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 #include <linux/pci-p2pdma.h>
  19
  20 #include "nvmet.h"
  21
  22 struct workqueue_struct *buffered_io_wq;
  23 static const struct nvmet_fabrics_ops *nvmet_transports[NVMF_TRTYPE_MAX];
  24 static DEFINE_IDA(cntlid_ida);
  25
  26 /*
  27  * This read/write semaphore is used to synchronize access to configuration
  28  * information on a target system that will result in discovery log page
  29  * information change for at least one host.
  30  * The full list of resources to protected by this semaphore is:
  31  *
  32  *  - subsystems list
  33  *  - per-subsystem allowed hosts list
  34  *  - allow_any_host subsystem attribute
  35  *  - nvmet_genctr
  36  *  - the nvmet_transports array
  37  *
  38  * When updating any of those lists/structures write lock should be obtained,
  39  * while when reading (popolating discovery log page or checking host-subsystem
  40  * link) read lock is obtained to allow concurrent reads.
  41  */
  42 DECLARE_RWSEM(nvmet_config_sem);
  43
  44 u32 nvmet_ana_group_enabled[NVMET_MAX_ANAGRPS + 1];
  45 u64 nvmet_ana_chgcnt;
  46 DECLARE_RWSEM(nvmet_ana_sem);
  47
  48 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
  49                 const char *subsysnqn);
  50
  51 u16 nvmet_copy_to_sgl(struct nvmet_req *req, off_t off, const void *buf,
  52                 size_t len)
  53 {
  54         if (sg_pcopy_from_buffer(req->sg, req->sg_cnt, buf, len, off) != len)
  55                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
  56         return 0;
  57 }
  58
  59 u16 nvmet_copy_from_sgl(struct nvmet_req *req, off_t off, void *buf, size_t len)
  60 {
  61         if (sg_pcopy_to_buffer(req->sg, req->sg_cnt, buf, len, off) != len)
  62                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
  63         return 0;
  64 }
  65
  66 u16 nvmet_zero_sgl(struct nvmet_req *req, off_t off, size_t len)
  67 {
  68         if (sg_zero_buffer(req->sg, req->sg_cnt, len, off) != len)
  69                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
  70         return 0;
  71 }
  72
  73 static unsigned int nvmet_max_nsid(struct nvmet_subsys *subsys)
  74 {
  75         struct nvmet_ns *ns;
  76
  77         if (list_empty(&subsys->namespaces))
  78                 return 0;
  79
  80         ns = list_last_entry(&subsys->namespaces, struct nvmet_ns, dev_link);
  81         return ns->nsid;
  82 }
  83
  84 static u32 nvmet_async_event_result(struct nvmet_async_event *aen)
  85 {
  86         return aen->event_type | (aen->event_info << 8) | (aen->log_page << 16);
  87 }
  88
  89 static void nvmet_async_events_free(struct nvmet_ctrl *ctrl)
  90 {
  91         struct nvmet_req *req;
  92
  93         while (1) {
  94                 mutex_lock(&ctrl->lock);
  95                 if (!ctrl->nr_async_event_cmds) {
  96                         mutex_unlock(&ctrl->lock);
  97                         return;
  98                 }
  99
 100                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
 101                 mutex_unlock(&ctrl->lock);
 102                 nvmet_req_complete(req, NVME_SC_INTERNAL | NVME_SC_DNR);
 103         }
 104 }
 105
 106 static void nvmet_async_event_work(struct work_struct *work)
 107 {
 108         struct nvmet_ctrl *ctrl =
 109                 container_of(work, struct nvmet_ctrl, async_event_work);
 110         struct nvmet_async_event *aen;
 111         struct nvmet_req *req;
 112
 113         while (1) {
 114                 mutex_lock(&ctrl->lock);
 115                 aen = list_first_entry_or_null(&ctrl->async_events,
 116                                 struct nvmet_async_event, entry);
 117                 if (!aen || !ctrl->nr_async_event_cmds) {
 118                         mutex_unlock(&ctrl->lock);
 119                         return;
 120                 }
 121
 122                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
 123                 nvmet_set_result(req, nvmet_async_event_result(aen));
 124
 125                 list_del(&aen->entry);
 126                 kfree(aen);
 127
 128                 mutex_unlock(&ctrl->lock);
 129                 nvmet_req_complete(req, 0);
 130         }
 131 }
 132
 133 static void nvmet_add_async_event(struct nvmet_ctrl *ctrl, u8 event_type,
 134                 u8 event_info, u8 log_page)
 135 {
 136         struct nvmet_async_event *aen;
 137
 138         aen = kmalloc(sizeof(*aen), GFP_KERNEL);
 139         if (!aen)
 140                 return;
 141
 142         aen->event_type = event_type;
 143         aen->event_info = event_info;
 144         aen->log_page = log_page;
 145
 146         mutex_lock(&ctrl->lock);
 147         list_add_tail(&aen->entry, &ctrl->async_events);
 148         mutex_unlock(&ctrl->lock);
 149
 150         schedule_work(&ctrl->async_event_work);
 151 }
 152
 153 static void nvmet_add_to_changed_ns_log(struct nvmet_ctrl *ctrl, __le32 nsid)
 154 {
 155         u32 i;
 156
 157         mutex_lock(&ctrl->lock);
 158         if (ctrl->nr_changed_ns > NVME_MAX_CHANGED_NAMESPACES)
 159                 goto out_unlock;
 160
 161         for (i = 0; i < ctrl->nr_changed_ns; i++) {
 162                 if (ctrl->changed_ns_list[i] == nsid)
 163                         goto out_unlock;
 164         }
 165
 166         if (ctrl->nr_changed_ns == NVME_MAX_CHANGED_NAMESPACES) {
 167                 ctrl->changed_ns_list[0] = cpu_to_le32(0xffffffff);
 168                 ctrl->nr_changed_ns = U32_MAX;
 169                 goto out_unlock;
