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 #include <linux/scatterlist.h>
  12
  13 #define CREATE_TRACE_POINTS
  14 #include "trace.h"
  15
  16 #include "nvmet.h"
  17
  18 struct workqueue_struct *buffered_io_wq;
  19 static const struct nvmet_fabrics_ops *nvmet_transports[NVMF_TRTYPE_MAX];
  20 static DEFINE_IDA(cntlid_ida);
  21
  22 /*
  23  * This read/write semaphore is used to synchronize access to configuration
  24  * information on a target system that will result in discovery log page
  25  * information change for at least one host.
  26  * The full list of resources to protected by this semaphore is:
  27  *
  28  *  - subsystems list
  29  *  - per-subsystem allowed hosts list
  30  *  - allow_any_host subsystem attribute
  31  *  - nvmet_genctr
  32  *  - the nvmet_transports array
  33  *
  34  * When updating any of those lists/structures write lock should be obtained,
  35  * while when reading (popolating discovery log page or checking host-subsystem
  36  * link) read lock is obtained to allow concurrent reads.
  37  */
  38 DECLARE_RWSEM(nvmet_config_sem);
  39
  40 u32 nvmet_ana_group_enabled[NVMET_MAX_ANAGRPS + 1];
  41 u64 nvmet_ana_chgcnt;
  42 DECLARE_RWSEM(nvmet_ana_sem);
  43
  44 inline u16 errno_to_nvme_status(struct nvmet_req *req, int errno)
  45 {
  46         u16 status;
  47
  48         switch (errno) {
  49         case 0:
  50                 status = NVME_SC_SUCCESS;
  51                 break;
  52         case -ENOSPC:
  53                 req->error_loc = offsetof(struct nvme_rw_command, length);
  54                 status = NVME_SC_CAP_EXCEEDED | NVME_SC_DNR;
  55                 break;
  56         case -EREMOTEIO:
  57                 req->error_loc = offsetof(struct nvme_rw_command, slba);
  58                 status = NVME_SC_LBA_RANGE | NVME_SC_DNR;
  59                 break;
  60         case -EOPNOTSUPP:
  61                 req->error_loc = offsetof(struct nvme_common_command, opcode);
  62                 switch (req->cmd->common.opcode) {
  63                 case nvme_cmd_dsm:
  64                 case nvme_cmd_write_zeroes:
  65                         status = NVME_SC_ONCS_NOT_SUPPORTED | NVME_SC_DNR;
  66                         break;
  67                 default:
  68                         status = NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
  69                 }
  70                 break;
  71         case -ENODATA:
  72                 req->error_loc = offsetof(struct nvme_rw_command, nsid);
  73                 status = NVME_SC_ACCESS_DENIED;
  74                 break;
  75         case -EIO:
  76                 /* FALLTHRU */
  77         default:
  78                 req->error_loc = offsetof(struct nvme_common_command, opcode);
  79                 status = NVME_SC_INTERNAL | NVME_SC_DNR;
  80         }
  81
  82         return status;
  83 }
  84
  85 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
  86                 const char *subsysnqn);
  87
  88 u16 nvmet_copy_to_sgl(struct nvmet_req *req, off_t off, const void *buf,
  89                 size_t len)
  90 {
  91         if (sg_pcopy_from_buffer(req->sg, req->sg_cnt, buf, len, off) != len) {
  92                 req->error_loc = offsetof(struct nvme_common_command, dptr);
  93                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
  94         }
  95         return 0;
  96 }
  97
  98 u16 nvmet_copy_from_sgl(struct nvmet_req *req, off_t off, void *buf, size_t len)
  99 {
 100         if (sg_pcopy_to_buffer(req->sg, req->sg_cnt, buf, len, off) != len) {
 101                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 102                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
 103         }
 104         return 0;
 105 }
 106
 107 u16 nvmet_zero_sgl(struct nvmet_req *req, off_t off, size_t len)
 108 {
 109         if (sg_zero_buffer(req->sg, req->sg_cnt, len, off) != len) {
 110                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 111                 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
 112         }
 113         return 0;
 114 }
 115
 116 static unsigned int nvmet_max_nsid(struct nvmet_subsys *subsys)
 117 {
 118         struct nvmet_ns *ns;
 119
 120         if (list_empty(&subsys->namespaces))
 121                 return 0;
 122
 123         ns = list_last_entry(&subsys->namespaces, struct nvmet_ns, dev_link);
 124         return ns->nsid;
 125 }
 126
 127 static u32 nvmet_async_event_result(struct nvmet_async_event *aen)
 128 {
 129         return aen->event_type | (aen->event_info << 8) | (aen->log_page << 16);
 130 }
 131
 132 static void nvmet_async_events_free(struct nvmet_ctrl *ctrl)
 133 {
 134         struct nvmet_req *req;
 135
 136         while (1) {
 137                 mutex_lock(&ctrl->lock);
 138                 if (!ctrl->nr_async_event_cmds) {
 139                         mutex_unlock(&ctrl->lock);
 140                         return;
 141                 }
 142
 143                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
 144                 mutex_unlock(&ctrl->lock);
 145                 nvmet_req_complete(req, NVME_SC_INTERNAL | NVME_SC_DNR);
 146         }
 147 }
 148
 149 static void nvmet_async_event_work(struct work_struct *work)
 150 {
 151         struct nvmet_ctrl *ctrl =
