drivers/nvme/target/core.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Common code for the NVMe target.
   4  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
   5  */
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7 #include <linux/module.h>
   8 #include <linux/random.h>
   9 #include <linux/rculist.h>
  10 #include <linux/pci-p2pdma.h>
  11
  12 #include "nvmet.h"
  13
  14 struct workqueue_struct *buffered_io_wq;
  15 static const struct nvmet_fabrics_ops *nvmet_transports[NVMF_TRTYPE_MAX];
  16 static DEFINE_IDA(cntlid_ida);
  17
  18 /*
  19  * This read/write semaphore is used to synchronize access to configuration
  20  * information on a target system that will result in discovery log page
  21  * information change for at least one host.
  22  * The full list of resources to protected by this semaphore is:
  23  *
  24  *  - subsystems list
  25  *  - per-subsystem allowed hosts list
  26  *  - allow_any_host subsystem attribute
  27  *  - nvmet_genctr
  28  *  - the nvmet_transports array
  29  *
  30  * When updating any of those lists/structures write lock should be obtained,
  31  * while when reading (popolating discovery log page or checking host-subsystem
  32  * link) read lock is obtained to allow concurrent reads.
  33  */
  34 DECLARE_RWSEM(nvmet_config_sem);
  35
  36 u32 nvmet_ana_group_enabled[NVMET_MAX_ANAGRPS + 1];
  37 u64 nvmet_ana_chgcnt;
  38 DECLARE_RWSEM(nvmet_ana_sem);
  39
  40 inline u16 errno_to_nvme_status(struct nvmet_req *req, int errno)
  41 {
  42         u16 status;
  43
  44         switch (errno) {
  45         case -ENOSPC:
  46                 req->error_loc = offsetof(struct nvme_rw_command, length);
  47                 status = NVME_SC_CAP_EXCEEDED | NVME_SC_DNR;
  48                 break;
  49         case -EREMOTEIO:
  50                 req->error_loc = offsetof(struct nvme_rw_command, slba);
  51                 status = NVME_SC_LBA_RANGE | NVME_SC_DNR;
  52                 break;
  53         case -EOPNOTSUPP:
  54                 req->error_loc = offsetof(struct nvme_common_command, opcode);
  55                 switch (req->cmd->common.opcode) {
  56                 case nvme_cmd_dsm:
  57                 case nvme_cmd_write_zeroes:
  58                         status = NVME_SC_ONCS_NOT_SUPPORTED | NVME_SC_DNR;
  59                         break;
  60                 default:
  61                         status = NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
  62                 }
  63                 break;
  64         case -ENODATA:
  65                 req->error_loc = offsetof(struct nvme_rw_command, nsid);
  66                 status = NVME_SC_ACCESS_DENIED;
  67                 break;
  68         case -EIO:
  69                 /* FALLTHRU */
  70         default:
  71                 req->error_loc = offsetof(struct nvme_common_command, opcode);
  72                 status = NVME_SC_INTERNAL | NVME_SC_DNR;
  73         }
  74
  75         return status;
  76 }
  77
  78 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
  79                 const char *subsysnqn);
  80
  81 u16 nvmet_copy_to_sgl(struct nvmet_req *req, off_t off, const void *buf,
  82                 size_t len)
  83 {
  84         if (sg_pcopy_from_buffer(req->sg, req->sg_cnt, buf, len, off) != len) {
  85                 req->error_loc = offsetof(struct nvme_common_command, dptr);
  86                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
  87         }
  88         return 0;
  89 }
  90
  91 u16 nvmet_copy_from_sgl(struct nvmet_req *req, off_t off, void *buf, size_t len)
  92 {
  93         if (sg_pcopy_to_buffer(req->sg, req->sg_cnt, buf, len, off) != len) {
  94                 req->error_loc = offsetof(struct nvme_common_command, dptr);
  95                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
  96         }
  97         return 0;
  98 }
  99
 100 u16 nvmet_zero_sgl(struct nvmet_req *req, off_t off, size_t len)
 101 {
 102         if (sg_zero_buffer(req->sg, req->sg_cnt, len, off) != len) {
 103                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 104                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
 105         }
 106         return 0;
 107 }
 108
 109 static unsigned int nvmet_max_nsid(struct nvmet_subsys *subsys)
 110 {
 111         struct nvmet_ns *ns;
 112
 113         if (list_empty(&subsys->namespaces))
 114                 return 0;
 115
 116         ns = list_last_entry(&subsys->namespaces, struct nvmet_ns, dev_link);
 117         return ns->nsid;
 118 }
 119
 120 static u32 nvmet_async_event_result(struct nvmet_async_event *aen)
 121 {
 122         return aen->event_type | (aen->event_info << 8) | (aen->log_page << 16);
 123 }
 124
 125 static void nvmet_async_events_free(struct nvmet_ctrl *ctrl)
 126 {
 127         struct nvmet_req *req;
 128
 129         while (1) {
 130                 mutex_lock(&ctrl->lock);
 131                 if (!ctrl->nr_async_event_cmds) {
 132                         mutex_unlock(&ctrl->lock);
 133                         return;
 134                 }
 135
 136                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
 137                 mutex_unlock(&ctrl->lock);
 138                 nvmet_req_complete(req, NVME_SC_INTERNAL | NVME_SC_DNR);
 139         }
 140 }
 141
 142 static void nvmet_async_event_work(struct work_struct *work)
 143 {
 144         struct nvmet_ctrl *ctrl =
 145                 container_of(work, struct nvmet_ctrl, async_event_work);
 146         struct nvmet_async_event *aen;
 147         struct nvmet_req *req;
 148
 149         while (1) {
 150                 mutex_lock(&ctrl->lock);
 151                 aen = list_first_entry_or_null(&ctrl->async_events,
 152                                 struct nvmet_async_event, entry);
 153                 if (!aen || !ctrl->nr_async_event_cmds) {
 154                         mutex_unlock(&ctrl->lock);
 155                         return;
 156                 }
 157
 158                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
 159                 nvmet_set_result(req, nvmet_async_event_result(aen));
 160
 161                 list_del(&aen->entry);
 162                 kfree(aen);
 163
