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 <linux/rculist.h>
  18
  19 #include "nvmet.h"
  20
  21 static struct nvmet_fabrics_ops *nvmet_transports[NVMF_TRTYPE_MAX];
  22 static DEFINE_IDA(cntlid_ida);
  23
  24 /*
  25  * This read/write semaphore is used to synchronize access to configuration
  26  * information on a target system that will result in discovery log page
  27  * information change for at least one host.
  28  * The full list of resources to protected by this semaphore is:
  29  *
  30  *  - subsystems list
  31  *  - per-subsystem allowed hosts list
  32  *  - allow_any_host subsystem attribute
  33  *  - nvmet_genctr
  34  *  - the nvmet_transports array
  35  *
  36  * When updating any of those lists/structures write lock should be obtained,
  37  * while when reading (popolating discovery log page or checking host-subsystem
  38  * link) read lock is obtained to allow concurrent reads.
  39  */
  40 DECLARE_RWSEM(nvmet_config_sem);
  41
  42 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
  43                 const char *subsysnqn);
  44
  45 u16 nvmet_copy_to_sgl(struct nvmet_req *req, off_t off, const void *buf,
  46                 size_t len)
  47 {
  48         if (sg_pcopy_from_buffer(req->sg, req->sg_cnt, buf, len, off) != len)
  49                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
  50         return 0;
  51 }
  52
  53 u16 nvmet_copy_from_sgl(struct nvmet_req *req, off_t off, void *buf, size_t len)
  54 {
  55         if (sg_pcopy_to_buffer(req->sg, req->sg_cnt, buf, len, off) != len)
  56                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
  57         return 0;
  58 }
  59
  60 static u32 nvmet_async_event_result(struct nvmet_async_event *aen)
  61 {
  62         return aen->event_type | (aen->event_info << 8) | (aen->log_page << 16);
  63 }
  64
  65 static void nvmet_async_events_free(struct nvmet_ctrl *ctrl)
  66 {
  67         struct nvmet_req *req;
  68
  69         while (1) {
  70                 mutex_lock(&ctrl->lock);
  71                 if (!ctrl->nr_async_event_cmds) {
  72                         mutex_unlock(&ctrl->lock);
  73                         return;
  74                 }
  75
  76                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
  77                 mutex_unlock(&ctrl->lock);
  78                 nvmet_req_complete(req, NVME_SC_INTERNAL | NVME_SC_DNR);
  79         }
  80 }
  81
  82 static void nvmet_async_event_work(struct work_struct *work)
  83 {
  84         struct nvmet_ctrl *ctrl =
  85                 container_of(work, struct nvmet_ctrl, async_event_work);
  86         struct nvmet_async_event *aen;
  87         struct nvmet_req *req;
  88
  89         while (1) {
  90                 mutex_lock(&ctrl->lock);
  91                 aen = list_first_entry_or_null(&ctrl->async_events,
  92                                 struct nvmet_async_event, entry);
  93                 if (!aen || !ctrl->nr_async_event_cmds) {
  94                         mutex_unlock(&ctrl->lock);
  95                         return;
  96                 }
  97
  98                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
  99                 nvmet_set_result(req, nvmet_async_event_result(aen));
 100
 101                 list_del(&aen->entry);
 102                 kfree(aen);
 103
 104                 mutex_unlock(&ctrl->lock);
 105                 nvmet_req_complete(req, 0);
 106         }
 107 }
 108
 109 static void nvmet_add_async_event(struct nvmet_ctrl *ctrl, u8 event_type,
 110                 u8 event_info, u8 log_page)
 111 {
 112         struct nvmet_async_event *aen;
 113
 114         aen = kmalloc(sizeof(*aen), GFP_KERNEL);
 115         if (!aen)
 116                 return;
 117
 118         aen->event_type = event_type;
 119         aen->event_info = event_info;
 120         aen->log_page = log_page;