 170         }
 171
 172         ctrl->changed_ns_list[ctrl->nr_changed_ns++] = nsid;
 173 out_unlock:
 174         mutex_unlock(&ctrl->lock);
 175 }
 176
 177 void nvmet_ns_changed(struct nvmet_subsys *subsys, u32 nsid)
 178 {
 179         struct nvmet_ctrl *ctrl;
 180
 181         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 182                 nvmet_add_to_changed_ns_log(ctrl, cpu_to_le32(nsid));
 183                 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_NS_ATTR))
 184                         continue;
 185                 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
 186                                 NVME_AER_NOTICE_NS_CHANGED,
 187                                 NVME_LOG_CHANGED_NS);
 188         }
 189 }
 190
 191 void nvmet_send_ana_event(struct nvmet_subsys *subsys,
 192                 struct nvmet_port *port)
 193 {
 194         struct nvmet_ctrl *ctrl;
 195
 196         mutex_lock(&subsys->lock);
 197         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 198                 if (port && ctrl->port != port)
 199                         continue;
 200                 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_ANA_CHANGE))
 201                         continue;
 202                 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
 203                                 NVME_AER_NOTICE_ANA, NVME_LOG_ANA);
 204         }
 205         mutex_unlock(&subsys->lock);
 206 }
 207
 208 void nvmet_port_send_ana_event(struct nvmet_port *port)
 209 {
 210         struct nvmet_subsys_link *p;
 211
 212         down_read(&nvmet_config_sem);
 213         list_for_each_entry(p, &port->subsystems, entry)
 214                 nvmet_send_ana_event(p->subsys, port);
 215         up_read(&nvmet_config_sem);
 216 }
 217
 218 int nvmet_register_transport(const struct nvmet_fabrics_ops *ops)
 219 {
 220         int ret = 0;
 221
 222         down_write(&nvmet_config_sem);
 223         if (nvmet_transports[ops->type])
 224                 ret = -EINVAL;
 225         else
 226                 nvmet_transports[ops->type] = ops;
 227         up_write(&nvmet_config_sem);
 228
 229         return ret;
 230 }
 231 EXPORT_SYMBOL_GPL(nvmet_register_transport);
 232
 233 void nvmet_unregister_transport(const struct nvmet_fabrics_ops *ops)
 234 {
 235         down_write(&nvmet_config_sem);
 236         nvmet_transports[ops->type] = NULL;
 237         up_write(&nvmet_config_sem);
 238 }
 239 EXPORT_SYMBOL_GPL(nvmet_unregister_transport);
 240
 241 int nvmet_enable_port(struct nvmet_port *port)
 242 {
 243         const struct nvmet_fabrics_ops *ops;
 244         int ret;
 245
 246         lockdep_assert_held(&nvmet_config_sem);
 247
 248         ops = nvmet_transports[port->disc_addr.trtype];
 249         if (!ops) {
 250                 up_write(&nvmet_config_sem);
 251                 request_module("nvmet-transport-%d", port->disc_addr.trtype);
 252                 down_write(&nvmet_config_sem);
 253                 ops = nvmet_transports[port->disc_addr.trtype];
 254                 if (!ops) {
 255                         pr_err("transport type %d not supported\n",
 256                                 port->disc_addr.trtype);
 257                         return -EINVAL;
 258                 }
 259         }
 260
 261         if (!try_module_get(ops->owner))
 262                 return -EINVAL;
 263
 264         ret = ops->add_port(port);
 265         if (ret) {
 266                 module_put(ops->owner);
 267                 return ret;
 268         }
 269
 270         /* If the transport didn't set inline_data_size, then disable it. */
 271         if (port->inline_data_size < 0)
 272                 port->inline_data_size = 0;
 273
 274         port->enabled = true;
 275         return 0;
 276 }
 277
 278 void nvmet_disable_port(struct nvmet_port *port)
 279 {
 280         const struct nvmet_fabrics_ops *ops;
 281
 282         lockdep_assert_held(&nvmet_config_sem);
 283
 284         port->enabled = false;
 285
 286         ops = nvmet_transports[port->disc_addr.trtype];
 287         ops->remove_port(port);
 288         module_put(ops->owner);
 289 }
 290
 291 static void nvmet_keep_alive_timer(struct work_struct *work)
 292 {
 293         struct nvmet_ctrl *ctrl = container_of(to_delayed_work(work),
 294                         struct nvmet_ctrl, ka_work);
 295         bool cmd_seen = ctrl->cmd_seen;
 296
 297         ctrl->cmd_seen = false;
 298         if (cmd_seen) {
 299                 pr_debug("ctrl %d reschedule traffic based keep-alive timer\n",
 300                         ctrl->cntlid);
 301                 schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
 302                 return;
 303         }
 304
 305         pr_err("ctrl %d keep-alive timer (%d seconds) expired!\n",
 306                 ctrl->cntlid, ctrl->kato);
 307
 308         nvmet_ctrl_fatal_error(ctrl);
 309 }
 310
 311 static void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl)
 312 {
 313         pr_debug("ctrl %d start keep-alive timer for %d secs\n",
 314                 ctrl->cntlid, ctrl->kato);
 315
 316         INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
 317         schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
 318 }
 319
 320 static void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
 321 {
 322         pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
 323
 324         cancel_delayed_work_sync(&ctrl->ka_work);
 325 }
 326
 327 static struct nvmet_ns *__nvmet_find_namespace(struct nvmet_ctrl *ctrl,