 164                 mutex_unlock(&ctrl->lock);
 165                 nvmet_req_complete(req, 0);
 166         }
 167 }
 168
 169 void nvmet_add_async_event(struct nvmet_ctrl *ctrl, u8 event_type,
 170                 u8 event_info, u8 log_page)
 171 {
 172         struct nvmet_async_event *aen;
 173
 174         aen = kmalloc(sizeof(*aen), GFP_KERNEL);
 175         if (!aen)
 176                 return;
 177
 178         aen->event_type = event_type;
 179         aen->event_info = event_info;
 180         aen->log_page = log_page;
 181
 182         mutex_lock(&ctrl->lock);
 183         list_add_tail(&aen->entry, &ctrl->async_events);
 184         mutex_unlock(&ctrl->lock);
 185
 186         schedule_work(&ctrl->async_event_work);
 187 }
 188
 189 static void nvmet_add_to_changed_ns_log(struct nvmet_ctrl *ctrl, __le32 nsid)
 190 {
 191         u32 i;
 192
 193         mutex_lock(&ctrl->lock);
 194         if (ctrl->nr_changed_ns > NVME_MAX_CHANGED_NAMESPACES)
 195                 goto out_unlock;
 196
 197         for (i = 0; i < ctrl->nr_changed_ns; i++) {
 198                 if (ctrl->changed_ns_list[i] == nsid)
 199                         goto out_unlock;
 200         }
 201
 202         if (ctrl->nr_changed_ns == NVME_MAX_CHANGED_NAMESPACES) {
 203                 ctrl->changed_ns_list[0] = cpu_to_le32(0xffffffff);
 204                 ctrl->nr_changed_ns = U32_MAX;
 205                 goto out_unlock;
 206         }
 207
 208         ctrl->changed_ns_list[ctrl->nr_changed_ns++] = nsid;
 209 out_unlock:
 210         mutex_unlock(&ctrl->lock);
 211 }
 212
 213 void nvmet_ns_changed(struct nvmet_subsys *subsys, u32 nsid)
 214 {
 215         struct nvmet_ctrl *ctrl;
 216
 217         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 218                 nvmet_add_to_changed_ns_log(ctrl, cpu_to_le32(nsid));
 219                 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_NS_ATTR))
 220                         continue;
 221                 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
 222                                 NVME_AER_NOTICE_NS_CHANGED,
 223                                 NVME_LOG_CHANGED_NS);
 224         }
 225 }
 226
 227 void nvmet_send_ana_event(struct nvmet_subsys *subsys,
 228                 struct nvmet_port *port)
 229 {
 230         struct nvmet_ctrl *ctrl;
 231
 232         mutex_lock(&subsys->lock);
 233         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 234                 if (port && ctrl->port != port)
 235                         continue;
 236                 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_ANA_CHANGE))
 237                         continue;
 238                 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
 239                                 NVME_AER_NOTICE_ANA, NVME_LOG_ANA);
 240         }
 241         mutex_unlock(&subsys->lock);
 242 }
 243
 244 void nvmet_port_send_ana_event(struct nvmet_port *port)
 245 {
 246         struct nvmet_subsys_link *p;
 247
 248         down_read(&nvmet_config_sem);
 249         list_for_each_entry(p, &port->subsystems, entry)
 250                 nvmet_send_ana_event(p->subsys, port);
 251         up_read(&nvmet_config_sem);
 252 }
 253
 254 int nvmet_register_transport(const struct nvmet_fabrics_ops *ops)
 255 {
 256         int ret = 0;
 257
 258         down_write(&nvmet_config_sem);
 259         if (nvmet_transports[ops->type])
 260                 ret = -EINVAL;
 261         else
 262                 nvmet_transports[ops->type] = ops;
 263         up_write(&nvmet_config_sem);
 264
 265         return ret;
 266 }
 267 EXPORT_SYMBOL_GPL(nvmet_register_transport);
 268
 269 void nvmet_unregister_transport(const struct nvmet_fabrics_ops *ops)
 270 {
 271         down_write(&nvmet_config_sem);
 272         nvmet_transports[ops->type] = NULL;
 273         up_write(&nvmet_config_sem);
 274 }
 275 EXPORT_SYMBOL_GPL(nvmet_unregister_transport);
 276
 277 int nvmet_enable_port(struct nvmet_port *port)
 278 {
 279         const struct nvmet_fabrics_ops *ops;
 280         int ret;
 281
 282         lockdep_assert_held(&nvmet_config_sem);
 283
 284         ops = nvmet_transports[port->disc_addr.trtype];
 285         if (!ops) {
 286                 up_write(&nvmet_config_sem);
 287                 request_module("nvmet-transport-%d", port->disc_addr.trtype);
 288                 down_write(&nvmet_config_sem);
 289                 ops = nvmet_transports[port->disc_addr.trtype];
 290                 if (!ops) {
 291                         pr_err("transport type %d not supported\n",
 292                                 port->disc_addr.trtype);
 293                         return -EINVAL;
 294                 }
 295         }
 296
 297         if (!try_module_get(ops->owner))
 298                 return -EINVAL;
 299
 300         ret = ops->add_port(port);
 301         if (ret) {
 302                 module_put(ops->owner);
 303                 return ret;
 304         }
 305
 306         /* If the transport didn't set inline_data_size, then disable it. */
 307         if (port->inline_data_size < 0)
 308                 port->inline_data_size = 0;
 309
 310         port->enabled = true;
 311         return 0;
 312 }
 313
 314 void nvmet_disable_port(struct nvmet_port *port)
 315 {
 316         const struct nvmet_fabrics_ops *ops;
 317
 318         lockdep_assert_held(&nvmet_config_sem);
 319
 320         port->enabled = false;
 321
 322         ops = nvmet_transports[port->disc_addr.trtype];
 323         ops->remove_port(port);
 324         module_put(ops->owner);
 325 }
 326
 327 static void nvmet_keep_alive_timer(struct work_struct *work)
 328 {
 329         struct nvmet_ctrl *ctrl = container_of(to_delayed_work(work),
 330                         struct nvmet_ctrl, ka_work);
 331         bool cmd_seen = ctrl->cmd_seen;
 332
 333         ctrl->cmd_seen = false;
 334         if (cmd_seen) {
 335                 pr_debug("ctrl %d reschedule traffic based keep-alive timer\n",
 336                         ctrl->cntlid);
 337                 schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
 338                 return;
 339         }
 340
 341         pr_err("ctrl %d keep-alive timer (%d seconds) expired!\n",
 342                 ctrl->cntlid, ctrl->kato);