 152                 container_of(work, struct nvmet_ctrl, async_event_work);
 153         struct nvmet_async_event *aen;
 154         struct nvmet_req *req;
 155
 156         while (1) {
 157                 mutex_lock(&ctrl->lock);
 158                 aen = list_first_entry_or_null(&ctrl->async_events,
 159                                 struct nvmet_async_event, entry);
 160                 if (!aen || !ctrl->nr_async_event_cmds) {
 161                         mutex_unlock(&ctrl->lock);
 162                         return;
 163                 }
 164
 165                 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
 166                 nvmet_set_result(req, nvmet_async_event_result(aen));
 167
 168                 list_del(&aen->entry);
 169                 kfree(aen);
 170
 171                 mutex_unlock(&ctrl->lock);
 172                 nvmet_req_complete(req, 0);
 173         }
 174 }
 175
 176 void nvmet_add_async_event(struct nvmet_ctrl *ctrl, u8 event_type,
 177                 u8 event_info, u8 log_page)
 178 {
 179         struct nvmet_async_event *aen;
 180
 181         aen = kmalloc(sizeof(*aen), GFP_KERNEL);
 182         if (!aen)
 183                 return;
 184
 185         aen->event_type = event_type;
 186         aen->event_info = event_info;
 187         aen->log_page = log_page;
 188
 189         mutex_lock(&ctrl->lock);
 190         list_add_tail(&aen->entry, &ctrl->async_events);
 191         mutex_unlock(&ctrl->lock);
 192
 193         schedule_work(&ctrl->async_event_work);
 194 }
 195
 196 static void nvmet_add_to_changed_ns_log(struct nvmet_ctrl *ctrl, __le32 nsid)
 197 {
 198         u32 i;
 199
 200         mutex_lock(&ctrl->lock);
 201         if (ctrl->nr_changed_ns > NVME_MAX_CHANGED_NAMESPACES)
 202                 goto out_unlock;
 203
 204         for (i = 0; i < ctrl->nr_changed_ns; i++) {
 205                 if (ctrl->changed_ns_list[i] == nsid)
 206                         goto out_unlock;
 207         }
 208
 209         if (ctrl->nr_changed_ns == NVME_MAX_CHANGED_NAMESPACES) {
 210                 ctrl->changed_ns_list[0] = cpu_to_le32(0xffffffff);
 211                 ctrl->nr_changed_ns = U32_MAX;
 212                 goto out_unlock;
 213         }
 214
 215         ctrl->changed_ns_list[ctrl->nr_changed_ns++] = nsid;
 216 out_unlock:
 217         mutex_unlock(&ctrl->lock);
 218 }
 219
 220 void nvmet_ns_changed(struct nvmet_subsys *subsys, u32 nsid)
 221 {
 222         struct nvmet_ctrl *ctrl;
 223
 224         lockdep_assert_held(&subsys->lock);
 225
 226         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 227                 nvmet_add_to_changed_ns_log(ctrl, cpu_to_le32(nsid));
 228                 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_NS_ATTR))
 229                         continue;
 230                 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
 231                                 NVME_AER_NOTICE_NS_CHANGED,
 232                                 NVME_LOG_CHANGED_NS);
 233         }
 234 }
 235
 236 void nvmet_send_ana_event(struct nvmet_subsys *subsys,
 237                 struct nvmet_port *port)
 238 {
 239         struct nvmet_ctrl *ctrl;
 240
 241         mutex_lock(&subsys->lock);
 242         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 243                 if (port && ctrl->port != port)
 244                         continue;
 245                 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_ANA_CHANGE))
 246                         continue;
 247                 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
 248                                 NVME_AER_NOTICE_ANA, NVME_LOG_ANA);
 249         }
 250         mutex_unlock(&subsys->lock);
 251 }
 252
 253 void nvmet_port_send_ana_event(struct nvmet_port *port)
 254 {
 255         struct nvmet_subsys_link *p;
 256
 257         down_read(&nvmet_config_sem);
 258         list_for_each_entry(p, &port->subsystems, entry)
 259                 nvmet_send_ana_event(p->subsys, port);
 260         up_read(&nvmet_config_sem);
 261 }
 262
 263 int nvmet_register_transport(const struct nvmet_fabrics_ops *ops)
 264 {
 265         int ret = 0;
 266
 267         down_write(&nvmet_config_sem);
 268         if (nvmet_transports[ops->type])
 269                 ret = -EINVAL;
 270         else
 271                 nvmet_transports[ops->type] = ops;
 272         up_write(&nvmet_config_sem);
 273
 274         return ret;
 275 }
 276 EXPORT_SYMBOL_GPL(nvmet_register_transport);
 277
 278 void nvmet_unregister_transport(const struct nvmet_fabrics_ops *ops)
 279 {
 280         down_write(&nvmet_config_sem);
 281         nvmet_transports[ops->type] = NULL;
 282         up_write(&nvmet_config_sem);
 283 }
 284 EXPORT_SYMBOL_GPL(nvmet_unregister_transport);
 285
 286 void nvmet_port_del_ctrls(struct nvmet_port *port, struct nvmet_subsys *subsys)
 287 {
 288         struct nvmet_ctrl *ctrl;
 289
 290         mutex_lock(&subsys->lock);
 291         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
 292                 if (ctrl->port == port)
 293                         ctrl->ops->delete_ctrl(ctrl);
 294         }
 295         mutex_unlock(&subsys->lock);
 296 }
 297
 298 int nvmet_enable_port(struct nvmet_port *port)
 299 {
 300         const struct nvmet_fabrics_ops *ops;
 301         int ret;