 328                 __le32 nsid)
 329 {
 330         struct nvmet_ns *ns;
 331
 332         list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link) {
 333                 if (ns->nsid == le32_to_cpu(nsid))
 334                         return ns;
 335         }
 336
 337         return NULL;
 338 }
 339
 340 struct nvmet_ns *nvmet_find_namespace(struct nvmet_ctrl *ctrl, __le32 nsid)
 341 {
 342         struct nvmet_ns *ns;
 343
 344         rcu_read_lock();
 345         ns = __nvmet_find_namespace(ctrl, nsid);
 346         if (ns)
 347                 percpu_ref_get(&ns->ref);
 348         rcu_read_unlock();
 349
 350         return ns;
 351 }
 352
 353 static void nvmet_destroy_namespace(struct percpu_ref *ref)
 354 {
 355         struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
 356
 357         complete(&ns->disable_done);
 358 }
 359
 360 void nvmet_put_namespace(struct nvmet_ns *ns)
 361 {
 362         percpu_ref_put(&ns->ref);
 363 }
 364
 365 static void nvmet_ns_dev_disable(struct nvmet_ns *ns)
 366 {
 367         nvmet_bdev_ns_disable(ns);
 368         nvmet_file_ns_disable(ns);
 369 }
 370
 371 static int nvmet_p2pmem_ns_enable(struct nvmet_ns *ns)
 372 {
 373         int ret;
 374         struct pci_dev *p2p_dev;
 375
 376         if (!ns->use_p2pmem)
 377                 return 0;
 378
 379         if (!ns->bdev) {
 380                 pr_err("peer-to-peer DMA is not supported by non-block device namespaces\n");
 381                 return -EINVAL;
 382         }
 383
 384         if (!blk_queue_pci_p2pdma(ns->bdev->bd_queue)) {
 385                 pr_err("peer-to-peer DMA is not supported by the driver of %s\n",
 386                        ns->device_path);
 387                 return -EINVAL;
 388         }
 389
 390         if (ns->p2p_dev) {
 391                 ret = pci_p2pdma_distance(ns->p2p_dev, nvmet_ns_dev(ns), true);
 392                 if (ret < 0)
 393                         return -EINVAL;
 394         } else {
 395                 /*
 396                  * Right now we just check that there is p2pmem available so
 397                  * we can report an error to the user right away if there
 398                  * is not. We'll find the actual device to use once we
 399                  * setup the controller when the port's device is available.
 400                  */
 401
 402                 p2p_dev = pci_p2pmem_find(nvmet_ns_dev(ns));
 403                 if (!p2p_dev) {
 404                         pr_err("no peer-to-peer memory is available for %s\n",
 405                                ns->device_path);
 406                         return -EINVAL;
 407                 }
 408
 409                 pci_dev_put(p2p_dev);
 410         }
 411
 412         return 0;
 413 }
 414
 415 /*
 416  * Note: ctrl->subsys->lock should be held when calling this function
 417  */
 418 static void nvmet_p2pmem_ns_add_p2p(struct nvmet_ctrl *ctrl,
 419                                     struct nvmet_ns *ns)
 420 {
 421         struct device *clients[2];
 422         struct pci_dev *p2p_dev;
 423         int ret;
 424
 425         if (!ctrl->p2p_client || !ns->use_p2pmem)
 426                 return;
 427
 428         if (ns->p2p_dev) {
 429                 ret = pci_p2pdma_distance(ns->p2p_dev, ctrl->p2p_client, true);
 430                 if (ret < 0)
 431                         return;
 432
 433                 p2p_dev = pci_dev_get(ns->p2p_dev);
 434         } else {
 435                 clients[0] = ctrl->p2p_client;
 436                 clients[1] = nvmet_ns_dev(ns);
 437
 438                 p2p_dev = pci_p2pmem_find_many(clients, ARRAY_SIZE(clients));
 439                 if (!p2p_dev) {
 440                         pr_err("no peer-to-peer memory is available that's supported by %s and %s\n",
 441                                dev_name(ctrl->p2p_client), ns->device_path);
 442                         return;
 443                 }
 444         }
 445
 446         ret = radix_tree_insert(&ctrl->p2p_ns_map, ns->nsid, p2p_dev);
 447         if (ret < 0)
 448                 pci_dev_put(p2p_dev);
 449
 450         pr_info("using p2pmem on %s for nsid %d\n", pci_name(p2p_dev),
 451                 ns->nsid);
 452 }
 453
 454 int nvmet_ns_enable(struct nvmet_ns *ns)
 455 {
 456         struct nvmet_subsys *subsys = ns->subsys;
 457         struct nvmet_ctrl *ctrl;
 458         int ret;
 459
 460         mutex_lock(&subsys->lock);
 461         ret = -EMFILE;
 462         if (subsys->nr_namespaces == NVMET_MAX_NAMESPACES)
 463                 goto out_unlock;
 464         ret = 0;
 465         if (ns->enabled)
 466                 goto out_unlock;
 467
 468         ret = nvmet_bdev_ns_enable(ns);
 469         if (ret == -ENOTBLK)
 470                 ret = nvmet_file_ns_enable(ns);
 471         if (ret)
 472                 goto out_unlock;
 473
 474         ret = nvmet_p2pmem_ns_enable(ns);
 475         if (ret)
 476                 goto out_unlock;
 477
 478         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 479                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
 480
 481         ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
 482                                 0, GFP_KERNEL);
 483         if (ret)
 484                 goto out_dev_put;
 485
 486         if (ns->nsid > subsys->max_nsid)
 487                 subsys->max_nsid = ns->nsid;
 488
 489         /*
 490          * The namespaces list needs to be sorted to simplify the implementation
 491          * of the Identify Namepace List subcommand.