 343
 344         nvmet_ctrl_fatal_error(ctrl);
 345 }
 346
 347 static void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl)
 348 {
 349         pr_debug("ctrl %d start keep-alive timer for %d secs\n",
 350                 ctrl->cntlid, ctrl->kato);
 351
 352         INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
 353         schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
 354 }
 355
 356 static void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
 357 {
 358         pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
 359
 360         cancel_delayed_work_sync(&ctrl->ka_work);
 361 }
 362
 363 static struct nvmet_ns *__nvmet_find_namespace(struct nvmet_ctrl *ctrl,
 364                 __le32 nsid)
 365 {
 366         struct nvmet_ns *ns;
 367
 368         list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link) {
 369                 if (ns->nsid == le32_to_cpu(nsid))
 370                         return ns;
 371         }
 372
 373         return NULL;
 374 }
 375
 376 struct nvmet_ns *nvmet_find_namespace(struct nvmet_ctrl *ctrl, __le32 nsid)
 377 {
 378         struct nvmet_ns *ns;
 379
 380         rcu_read_lock();
 381         ns = __nvmet_find_namespace(ctrl, nsid);
 382         if (ns)
 383                 percpu_ref_get(&ns->ref);
 384         rcu_read_unlock();
 385
 386         return ns;
 387 }
 388
 389 static void nvmet_destroy_namespace(struct percpu_ref *ref)
 390 {
 391         struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
 392
 393         complete(&ns->disable_done);
 394 }
 395
 396 void nvmet_put_namespace(struct nvmet_ns *ns)
 397 {
 398         percpu_ref_put(&ns->ref);
 399 }
 400
 401 static void nvmet_ns_dev_disable(struct nvmet_ns *ns)
 402 {
 403         nvmet_bdev_ns_disable(ns);
 404         nvmet_file_ns_disable(ns);
 405 }
 406
 407 static int nvmet_p2pmem_ns_enable(struct nvmet_ns *ns)
 408 {
 409         int ret;
 410         struct pci_dev *p2p_dev;
 411
 412         if (!ns->use_p2pmem)
 413                 return 0;
 414
 415         if (!ns->bdev) {
 416                 pr_err("peer-to-peer DMA is not supported by non-block device namespaces\n");
 417                 return -EINVAL;
 418         }
 419
 420         if (!blk_queue_pci_p2pdma(ns->bdev->bd_queue)) {
 421                 pr_err("peer-to-peer DMA is not supported by the driver of %s\n",
 422                        ns->device_path);
 423                 return -EINVAL;
 424         }
 425
 426         if (ns->p2p_dev) {
 427                 ret = pci_p2pdma_distance(ns->p2p_dev, nvmet_ns_dev(ns), true);
 428                 if (ret < 0)
 429                         return -EINVAL;
 430         } else {
 431                 /*
 432                  * Right now we just check that there is p2pmem available so
 433                  * we can report an error to the user right away if there
 434                  * is not. We'll find the actual device to use once we
 435                  * setup the controller when the port's device is available.
 436                  */
 437
 438                 p2p_dev = pci_p2pmem_find(nvmet_ns_dev(ns));
 439                 if (!p2p_dev) {
 440                         pr_err("no peer-to-peer memory is available for %s\n",
 441                                ns->device_path);
 442                         return -EINVAL;
 443                 }
 444
 445                 pci_dev_put(p2p_dev);
 446         }
 447
 448         return 0;
 449 }
 450
 451 /*
 452  * Note: ctrl->subsys->lock should be held when calling this function
 453  */
 454 static void nvmet_p2pmem_ns_add_p2p(struct nvmet_ctrl *ctrl,
 455                                     struct nvmet_ns *ns)
 456 {
 457         struct device *clients[2];
 458         struct pci_dev *p2p_dev;
 459         int ret;
 460
 461         if (!ctrl->p2p_client || !ns->use_p2pmem)
 462                 return;
 463
 464         if (ns->p2p_dev) {
 465                 ret = pci_p2pdma_distance(ns->p2p_dev, ctrl->p2p_client, true);
 466                 if (ret < 0)
 467                         return;
 468
 469                 p2p_dev = pci_dev_get(ns->p2p_dev);
 470         } else {
 471                 clients[0] = ctrl->p2p_client;
 472                 clients[1] = nvmet_ns_dev(ns);
 473
 474                 p2p_dev = pci_p2pmem_find_many(clients, ARRAY_SIZE(clients));
 475                 if (!p2p_dev) {
 476                         pr_err("no peer-to-peer memory is available that's supported by %s and %s\n",
 477                                dev_name(ctrl->p2p_client), ns->device_path);
 478                         return;
 479                 }
 480         }
 481
 482         ret = radix_tree_insert(&ctrl->p2p_ns_map, ns->nsid, p2p_dev);
 483         if (ret < 0)
 484                 pci_dev_put(p2p_dev);
 485
 486         pr_info("using p2pmem on %s for nsid %d\n", pci_name(p2p_dev),
 487                 ns->nsid);
 488 }
 489
 490 int nvmet_ns_enable(struct nvmet_ns *ns)
 491 {
 492         struct nvmet_subsys *subsys = ns->subsys;
 493         struct nvmet_ctrl *ctrl;
 494         int ret;
 495
 496         mutex_lock(&subsys->lock);
 497         ret = -EMFILE;
 498         if (subsys->nr_namespaces == NVMET_MAX_NAMESPACES)
 499                 goto out_unlock;
 500         ret = 0;
 501         if (ns->enabled)
 502                 goto out_unlock;
 503
 504         ret = nvmet_bdev_ns_enable(ns);
 505         if (ret == -ENOTBLK)
 506                 ret = nvmet_file_ns_enable(ns);
 507         if (ret)
 508                 goto out_unlock;
 509
 510         ret = nvmet_p2pmem_ns_enable(ns);
 511         if (ret)
 512                 goto out_unlock;
 513
 514         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 515                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
 516
 517         ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
 518                                 0, GFP_KERNEL);
 519         if (ret)
 520                 goto out_dev_put;
 521
 522         if (ns->nsid > subsys->max_nsid)
 523                 subsys->max_nsid = ns->nsid;
 524
 525         /*
 526          * The namespaces list needs to be sorted to simplify the implementation
 527          * of the Identify Namepace List subcommand.