 302
 303         lockdep_assert_held(&nvmet_config_sem);
 304
 305         ops = nvmet_transports[port->disc_addr.trtype];
 306         if (!ops) {
 307                 up_write(&nvmet_config_sem);
 308                 request_module("nvmet-transport-%d", port->disc_addr.trtype);
 309                 down_write(&nvmet_config_sem);
 310                 ops = nvmet_transports[port->disc_addr.trtype];
 311                 if (!ops) {
 312                         pr_err("transport type %d not supported\n",
 313                                 port->disc_addr.trtype);
 314                         return -EINVAL;
 315                 }
 316         }
 317
 318         if (!try_module_get(ops->owner))
 319                 return -EINVAL;
 320
 321         ret = ops->add_port(port);
 322         if (ret) {
 323                 module_put(ops->owner);
 324                 return ret;
 325         }
 326
 327         /* If the transport didn't set inline_data_size, then disable it. */
 328         if (port->inline_data_size < 0)
 329                 port->inline_data_size = 0;
 330
 331         port->enabled = true;
 332         port->tr_ops = ops;
 333         return 0;
 334 }
 335
 336 void nvmet_disable_port(struct nvmet_port *port)
 337 {
 338         const struct nvmet_fabrics_ops *ops;
 339
 340         lockdep_assert_held(&nvmet_config_sem);
 341
 342         port->enabled = false;
 343         port->tr_ops = NULL;
 344
 345         ops = nvmet_transports[port->disc_addr.trtype];
 346         ops->remove_port(port);
 347         module_put(ops->owner);
 348 }
 349
 350 static void nvmet_keep_alive_timer(struct work_struct *work)
 351 {
 352         struct nvmet_ctrl *ctrl = container_of(to_delayed_work(work),
 353                         struct nvmet_ctrl, ka_work);
 354         bool cmd_seen = ctrl->cmd_seen;
 355
 356         ctrl->cmd_seen = false;
 357         if (cmd_seen) {
 358                 pr_debug("ctrl %d reschedule traffic based keep-alive timer\n",
 359                         ctrl->cntlid);
 360                 schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
 361                 return;
 362         }
 363
 364         pr_err("ctrl %d keep-alive timer (%d seconds) expired!\n",
 365                 ctrl->cntlid, ctrl->kato);
 366
 367         nvmet_ctrl_fatal_error(ctrl);
 368 }
 369
 370 static void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl)
 371 {
 372         pr_debug("ctrl %d start keep-alive timer for %d secs\n",
 373                 ctrl->cntlid, ctrl->kato);
 374
 375         INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
 376         schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
 377 }
 378
 379 static void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
 380 {
 381         pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
 382
 383         cancel_delayed_work_sync(&ctrl->ka_work);
 384 }
 385
 386 static struct nvmet_ns *__nvmet_find_namespace(struct nvmet_ctrl *ctrl,
 387                 __le32 nsid)
 388 {
 389         struct nvmet_ns *ns;
 390
 391         list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link) {
 392                 if (ns->nsid == le32_to_cpu(nsid))
 393                         return ns;
 394         }
 395
 396         return NULL;
 397 }
 398
 399 struct nvmet_ns *nvmet_find_namespace(struct nvmet_ctrl *ctrl, __le32 nsid)
 400 {
 401         struct nvmet_ns *ns;
 402
 403         rcu_read_lock();
 404         ns = __nvmet_find_namespace(ctrl, nsid);
 405         if (ns)
 406                 percpu_ref_get(&ns->ref);
 407         rcu_read_unlock();
 408
 409         return ns;
 410 }
 411
 412 static void nvmet_destroy_namespace(struct percpu_ref *ref)
 413 {
 414         struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
 415
 416         complete(&ns->disable_done);
 417 }
 418
 419 void nvmet_put_namespace(struct nvmet_ns *ns)
 420 {
 421         percpu_ref_put(&ns->ref);
 422 }
 423
 424 static void nvmet_ns_dev_disable(struct nvmet_ns *ns)
 425 {
 426         nvmet_bdev_ns_disable(ns);
 427         nvmet_file_ns_disable(ns);
 428 }
 429
 430 static int nvmet_p2pmem_ns_enable(struct nvmet_ns *ns)
 431 {
 432         int ret;
 433         struct pci_dev *p2p_dev;
 434
 435         if (!ns->use_p2pmem)
 436                 return 0;
 437
 438         if (!ns->bdev) {
 439                 pr_err("peer-to-peer DMA is not supported by non-block device namespaces\n");
 440                 return -EINVAL;
 441         }
 442
 443         if (!blk_queue_pci_p2pdma(ns->bdev->bd_queue)) {
 444                 pr_err("peer-to-peer DMA is not supported by the driver of %s\n",
 445                        ns->device_path);
 446                 return -EINVAL;
 447         }
 448
 449         if (ns->p2p_dev) {
 450                 ret = pci_p2pdma_distance(ns->p2p_dev, nvmet_ns_dev(ns), true);
 451                 if (ret < 0)
 452                         return -EINVAL;
 453         } else {
 454                 /*
 455                  * Right now we just check that there is p2pmem available so
 456                  * we can report an error to the user right away if there
 457                  * is not. We'll find the actual device to use once we
 458                  * setup the controller when the port's device is available.