 121
 122         mutex_lock(&ctrl->lock);
 123         list_add_tail(&aen->entry, &ctrl->async_events);
 124         mutex_unlock(&ctrl->lock);
 125
 126         schedule_work(&ctrl->async_event_work);
 127 }
 128
 129 int nvmet_register_transport(struct nvmet_fabrics_ops *ops)
 130 {
 131         int ret = 0;
 132
 133         down_write(&nvmet_config_sem);
 134         if (nvmet_transports[ops->type])
 135                 ret = -EINVAL;
 136         else
 137                 nvmet_transports[ops->type] = ops;
 138         up_write(&nvmet_config_sem);
 139
 140         return ret;
 141 }
 142 EXPORT_SYMBOL_GPL(nvmet_register_transport);
 143
 144 void nvmet_unregister_transport(struct nvmet_fabrics_ops *ops)
 145 {
 146         down_write(&nvmet_config_sem);
 147         nvmet_transports[ops->type] = NULL;
 148         up_write(&nvmet_config_sem);
 149 }
 150 EXPORT_SYMBOL_GPL(nvmet_unregister_transport);
 151
 152 int nvmet_enable_port(struct nvmet_port *port)
 153 {
 154         struct nvmet_fabrics_ops *ops;
 155         int ret;
 156
 157         lockdep_assert_held(&nvmet_config_sem);
 158
 159         ops = nvmet_transports[port->disc_addr.trtype];
 160         if (!ops) {
 161                 up_write(&nvmet_config_sem);
 162                 request_module("nvmet-transport-%d", port->disc_addr.trtype);
 163                 down_write(&nvmet_config_sem);
 164                 ops = nvmet_transports[port->disc_addr.trtype];
 165                 if (!ops) {
 166                         pr_err("transport type %d not supported\n",
 167                                 port->disc_addr.trtype);
 168                         return -EINVAL;
 169                 }
 170         }
 171
 172         if (!try_module_get(ops->owner))
 173                 return -EINVAL;
 174
 175         ret = ops->add_port(port);
 176         if (ret) {
 177                 module_put(ops->owner);
 178                 return ret;
 179         }
 180
 181         port->enabled = true;
 182         return 0;
 183 }
 184
 185 void nvmet_disable_port(struct nvmet_port *port)
 186 {
 187         struct nvmet_fabrics_ops *ops;
 188
 189         lockdep_assert_held(&nvmet_config_sem);
 190
 191         port->enabled = false;
 192
 193         ops = nvmet_transports[port->disc_addr.trtype];
 194         ops->remove_port(port);
 195         module_put(ops->owner);
 196 }
 197
 198 static void nvmet_keep_alive_timer(struct work_struct *work)
 199 {
 200         struct nvmet_ctrl *ctrl = container_of(to_delayed_work(work),
 201                         struct nvmet_ctrl, ka_work);
 202
 203         pr_err("ctrl %d keep-alive timer (%d seconds) expired!\n",
 204                 ctrl->cntlid, ctrl->kato);
 205
 206         nvmet_ctrl_fatal_error(ctrl);
 207 }
 208
 209 static void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl)
 210 {
 211         if (unlikely(ctrl->kato == 0))
 212                 return;
 213
 214         pr_debug("ctrl %d start keep-alive timer for %d secs\n",
 215                 ctrl->cntlid, ctrl->kato);
 216
 217         INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
 218         schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
 219 }
 220
 221 static void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
 222 {
 223         if (unlikely(ctrl->kato == 0))
 224                 return;
 225
 226         pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
 227
 228         cancel_delayed_work_sync(&ctrl->ka_work);
 229 }
 230
 231 static struct nvmet_ns *__nvmet_find_namespace(struct nvmet_ctrl *ctrl,
 232                 __le32 nsid)
 233 {
 234         struct nvmet_ns *ns;
 235
 236         list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link) {
 237                 if (ns->nsid == le32_to_cpu(nsid))
 238                         return ns;
 239         }
 240
 241         return NULL;
