_init(&subsys->cntlid_ida);
 923
 924         INIT_LIST_HEAD(&subsys->hosts);
 925
 926         return subsys;
 927 }
 928
 929 static void nvmet_subsys_free(struct kref *ref)
 930 {
 931         struct nvmet_subsys *subsys =
 932                 container_of(ref, struct nvmet_subsys, ref);
 933
 934         WARN_ON_ONCE(!list_empty(&subsys->namespaces));
 935
 936         ida_destroy(&subsys->cntlid_ida);
 937         kfree(subsys->subsysnqn);
 938         kfree(subsys);
 939 }
 940
 941 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
 942 {
 943         struct nvmet_ctrl *ctrl;
 944
 945         mutex_lock(&subsys->lock);
 946         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 947                 ctrl->ops->delete_ctrl(ctrl);
 948         mutex_unlock(&subsys->lock);
 949 }
 950
 951 void nvmet_subsys_put(struct nvmet_subsys *subsys)
 952 {
 953         kref_put(&subsys->ref, nvmet_subsys_free);
 954 }
 955
 956 static int __init nvmet_init(void)
 957 {
 958         int error;
 959
 960         error = nvmet_init_discovery();
 961         if (error)
 962                 goto out;
 963
 964         error = nvmet_init_configfs();
 965         if (error)
 966                 goto out_exit_discovery;
 967         return 0;
 968
 969 out_exit_discovery:
 970         nvmet_exit_discovery();
 971 out:
 972         return error;
 973 }
 974
 975 static void __exit nvmet_exit(void)
 976 {
 977         nvmet_exit_configfs();
 978         nvmet_exit_discovery();
 979
 980         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
 981         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
 982 }
 983
 984 module_init(nvmet_init);
 985 module_exit(nvmet_exit);
 986
 987 MODULE_LICENSE("GPL v2");