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 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_update_sq_head(struct nvmet_req *req)
 601 {
 602         if (req->sq->size) {
 603                 u32 old_sqhd, new_sqhd;
 604
 605                 do {
 606                         old_sqhd = req->sq->sqhd;
 607                         new_sqhd = (old_sqhd + 1) % req->sq->size;
 608                 } while (cmpxchg(&req->sq->sqhd, old_sqhd, new_sqhd) !=
 609                                         old_sqhd);
 610         }
 611         req->rsp->sq_head = cpu_to_le16(req->sq->sqhd & 0x0000FFFF);
 612 }
 613
 614 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
 615 {
 616         if (!req->sq->sqhd_disabled)
 617                 nvmet_update_sq_head(req);
 618         req->rsp->sq_id = cpu_to_le16(req->sq->qid);
 619         req->rsp->command_id = req->cmd->common.command_id;
 620         if (status)
 621                 nvmet_set_status(req, status);
 622         if (req->ns)
 623                 nvmet_put_namespace(req->ns);
 624         req->ops->queue_response(req);
 625 }
 626
 627 void nvmet_req_complete(struct nvmet_req *req, u16 status)
 628 {
 629         __nvmet_req_complete(req, status);
 630         percpu_ref_put(&req->sq->ref);
 631 }
 632 EXPORT_SYMBOL_GPL(nvmet_req_complete);
 633
 634 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
 635                 u16 qid, u16 size)
 636 {
 637         cq->qid = qid;
 638         cq->size = size;
 639
 640         ctrl->cqs[qid] = cq;
 641 }
 642
 643 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
 644                 u16 qid, u16 size)
 645 {
 646         sq->sqhd = 0;
 647         sq->qid = qid;
 648         sq->size = size;
 649
 650         ctrl->sqs[qid] = sq;
 651 }
 652
 653 static void nvmet_confirm_sq(struct percpu_ref *ref)
 654 {
 655         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 656
 657         complete(&sq->confirm_done);
 658 }
 659
 660 void nvmet_sq_destroy(struct nvmet_sq *sq)
 661 {
 662         /*
 663          * If this is the admin queue, complete all AERs so that our
 664          * queue doesn't have outstanding requests on it.
 665          */
 666         if (sq->ctrl && sq->ctrl->sqs && sq->ctrl->sqs[0] == sq)
 667                 nvmet_async_events_free(sq->ctrl);
 668         percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
 669         wait_for_completion(&sq->confirm_done);
 670         wait_for_completion(&sq->free_done);
 671         percpu_ref_exit(&sq->ref);
 672
 673         if (sq->ctrl) {
 674                 nvmet_ctrl_put(sq->ctrl);
 675                 sq->ctrl = NULL; /* allows reusing the queue later */
 676         }
 677 }
 678 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
 679
 680 static void nvmet_sq_free(struct percpu_ref *ref)
 681 {
 682         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 683
 684         complete(&sq->free_done);
 685 }
 686
 687 int nvmet_sq_init(struct nvmet_sq *sq)
 688 {
 689         int ret;
 690
 691         ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
 692         if (ret) {
 693                 pr_err("percpu_ref init failed!\n");
 694                 return ret;
 695         }
 696         init_completion(&sq->free_done);
 697         init_completion(&sq->confirm_done);
 698
 699         return 0;
 700 }
 701 EXPORT_SYMBOL_GPL(nvmet_sq_init);
 702
 703 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
 704                 struct nvmet_ns *ns)
 705 {
 706         enum nvme_ana_state state = port->ana_state[ns->anagrpid];
 707
 708         if (unlikely(state == NVME_ANA_INACCESSIBLE))
 709                 return NVME_SC_ANA_INACCESSIBLE;
 710         if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
 711                 return NVME_SC_ANA_PERSISTENT_LOSS;
 712         if (unlikely(state == NVME_ANA_CHANGE))
 713                 return NVME_SC_ANA_TRANSITION;
 714         return 0;
 715 }
 716
 717 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
 718 {
 719         if (unlikely(req->ns->readonly)) {
 720                 switch (req->cmd->common.opcode) {
 721                 case nvme_cmd_read:
 722                 case nvme_cmd_flush:
 723                         break;
 724                 default:
 725                         return NVME_SC_NS_WRITE_PROTECTED;
 726                 }
 727         }
 728
 729         return 0;
 730 }
 731
 732 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
 733 {
 734         struct nvme_command *cmd = req->cmd;
 735         u16 ret;
 736