 242 }
 243
 244 struct nvmet_ns *nvmet_find_namespace(struct nvmet_ctrl *ctrl, __le32 nsid)
 245 {
 246         struct nvmet_ns *ns;
 247
 248         rcu_read_lock();
 249         ns = __nvmet_find_namespace(ctrl, nsid);
 250         if (ns)
 251                 percpu_ref_get(&ns->ref);
 252         rcu_read_unlock();
 253
 254         return ns;
 255 }
 256
 257 static void nvmet_destroy_namespace(struct percpu_ref *ref)
 258 {
 259         struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
 260
 261         complete(&ns->disable_done);
 262 }
 263
 264 void nvmet_put_namespace(struct nvmet_ns *ns)
 265 {
 266         percpu_ref_put(&ns->ref);
 267 }
 268
 269 int nvmet_ns_enable(struct nvmet_ns *ns)
 270 {
 271         struct nvmet_subsys *subsys = ns->subsys;
 272         struct nvmet_ctrl *ctrl;
 273         int ret = 0;
 274
 275         mutex_lock(&subsys->lock);
 276         if (ns->enabled)
 277                 goto out_unlock;
 278
 279         ns->bdev = blkdev_get_by_path(ns->device_path, FMODE_READ | FMODE_WRITE,
 280                         NULL);
 281         if (IS_ERR(ns->bdev)) {
 282                 pr_err("failed to open block device %s: (%ld)\n",
 283                        ns->device_path, PTR_ERR(ns->bdev));
 284                 ret = PTR_ERR(ns->bdev);
 285                 ns->bdev = NULL;
 286                 goto out_unlock;
 287         }
 288
 289         ns->size = i_size_read(ns->bdev->bd_inode);
 290         ns->blksize_shift = blksize_bits(bdev_logical_block_size(ns->bdev));
 291
 292         ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
 293                                 0, GFP_KERNEL);
 294         if (ret)
 295                 goto out_blkdev_put;
 296
 297         if (ns->nsid > subsys->max_nsid)
 298                 subsys->max_nsid = ns->nsid;
 299
 300         /*
 301          * The namespaces list needs to be sorted to simplify the implementation
 302          * of the Identify Namepace List subcommand.
 303          */
 304         if (list_empty(&subsys->namespaces)) {
 305                 list_add_tail_rcu(&ns->dev_link, &subsys->namespaces);
 306         } else {
 307                 struct nvmet_ns *old;
 308
 309                 list_for_each_entry_rcu(old, &subsys->namespaces, dev_link) {
 310                         BUG_ON(ns->nsid == old->nsid);
 311                         if (ns->nsid < old->nsid)
 312                                 break;
 313                 }
 314
 315                 list_add_tail_rcu(&ns->dev_link, &old->dev_link);
 316         }
 317
 318         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 319                 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE, 0, 0);
 320
 321         ns->enabled = true;
 322         ret = 0;
 323 out_unlock:
 324         mutex_unlock(&subsys->lock);
 325         return ret;
 326 out_blkdev_put:
 327         blkdev_put(ns->bdev, FMODE_WRITE|FMODE_READ);
 328         ns->bdev = NULL;
 329         goto out_unlock;
 330 }
 331
 332 void nvmet_ns_disable(struct nvmet_ns *ns)
 333 {
 334         struct nvmet_subsys *subsys = ns->subsys;
 335         struct nvmet_ctrl *ctrl;
 336
 337         mutex_lock(&subsys->lock);
 338         if (!ns->enabled)
 339                 goto out_unlock;
 340
 341         ns->enabled = false;
 342         list_del_rcu(&ns->dev_link);
 343         mutex_unlock(&subsys->lock);
 344
 345         /*
 346          * Now that we removed the namespaces from the lookup list, we
 347          * can kill the per_cpu ref and wait for any remaining references
 348          * to be dropped, as well as a RCU grace period for anyone only
 349          * using the namepace under rcu_read_lock().  Note that we can't
 350          * use call_rcu here as we need to ensure the namespaces have
 351          * been fully destroyed before unloading the module.