 528          */
 529         if (list_empty(&subsys->namespaces)) {
 530                 list_add_tail_rcu(&ns->dev_link, &subsys->namespaces);
 531         } else {
 532                 struct nvmet_ns *old;
 533
 534                 list_for_each_entry_rcu(old, &subsys->namespaces, dev_link) {
 535                         BUG_ON(ns->nsid == old->nsid);
 536                         if (ns->nsid < old->nsid)
 537                                 break;
 538                 }
 539
 540                 list_add_tail_rcu(&ns->dev_link, &old->dev_link);
 541         }
 542         subsys->nr_namespaces++;
 543
 544         nvmet_ns_changed(subsys, ns->nsid);
 545         ns->enabled = true;
 546         ret = 0;
 547 out_unlock:
 548         mutex_unlock(&subsys->lock);
 549         return ret;
 550 out_dev_put:
 551         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 552                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 553
 554         nvmet_ns_dev_disable(ns);
 555         goto out_unlock;
 556 }
 557
 558 void nvmet_ns_disable(struct nvmet_ns *ns)
 559 {
 560         struct nvmet_subsys *subsys = ns->subsys;
 561         struct nvmet_ctrl *ctrl;
 562
 563         mutex_lock(&subsys->lock);
 564         if (!ns->enabled)
 565                 goto out_unlock;
 566
 567         ns->enabled = false;
 568         list_del_rcu(&ns->dev_link);
 569         if (ns->nsid == subsys->max_nsid)
 570                 subsys->max_nsid = nvmet_max_nsid(subsys);
 571
 572         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 573                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 574
 575         mutex_unlock(&subsys->lock);
 576
 577         /*
 578          * Now that we removed the namespaces from the lookup list, we
 579          * can kill the per_cpu ref and wait for any remaining references
 580          * to be dropped, as well as a RCU grace period for anyone only
 581          * using the namepace under rcu_read_lock().  Note that we can't
 582          * use call_rcu here as we need to ensure the namespaces have
 583          * been fully destroyed before unloading the module.
 584          */
 585         percpu_ref_kill(&ns->ref);
 586         synchronize_rcu();
 587         wait_for_completion(&ns->disable_done);
 588         percpu_ref_exit(&ns->ref);
 589
 590         mutex_lock(&subsys->lock);
 591
 592         subsys->nr_namespaces--;
 593         nvmet_ns_changed(subsys, ns->nsid);
 594         nvmet_ns_dev_disable(ns);
 595 out_unlock:
 596         mutex_unlock(&subsys->lock);
 597 }
 598
 599 void nvmet_ns_free(struct nvmet_ns *ns)
 600 {
 601         nvmet_ns_disable(ns);
 602
 603         down_write(&nvmet_ana_sem);
 604         nvmet_ana_group_enabled[ns->anagrpid]--;
 605         up_write(&nvmet_ana_sem);
 606
 607         kfree(ns->device_path);
 608         kfree(ns);
 609 }
 610
 611 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
 612 {
 613         struct nvmet_ns *ns;
 614
 615         ns = kzalloc(sizeof(*ns), GFP_KERNEL);
 616         if (!ns)
 617                 return NULL;
 618
 619         INIT_LIST_HEAD(&ns->dev_link);
 620         init_completion(&ns->disable_done);
 621
 622         ns->nsid = nsid;
 623         ns->subsys = subsys;
 624
 625         down_write(&nvmet_ana_sem);
 626         ns->anagrpid = NVMET_DEFAULT_ANA_GRPID;
 627         nvmet_ana_group_enabled[ns->anagrpid]++;
 628         up_write(&nvmet_ana_sem);
 629
 630         uuid_gen(&ns->uuid);
 631         ns->buffered_io = false;
 632
 633         return ns;
 634 }
 635
 636 static void nvmet_update_sq_head(struct nvmet_req *req)
 637 {
 638         if (req->sq->size) {
 639                 u32 old_sqhd, new_sqhd;
 640
 641                 do {
 642                         old_sqhd = req->sq->sqhd;
 643                         new_sqhd = (old_sqhd + 1) % req->sq->size;
 644                 } while (cmpxchg(&req->sq->sqhd, old_sqhd, new_sqhd) !=
 645                                         old_sqhd);
 646         }
 647         req->rsp->sq_head = cpu_to_le16(req->sq->sqhd & 0x0000FFFF);
 648 }
 649
 650 static void nvmet_set_error(struct nvmet_req *req, u16 status)
 651 {
 652         struct nvmet_ctrl *ctrl = req->sq->ctrl;
 653         struct nvme_error_slot *new_error_slot;
 654         unsigned long flags;
 655
 656         req->rsp->status = cpu_to_le16(status << 1);
 657
 658         if (!ctrl || req->error_loc == NVMET_NO_ERROR_LOC)
 659                 return;
 660
 661         spin_lock_irqsave(&ctrl->error_lock, flags);
 662         ctrl->err_counter++;
 663         new_error_slot =
 664                 &ctrl->slots[ctrl->err_counter % NVMET_ERROR_LOG_SLOTS];
 665
 666         new_error_slot->error_count = cpu_to_le64(ctrl->err_counter);
 667         new_error_slot->sqid = cpu_to_le16(req->sq->qid);
 668         new_error_slot->cmdid = cpu_to_le16(req->cmd->common.command_id);
 669         new_error_slot->status_field = cpu_to_le16(status << 1);
 670         new_error_slot->param_error_location = cpu_to_le16(req->error_loc);
 671         new_error_slot->lba = cpu_to_le64(req->error_slba);
 672         new_error_slot->nsid = req->cmd->common.nsid;
 673         spin_unlock_irqrestore(&ctrl->error_lock, flags);
 674
 675         /* set the more bit for this request */
 676         req->rsp->status |= cpu_to_le16(1 << 14);
 677 }
 678
 679 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
 680 {
 681         if (!req->sq->sqhd_disabled)
 682                 nvmet_update_sq_head(req);
 683         req->rsp->sq_id = cpu_to_le16(req->sq->qid);
 684         req->rsp->command_id = req->cmd->common.command_id;
 685
 686         if (unlikely(status))
 687                 nvmet_set_error(req, status);
 688         if (req->ns)
 689                 nvmet_put_namespace(req->ns);
 690         req->ops->queue_response(req);
 691 }
 692
 693 void nvmet_req_complete(struct nvmet_req *req, u16 status)
 694 {
 695         __nvmet_req_complete(req, status);
 696         percpu_ref_put(&req->sq->ref);
 697 }
 698 EXPORT_SYMBOL_GPL(nvmet_req_complete);
 699
 700 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
 701                 u16 qid, u16 size)
 702 {
 703         cq->qid = qid;
 704         cq->size = size;
 705
 706         ctrl->cqs[qid] = cq;
 707 }
 708
 709 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
 710                 u16 qid, u16 size)
 711 {
 712         sq->sqhd = 0;
 713         sq->qid = qid;
 714         sq->size = size;
 715
 716         ctrl->sqs[qid] = sq;
 717 }
 718
 719 static void nvmet_confirm_sq(struct percpu_ref *ref)
 720 {
 721         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 722
 723         complete(&sq->confirm_done);
 724 }
 725
 726 void nvmet_sq_destroy(struct nvmet_sq *sq)
 727 {
 728         /*
 729          * If this is the admin queue, complete all AERs so that our
 730          * queue doesn't have outstanding requests on it.