 492          */
 493         if (list_empty(&subsys->namespaces)) {
 494                 list_add_tail_rcu(&ns->dev_link, &subsys->namespaces);
 495         } else {
 496                 struct nvmet_ns *old;
 497
 498                 list_for_each_entry_rcu(old, &subsys->namespaces, dev_link) {
 499                         BUG_ON(ns->nsid == old->nsid);
 500                         if (ns->nsid < old->nsid)
 501                                 break;
 502                 }
 503
 504                 list_add_tail_rcu(&ns->dev_link, &old->dev_link);
 505         }
 506         subsys->nr_namespaces++;
 507
 508         nvmet_ns_changed(subsys, ns->nsid);
 509         ns->enabled = true;
 510         ret = 0;
 511 out_unlock:
 512         mutex_unlock(&subsys->lock);
 513         return ret;
 514 out_dev_put:
 515         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 516                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 517
 518         nvmet_ns_dev_disable(ns);
 519         goto out_unlock;
 520 }
 521
 522 void nvmet_ns_disable(struct nvmet_ns *ns)
 523 {
 524         struct nvmet_subsys *subsys = ns->subsys;
 525         struct nvmet_ctrl *ctrl;
 526
 527         mutex_lock(&subsys->lock);
 528         if (!ns->enabled)
 529                 goto out_unlock;
 530
 531         ns->enabled = false;
 532         list_del_rcu(&ns->dev_link);
 533         if (ns->nsid == subsys->max_nsid)
 534                 subsys->max_nsid = nvmet_max_nsid(subsys);
 535
 536         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 537                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 538
 539         mutex_unlock(&subsys->lock);
 540
 541         /*
 542          * Now that we removed the namespaces from the lookup list, we
 543          * can kill the per_cpu ref and wait for any remaining references
 544          * to be dropped, as well as a RCU grace period for anyone only
 545          * using the namepace under rcu_read_lock().  Note that we can't
 546          * use call_rcu here as we need to ensure the namespaces have
 547          * been fully destroyed before unloading the module.
 548          */
 549         percpu_ref_kill(&ns->ref);
 550         synchronize_rcu();
 551         wait_for_completion(&ns->disable_done);
 552         percpu_ref_exit(&ns->ref);
 553
 554         mutex_lock(&subsys->lock);
 555
 556         subsys->nr_namespaces--;
 557         nvmet_ns_changed(subsys, ns->nsid);
 558         nvmet_ns_dev_disable(ns);
 559 out_unlock:
 560         mutex_unlock(&subsys->lock);
 561 }
 562
 563 void nvmet_ns_free(struct nvmet_ns *ns)
 564 {
 565         nvmet_ns_disable(ns);
 566
 567         down_write(&nvmet_ana_sem);
 568         nvmet_ana_group_enabled[ns->anagrpid]--;
 569         up_write(&nvmet_ana_sem);
 570
 571         kfree(ns->device_path);
 572         kfree(ns);
 573 }
 574
 575 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
 576 {
 577         struct nvmet_ns *ns;
 578
 579         ns = kzalloc(sizeof(*ns), GFP_KERNEL);
 580         if (!ns)
 581                 return NULL;
 582
 583         INIT_LIST_HEAD(&ns->dev_link);
 584         init_completion(&ns->disable_done);
 585
 586         ns->nsid = nsid;
 587         ns->subsys = subsys;
 588
 589         down_write(&nvmet_ana_sem);
 590         ns->anagrpid = NVMET_DEFAULT_ANA_GRPID;
 591         nvmet_ana_group_enabled[ns->anagrpid]++;
 592         up_write(&nvmet_ana_sem);
 593
 594         uuid_gen(&ns->uuid);
 595         ns->buffered_io = false;
 596
 597         return ns;
 598 }
 599
 600 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
 601 {
 602         u32 old_sqhd, new_sqhd;
 603         u16 sqhd;
 604
 605         if (status)
 606                 nvmet_set_status(req, status);
 607
 608         if (req->sq->size) {
 609                 do {
 610                         old_sqhd = req->sq->sqhd;
 611                         new_sqhd = (old_sqhd + 1) % req->sq->size;
 612                 } while (cmpxchg(&req->sq->sqhd, old_sqhd, new_sqhd) !=
 613                                         old_sqhd);
 614         }
 615         sqhd = req->sq->sqhd & 0x0000FFFF;
 616         req->rsp->sq_head = cpu_to_le16(sqhd);
 617         req->rsp->sq_id = cpu_to_le16(req->sq->qid);
 618         req->rsp->command_id = req->cmd->common.command_id;
 619
 620         if (req->ns)
 621                 nvmet_put_namespace(req->ns);
 622         req->ops->queue_response(req);
 623 }
 624
 625 void nvmet_req_complete(struct nvmet_req *req, u16 status)
 626 {
 627         __nvmet_req_complete(req, status);
 628         percpu_ref_put(&req->sq->ref);
 629 }
 630 EXPORT_SYMBOL_GPL(nvmet_req_complete);
 631
 632 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
 633                 u16 qid, u16 size)
 634 {
 635         cq->qid = qid;
 636         cq->size = size;
 637
 638         ctrl->cqs[qid] = cq;
 639 }
 640
 641 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
 642                 u16 qid, u16 size)
 643 {
 644         sq->sqhd = 0;
 645         sq->qid = qid;
 646         sq->size = size;
 647
 648         ctrl->sqs[qid] = sq;
 649 }
 650
 651 static void nvmet_confirm_sq(struct percpu_ref *ref)
 652 {
 653         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 654
 655         complete(&sq->confirm_done);
 656 }
 657
 658 void nvmet_sq_destroy(struct nvmet_sq *sq)
 659 {
 660         /*
 661          * If this is the admin queue, complete all AERs so that our
 662          * queue doesn't have outstanding requests on it.