 352          */
 353         percpu_ref_kill(&ns->ref);
 354         synchronize_rcu();
 355         wait_for_completion(&ns->disable_done);
 356         percpu_ref_exit(&ns->ref);
 357
 358         mutex_lock(&subsys->lock);
 359         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 360                 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE, 0, 0);
 361
 362         if (ns->bdev)
 363                 blkdev_put(ns->bdev, FMODE_WRITE|FMODE_READ);
 364 out_unlock:
 365         mutex_unlock(&subsys->lock);
 366 }
 367
 368 void nvmet_ns_free(struct nvmet_ns *ns)
 369 {
 370         nvmet_ns_disable(ns);
 371
 372         kfree(ns->device_path);
 373         kfree(ns);
 374 }
 375
 376 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
 377 {
 378         struct nvmet_ns *ns;
 379
 380         ns = kzalloc(sizeof(*ns), GFP_KERNEL);
 381         if (!ns)
 382                 return NULL;
 383
 384         INIT_LIST_HEAD(&ns->dev_link);
 385         init_completion(&ns->disable_done);
 386
 387         ns->nsid = nsid;
 388         ns->subsys = subsys;
 389         uuid_gen(&ns->uuid);
 390
 391         return ns;
 392 }
 393
 394 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
 395 {
 396         u32 old_sqhd, new_sqhd;
 397         u16 sqhd;
 398
 399         if (status)
 400                 nvmet_set_status(req, status);
 401
 402         if (req->sq->size) {
 403                 do {
 404                         old_sqhd = req->sq->sqhd;
 405                         new_sqhd = (old_sqhd + 1) % req->sq->size;
 406                 } while (cmpxchg(&req->sq->sqhd, old_sqhd, new_sqhd) !=
 407                                         old_sqhd);
 408         }
 409         sqhd = req->sq->sqhd & 0x0000FFFF;
 410         req->rsp->sq_head = cpu_to_le16(sqhd);
 411         req->rsp->sq_id = cpu_to_le16(req->sq->qid);
 412         req->rsp->command_id = req->cmd->common.command_id;
 413
 414         if (req->ns)
 415                 nvmet_put_namespace(req->ns);
 416         req->ops->queue_response(req);
 417 }
 418
 419 void nvmet_req_complete(struct nvmet_req *req, u16 status)
 420 {
 421         __nvmet_req_complete(req, status);
 422         percpu_ref_put(&req->sq->ref);
 423 }
 424 EXPORT_SYMBOL_GPL(nvmet_req_complete);
 425
 426 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
 427                 u16 qid, u16 size)
 428 {
 429         cq->qid = qid;
 430         cq->size = size;
 431
 432         ctrl->cqs[qid] = cq;
 433 }
 434
 435 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
 436                 u16 qid, u16 size)
 437 {
 438         sq->sqhd = 0;
 439         sq->qid = qid;
 440         sq->size = size;
 441
 442         ctrl->sqs[qid] = sq;
 443 }
 444
 445 static void nvmet_confirm_sq(struct percpu_ref *ref)
 446 {
 447         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 448
 449         complete(&sq->confirm_done);
 450 }
 451
 452 void nvmet_sq_destroy(struct nvmet_sq *sq)
 453 {
 454         /*
 455          * If this is the admin queue, complete all AERs so that our
 456          * queue doesn't have outstanding requests on it.
 457          */
 458         if (sq->ctrl && sq->ctrl->sqs && sq->ctrl->sqs[0] == sq)
 459                 nvmet_async_events_free(sq->ctrl);
 460         percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
 461         wait_for_completion(&sq->confirm_done);
 462         wait_for_completion(&sq->free_done);
 463         percpu_ref_exit(&sq->ref);
 464
 465         if (sq->ctrl) {
 466                 nvmet_ctrl_put(sq->ctrl);
 467                 sq->ctrl = NULL; /* allows reusing the queue later */
 468         }
 469 }
 470 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
 471
 472 static void nvmet_sq_free(struct percpu_ref *ref)
 473 {
 474         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 475
 476         complete(&sq->free_done);
 477 }
 478
 479 int nvmet_sq_init(struct nvmet_sq *sq)
 480 {
 481         int ret;
 482
 483         ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
 484         if (ret) {
 485                 pr_err("percpu_ref init failed!\n");
 486                 return ret;
 487         }
 488         init_completion(&sq->free_done);
 489         init_completion(&sq->confirm_done);
 490
 491         return 0;
 492 }
 493 EXPORT_SYMBOL_GPL(nvmet_sq_init);
 494
 495 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
 496                 struct nvmet_sq *sq, struct nvmet_fabrics_ops *ops)
 497 {
 498         u8 flags = req->cmd->common.flags;
 499         u16 status;
 500
 501         req->cq = cq;
 502         req->sq = sq;
 503         req->ops = ops;
 504         req->sg = NULL;
 505         req->sg_cnt = 0;
 506         req->rsp->status = 0;
 507
 508         /* no support for fused commands yet */
 509         if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
 510                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 511                 goto fail;
 512         }
 513
 514         /*
 515          * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
 516          * contains an address of a single contiguous physical buffer that is
 517          * byte aligned.