 459                  */
 460
 461                 p2p_dev = pci_p2pmem_find(nvmet_ns_dev(ns));
 462                 if (!p2p_dev) {
 463                         pr_err("no peer-to-peer memory is available for %s\n",
 464                                ns->device_path);
 465                         return -EINVAL;
 466                 }
 467
 468                 pci_dev_put(p2p_dev);
 469         }
 470
 471         return 0;
 472 }
 473
 474 /*
 475  * Note: ctrl->subsys->lock should be held when calling this function
 476  */
 477 static void nvmet_p2pmem_ns_add_p2p(struct nvmet_ctrl *ctrl,
 478                                     struct nvmet_ns *ns)
 479 {
 480         struct device *clients[2];
 481         struct pci_dev *p2p_dev;
 482         int ret;
 483
 484         if (!ctrl->p2p_client || !ns->use_p2pmem)
 485                 return;
 486
 487         if (ns->p2p_dev) {
 488                 ret = pci_p2pdma_distance(ns->p2p_dev, ctrl->p2p_client, true);
 489                 if (ret < 0)
 490                         return;
 491
 492                 p2p_dev = pci_dev_get(ns->p2p_dev);
 493         } else {
 494                 clients[0] = ctrl->p2p_client;
 495                 clients[1] = nvmet_ns_dev(ns);
 496
 497                 p2p_dev = pci_p2pmem_find_many(clients, ARRAY_SIZE(clients));
 498                 if (!p2p_dev) {
 499                         pr_err("no peer-to-peer memory is available that's supported by %s and %s\n",
 500                                dev_name(ctrl->p2p_client), ns->device_path);
 501                         return;
 502                 }
 503         }
 504
 505         ret = radix_tree_insert(&ctrl->p2p_ns_map, ns->nsid, p2p_dev);
 506         if (ret < 0)
 507                 pci_dev_put(p2p_dev);
 508
 509         pr_info("using p2pmem on %s for nsid %d\n", pci_name(p2p_dev),
 510                 ns->nsid);
 511 }
 512
 513 int nvmet_ns_enable(struct nvmet_ns *ns)
 514 {
 515         struct nvmet_subsys *subsys = ns->subsys;
 516         struct nvmet_ctrl *ctrl;
 517         int ret;
 518
 519         mutex_lock(&subsys->lock);
 520         ret = 0;
 521         if (ns->enabled)
 522                 goto out_unlock;
 523
 524         ret = -EMFILE;
 525         if (subsys->nr_namespaces == NVMET_MAX_NAMESPACES)
 526                 goto out_unlock;
 527
 528         ret = nvmet_bdev_ns_enable(ns);
 529         if (ret == -ENOTBLK)
 530                 ret = nvmet_file_ns_enable(ns);
 531         if (ret)
 532                 goto out_unlock;
 533
 534         ret = nvmet_p2pmem_ns_enable(ns);
 535         if (ret)
 536                 goto out_dev_disable;
 537
 538         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 539                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
 540
 541         ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
 542                                 0, GFP_KERNEL);
 543         if (ret)
 544                 goto out_dev_put;
 545
 546         if (ns->nsid > subsys->max_nsid)
 547                 subsys->max_nsid = ns->nsid;
 548
 549         /*
 550          * The namespaces list needs to be sorted to simplify the implementation
 551          * of the Identify Namepace List subcommand.
 552          */
 553         if (list_empty(&subsys->namespaces)) {
 554                 list_add_tail_rcu(&ns->dev_link, &subsys->namespaces);
 555         } else {
 556                 struct nvmet_ns *old;
 557
 558                 list_for_each_entry_rcu(old, &subsys->namespaces, dev_link,
 559                                         lockdep_is_held(&subsys->lock)) {
 560                         BUG_ON(ns->nsid == old->nsid);
 561                         if (ns->nsid < old->nsid)
 562                                 break;
 563                 }
 564
 565                 list_add_tail_rcu(&ns->dev_link, &old->dev_link);
 566         }
 567         subsys->nr_namespaces++;
 568
 569         nvmet_ns_changed(subsys, ns->nsid);
 570         ns->enabled = true;
 571         ret = 0;
 572 out_unlock:
 573         mutex_unlock(&subsys->lock);
 574         return ret;
 575 out_dev_put:
 576         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 577                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 578 out_dev_disable:
 579         nvmet_ns_dev_disable(ns);
 580         goto out_unlock;
 581 }
 582
 583 void nvmet_ns_disable(struct nvmet_ns *ns)
 584 {
 585         struct nvmet_subsys *subsys = ns->subsys;
 586         struct nvmet_ctrl *ctrl;
 587
 588         mutex_lock(&subsys->lock);
 589         if (!ns->enabled)
 590                 goto out_unlock;
 591
 592         ns->enabled = false;
 593         list_del_rcu(&ns->dev_link);
 594         if (ns->nsid == subsys->max_nsid)
 595                 subsys->max_nsid = nvmet_max_nsid(subsys);
 596
 597         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
 598                 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
 599
 600         mutex_unlock(&subsys->lock);
 601
 602         /*
 603          * Now that we removed the namespaces from the lookup list, we
 604          * can kill the per_cpu ref and wait for any remaining references
 605          * to be dropped, as well as a RCU grace period for anyone only
 606          * using the namepace under rcu_read_lock().  Note that we can't
 607          * use call_rcu here as we need to ensure the namespaces have
 608          * been fully destroyed before unloading the module.