 731          */
 732         if (sq->ctrl && sq->ctrl->sqs && sq->ctrl->sqs[0] == sq)
 733                 nvmet_async_events_free(sq->ctrl);
 734         percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
 735         wait_for_completion(&sq->confirm_done);
 736         wait_for_completion(&sq->free_done);
 737         percpu_ref_exit(&sq->ref);
 738
 739         if (sq->ctrl) {
 740                 nvmet_ctrl_put(sq->ctrl);
 741                 sq->ctrl = NULL; /* allows reusing the queue later */
 742         }
 743 }
 744 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
 745
 746 static void nvmet_sq_free(struct percpu_ref *ref)
 747 {
 748         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 749
 750         complete(&sq->free_done);
 751 }
 752
 753 int nvmet_sq_init(struct nvmet_sq *sq)
 754 {
 755         int ret;
 756
 757         ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
 758         if (ret) {
 759                 pr_err("percpu_ref init failed!\n");
 760                 return ret;
 761         }
 762         init_completion(&sq->free_done);
 763         init_completion(&sq->confirm_done);
 764
 765         return 0;
 766 }
 767 EXPORT_SYMBOL_GPL(nvmet_sq_init);
 768
 769 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
 770                 struct nvmet_ns *ns)
 771 {
 772         enum nvme_ana_state state = port->ana_state[ns->anagrpid];
 773
 774         if (unlikely(state == NVME_ANA_INACCESSIBLE))
 775                 return NVME_SC_ANA_INACCESSIBLE;
 776         if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
 777                 return NVME_SC_ANA_PERSISTENT_LOSS;
 778         if (unlikely(state == NVME_ANA_CHANGE))
 779                 return NVME_SC_ANA_TRANSITION;
 780         return 0;
 781 }
 782
 783 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
 784 {
 785         if (unlikely(req->ns->readonly)) {
 786                 switch (req->cmd->common.opcode) {
 787                 case nvme_cmd_read:
 788                 case nvme_cmd_flush:
 789                         break;
 790                 default:
 791                         return NVME_SC_NS_WRITE_PROTECTED;
 792                 }
 793         }
 794
 795         return 0;
 796 }
 797
 798 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
 799 {
 800         struct nvme_command *cmd = req->cmd;
 801         u16 ret;
 802
 803         ret = nvmet_check_ctrl_status(req, cmd);
 804         if (unlikely(ret))
 805                 return ret;
 806
 807         req->ns = nvmet_find_namespace(req->sq->ctrl, cmd->rw.nsid);
 808         if (unlikely(!req->ns)) {
 809                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 810                 return NVME_SC_INVALID_NS | NVME_SC_DNR;
 811         }
 812         ret = nvmet_check_ana_state(req->port, req->ns);
 813         if (unlikely(ret)) {
 814                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 815                 return ret;
 816         }
 817         ret = nvmet_io_cmd_check_access(req);
 818         if (unlikely(ret)) {
 819                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 820                 return ret;
 821         }
 822
 823         if (req->ns->file)
 824                 return nvmet_file_parse_io_cmd(req);
 825         else
 826                 return nvmet_bdev_parse_io_cmd(req);
 827 }
 828
 829 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
 830                 struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
 831 {
 832         u8 flags = req->cmd->common.flags;
 833         u16 status;
 834
 835         req->cq = cq;
 836         req->sq = sq;
 837         req->ops = ops;
 838         req->sg = NULL;
 839         req->sg_cnt = 0;
 840         req->transfer_len = 0;
 841         req->rsp->status = 0;
 842         req->rsp->sq_head = 0;
 843         req->ns = NULL;
 844         req->error_loc = NVMET_NO_ERROR_LOC;
 845         req->error_slba = 0;
 846
 847         /* no support for fused commands yet */
 848         if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
 849                 req->error_loc = offsetof(struct nvme_common_command, flags);
 850                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 851                 goto fail;
 852         }
 853
 854         /*
 855          * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
 856          * contains an address of a single contiguous physical buffer that is
 857          * byte aligned.