 518          */
 519         if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
 520                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 521                 goto fail;
 522         }
 523
 524         if (unlikely(!req->sq->ctrl))
 525                 /* will return an error for any Non-connect command: */
 526                 status = nvmet_parse_connect_cmd(req);
 527         else if (likely(req->sq->qid != 0))
 528                 status = nvmet_parse_io_cmd(req);
 529         else if (req->cmd->common.opcode == nvme_fabrics_command)
 530                 status = nvmet_parse_fabrics_cmd(req);
 531         else if (req->sq->ctrl->subsys->type == NVME_NQN_DISC)
 532                 status = nvmet_parse_discovery_cmd(req);
 533         else
 534                 status = nvmet_parse_admin_cmd(req);
 535
 536         if (status)
 537                 goto fail;
 538
 539         if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
 540                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 541                 goto fail;
 542         }
 543
 544         return true;
 545
 546 fail:
 547         __nvmet_req_complete(req, status);
 548         return false;
 549 }
 550 EXPORT_SYMBOL_GPL(nvmet_req_init);
 551
 552 void nvmet_req_uninit(struct nvmet_req *req)
 553 {
 554         percpu_ref_put(&req->sq->ref);
 555 }
 556 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
 557
 558 static inline bool nvmet_cc_en(u32 cc)
 559 {
 560         return (cc >> NVME_CC_EN_SHIFT) & 0x1;
 561 }
 562
 563 static inline u8 nvmet_cc_css(u32 cc)
 564 {
 565         return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
 566 }
 567
 568 static inline u8 nvmet_cc_mps(u32 cc)
 569 {
 570         return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
 571 }
 572
 573 static inline u8 nvmet_cc_ams(u32 cc)
 574 {
 575         return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
 576 }
 577
 578 static inline u8 nvmet_cc_shn(u32 cc)
 579 {
 580         return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
 581 }
 582
 583 static inline u8 nvmet_cc_iosqes(u32 cc)
 584 {
 585         return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
 586 }
 587
 588 static inline u8 nvmet_cc_iocqes(u32 cc)
 589 {
 590         return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
 591 }
 592
 593 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
 594 {
 595         lockdep_assert_held(&ctrl->lock);
 596
 597         /*
 598          * Only I/O controllers should verify iosqes,iocqes.
 599          * Strictly speaking, the spec says a discovery controller
 600          * should verify iosqes,iocqes are zeroed, however that
 601          * would break backwards compatibility, so don't enforce it.
 602          */
 603         if (ctrl->subsys->type != NVME_NQN_DISC &&
 604             (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
 605              nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES)) {
 606                 ctrl->csts = NVME_CSTS_CFS;
 607                 return;
 608         }
 609
 610         if (nvmet_cc_mps(ctrl->cc) != 0 ||
 611             nvmet_cc_ams(ctrl->cc) != 0 ||
 612             nvmet_cc_css(ctrl->cc) != 0) {
 613                 ctrl->csts = NVME_CSTS_CFS;
 614                 return;
 615         }
 616
 617         ctrl->csts = NVME_CSTS_RDY;
 618
 619         /*
 620          * Controllers that are not yet enabled should not really enforce the
 621          * keep alive timeout, but we still want to track a timeout and cleanup
 622          * in case a host died before it enabled the controller.  Hence, simply
 623          * reset the keep alive timer when the controller is enabled.