 609          */
 610         percpu_ref_kill(&ns->ref);
 611         synchronize_rcu();
 612         wait_for_completion(&ns->disable_done);
 613         percpu_ref_exit(&ns->ref);
 614
 615         mutex_lock(&subsys->lock);
 616
 617         subsys->nr_namespaces--;
 618         nvmet_ns_changed(subsys, ns->nsid);
 619         nvmet_ns_dev_disable(ns);
 620 out_unlock:
 621         mutex_unlock(&subsys->lock);
 622 }
 623
 624 void nvmet_ns_free(struct nvmet_ns *ns)
 625 {
 626         nvmet_ns_disable(ns);
 627
 628         down_write(&nvmet_ana_sem);
 629         nvmet_ana_group_enabled[ns->anagrpid]--;
 630         up_write(&nvmet_ana_sem);
 631
 632         kfree(ns->device_path);
 633         kfree(ns);
 634 }
 635
 636 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
 637 {
 638         struct nvmet_ns *ns;
 639
 640         ns = kzalloc(sizeof(*ns), GFP_KERNEL);
 641         if (!ns)
 642                 return NULL;
 643
 644         INIT_LIST_HEAD(&ns->dev_link);
 645         init_completion(&ns->disable_done);
 646
 647         ns->nsid = nsid;
 648         ns->subsys = subsys;
 649
 650         down_write(&nvmet_ana_sem);
 651         ns->anagrpid = NVMET_DEFAULT_ANA_GRPID;
 652         nvmet_ana_group_enabled[ns->anagrpid]++;
 653         up_write(&nvmet_ana_sem);
 654
 655         uuid_gen(&ns->uuid);
 656         ns->buffered_io = false;
 657
 658         return ns;
 659 }
 660
 661 static void nvmet_update_sq_head(struct nvmet_req *req)
 662 {
 663         if (req->sq->size) {
 664                 u32 old_sqhd, new_sqhd;
 665
 666                 do {
 667                         old_sqhd = req->sq->sqhd;
 668                         new_sqhd = (old_sqhd + 1) % req->sq->size;
 669                 } while (cmpxchg(&req->sq->sqhd, old_sqhd, new_sqhd) !=
 670                                         old_sqhd);
 671         }
 672         req->cqe->sq_head = cpu_to_le16(req->sq->sqhd & 0x0000FFFF);
 673 }
 674
 675 static void nvmet_set_error(struct nvmet_req *req, u16 status)
 676 {
 677         struct nvmet_ctrl *ctrl = req->sq->ctrl;
 678         struct nvme_error_slot *new_error_slot;
 679         unsigned long flags;
 680
 681         req->cqe->status = cpu_to_le16(status << 1);
 682
 683         if (!ctrl || req->error_loc == NVMET_NO_ERROR_LOC)
 684                 return;
 685
 686         spin_lock_irqsave(&ctrl->error_lock, flags);
 687         ctrl->err_counter++;
 688         new_error_slot =
 689                 &ctrl->slots[ctrl->err_counter % NVMET_ERROR_LOG_SLOTS];
 690
 691         new_error_slot->error_count = cpu_to_le64(ctrl->err_counter);
 692         new_error_slot->sqid = cpu_to_le16(req->sq->qid);
 693         new_error_slot->cmdid = cpu_to_le16(req->cmd->common.command_id);
 694         new_error_slot->status_field = cpu_to_le16(status << 1);
 695         new_error_slot->param_error_location = cpu_to_le16(req->error_loc);
 696         new_error_slot->lba = cpu_to_le64(req->error_slba);
 697         new_error_slot->nsid = req->cmd->common.nsid;
 698         spin_unlock_irqrestore(&ctrl->error_lock, flags);
 699
 700         /* set the more bit for this request */
 701         req->cqe->status |= cpu_to_le16(1 << 14);
 702 }
 703
 704 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
 705 {
 706         if (!req->sq->sqhd_disabled)
 707                 nvmet_update_sq_head(req);
 708         req->cqe->sq_id = cpu_to_le16(req->sq->qid);
 709         req->cqe->command_id = req->cmd->common.command_id;
 710
 711         if (unlikely(status))
 712                 nvmet_set_error(req, status);
 713
 714         trace_nvmet_req_complete(req);
 715
 716         if (req->ns)
 717                 nvmet_put_namespace(req->ns);
 718         req->ops->queue_response(req);
 719 }
 720
 721 void nvmet_req_complete(struct nvmet_req *req, u16 status)
 722 {
 723         __nvmet_req_complete(req, status);
 724         percpu_ref_put(&req->sq->ref);
 725 }
 726 EXPORT_SYMBOL_GPL(nvmet_req_complete);
 727
 728 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
 729                 u16 qid, u16 size)
 730 {
 731         cq->qid = qid;
 732         cq->size = size;
 733
 734         ctrl->cqs[qid] = cq;
 735 }
 736
 737 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
 738                 u16 qid, u16 size)
 739 {
 740         sq->sqhd = 0;
 741         sq->qid = qid;
 742         sq->size = size;
 743
 744         ctrl->sqs[qid] = sq;
 745 }
 746
 747 static void nvmet_confirm_sq(struct percpu_ref *ref)
 748 {
 749         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 750
 751         complete(&sq->confirm_done);
 752 }
 753
 754 void nvmet_sq_destroy(struct nvmet_sq *sq)
 755 {
 756         /*
 757          * If this is the admin queue, complete all AERs so that our
 758          * queue doesn't have outstanding requests on it.