 858          */
 859         if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
 860                 req->error_loc = offsetof(struct nvme_common_command, flags);
 861                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 862                 goto fail;
 863         }
 864
 865         if (unlikely(!req->sq->ctrl))
 866                 /* will return an error for any Non-connect command: */
 867                 status = nvmet_parse_connect_cmd(req);
 868         else if (likely(req->sq->qid != 0))
 869                 status = nvmet_parse_io_cmd(req);
 870         else if (req->cmd->common.opcode == nvme_fabrics_command)
 871                 status = nvmet_parse_fabrics_cmd(req);
 872         else if (req->sq->ctrl->subsys->type == NVME_NQN_DISC)
 873                 status = nvmet_parse_discovery_cmd(req);
 874         else
 875                 status = nvmet_parse_admin_cmd(req);
 876
 877         if (status)
 878                 goto fail;
 879
 880         if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
 881                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 882                 goto fail;
 883         }
 884
 885         if (sq->ctrl)
 886                 sq->ctrl->cmd_seen = true;
 887
 888         return true;
 889
 890 fail:
 891         __nvmet_req_complete(req, status);
 892         return false;
 893 }
 894 EXPORT_SYMBOL_GPL(nvmet_req_init);
 895
 896 void nvmet_req_uninit(struct nvmet_req *req)
 897 {
 898         percpu_ref_put(&req->sq->ref);
 899         if (req->ns)
 900                 nvmet_put_namespace(req->ns);
 901 }
 902 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
 903
 904 void nvmet_req_execute(struct nvmet_req *req)
 905 {
 906         if (unlikely(req->data_len != req->transfer_len)) {
 907                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 908                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
 909         } else
 910                 req->execute(req);
 911 }
 912 EXPORT_SYMBOL_GPL(nvmet_req_execute);
 913
 914 int nvmet_req_alloc_sgl(struct nvmet_req *req)
 915 {
 916         struct pci_dev *p2p_dev = NULL;
 917
 918         if (IS_ENABLED(CONFIG_PCI_P2PDMA)) {
 919                 if (req->sq->ctrl && req->ns)
 920                         p2p_dev = radix_tree_lookup(&req->sq->ctrl->p2p_ns_map,
 921                                                     req->ns->nsid);
 922
 923                 req->p2p_dev = NULL;
 924                 if (req->sq->qid && p2p_dev) {
 925                         req->sg = pci_p2pmem_alloc_sgl(p2p_dev, &req->sg_cnt,
 926                                                        req->transfer_len);
 927                         if (req->sg) {
 928                                 req->p2p_dev = p2p_dev;
 929                                 return 0;
 930                         }
 931                 }
 932
 933                 /*
 934                  * If no P2P memory was available we fallback to using
 935                  * regular memory
 936                  */
 937         }
 938
 939         req->sg = sgl_alloc(req->transfer_len, GFP_KERNEL, &req->sg_cnt);
 940         if (!req->sg)
 941                 return -ENOMEM;
 942
 943         return 0;
 944 }
 945 EXPORT_SYMBOL_GPL(nvmet_req_alloc_sgl);
 946
 947 void nvmet_req_free_sgl(struct nvmet_req *req)
 948 {
 949         if (req->p2p_dev)
 950                 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
 951         else
 952                 sgl_free(req->sg);
 953
 954         req->sg = NULL;
 955         req->sg_cnt = 0;
 956 }
 957 EXPORT_SYMBOL_GPL(nvmet_req_free_sgl);
 958
 959 static inline bool nvmet_cc_en(u32 cc)
 960 {
 961         return (cc >> NVME_CC_EN_SHIFT) & 0x1;
 962 }
 963
 964 static inline u8 nvmet_cc_css(u32 cc)
 965 {
 966         return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
 967 }
 968
 969 static inline u8 nvmet_cc_mps(u32 cc)
 970 {
 971         return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
 972 }
 973
 974 static inline u8 nvmet_cc_ams(u32 cc)
 975 {
 976         return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
 977 }
 978
 979 static inline u8 nvmet_cc_shn(u32 cc)
 980 {
 981         return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
 982 }
 983
 984 static inline u8 nvmet_cc_iosqes(u32 cc)
 985 {
 986         return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
 987 }
 988
 989 static inline u8 nvmet_cc_iocqes(u32 cc)
 990 {
 991         return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
 992 }
 993
 994 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
 995 {
 996         lockdep_assert_held(&ctrl->lock);
 997
 998         if (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
 999             nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES ||
1000             nvmet_cc_mps(ctrl->cc) != 0 ||
1001             nvmet_cc_ams(ctrl->cc) != 0 ||
1002             nvmet_cc_css(ctrl->cc) != 0) {
1003                 ctrl->csts = NVME_CSTS_CFS;
1004                 return;
1005         }
1006
1007         ctrl->csts = NVME_CSTS_RDY;
1008
1009         /*
1010          * Controllers that are not yet enabled should not really enforce the
1011          * keep alive timeout, but we still want to track a timeout and cleanup
1012          * in case a host died before it enabled the controller.  Hence, simply
1013          * reset the keep alive timer when the controller is enabled.