 624          */
 625         if (ctrl->kato)
 626                 mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
 627 }
 628
 629 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
 630 {
 631         lockdep_assert_held(&ctrl->lock);
 632
 633         /* XXX: tear down queues? */
 634         ctrl->csts &= ~NVME_CSTS_RDY;
 635         ctrl->cc = 0;
 636 }
 637
 638 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
 639 {
 640         u32 old;
 641
 642         mutex_lock(&ctrl->lock);
 643         old = ctrl->cc;
 644         ctrl->cc = new;
 645
 646         if (nvmet_cc_en(new) && !nvmet_cc_en(old))
 647                 nvmet_start_ctrl(ctrl);
 648         if (!nvmet_cc_en(new) && nvmet_cc_en(old))
 649                 nvmet_clear_ctrl(ctrl);
 650         if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
 651                 nvmet_clear_ctrl(ctrl);
 652                 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
 653         }
 654         if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
 655                 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
 656         mutex_unlock(&ctrl->lock);
 657 }
 658
 659 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
 660 {
 661         /* command sets supported: NVMe command set: */
 662         ctrl->cap = (1ULL << 37);
 663         /* CC.EN timeout in 500msec units: */
 664         ctrl->cap |= (15ULL << 24);
 665         /* maximum queue entries supported: */
 666         ctrl->cap |= NVMET_QUEUE_SIZE - 1;
 667 }
 668
 669 u16 nvmet_ctrl_find_get(const char *subsysnqn, const char *hostnqn, u16 cntlid,
 670                 struct nvmet_req *req, struct nvmet_ctrl **ret)
 671 {
 672         struct nvmet_subsys *subsys;
 673         struct nvmet_ctrl *ctrl;
 674         u16 status = 0;
 675
 676         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
 677         if (!subsys) {
 678                 pr_warn("connect request for invalid subsystem %s!\n",
 679                         subsysnqn);
 680                 req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
 681                 return NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
 682         }
 683
 684         mutex_lock(&subsys->lock);
 685         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 686                 if (ctrl->cntlid == cntlid) {
 687                         if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
 688                                 pr_warn("hostnqn mismatch.\n");
 689                                 continue;
 690                         }
 691                         if (!kref_get_unless_zero(&ctrl->ref))
 692                                 continue;
 693
 694                         *ret = ctrl;
 695                         goto out;
 696                 }
 697         }
 698
 699         pr_warn("could not find controller %d for subsys %s / host %s\n",
 700                 cntlid, subsysnqn, hostnqn);
 701         req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
 702         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
 703
 704 out:
 705         mutex_unlock(&subsys->lock);
 706         nvmet_subsys_put(subsys);
 707         return status;
 708 }
 709
 710 u16 nvmet_check_ctrl_status(struct nvmet_req *req, struct nvme_command *cmd)
 711 {
 712         if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
 713                 pr_err("got io cmd %d while CC.EN == 0 on qid = %d\n",
 714                        cmd->common.opcode, req->sq->qid);
 715                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
 716         }
 717
 718         if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
 719                 pr_err("got io cmd %d while CSTS.RDY == 0 on qid = %d\n",
 720                        cmd->common.opcode, req->sq->qid);
 721                 req->ns = NULL;
 722                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
 723         }
 724         return 0;
 725 }
 726
 727 static bool __nvmet_host_allowed(struct nvmet_subsys *subsys,
 728                 const char *hostnqn)
 729 {
 730         struct nvmet_host_link *p;
 731
 732         if (subsys->allow_any_host)
 733                 return true;
 734
 735         list_for_each_entry(p, &subsys->hosts, entry) {
 736                 if (!strcmp(nvmet_host_name(p->host), hostnqn))
 737                         return true;
 738         }
 739
 740         return false;
 741 }
 742
 743 static bool nvmet_host_discovery_allowed(struct nvmet_req *req,
 744                 const char *hostnqn)
 745 {
 746         struct nvmet_subsys_link *s;
 747
 748         list_for_each_entry(s, &req->port->subsystems, entry) {
 749                 if (__nvmet_host_allowed(s->subsys, hostnqn))
 750                         return true;
 751         }
 752
 753         return false;
 754 }
 755
 756 bool nvmet_host_allowed(struct nvmet_req *req, struct nvmet_subsys *subsys,
 757                 const char *hostnqn)
 758 {
 759         lockdep_assert_held(&nvmet_config_sem);
 760
 761         if (subsys->type == NVME_NQN_DISC)
 762                 return nvmet_host_discovery_allowed(req, hostnqn);
 763         else