 737         ret = nvmet_check_ctrl_status(req, cmd);
 738         if (unlikely(ret))
 739                 return ret;
 740
 741         req->ns = nvmet_find_namespace(req->sq->ctrl, cmd->rw.nsid);
 742         if (unlikely(!req->ns))
 743                 return NVME_SC_INVALID_NS | NVME_SC_DNR;
 744         ret = nvmet_check_ana_state(req->port, req->ns);
 745         if (unlikely(ret))
 746                 return ret;
 747         ret = nvmet_io_cmd_check_access(req);
 748         if (unlikely(ret))
 749                 return ret;
 750
 751         if (req->ns->file)
 752                 return nvmet_file_parse_io_cmd(req);
 753         else
 754                 return nvmet_bdev_parse_io_cmd(req);
 755 }
 756
 757 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
 758                 struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
 759 {
 760         u8 flags = req->cmd->common.flags;
 761         u16 status;
 762
 763         req->cq = cq;
 764         req->sq = sq;
 765         req->ops = ops;
 766         req->sg = NULL;
 767         req->sg_cnt = 0;
 768         req->transfer_len = 0;
 769         req->rsp->status = 0;
 770         req->rsp->sq_head = 0;
 771         req->ns = NULL;
 772
 773         /* no support for fused commands yet */
 774         if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
 775                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 776                 goto fail;
 777         }
 778
 779         /*
 780          * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
 781          * contains an address of a single contiguous physical buffer that is
 782          * byte aligned.
 783          */
 784         if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
 785                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 786                 goto fail;
 787         }
 788
 789         if (unlikely(!req->sq->ctrl))
 790                 /* will return an error for any Non-connect command: */
 791                 status = nvmet_parse_connect_cmd(req);
 792         else if (likely(req->sq->qid != 0))
 793                 status = nvmet_parse_io_cmd(req);
 794         else if (req->cmd->common.opcode == nvme_fabrics_command)
 795                 status = nvmet_parse_fabrics_cmd(req);
 796         else if (req->sq->ctrl->subsys->type == NVME_NQN_DISC)
 797                 status = nvmet_parse_discovery_cmd(req);
 798         else
 799                 status = nvmet_parse_admin_cmd(req);
 800
 801         if (status)
 802                 goto fail;
 803
 804         if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
 805                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 806                 goto fail;
 807         }
 808
 809         if (sq->ctrl)
 810                 sq->ctrl->cmd_seen = true;
 811
 812         return true;
 813
 814 fail:
 815         __nvmet_req_complete(req, status);
 816         return false;
 817 }
 818 EXPORT_SYMBOL_GPL(nvmet_req_init);
 819
 820 void nvmet_req_uninit(struct nvmet_req *req)
 821 {
 822         percpu_ref_put(&req->sq->ref);
 823         if (req->ns)
 824                 nvmet_put_namespace(req->ns);
 825 }
 826 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
 827
 828 void nvmet_req_execute(struct nvmet_req *req)
 829 {
 830         if (unlikely(req->data_len != req->transfer_len))
 831                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
 832         else
 833                 req->execute(req);
 834 }
 835 EXPORT_SYMBOL_GPL(nvmet_req_execute);
 836
 837 int nvmet_req_alloc_sgl(struct nvmet_req *req)
 838 {
 839         struct pci_dev *p2p_dev = NULL;
 840
 841         if (IS_ENABLED(CONFIG_PCI_P2PDMA)) {
 842                 if (req->sq->ctrl && req->ns)
 843                         p2p_dev = radix_tree_lookup(&req->sq->ctrl->p2p_ns_map,
 844                                                     req->ns->nsid);
 845
 846                 req->p2p_dev = NULL;
 847                 if (req->sq->qid && p2p_dev) {
 848                         req->sg = pci_p2pmem_alloc_sgl(p2p_dev, &req->sg_cnt,
 849                                                        req->transfer_len);
 850                         if (req->sg) {
 851                                 req->p2p_dev = p2p_dev;
 852                                 return 0;
 853                         }
 854                 }
 855
 856                 /*
 857                  * If no P2P memory was available we fallback to using
 858                  * regular memory
 859                  */
 860         }
 861