1014          */
1015         mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
1016 }
1017
1018 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
1019 {
1020         lockdep_assert_held(&ctrl->lock);
1021
1022         /* XXX: tear down queues? */
1023         ctrl->csts &= ~NVME_CSTS_RDY;
1024         ctrl->cc = 0;
1025 }
1026
1027 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
1028 {
1029         u32 old;
1030
1031         mutex_lock(&ctrl->lock);
1032         old = ctrl->cc;
1033         ctrl->cc = new;
1034
1035         if (nvmet_cc_en(new) && !nvmet_cc_en(old))
1036                 nvmet_start_ctrl(ctrl);
1037         if (!nvmet_cc_en(new) && nvmet_cc_en(old))
1038                 nvmet_clear_ctrl(ctrl);
1039         if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
1040                 nvmet_clear_ctrl(ctrl);
1041                 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
1042         }
1043         if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
1044                 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
1045         mutex_unlock(&ctrl->lock);
1046 }
1047
1048 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
1049 {
1050         /* command sets supported: NVMe command set: */
1051         ctrl->cap = (1ULL << 37);
1052         /* CC.EN timeout in 500msec units: */
1053         ctrl->cap |= (15ULL << 24);
1054         /* maximum queue entries supported: */
1055         ctrl->cap |= NVMET_QUEUE_SIZE - 1;
1056 }
1057
1058 u16 nvmet_ctrl_find_get(const char *subsysnqn, const char *hostnqn, u16 cntlid,
1059                 struct nvmet_req *req, struct nvmet_ctrl **ret)
1060 {
1061         struct nvmet_subsys *subsys;
1062         struct nvmet_ctrl *ctrl;
1063         u16 status = 0;
1064
1065         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1066         if (!subsys) {
1067                 pr_warn("connect request for invalid subsystem %s!\n",
1068                         subsysnqn);
1069                 req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1070                 return NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1071         }
1072
1073         mutex_lock(&subsys->lock);
1074         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
1075                 if (ctrl->cntlid == cntlid) {
1076                         if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
1077                                 pr_warn("hostnqn mismatch.\n");
1078                                 continue;
1079                         }
1080                         if (!kref_get_unless_zero(&ctrl->ref))
1081                                 continue;
1082
1083                         *ret = ctrl;
1084                         goto out;
1085                 }
1086         }
1087
1088         pr_warn("could not find controller %d for subsys %s / host %s\n",
1089                 cntlid, subsysnqn, hostnqn);
1090         req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1091         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1092
1093 out:
1094         mutex_unlock(&subsys->lock);
1095         nvmet_subsys_put(subsys);
1096         return status;
1097 }
1098
1099 u16 nvmet_check_ctrl_status(struct nvmet_req *req, struct nvme_command *cmd)
1100 {
1101         if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1102                 pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1103                        cmd->common.opcode, req->sq->qid);
1104                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1105         }
1106
1107         if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1108                 pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1109                        cmd->common.opcode, req->sq->qid);
1110                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1111         }
1112         return 0;
1113 }
1114
1115 bool nvmet_host_allowed(struct nvmet_subsys *subsys, const char *hostnqn)
1116 {
1117         struct nvmet_host_link *p;
1118
1119         lockdep_assert_held(&nvmet_config_sem);
1120
1121         if (subsys->allow_any_host)
1122                 return true;
1123
1124         if (subsys->type == NVME_NQN_DISC) /* allow all access to disc subsys */
1125                 return true;
1126
1127         list_for_each_entry(p, &subsys->hosts, entry) {
1128                 if (!strcmp(nvmet_host_name(p->host), hostnqn))
1129                         return true;
1130         }
1131
1132         return false;
1133 }
1134
1135 /*
1136  * Note: ctrl->subsys->lock should be held when calling this function
1137  */
1138 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1139                 struct nvmet_req *req)
1140 {
1141         struct nvmet_ns *ns;
1142
1143         if (!req->p2p_client)
1144                 return;
1145
1146         ctrl->p2p_client = get_device(req->p2p_client);
1147
1148         list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link)
1149                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1150 }
1151
1152 /*
1153  * Note: ctrl->subsys->lock should be held when calling this function
1154  */
1155 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1156 {
1157         struct radix_tree_iter iter;
1158         void __rcu **slot;
1159
1160         radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1161                 pci_dev_put(radix_tree_deref_slot(slot));
1162
1163         put_device(ctrl->p2p_client);
1164 }
1165
1166 static void nvmet_fatal_error_handler(struct work_struct *work)
1167 {
1168         struct nvmet_ctrl *ctrl =
1169                         container_of(work, struct nvmet_ctrl, fatal_err_work);
1170
1171         pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1172         ctrl->ops->delete_ctrl(ctrl);
1173 }
1174
1175 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1176                 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
1177 {
1178         struct nvmet_subsys *subsys;
1179         struct nvmet_ctrl *ctrl;
1180         int ret;
1181         u16 status;
1182
1183         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1184         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1185         if (!subsys) {
1186                 pr_warn("connect request for invalid subsystem %s!\n",
1187                         subsysnqn);
1188                 req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1189                 goto out;
1190         }
1191
1192         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1193         down_read(&nvmet_config_sem);
1194         if (!nvmet_host_allowed(subsys, hostnqn)) {
1195                 pr_info("connect by host %s for subsystem %s not allowed\n",
1196                         hostnqn, subsysnqn);
1197                 req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1198                 up_read(&nvmet_config_sem);
1199                 status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
1200                 goto out_put_subsystem;
1201         }
1202         up_read(&nvmet_config_sem);
1203
1204         status = NVME_SC_INTERNAL;
1205         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1206         if (!ctrl)
1207                 goto out_put_subsystem;
1208         mutex_init(&ctrl->lock);
1209
1210         nvmet_init_cap(ctrl);
1211
1212         ctrl->port = req->port;
1213
1214         INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
1215         INIT_LIST_HEAD(&ctrl->async_events);
1216         INIT_RADIX_TREE(&ctrl->p2p_ns_map, GFP_KERNEL);
1217         INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1218
1219         memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1220         memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1221
1222         kref_init(&ctrl->ref);
1223         ctrl->subsys = subsys;
1224         WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1225
1226         ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1227                         sizeof(__le32), GFP_KERNEL);