 764                 return __nvmet_host_allowed(subsys, hostnqn);
 765 }
 766
 767 static void nvmet_fatal_error_handler(struct work_struct *work)
 768 {
 769         struct nvmet_ctrl *ctrl =
 770                         container_of(work, struct nvmet_ctrl, fatal_err_work);
 771
 772         pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
 773         ctrl->ops->delete_ctrl(ctrl);
 774 }
 775
 776 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
 777                 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
 778 {
 779         struct nvmet_subsys *subsys;
 780         struct nvmet_ctrl *ctrl;
 781         int ret;
 782         u16 status;
 783
 784         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
 785         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
 786         if (!subsys) {
 787                 pr_warn("connect request for invalid subsystem %s!\n",
 788                         subsysnqn);
 789                 req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
 790                 goto out;
 791         }
 792
 793         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
 794         down_read(&nvmet_config_sem);
 795         if (!nvmet_host_allowed(req, subsys, hostnqn)) {
 796                 pr_info("connect by host %s for subsystem %s not allowed\n",
 797                         hostnqn, subsysnqn);
 798                 req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
 799                 up_read(&nvmet_config_sem);
 800                 status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
 801                 goto out_put_subsystem;
 802         }
 803         up_read(&nvmet_config_sem);
 804
 805         status = NVME_SC_INTERNAL;
 806         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
 807         if (!ctrl)
 808                 goto out_put_subsystem;
 809         mutex_init(&ctrl->lock);
 810
 811         nvmet_init_cap(ctrl);
 812
 813         INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
 814         INIT_LIST_HEAD(&ctrl->async_events);
 815         INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
 816
 817         memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
 818         memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
 819
 820         kref_init(&ctrl->ref);
 821         ctrl->subsys = subsys;
 822
 823         ctrl->cqs = kcalloc(subsys->max_qid + 1,
 824                         sizeof(struct nvmet_cq *),
 825                         GFP_KERNEL);
 826         if (!ctrl->cqs)
 827                 goto out_free_ctrl;
 828
 829         ctrl->sqs = kcalloc(subsys->max_qid + 1,
 830                         sizeof(struct nvmet_sq *),
 831                         GFP_KERNEL);
 832         if (!ctrl->sqs)
 833                 goto out_free_cqs;
 834
 835         ret = ida_simple_get(&cntlid_ida,
 836                              NVME_CNTLID_MIN, NVME_CNTLID_MAX,
 837                              GFP_KERNEL);
 838         if (ret < 0) {
 839                 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
 840                 goto out_free_sqs;
 841         }
 842         ctrl->cntlid = ret;
 843
 844         ctrl->ops = req->ops;
 845         if (ctrl->subsys->type == NVME_NQN_DISC) {
 846                 /* Don't accept keep-alive timeout for discovery controllers */
 847                 if (kato) {
 848                         status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 849                         goto out_free_sqs;
 850                 }
 851
 852                 /*
 853                  * Discovery controllers use some arbitrary high value in order
 854                  * to cleanup stale discovery sessions
 855                  *
 856                  * From the latest base diff RC:
 857                  * "The Keep Alive command is not supported by
 858                  * Discovery controllers. A transport may specify a
 859                  * fixed Discovery controller activity timeout value
 860                  * (e.g., 2 minutes).  If no commands are received
 861                  * by a Discovery controller within that time
 862                  * period, the controller may perform the
 863                  * actions for Keep Alive Timer expiration".
 864                  */
 865                 ctrl->kato = NVMET_DISC_KATO;
 866         } else {
 867                 /* keep-alive timeout in seconds */
 868                 ctrl->kato = DIV_ROUND_UP(kato, 1000);
 869         }
 870         nvmet_start_keep_alive_timer(ctrl);
 871
 872         mutex_lock(&subsys->lock);
 873         list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
 874         mutex_unlock(&subsys->lock);
 875
 876         *ctrlp = ctrl;
 877         return 0;
 878
 879 out_free_sqs:
 880         kfree(ctrl->sqs);
 881 out_free_cqs:
 882         kfree(ctrl->cqs);
 883 out_free_ctrl:
 884         kfree(ctrl);
 885 out_put_subsystem:
 886         nvmet_subsys_put(subsys);
 887 out:
 888         return status;
 889 }
 890
 891 static void nvmet_ctrl_free(struct kref *ref)
 892 {
 893         struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
 894         struct nvmet_subsys *subsys = ctrl->subsys;