 862         req->sg = sgl_alloc(req->transfer_len, GFP_KERNEL, &req->sg_cnt);
 863         if (!req->sg)
 864                 return -ENOMEM;
 865
 866         return 0;
 867 }
 868 EXPORT_SYMBOL_GPL(nvmet_req_alloc_sgl);
 869
 870 void nvmet_req_free_sgl(struct nvmet_req *req)
 871 {
 872         if (req->p2p_dev)
 873                 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
 874         else
 875                 sgl_free(req->sg);
 876
 877         req->sg = NULL;
 878         req->sg_cnt = 0;
 879 }
 880 EXPORT_SYMBOL_GPL(nvmet_req_free_sgl);
 881
 882 static inline bool nvmet_cc_en(u32 cc)
 883 {
 884         return (cc >> NVME_CC_EN_SHIFT) & 0x1;
 885 }
 886
 887 static inline u8 nvmet_cc_css(u32 cc)
 888 {
 889         return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
 890 }
 891
 892 static inline u8 nvmet_cc_mps(u32 cc)
 893 {
 894         return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
 895 }
 896
 897 static inline u8 nvmet_cc_ams(u32 cc)
 898 {
 899         return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
 900 }
 901
 902 static inline u8 nvmet_cc_shn(u32 cc)
 903 {
 904         return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
 905 }
 906
 907 static inline u8 nvmet_cc_iosqes(u32 cc)
 908 {
 909         return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
 910 }
 911
 912 static inline u8 nvmet_cc_iocqes(u32 cc)
 913 {
 914         return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
 915 }
 916
 917 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
 918 {
 919         lockdep_assert_held(&ctrl->lock);
 920
 921         if (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
 922             nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES ||
 923             nvmet_cc_mps(ctrl->cc) != 0 ||
 924             nvmet_cc_ams(ctrl->cc) != 0 ||
 925             nvmet_cc_css(ctrl->cc) != 0) {
 926                 ctrl->csts = NVME_CSTS_CFS;
 927                 return;
 928         }
 929
 930         ctrl->csts = NVME_CSTS_RDY;
 931
 932         /*
 933          * Controllers that are not yet enabled should not really enforce the
 934          * keep alive timeout, but we still want to track a timeout and cleanup
 935          * in case a host died before it enabled the controller.  Hence, simply
 936          * reset the keep alive timer when the controller is enabled.
 937          */
 938         mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
 939 }
 940
 941 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
 942 {
 943         lockdep_assert_held(&ctrl->lock);
 944
 945         /* XXX: tear down queues? */
 946         ctrl->csts &= ~NVME_CSTS_RDY;
 947         ctrl->cc = 0;
 948 }
 949
 950 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
 951 {
 952         u32 old;
 953
 954         mutex_lock(&ctrl->lock);
 955         old = ctrl->cc;
 956         ctrl->cc = new;
 957
 958         if (nvmet_cc_en(new) && !nvmet_cc_en(old))
 959                 nvmet_start_ctrl(ctrl);
 960         if (!nvmet_cc_en(new) && nvmet_cc_en(old))
 961                 nvmet_clear_ctrl(ctrl);
 962         if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
 963                 nvmet_clear_ctrl(ctrl);
 964                 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
 965         }
 966         if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
 967                 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
 968         mutex_unlock(&ctrl->lock);
 969 }
 970
 971 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
 972 {
 973         /* command sets supported: NVMe command set: */
 974         ctrl->cap = (1ULL << 37);
 975         /* CC.EN timeout in 500msec units: */
 976         ctrl->cap |= (15ULL << 24);
 977         /* maximum queue entries supported: */
 978         ctrl->cap |= NVMET_QUEUE_SIZE - 1;
 979 }
 980
 981 u16 nvmet_ctrl_find_get(const char *subsysnqn, const char *hostnqn, u16 cntlid,
 982                 struct nvmet_req *req, struct nvmet_ctrl **ret)
 983 {
 984         struct nvmet_subsys *subsys;
 985         struct nvmet_ctrl *ctrl;
 986         u16 status = 0;
 987
 988         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
 989         if (!subsys) {
 990                 pr_warn("connect request for invalid subsystem %s!\n",
 991                         subsysnqn);
 992                 req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
 993                 return NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
 994         }
 995
 996         mutex_lock(&subsys->lock);