 759          */
 760         if (sq->ctrl && sq->ctrl->sqs && sq->ctrl->sqs[0] == sq)
 761                 nvmet_async_events_free(sq->ctrl);
 762         percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
 763         wait_for_completion(&sq->confirm_done);
 764         wait_for_completion(&sq->free_done);
 765         percpu_ref_exit(&sq->ref);
 766
 767         if (sq->ctrl) {
 768                 nvmet_ctrl_put(sq->ctrl);
 769                 sq->ctrl = NULL; /* allows reusing the queue later */
 770         }
 771 }
 772 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
 773
 774 static void nvmet_sq_free(struct percpu_ref *ref)
 775 {
 776         struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
 777
 778         complete(&sq->free_done);
 779 }
 780
 781 int nvmet_sq_init(struct nvmet_sq *sq)
 782 {
 783         int ret;
 784
 785         ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
 786         if (ret) {
 787                 pr_err("percpu_ref init failed!\n");
 788                 return ret;
 789         }
 790         init_completion(&sq->free_done);
 791         init_completion(&sq->confirm_done);
 792
 793         return 0;
 794 }
 795 EXPORT_SYMBOL_GPL(nvmet_sq_init);
 796
 797 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
 798                 struct nvmet_ns *ns)
 799 {
 800         enum nvme_ana_state state = port->ana_state[ns->anagrpid];
 801
 802         if (unlikely(state == NVME_ANA_INACCESSIBLE))
 803                 return NVME_SC_ANA_INACCESSIBLE;
 804         if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
 805                 return NVME_SC_ANA_PERSISTENT_LOSS;
 806         if (unlikely(state == NVME_ANA_CHANGE))
 807                 return NVME_SC_ANA_TRANSITION;
 808         return 0;
 809 }
 810
 811 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
 812 {
 813         if (unlikely(req->ns->readonly)) {
 814                 switch (req->cmd->common.opcode) {
 815                 case nvme_cmd_read:
 816                 case nvme_cmd_flush:
 817                         break;
 818                 default:
 819                         return NVME_SC_NS_WRITE_PROTECTED;
 820                 }
 821         }
 822
 823         return 0;
 824 }
 825
 826 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
 827 {
 828         struct nvme_command *cmd = req->cmd;
 829         u16 ret;
 830
 831         ret = nvmet_check_ctrl_status(req, cmd);
 832         if (unlikely(ret))
 833                 return ret;
 834
 835         req->ns = nvmet_find_namespace(req->sq->ctrl, cmd->rw.nsid);
 836         if (unlikely(!req->ns)) {
 837                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 838                 return NVME_SC_INVALID_NS | NVME_SC_DNR;
 839         }
 840         ret = nvmet_check_ana_state(req->port, req->ns);
 841         if (unlikely(ret)) {
 842                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 843                 return ret;
 844         }
 845         ret = nvmet_io_cmd_check_access(req);
 846         if (unlikely(ret)) {
 847                 req->error_loc = offsetof(struct nvme_common_command, nsid);
 848                 return ret;
 849         }
 850
 851         if (req->ns->file)
 852                 return nvmet_file_parse_io_cmd(req);
 853         else
 854                 return nvmet_bdev_parse_io_cmd(req);
 855 }
 856
 857 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
 858                 struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
 859 {
 860         u8 flags = req->cmd->common.flags;
 861         u16 status;
 862
 863         req->cq = cq;
 864         req->sq = sq;
 865         req->ops = ops;
 866         req->sg = NULL;
 867         req->sg_cnt = 0;
 868         req->transfer_len = 0;
 869         req->cqe->status = 0;
 870         req->cqe->sq_head = 0;
 871         req->ns = NULL;
 872         req->error_loc = NVMET_NO_ERROR_LOC;
 873         req->error_slba = 0;
 874
 875         trace_nvmet_req_init(req, req->cmd);
 876
 877         /* no support for fused commands yet */
 878         if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
 879                 req->error_loc = offsetof(struct nvme_common_command, flags);
 880                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 881                 goto fail;
 882         }
 883
 884         /*
 885          * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
 886          * contains an address of a single contiguous physical buffer that is
 887          * byte aligned.