 663          */
 664         if (sq->ctrl && sq->ctrl->sqs && sq->ctrl->sqs[0] == sq)
 665                 nvmet_async_events_free(sq->ctrl);
 666         percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
 667         wait_for_completion(&sq->confirm_done);
 668         wait_for_completion(&sq->free_done);
 669         percpu_ref_exit(&sq->ref);
 670
 671         if (sq->ctrl) {
 672                 nvmet_ctrl_put(sq->ctrl);
 673                 sq->ctrl = NULL; /* allows reusing the queue later */
 674         }
 675 }
 676 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
 677
 678 static void nvmet_sq_free(struct percpu_ref *ref)
 679 {
 680         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 681
 682         complete(&sq->free_done);
 683 }
 684
 685 int nvmet_sq_init(struct nvmet_sq *sq)
 686 {
 687         int ret;
 688
 689         ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
 690         if (ret) {
 691                 pr_err("percpu_ref init failed!\n");
 692                 return ret;
 693         }
 694         init_completion(&sq->free_done);
 695         init_completion(&sq->confirm_done);
 696
 697         return 0;
 698 }
 699 EXPORT_SYMBOL_GPL(nvmet_sq_init);
 700
 701 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
 702                 struct nvmet_ns *ns)
 703 {
 704         enum nvme_ana_state state = port->ana_state[ns->anagrpid];
 705
 706         if (unlikely(state == NVME_ANA_INACCESSIBLE))
 707                 return NVME_SC_ANA_INACCESSIBLE;
 708         if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
 709                 return NVME_SC_ANA_PERSISTENT_LOSS;
 710         if (unlikely(state == NVME_ANA_CHANGE))
 711                 return NVME_SC_ANA_TRANSITION;
 712         return 0;
 713 }
 714
 715 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
 716 {
 717         if (unlikely(req->ns->readonly)) {
 718                 switch (req->cmd->common.opcode) {
 719                 case nvme_cmd_read:
 720                 case nvme_cmd_flush:
 721                         break;
 722                 default:
 723                         return NVME_SC_NS_WRITE_PROTECTED;
 724                 }
 725         }
 726
 727         return 0;
 728 }
 729
 730 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
 731 {
 732         struct nvme_command *cmd = req->cmd;
 733         u16 ret;
 734
 735         ret = nvmet_check_ctrl_status(req, cmd);
 736         if (unlikely(ret))
 737                 return ret;
 738
 739         req->ns = nvmet_find_namespace(req->sq->ctrl, cmd->rw.nsid);
 740         if (unlikely(!req->ns))
 741                 return NVME_SC_INVALID_NS | NVME_SC_DNR;
 742         ret = nvmet_check_ana_state(req->port, req->ns);
 743         if (unlikely(ret))
 744                 return ret;
 745         ret = nvmet_io_cmd_check_access(req);
 746         if (unlikely(ret))
 747                 return ret;
 748
 749         if (req->ns->file)
 750                 return nvmet_file_parse_io_cmd(req);
 751         else
 752                 return nvmet_bdev_parse_io_cmd(req);
 753 }
 754
 755 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
 756                 struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
 757 {
 758         u8 flags = req->cmd->common.flags;
 759         u16 status;
 760
 761         req->cq = cq;
 762         req->sq = sq;
 763         req->ops = ops;
 764         req->sg = NULL;
 765         req->sg_cnt = 0;
 766         req->transfer_len = 0;
 767         req->rsp->status = 0;
 768         req->ns = NULL;
 769
 770         /* no support for fused commands yet */
 771         if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
 772                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 773                 goto fail;
 774         }
 775
 776         /*
 777          * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
 778          * contains an address of a single contiguous physical buffer that is
 779          * byte aligned.
 780          */
 781         if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
 782                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 783                 goto fail;
 784         }
 785
 786         if (unlikely(!req->sq->ctrl))
 787                 /* will return an error for any Non-connect command: */
 788                 status = nvmet_parse_connect_cmd(req);
 789         else if (likely(req->sq->qid != 0))
 790                 status = nvmet_parse_io_cmd(req);
 791         else if (req->cmd->common.opcode == nvme_fabrics_command)
 792                 status = nvmet_parse_fabrics_cmd(req);
 793         else if (req->sq->ctrl->subsys->type == NVME_NQN_DISC)
 794                 status = nvmet_parse_discovery_cmd(req);
 795         else
 796                 status = nvmet_parse_admin_cmd(req);
 797
 798         if (status)
 799                 goto fail;
 800
 801         if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
 802                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 803                 goto fail;
 804         }
 805
 806         if (sq->ctrl)
 807                 sq->ctrl->cmd_seen = true;
 808
 809         return true;
 810
 811 fail:
 812         __nvmet_req_complete(req, status);
 813         return false;
 814 }
 815 EXPORT_SYMBOL_GPL(nvmet_req_init);
 816
 817 void nvmet_req_uninit(struct nvmet_req *req)
 818 {
 819         percpu_ref_put(&req->sq->ref);
 820         if (req->ns)
 821                 nvmet_put_namespace(req->ns);
 822 }
 823 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
 824
 825 void nvmet_req_execute(struct nvmet_req *req)
 826 {
 827         if (unlikely(req->data_len != req->transfer_len))
 828                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
 829         else
 830                 req->execute(req);
 831 }
 832 EXPORT_SYMBOL_GPL(nvmet_req_execute);
 833
 834 int nvmet_req_alloc_sgl(struct nvmet_req *req)
 835 {
 836         struct pci_dev *p2p_dev = NULL;
 837
 838         if (IS_ENABLED(CONFIG_PCI_P2PDMA)) {
 839                 if (req->sq->ctrl && req->ns)
 840                         p2p_dev = radix_tree_lookup(&req->sq->ctrl->p2p_ns_map,
 841                                                     req->ns->nsid);
 842
 843                 req->p2p_dev = NULL;
 844                 if (req->sq->qid && p2p_dev) {
 845                         req->sg = pci_p2pmem_alloc_sgl(p2p_dev, &req->sg_cnt,