 997         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 998                 if (ctrl->cntlid == cntlid) {
 999                         if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
1000                                 pr_warn("hostnqn mismatch.\n");
1001                                 continue;
1002                         }
1003                         if (!kref_get_unless_zero(&ctrl->ref))
1004                                 continue;
1005
1006                         *ret = ctrl;
1007                         goto out;
1008                 }
1009         }
1010
1011         pr_warn("could not find controller %d for subsys %s / host %s\n",
1012                 cntlid, subsysnqn, hostnqn);
1013         req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1014         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1015
1016 out:
1017         mutex_unlock(&subsys->lock);
1018         nvmet_subsys_put(subsys);
1019         return status;
1020 }
1021
1022 u16 nvmet_check_ctrl_status(struct nvmet_req *req, struct nvme_command *cmd)
1023 {
1024         if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1025                 pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1026                        cmd->common.opcode, req->sq->qid);
1027                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1028         }
1029
1030         if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1031                 pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1032                        cmd->common.opcode, req->sq->qid);
1033                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1034         }
1035         return 0;
1036 }
1037
1038 bool nvmet_host_allowed(struct nvmet_subsys *subsys, const char *hostnqn)
1039 {
1040         struct nvmet_host_link *p;
1041
1042         lockdep_assert_held(&nvmet_config_sem);
1043
1044         if (subsys->allow_any_host)
1045                 return true;
1046
1047         if (subsys->type == NVME_NQN_DISC) /* allow all access to disc subsys */
1048                 return true;
1049
1050         list_for_each_entry(p, &subsys->hosts, entry) {
1051                 if (!strcmp(nvmet_host_name(p->host), hostnqn))
1052                         return true;
1053         }
1054
1055         return false;
1056 }
1057
1058 /*
1059  * Note: ctrl->subsys->lock should be held when calling this function
1060  */
1061 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1062                 struct nvmet_req *req)
1063 {
1064         struct nvmet_ns *ns;
1065
1066         if (!req->p2p_client)
1067                 return;
1068
1069         ctrl->p2p_client = get_device(req->p2p_client);
1070
1071         list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link)
1072                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1073 }
1074
1075 /*
1076  * Note: ctrl->subsys->lock should be held when calling this function
1077  */
1078 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1079 {
1080         struct radix_tree_iter iter;
1081         void __rcu **slot;
1082
1083         radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1084                 pci_dev_put(radix_tree_deref_slot(slot));
1085
1086         put_device(ctrl->p2p_client);
1087 }
1088
1089 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1090                 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
1091 {
1092         struct nvmet_subsys *subsys;
1093         struct nvmet_ctrl *ctrl;
1094         int ret;
1095         u16 status;
1096
1097         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1098         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1099         if (!subsys) {
1100                 pr_warn("connect request for invalid subsystem %s!\n",
1101                         subsysnqn);
1102                 req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1103                 goto out;
1104         }
1105
1106         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1107         down_read(&nvmet_config_sem);
1108         if (!nvmet_host_allowed(subsys, hostnqn)) {
1109                 pr_info("connect by host %s for subsystem %s not allowed\n",
1110                         hostnqn, subsysnqn);
1111                 req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1112                 up_read(&nvmet_config_sem);
1113                 status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
1114                 goto out_put_subsystem;
1115         }
1116         up_read(&nvmet_config_sem);