1228         if (!ctrl->changed_ns_list)
1229                 goto out_free_ctrl;
1230
1231         ctrl->cqs = kcalloc(subsys->max_qid + 1,
1232                         sizeof(struct nvmet_cq *),
1233                         GFP_KERNEL);
1234         if (!ctrl->cqs)
1235                 goto out_free_changed_ns_list;
1236
1237         ctrl->sqs = kcalloc(subsys->max_qid + 1,
1238                         sizeof(struct nvmet_sq *),
1239                         GFP_KERNEL);
1240         if (!ctrl->sqs)
1241                 goto out_free_cqs;
1242
1243         ret = ida_simple_get(&cntlid_ida,
1244                              NVME_CNTLID_MIN, NVME_CNTLID_MAX,
1245                              GFP_KERNEL);
1246         if (ret < 0) {
1247                 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
1248                 goto out_free_sqs;
1249         }
1250         ctrl->cntlid = ret;
1251
1252         ctrl->ops = req->ops;
1253
1254         /*
1255          * Discovery controllers may use some arbitrary high value
1256          * in order to cleanup stale discovery sessions
1257          */
1258         if ((ctrl->subsys->type == NVME_NQN_DISC) && !kato)
1259                 kato = NVMET_DISC_KATO_MS;
1260
1261         /* keep-alive timeout in seconds */
1262         ctrl->kato = DIV_ROUND_UP(kato, 1000);
1263
1264         ctrl->err_counter = 0;
1265         spin_lock_init(&ctrl->error_lock);
1266
1267         nvmet_start_keep_alive_timer(ctrl);
1268
1269         mutex_lock(&subsys->lock);
1270         list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1271         nvmet_setup_p2p_ns_map(ctrl, req);
1272         mutex_unlock(&subsys->lock);
1273
1274         *ctrlp = ctrl;
1275         return 0;
1276
1277 out_free_sqs:
1278         kfree(ctrl->sqs);
1279 out_free_cqs:
1280         kfree(ctrl->cqs);
1281 out_free_changed_ns_list:
1282         kfree(ctrl->changed_ns_list);
1283 out_free_ctrl:
1284         kfree(ctrl);
1285 out_put_subsystem:
1286         nvmet_subsys_put(subsys);
1287 out:
1288         return status;
1289 }
1290
1291 static void nvmet_ctrl_free(struct kref *ref)
1292 {
1293         struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1294         struct nvmet_subsys *subsys = ctrl->subsys;
1295
1296         mutex_lock(&subsys->lock);
1297         nvmet_release_p2p_ns_map(ctrl);
1298         list_del(&ctrl->subsys_entry);
1299         mutex_unlock(&subsys->lock);
1300
1301         nvmet_stop_keep_alive_timer(ctrl);
1302
1303         flush_work(&ctrl->async_event_work);
1304         cancel_work_sync(&ctrl->fatal_err_work);
1305
1306         ida_simple_remove(&cntlid_ida, ctrl->cntlid);
1307
1308         kfree(ctrl->sqs);
1309         kfree(ctrl->cqs);
1310         kfree(ctrl->changed_ns_list);
1311         kfree(ctrl);
1312
1313         nvmet_subsys_put(subsys);
1314 }
1315
1316 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1317 {
1318         kref_put(&ctrl->ref, nvmet_ctrl_free);
1319 }
1320
1321 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1322 {
1323         mutex_lock(&ctrl->lock);
1324         if (!(ctrl->csts & NVME_CSTS_CFS)) {
1325                 ctrl->csts |= NVME_CSTS_CFS;
1326                 schedule_work(&ctrl->fatal_err_work);
1327         }
1328         mutex_unlock(&ctrl->lock);
1329 }
1330 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1331
1332 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1333                 const char *subsysnqn)
1334 {
1335         struct nvmet_subsys_link *p;
1336
1337         if (!port)
1338                 return NULL;
1339
1340         if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1341                 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1342                         return NULL;
1343                 return nvmet_disc_subsys;
1344         }
1345
1346         down_read(&nvmet_config_sem);
1347         list_for_each_entry(p, &port->subsystems, entry) {
1348                 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1349                                 NVMF_NQN_SIZE)) {
1350                         if (!kref_get_unless_zero(&p->subsys->ref))
1351                                 break;
1352                         up_read(&nvmet_config_sem);
1353                         return p->subsys;
1354                 }
1355         }
1356         up_read(&nvmet_config_sem);
1357         return NULL;
1358 }
1359
1360 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1361                 enum nvme_subsys_type type)
1362 {
1363         struct nvmet_subsys *subsys;
1364
1365         subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1366         if (!subsys)
1367                 return NULL;
1368
1369         subsys->ver = NVME_VS(1, 3, 0); /* NVMe 1.3.0 */
1370         /* generate a random serial number as our controllers are ephemeral: */
1371         get_random_bytes(&subsys->serial, sizeof(subsys->serial));
1372
1373         switch (type) {
1374         case NVME_NQN_NVME:
1375                 subsys->max_qid = NVMET_NR_QUEUES;
1376                 break;
1377         case NVME_NQN_DISC:
1378                 subsys->max_qid = 0;
1379                 break;
1380         default:
1381                 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1382                 kfree(subsys);
1383                 return NULL;
1384         }
1385         subsys->type = type;
1386         subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1387                         GFP_KERNEL);
1388         if (!subsys->subsysnqn) {
1389                 kfree(subsys);
1390                 return NULL;
1391         }
1392
1393         kref_init(&subsys->ref);
1394
1395         mutex_init(&subsys->lock);
1396         INIT_LIST_HEAD(&subsys->namespaces);
1397         INIT_LIST_HEAD(&subsys->ctrls);
1398         INIT_LIST_HEAD(&subsys->hosts);
1399
1400         return subsys;
1401 }
1402
1403 static void nvmet_subsys_free(struct kref *ref)
1404 {
1405         struct nvmet_subsys *subsys =
1406                 container_of(ref, struct nvmet_subsys, ref);
1407
1408         WARN_ON_ONCE(!list_empty(&subsys->namespaces));
1409
1410         kfree(subsys->subsysnqn);
1411         kfree(subsys);
1412 }
1413
1414 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1415 {
1416         struct nvmet_ctrl *ctrl;
1417
1418         mutex_lock(&subsys->lock);
1419         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1420                 ctrl->ops->delete_ctrl(ctrl);
1421         mutex_unlock(&subsys->lock);
1422 }
1423
1424 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1425 {
1426         kref_put(&subsys->ref, nvmet_subsys_free);
1427 }
1428
1429 static int __init nvmet_init(void)
1430 {
1431         int error;
1432
1433         nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1434
1435         buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1436                         WQ_MEM_RECLAIM, 0);
1437         if (!buffered_io_wq) {
1438                 error = -ENOMEM;
1439                 goto out;
1440         }
1441
1442         error = nvmet_init_discovery();
1443         if (error)
1444                 goto out_free_work_queue;
1445
1446         error = nvmet_init_configfs();
1447         if (error)
1448                 goto out_exit_discovery;
1449         return 0;
1450
1451 out_exit_discovery:
1452         nvmet_exit_discovery();
1453 out_free_work_queue:
1454         destroy_workqueue(buffered_io_wq);
1455 out:
1456         return error;
1457 }
1458
1459 static void __exit nvmet_exit(void)
1460 {
1461         nvmet_exit_configfs();
1462         nvmet_exit_discovery();
1463         ida_destroy(&cntlid_ida);
1464         destroy_workqueue(buffered_io_wq);
1465
1466         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1467         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1468 }
1469
1470 module_init(nvmet_init);
1471 module_exit(nvmet_exit);
1472
1473 MODULE_LICENSE("GPL v2");