 846                                                        req->transfer_len);
 847                         if (req->sg) {
 848                                 req->p2p_dev = p2p_dev;
 849                                 return 0;
 850                         }
 851                 }
 852
 853                 /*
 854                  * If no P2P memory was available we fallback to using
 855                  * regular memory
 856                  */
 857         }
 858
 859         req->sg = sgl_alloc(req->transfer_len, GFP_KERNEL, &req->sg_cnt);
 860         if (!req->sg)
 861                 return -ENOMEM;
 862
 863         return 0;
 864 }
 865 EXPORT_SYMBOL_GPL(nvmet_req_alloc_sgl);
 866
 867 void nvmet_req_free_sgl(struct nvmet_req *req)
 868 {
 869         if (req->p2p_dev)
 870                 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
 871         else
 872                 sgl_free(req->sg);
 873
 874         req->sg = NULL;
 875         req->sg_cnt = 0;
 876 }
 877 EXPORT_SYMBOL_GPL(nvmet_req_free_sgl);
 878
 879 static inline bool nvmet_cc_en(u32 cc)
 880 {
 881         return (cc >> NVME_CC_EN_SHIFT) & 0x1;
 882 }
 883
 884 static inline u8 nvmet_cc_css(u32 cc)
 885 {
 886         return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
 887 }
 888
 889 static inline u8 nvmet_cc_mps(u32 cc)
 890 {
 891         return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
 892 }
 893
 894 static inline u8 nvmet_cc_ams(u32 cc)
 895 {
 896         return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
 897 }
 898
 899 static inline u8 nvmet_cc_shn(u32 cc)
 900 {
 901         return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
 902 }
 903
 904 static inline u8 nvmet_cc_iosqes(u32 cc)
 905 {
 906         return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
 907 }
 908
 909 static inline u8 nvmet_cc_iocqes(u32 cc)
 910 {
 911         return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
 912 }
 913
 914 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
 915 {
 916         lockdep_assert_held(&ctrl->lock);
 917
 918         if (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
 919             nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES ||
 920             nvmet_cc_mps(ctrl->cc) != 0 ||
 921             nvmet_cc_ams(ctrl->cc) != 0 ||
 922             nvmet_cc_css(ctrl->cc) != 0) {
 923                 ctrl->csts = NVME_CSTS_CFS;
 924                 return;
 925         }
 926
 927         ctrl->csts = NVME_CSTS_RDY;
 928
 929         /*
 930          * Controllers that are not yet enabled should not really enforce the
 931          * keep alive timeout, but we still want to track a timeout and cleanup
 932          * in case a host died before it enabled the controller.  Hence, simply
 933          * reset the keep alive timer when the controller is enabled.
 934          */
 935         mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
 936 }
 937
 938 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
 939 {
 940         lockdep_assert_held(&ctrl->lock);
 941
 942         /* XXX: tear down queues? */
 943         ctrl->csts &= ~NVME_CSTS_RDY;
 944         ctrl->cc = 0;
 945 }
 946
 947 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
 948 {
 949         u32 old;
 950
 951         mutex_lock(&ctrl->lock);
 952         old = ctrl->cc;
 953         ctrl->cc = new;
 954
 955         if (nvmet_cc_en(new) && !nvmet_cc_en(old))
 956                 nvmet_start_ctrl(ctrl);
 957         if (!nvmet_cc_en(new) && nvmet_cc_en(old))
 958                 nvmet_clear_ctrl(ctrl);
 959         if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
 960                 nvmet_clear_ctrl(ctrl);
 961                 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
 962         }
 963         if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
 964                 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
 965         mutex_unlock(&ctrl->lock);
 966 }
 967
 968 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
 969 {
 970         /* command sets supported: NVMe command set: */
 971         ctrl->cap = (1ULL << 37);
 972         /* CC.EN timeout in 500msec units: */
 973         ctrl->cap |= (15ULL << 24);
 974         /* maximum queue entries supported: */
 975         ctrl->cap |= NVMET_QUEUE_SIZE - 1;
 976 }
 977
 978 u16 nvmet_ctrl_find_get(const char *subsysnqn, const char *hostnqn, u16 cntlid,
 979                 struct nvmet_req *req, struct nvmet_ctrl **ret)
 980 {
 981         struct nvmet_subsys *subsys;
 982         struct nvmet_ctrl *ctrl;
 983         u16 status = 0;
 984
 985         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
 986         if (!subsys) {
 987                 pr_warn("connect request for invalid subsystem %s!\n",
 988                         subsysnqn);
 989                 req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
 990                 return NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
 991         }
 992
 993         mutex_lock(&subsys->lock);
 994         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 995                 if (ctrl->cntlid == cntlid) {
 996                         if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
 997                                 pr_warn("hostnqn mismatch.\n");
 998                                 continue;
 999                         }
1000                         if (!kref_get_unless_zero(&ctrl->ref))
1001                                 continue;
1002
1003                         *ret = ctrl;
1004                         goto out;
1005                 }
1006         }
1007
1008         pr_warn("could not find controller %d for subsys %s / host %s\n",
1009                 cntlid, subsysnqn, hostnqn);
1010         req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1011         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1012
1013 out:
1014         mutex_unlock(&subsys->lock);
1015         nvmet_subsys_put(subsys);
1016         return status;
1017 }
1018
1019 u16 nvmet_check_ctrl_status(struct nvmet_req *req, struct nvme_command *cmd)
1020 {
1021         if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1022                 pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1023                        cmd->common.opcode, req->sq->qid);
1024                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1025         }
1026