1117
1118         status = NVME_SC_INTERNAL;
1119         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1120         if (!ctrl)
1121                 goto out_put_subsystem;
1122         mutex_init(&ctrl->lock);
1123
1124         nvmet_init_cap(ctrl);
1125
1126         ctrl->port = req->port;
1127
1128         INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
1129         INIT_LIST_HEAD(&ctrl->async_events);
1130         INIT_RADIX_TREE(&ctrl->p2p_ns_map, GFP_KERNEL);
1131
1132         memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1133         memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1134
1135         kref_init(&ctrl->ref);
1136         ctrl->subsys = subsys;
1137         WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1138
1139         ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1140                         sizeof(__le32), GFP_KERNEL);
1141         if (!ctrl->changed_ns_list)
1142                 goto out_free_ctrl;
1143
1144         ctrl->cqs = kcalloc(subsys->max_qid + 1,
1145                         sizeof(struct nvmet_cq *),
1146                         GFP_KERNEL);
1147         if (!ctrl->cqs)
1148                 goto out_free_changed_ns_list;
1149
1150         ctrl->sqs = kcalloc(subsys->max_qid + 1,
1151                         sizeof(struct nvmet_sq *),
1152                         GFP_KERNEL);
1153         if (!ctrl->sqs)
1154                 goto out_free_cqs;
1155
1156         ret = ida_simple_get(&cntlid_ida,
1157                              NVME_CNTLID_MIN, NVME_CNTLID_MAX,
1158                              GFP_KERNEL);
1159         if (ret < 0) {
1160                 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
1161                 goto out_free_sqs;
1162         }
1163         ctrl->cntlid = ret;
1164
1165         ctrl->ops = req->ops;
1166
1167         /*
1168          * Discovery controllers may use some arbitrary high value
1169          * in order to cleanup stale discovery sessions
1170          */
1171         if ((ctrl->subsys->type == NVME_NQN_DISC) && !kato)
1172                 kato = NVMET_DISC_KATO_MS;
1173
1174         /* keep-alive timeout in seconds */
1175         ctrl->kato = DIV_ROUND_UP(kato, 1000);
1176
1177         nvmet_start_keep_alive_timer(ctrl);
1178
1179         mutex_lock(&subsys->lock);
1180         list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1181         nvmet_setup_p2p_ns_map(ctrl, req);
1182         mutex_unlock(&subsys->lock);
1183
1184         *ctrlp = ctrl;
1185         return 0;
1186
1187 out_free_sqs:
1188         kfree(ctrl->sqs);
1189 out_free_cqs:
1190         kfree(ctrl->cqs);
1191 out_free_changed_ns_list:
1192         kfree(ctrl->changed_ns_list);
1193 out_free_ctrl:
1194         kfree(ctrl);
1195 out_put_subsystem:
1196         nvmet_subsys_put(subsys);
1197 out:
1198         return status;
1199 }
1200
1201 static void nvmet_ctrl_free(struct kref *ref)
1202 {
1203         struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1204         struct nvmet_subsys *subsys = ctrl->subsys;
1205
1206         mutex_lock(&subsys->lock);
1207         nvmet_release_p2p_ns_map(ctrl);
1208         list_del(&ctrl->subsys_entry);
1209         mutex_unlock(&subsys->lock);
1210
1211         nvmet_stop_keep_alive_timer(ctrl);
1212
1213         flush_work(&ctrl->async_event_work);
1214         cancel_work_sync(&ctrl->fatal_err_work);
1215
1216         ida_simple_remove(&cntlid_ida, ctrl->cntlid);
1217
1218         kfree(ctrl->sqs);
1219         kfree(ctrl->cqs);
1220         kfree(ctrl->changed_ns_list);
1221         kfree(ctrl);
1222
1223         nvmet_subsys_put(subsys);
1224 }
1225
1226 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1227 {
1228         kref_put(&ctrl->ref, nvmet_ctrl_free);
1229 }
1230
1231 static void nvmet_fatal_error_handler(struct work_struct *work)
1232 {
1233         struct nvmet_ctrl *ctrl =
1234                         container_of(work, struct nvmet_ctrl, fatal_err_work);
1235
1236         pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1237         ctrl->ops->delete_ctrl(ctrl);
1238 }
1239
1240 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1241 {
1242         mutex_lock(&ctrl->lock);
1243         if (!(ctrl->csts & NVME_CSTS_CFS)) {
1244                 ctrl->csts |= NVME_CSTS_CFS;
1245                 INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1246                 schedule_work(&ctrl->fatal_err_work);
1247         }
1248         mutex_unlock(&ctrl->lock);
1249 }
1250 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1251