 888          */
 889         if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
 890                 req->error_loc = offsetof(struct nvme_common_command, flags);
 891                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 892                 goto fail;
 893         }
 894
 895         if (unlikely(!req->sq->ctrl))
 896                 /* will return an error for any Non-connect command: */
 897                 status = nvmet_parse_connect_cmd(req);
 898         else if (likely(req->sq->qid != 0))
 899                 status = nvmet_parse_io_cmd(req);
 900         else if (nvme_is_fabrics(req->cmd))
 901                 status = nvmet_parse_fabrics_cmd(req);
 902         else if (req->sq->ctrl->subsys->type == NVME_NQN_DISC)
 903                 status = nvmet_parse_discovery_cmd(req);
 904         else
 905                 status = nvmet_parse_admin_cmd(req);
 906
 907         if (status)
 908                 goto fail;
 909
 910         if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
 911                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 912                 goto fail;
 913         }
 914
 915         if (sq->ctrl)
 916                 sq->ctrl->cmd_seen = true;
 917
 918         return true;
 919
 920 fail:
 921         __nvmet_req_complete(req, status);
 922         return false;
 923 }
 924 EXPORT_SYMBOL_GPL(nvmet_req_init);
 925
 926 void nvmet_req_uninit(struct nvmet_req *req)
 927 {
 928         percpu_ref_put(&req->sq->ref);
 929         if (req->ns)
 930                 nvmet_put_namespace(req->ns);
 931 }
 932 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
 933
 934 void nvmet_req_execute(struct nvmet_req *req)
 935 {
 936         if (unlikely(req->data_len != req->transfer_len)) {
 937                 req->error_loc = offsetof(struct nvme_common_command, dptr);
 938                 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
 939         } else
 940                 req->execute(req);
 941 }
 942 EXPORT_SYMBOL_GPL(nvmet_req_execute);
 943
 944 int nvmet_req_alloc_sgl(struct nvmet_req *req)
 945 {
 946         struct pci_dev *p2p_dev = NULL;
 947
 948         if (IS_ENABLED(CONFIG_PCI_P2PDMA)) {
 949                 if (req->sq->ctrl && req->ns)
 950                         p2p_dev = radix_tree_lookup(&req->sq->ctrl->p2p_ns_map,
 951                                                     req->ns->nsid);
 952
 953                 req->p2p_dev = NULL;
 954                 if (req->sq->qid && p2p_dev) {
 955                         req->sg = pci_p2pmem_alloc_sgl(p2p_dev, &req->sg_cnt,
 956                                                        req->transfer_len);
 957                         if (req->sg) {
 958                                 req->p2p_dev = p2p_dev;
 959                                 return 0;
 960                         }
 961                 }
 962
 963                 /*
 964                  * If no P2P memory was available we fallback to using
 965                  * regular memory
 966                  */
 967         }
 968
 969         req->sg = sgl_alloc(req->transfer_len, GFP_KERNEL, &req->sg_cnt);
 970         if (!req->sg)
 971                 return -ENOMEM;
 972
 973         return 0;
 974 }
 975 EXPORT_SYMBOL_GPL(nvmet_req_alloc_sgl);
 976
 977 void nvmet_req_free_sgl(struct nvmet_req *req)
 978 {
 979         if (req->p2p_dev)
 980                 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
 981         else
 982                 sgl_free(req->sg);
 983
 984         req->sg = NULL;
 985         req->sg_cnt = 0;
 986 }
 987 EXPORT_SYMBOL_GPL(nvmet_req_free_sgl);
 988
 989 static inline bool nvmet_cc_en(u32 cc)
 990 {
 991         return (cc >> NVME_CC_EN_SHIFT) & 0x1;
 992 }
 993
 994 static inline u8 nvmet_cc_css(u32 cc)
 995 {
 996         return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
 997 }
 998
 999 static inline u8 nvmet_cc_mps(u32 cc)
1000 {
1001         return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
1002 }
1003
1004 static inline u8 nvmet_cc_ams(u32 cc)
1005 {
1006         return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
1007 }
1008
1009 static inline u8 nvmet_cc_shn(u32 cc)
1010 {
1011         return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
1012 }
1013
1014 static inline u8 nvmet_cc_iosqes(u32 cc)
1015 {
1016         return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
1017 }
1018
1019 static inline u8 nvmet_cc_iocqes(u32 cc)
1020 {
1021         return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
1022 }
1023
1024 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
1025 {
1026         lockdep_assert_held(&ctrl->lock);
1027
1028         if (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
1029             nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES ||
1030             nvmet_cc_mps(ctrl->cc) != 0 ||
1031             nvmet_cc_ams(ctrl->cc) != 0 ||
1032             nvmet_cc_css(ctrl->cc) != 0) {
1033                 ctrl->csts = NVME_CSTS_CFS;
1034                 return;
1035         }
1036
1037         ctrl->csts = NVME_CSTS_RDY;
1038
1039         /*
1040          * Controllers that are not yet enabled should not really enforce the
1041          * keep alive timeout, but we still want to track a timeout and cleanup
1042          * in case a host died before it enabled the controller.  Hence, simply
1043          * reset the keep alive timer when the controller is enabled.