1027         if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1028                 pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1029                        cmd->common.opcode, req->sq->qid);
1030                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1031         }
1032         return 0;
1033 }
1034
1035 static bool __nvmet_host_allowed(struct nvmet_subsys *subsys,
1036                 const char *hostnqn)
1037 {
1038         struct nvmet_host_link *p;
1039
1040         if (subsys->allow_any_host)
1041                 return true;
1042
1043         list_for_each_entry(p, &subsys->hosts, entry) {
1044                 if (!strcmp(nvmet_host_name(p->host), hostnqn))
1045                         return true;
1046         }
1047
1048         return false;
1049 }
1050
1051 static bool nvmet_host_discovery_allowed(struct nvmet_req *req,
1052                 const char *hostnqn)
1053 {
1054         struct nvmet_subsys_link *s;
1055
1056         list_for_each_entry(s, &req->port->subsystems, entry) {
1057                 if (__nvmet_host_allowed(s->subsys, hostnqn))
1058                         return true;
1059         }
1060
1061         return false;
1062 }
1063
1064 bool nvmet_host_allowed(struct nvmet_req *req, struct nvmet_subsys *subsys,
1065                 const char *hostnqn)
1066 {
1067         lockdep_assert_held(&nvmet_config_sem);
1068
1069         if (subsys->type == NVME_NQN_DISC)
1070                 return nvmet_host_discovery_allowed(req, hostnqn);
1071         else
1072                 return __nvmet_host_allowed(subsys, hostnqn);
1073 }
1074
1075 /*
1076  * Note: ctrl->subsys->lock should be held when calling this function
1077  */
1078 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1079                 struct nvmet_req *req)
1080 {
1081         struct nvmet_ns *ns;
1082
1083         if (!req->p2p_client)
1084                 return;
1085
1086         ctrl->p2p_client = get_device(req->p2p_client);
1087
1088         list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link)
1089                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1090 }
1091
1092 /*
1093  * Note: ctrl->subsys->lock should be held when calling this function
1094  */
1095 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1096 {
1097         struct radix_tree_iter iter;
1098         void __rcu **slot;
1099
1100         radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1101                 pci_dev_put(radix_tree_deref_slot(slot));
1102
1103         put_device(ctrl->p2p_client);
1104 }
1105
1106 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1107                 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
1108 {
1109         struct nvmet_subsys *subsys;
1110         struct nvmet_ctrl *ctrl;
1111         int ret;
1112         u16 status;
1113
1114         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1115         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1116         if (!subsys) {
1117                 pr_warn("connect request for invalid subsystem %s!\n",
1118                         subsysnqn);
1119                 req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1120                 goto out;
1121         }
1122
1123         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1124         down_read(&nvmet_config_sem);
1125         if (!nvmet_host_allowed(req, subsys, hostnqn)) {
1126                 pr_info("connect by host %s for subsystem %s not allowed\n",
1127                         hostnqn, subsysnqn);
1128                 req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1129                 up_read(&nvmet_config_sem);
1130                 status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
1131                 goto out_put_subsystem;
1132         }
1133         up_read(&nvmet_config_sem);
1134
1135         status = NVME_SC_INTERNAL;
1136         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1137         if (!ctrl)
1138                 goto out_put_subsystem;
1139         mutex_init(&ctrl->lock);
1140
1141         nvmet_init_cap(ctrl);
1142
1143         ctrl->port = req->port;
1144
1145         INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
1146         INIT_LIST_HEAD(&ctrl->async_events);
1147         INIT_RADIX_TREE(&ctrl->p2p_ns_map, GFP_KERNEL);
1148
1149         memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1150         memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1151
1152         kref_init(&ctrl->ref);
1153         ctrl->subsys = subsys;
1154         WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1155
1156         ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1157                         sizeof(__le32), GFP_KERNEL);
1158         if (!ctrl->changed_ns_list)
1159                 goto out_free_ctrl;
1160
1161         ctrl->cqs = kcalloc(subsys->max_qid + 1,
1162                         sizeof(struct nvmet_cq *),
1163                         GFP_KERNEL);
1164         if (!ctrl->cqs)
1165                 goto out_free_changed_ns_list;
1166
1167         ctrl->sqs = kcalloc(subsys->max_qid + 1,
1168                         sizeof(struct nvmet_sq *),
1169                         GFP_KERNEL);
1170         if (!ctrl->sqs)
1171                 goto out_free_cqs;
1172
1173         ret = ida_simple_get(&cntlid_ida,
1174                              NVME_CNTLID_MIN, NVME_CNTLID_MAX,
1175                              GFP_KERNEL);
1176         if (ret < 0) {
1177                 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
1178                 goto out_free_sqs;
1179         }
1180         ctrl->cntlid = ret;
1181
1182         ctrl->ops = req->ops;
1183
1184         /*
1185          * Discovery controllers may use some arbitrary high value
1186          * in order to cleanup stale discovery sessions
1187          */
1188         if ((ctrl->subsys->type == NVME_NQN_DISC) && !kato)
1189                 kato = NVMET_DISC_KATO_MS;
1190
1191         /* keep-alive timeout in seconds */
1192         ctrl->kato = DIV_ROUND_UP(kato, 1000);
1193
1194         nvmet_start_keep_alive_timer(ctrl);
1195
1196         mutex_lock(&subsys->lock);
1197         list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1198         nvmet_setup_p2p_ns_map(ctrl, req);
1199         mutex_unlock(&subsys->lock);
1200
1201         *ctrlp = ctrl;
1202         return 0;
1203
1204 out_free_sqs:
1205         kfree(ctrl->sqs);
1206 out_free_cqs:
1207         kfree(ctrl->cqs);
1208 out_free_changed_ns_list:
1209         kfree(ctrl->changed_ns_list);
1210 out_free_ctrl:
1211         kfree(ctrl);
1212 out_put_subsystem:
1213         nvmet_subsys_put(subsys);
1214 out:
1215         return status;
1216 }
1217
1218 static void nvmet_ctrl_free(struct kref *ref)
1219 {
1220         struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1221         struct nvmet_subsys *subsys = ctrl->subsys;
1222
1223         mutex_lock(&subsys->lock);
1224         nvmet_release_p2p_ns_map(ctrl);
1225         list_del(&ctrl->subsys_entry);
1226         mutex_unlock(&subsys->lock);
1227
1228         nvmet_stop_keep_alive_timer(ctrl);
1229
1230         flush_work(&ctrl->async_event_work);
1231         cancel_work_sync(&ctrl->fatal_err_work);
1232
1233         ida_simple_remove(&cntlid_ida, ctrl->cntlid);
1234
1235         kfree(ctrl->sqs);
1236         kfree(ctrl->cqs);
1237         kfree(ctrl->changed_ns_list);
1238         kfree(ctrl);
1239
1240         nvmet_subsys_put(subsys);
1241 }
1242
1243 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1244 {
1245         kref_put(&ctrl->ref, nvmet_ctrl_free);
1246 }
1247
1248 static void nvmet_fatal_error_handler(struct work_struct *work)
1249 {
1250         struct nvmet_ctrl *ctrl =
1251                         container_of(work, struct nvmet_ctrl, fatal_err_work);
1252
1253         pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1254         ctrl->ops->delete_ctrl(ctrl);
1255 }
1256
1257 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1258 {
1259         mutex_lock(&ctrl->lock);
1260         if (!(ctrl->csts & NVME_CSTS_CFS)) {
1261                 ctrl->csts |= NVME_CSTS_CFS;
1262                 INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1263                 schedule_work(&ctrl->fatal_err_work);
1264         }
1265         mutex_unlock(&ctrl->lock);
1266 }
1267 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1268
1269 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1270                 const char *subsysnqn)
1271 {
1272         struct nvmet_subsys_link *p;
1273
1274         if (!port)
1275                 return NULL;
1276
1277         if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1278                 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1279                         return NULL;
1280                 return nvmet_disc_subsys;
1281         }
1282
1283         down_read(&nvmet_config_sem);
1284         list_for_each_entry(p, &port->subsystems, entry) {
1285                 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1286                                 NVMF_NQN_SIZE)) {
1287                         if (!kref_get_unless_zero(&p->subsys->ref))
1288                                 break;
1289                         up_read(&nvmet_config_sem);
1290                         return p->subsys;
1291                 }
1292         }
1293         up_read(&nvmet_config_sem);
1294         return NULL;
1295 }
1296
1297 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1298                 enum nvme_subsys_type type)
1299 {
1300         struct nvmet_subsys *subsys;
1301
1302         subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1303         if (!subsys)
1304                 return NULL;
1305
1306         subsys->ver = NVME_VS(1, 3, 0); /* NVMe 1.3.0 */
1307         /* generate a random serial number as our controllers are ephemeral: */
1308         get_random_bytes(&subsys->serial, sizeof(subsys->serial));
1309
1310         switch (type) {
1311         case NVME_NQN_NVME:
1312                 subsys->max_qid = NVMET_NR_QUEUES;
1313                 break;
1314         case NVME_NQN_DISC:
1315                 subsys->max_qid = 0;
1316                 break;
1317         default:
1318                 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1319                 kfree(subsys);
1320                 return NULL;
1321         }
1322         subsys->type = type;
1323         subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1324                         GFP_KERNEL);
1325         if (!subsys->subsysnqn) {
1326                 kfree(subsys);
1327                 return NULL;
1328         }
1329
1330         kref_init(&subsys->ref);
1331
1332         mutex_init(&subsys->lock);
1333         INIT_LIST_HEAD(&subsys->namespaces);
1334         INIT_LIST_HEAD(&subsys->ctrls);
1335         INIT_LIST_HEAD(&subsys->hosts);
1336
1337         return subsys;
1338 }
1339
1340 static void nvmet_subsys_free(struct kref *ref)
1341 {
1342         struct nvmet_subsys *subsys =
1343                 container_of(ref, struct nvmet_subsys, ref);
1344
1345         WARN_ON_ONCE(!list_empty(&subsys->namespaces));
1346
1347         kfree(subsys->subsysnqn);
1348         kfree(subsys);
1349 }
1350
1351 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1352 {
1353         struct nvmet_ctrl *ctrl;
1354
1355         mutex_lock(&subsys->lock);
1356         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1357                 ctrl->ops->delete_ctrl(ctrl);
1358         mutex_unlock(&subsys->lock);
1359 }
1360
1361 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1362 {
1363         kref_put(&subsys->ref, nvmet_subsys_free);
1364 }
1365
1366 static int __init nvmet_init(void)
1367 {
1368         int error;
1369
1370         nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1371
1372         buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1373                         WQ_MEM_RECLAIM, 0);
1374         if (!buffered_io_wq) {
1375                 error = -ENOMEM;
1376                 goto out;
1377         }
1378
1379         error = nvmet_init_discovery();
1380         if (error)
1381                 goto out_free_work_queue;
1382
1383         error = nvmet_init_configfs();
1384         if (error)
1385                 goto out_exit_discovery;
1386         return 0;
1387
1388 out_exit_discovery:
1389         nvmet_exit_discovery();
1390 out_free_work_queue:
1391         destroy_workqueue(buffered_io_wq);
1392 out:
1393         return error;
1394 }
1395
1396 static void __exit nvmet_exit(void)
1397 {
1398         nvmet_exit_configfs();
1399         nvmet_exit_discovery();
1400         ida_destroy(&cntlid_ida);
1401         destroy_workqueue(buffered_io_wq);
1402
1403         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1404         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1405 }
1406
1407 module_init(nvmet_init);
1408 module_exit(nvmet_exit);
1409
1410 MODULE_LICENSE("GPL v2");