1252 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1253                 const char *subsysnqn)
1254 {
1255         struct nvmet_subsys_link *p;
1256
1257         if (!port)
1258                 return NULL;
1259
1260         if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1261                 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1262                         return NULL;
1263                 return nvmet_disc_subsys;
1264         }
1265
1266         down_read(&nvmet_config_sem);
1267         list_for_each_entry(p, &port->subsystems, entry) {
1268                 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1269                                 NVMF_NQN_SIZE)) {
1270                         if (!kref_get_unless_zero(&p->subsys->ref))
1271                                 break;
1272                         up_read(&nvmet_config_sem);
1273                         return p->subsys;
1274                 }
1275         }
1276         up_read(&nvmet_config_sem);
1277         return NULL;
1278 }
1279
1280 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1281                 enum nvme_subsys_type type)
1282 {
1283         struct nvmet_subsys *subsys;
1284
1285         subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1286         if (!subsys)
1287                 return NULL;
1288
1289         subsys->ver = NVME_VS(1, 3, 0); /* NVMe 1.3.0 */
1290         /* generate a random serial number as our controllers are ephemeral: */
1291         get_random_bytes(&subsys->serial, sizeof(subsys->serial));
1292
1293         switch (type) {
1294         case NVME_NQN_NVME:
1295                 subsys->max_qid = NVMET_NR_QUEUES;
1296                 break;
1297         case NVME_NQN_DISC:
1298                 subsys->max_qid = 0;
1299                 break;
1300         default:
1301                 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1302                 kfree(subsys);
1303                 return NULL;
1304         }
1305         subsys->type = type;
1306         subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1307                         GFP_KERNEL);
1308         if (!subsys->subsysnqn) {
1309                 kfree(subsys);
1310                 return NULL;
1311         }
1312
1313         kref_init(&subsys->ref);
1314
1315         mutex_init(&subsys->lock);
1316         INIT_LIST_HEAD(&subsys->namespaces);
1317         INIT_LIST_HEAD(&subsys->ctrls);
1318         INIT_LIST_HEAD(&subsys->hosts);
1319
1320         return subsys;
1321 }
1322
1323 static void nvmet_subsys_free(struct kref *ref)
1324 {
1325         struct nvmet_subsys *subsys =
1326                 container_of(ref, struct nvmet_subsys, ref);
1327
1328         WARN_ON_ONCE(!list_empty(&subsys->namespaces));
1329
1330         kfree(subsys->subsysnqn);
1331         kfree(subsys);
1332 }
1333
1334 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1335 {
1336         struct nvmet_ctrl *ctrl;
1337
1338         mutex_lock(&subsys->lock);
1339         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1340                 ctrl->ops->delete_ctrl(ctrl);
1341         mutex_unlock(&subsys->lock);
1342 }
1343
1344 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1345 {
1346         kref_put(&subsys->ref, nvmet_subsys_free);
1347 }
1348
1349 static int __init nvmet_init(void)
1350 {
1351         int error;
1352
1353         nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1354
1355         buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1356                         WQ_MEM_RECLAIM, 0);
1357         if (!buffered_io_wq) {
1358                 error = -ENOMEM;
1359                 goto out;
1360         }
1361
1362         error = nvmet_init_discovery();
1363         if (error)
1364                 goto out_free_work_queue;
1365
1366         error = nvmet_init_configfs();
1367         if (error)
1368                 goto out_exit_discovery;
1369         return 0;
1370
1371 out_exit_discovery:
1372         nvmet_exit_discovery();
1373 out_free_work_queue:
1374         destroy_workqueue(buffered_io_wq);
1375 out:
1376         return error;
1377 }
1378
1379 static void __exit nvmet_exit(void)
1380 {
1381         nvmet_exit_configfs();
1382         nvmet_exit_discovery();
1383         ida_destroy(&cntlid_ida);
1384         destroy_workqueue(buffered_io_wq);
1385
1386         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1387         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1388 }
1389
1390 module_init(nvmet_init);
1391 module_exit(nvmet_exit);
1392
1393 MODULE_LICENSE("GPL v2");