 895
 896         nvmet_stop_keep_alive_timer(ctrl);
 897
 898         mutex_lock(&subsys->lock);
 899         list_del(&ctrl->subsys_entry);
 900         mutex_unlock(&subsys->lock);
 901
 902         flush_work(&ctrl->async_event_work);
 903         cancel_work_sync(&ctrl->fatal_err_work);
 904
 905         ida_simple_remove(&cntlid_ida, ctrl->cntlid);
 906         nvmet_subsys_put(subsys);
 907
 908         kfree(ctrl->sqs);
 909         kfree(ctrl->cqs);
 910         kfree(ctrl);
 911 }
 912
 913 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
 914 {
 915         kref_put(&ctrl->ref, nvmet_ctrl_free);
 916 }
 917
 918 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
 919 {
 920         mutex_lock(&ctrl->lock);
 921         if (!(ctrl->csts & NVME_CSTS_CFS)) {
 922                 ctrl->csts |= NVME_CSTS_CFS;
 923                 schedule_work(&ctrl->fatal_err_work);
 924         }
 925         mutex_unlock(&ctrl->lock);
 926 }
 927 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
 928
 929 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
 930                 const char *subsysnqn)
 931 {
 932         struct nvmet_subsys_link *p;
 933
 934         if (!port)
 935                 return NULL;
 936
 937         if (!strncmp(NVME_DISC_SUBSYS_NAME, subsysnqn,
 938                         NVMF_NQN_SIZE)) {
 939                 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
 940                         return NULL;
 941                 return nvmet_disc_subsys;
 942         }
 943
 944         down_read(&nvmet_config_sem);
 945         list_for_each_entry(p, &port->subsystems, entry) {
 946                 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
 947                                 NVMF_NQN_SIZE)) {
 948                         if (!kref_get_unless_zero(&p->subsys->ref))
 949                                 break;
 950                         up_read(&nvmet_config_sem);
 951                         return p->subsys;
 952                 }
 953         }
 954         up_read(&nvmet_config_sem);
 955         return NULL;
 956 }
 957
 958 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
 959                 enum nvme_subsys_type type)
 960 {
 961         struct nvmet_subsys *subsys;
 962
 963         subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
 964         if (!subsys)
 965                 return NULL;
 966
 967         subsys->ver = NVME_VS(1, 3, 0); /* NVMe 1.3.0 */
 968         /* generate a random serial number as our controllers are ephemeral: */
 969         get_random_bytes(&subsys->serial, sizeof(subsys->serial));
 970
 971         switch (type) {
 972         case NVME_NQN_NVME:
 973                 subsys->max_qid = NVMET_NR_QUEUES;
 974                 break;
 975         case NVME_NQN_DISC:
 976                 subsys->max_qid = 0;
 977                 break;
 978         default:
 979                 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
 980                 kfree(subsys);
 981                 return NULL;
 982         }
 983         subsys->type = type;
 984         subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
 985                         GFP_KERNEL);
 986         if (!subsys->subsysnqn) {
 987                 kfree(subsys);
 988                 return NULL;
 989         }
 990
 991         kref_init(&subsys->ref);
 992
 993         mutex_init(&subsys->lock);
 994         INIT_LIST_HEAD(&subsys->namespaces);
 995         INIT_LIST_HEAD(&subsys->ctrls);
 996         INIT_LIST_HEAD(&subsys->hosts);
 997
 998         return subsys;
 999 }
1000
1001 static void nvmet_subsys_free(struct kref *ref)
1002 {
1003         struct nvmet_subsys *subsys =
1004                 container_of(ref, struct nvmet_subsys, ref);
1005
1006         WARN_ON_ONCE(!list_empty(&subsys->namespaces));
1007
1008         kfree(subsys->subsysnqn);
1009         kfree(subsys);
1010 }
1011
1012 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1013 {
1014         struct nvmet_ctrl *ctrl;
1015
1016         mutex_lock(&subsys->lock);
1017         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1018                 ctrl->ops->delete_ctrl(ctrl);
1019         mutex_unlock(&subsys->lock);
1020 }
1021
1022 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1023 {
1024         kref_put(&subsys->ref, nvmet_subsys_free);
1025 }
1026
1027 static int __init nvmet_init(void)
1028 {
1029         int error;
1030
1031         error = nvmet_init_discovery();
1032         if (error)
1033                 goto out;
1034
1035         error = nvmet_init_configfs();
1036         if (error)
1037                 goto out_exit_discovery;
1038         return 0;
1039
1040 out_exit_discovery:
1041         nvmet_exit_discovery();
1042 out:
1043         return error;
1044 }
1045
1046 static void __exit nvmet_exit(void)
1047 {
1048         nvmet_exit_configfs();
1049         nvmet_exit_discovery();
1050         ida_destroy(&cntlid_ida);
1051
1052         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1053         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1054 }
1055
1056 module_init(nvmet_init);
1057 module_exit(nvmet_exit);
1058
1059 MODULE_LICENSE("GPL v2");