1044          */
1045         mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
1046 }
1047
1048 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
1049 {
1050         lockdep_assert_held(&ctrl->lock);
1051
1052         /* XXX: tear down queues? */
1053         ctrl->csts &= ~NVME_CSTS_RDY;
1054         ctrl->cc = 0;
1055 }
1056
1057 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
1058 {
1059         u32 old;
1060
1061         mutex_lock(&ctrl->lock);
1062         old = ctrl->cc;
1063         ctrl->cc = new;
1064
1065         if (nvmet_cc_en(new) && !nvmet_cc_en(old))
1066                 nvmet_start_ctrl(ctrl);
1067         if (!nvmet_cc_en(new) && nvmet_cc_en(old))
1068                 nvmet_clear_ctrl(ctrl);
1069         if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
1070                 nvmet_clear_ctrl(ctrl);
1071                 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
1072         }
1073         if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
1074                 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
1075         mutex_unlock(&ctrl->lock);
1076 }
1077
1078 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
1079 {
1080         /* command sets supported: NVMe command set: */
1081         ctrl->cap = (1ULL << 37);
1082         /* CC.EN timeout in 500msec units: */
1083         ctrl->cap |= (15ULL << 24);
1084         /* maximum queue entries supported: */
1085         ctrl->cap |= NVMET_QUEUE_SIZE - 1;
1086 }
1087
1088 u16 nvmet_ctrl_find_get(const char *subsysnqn, const char *hostnqn, u16 cntlid,
1089                 struct nvmet_req *req, struct nvmet_ctrl **ret)
1090 {
1091         struct nvmet_subsys *subsys;
1092         struct nvmet_ctrl *ctrl;
1093         u16 status = 0;
1094
1095         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1096         if (!subsys) {
1097                 pr_warn("connect request for invalid subsystem %s!\n",
1098                         subsysnqn);
1099                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1100                 return NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1101         }
1102
1103         mutex_lock(&subsys->lock);
1104         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
1105                 if (ctrl->cntlid == cntlid) {
1106                         if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
1107                                 pr_warn("hostnqn mismatch.\n");
1108                                 continue;
1109                         }
1110                         if (!kref_get_unless_zero(&ctrl->ref))
1111                                 continue;
1112
1113                         *ret = ctrl;
1114                         goto out;
1115                 }
1116         }
1117
1118         pr_warn("could not find controller %d for subsys %s / host %s\n",
1119                 cntlid, subsysnqn, hostnqn);
1120         req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1121         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1122
1123 out:
1124         mutex_unlock(&subsys->lock);
1125         nvmet_subsys_put(subsys);
1126         return status;
1127 }
1128
1129 u16 nvmet_check_ctrl_status(struct nvmet_req *req, struct nvme_command *cmd)
1130 {
1131         if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1132                 pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1133                        cmd->common.opcode, req->sq->qid);
1134                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1135         }
1136
1137         if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1138                 pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1139                        cmd->common.opcode, req->sq->qid);
1140                 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1141         }
1142         return 0;
1143 }
1144
1145 bool nvmet_host_allowed(struct nvmet_subsys *subsys, const char *hostnqn)
1146 {
1147         struct nvmet_host_link *p;
1148
1149         lockdep_assert_held(&nvmet_config_sem);
1150
1151         if (subsys->allow_any_host)
1152                 return true;
1153
1154         if (subsys->type == NVME_NQN_DISC) /* allow all access to disc subsys */
1155                 return true;
1156
1157         list_for_each_entry(p, &subsys->hosts, entry) {
1158                 if (!strcmp(nvmet_host_name(p->host), hostnqn))
1159                         return true;
1160         }
1161
1162         return false;
1163 }
1164
1165 /*
1166  * Note: ctrl->subsys->lock should be held when calling this function
1167  */
1168 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1169                 struct nvmet_req *req)
1170 {
1171         struct nvmet_ns *ns;
1172
1173         if (!req->p2p_client)
1174                 return;
1175
1176         ctrl->p2p_client = get_device(req->p2p_client);
1177
1178         list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link,
1179                                 lockdep_is_held(&ctrl->subsys->lock))
1180                 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1181 }
1182
1183 /*
1184  * Note: ctrl->subsys->lock should be held when calling this function
1185  */
1186 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1187 {
1188         struct radix_tree_iter iter;
1189         void __rcu **slot;
1190
1191         radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1192                 pci_dev_put(radix_tree_deref_slot(slot));
1193
1194         put_device(ctrl->p2p_client);
1195 }
1196
1197 static void nvmet_fatal_error_handler(struct work_struct *work)
1198 {
1199         struct nvmet_ctrl *ctrl =
1200                         container_of(work, struct nvmet_ctrl, fatal_err_work);
1201
1202         pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1203         ctrl->ops->delete_ctrl(ctrl);
1204 }
1205
1206 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1207                 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
1208 {
1209         struct nvmet_subsys *subsys;
1210         struct nvmet_ctrl *ctrl;
1211         int ret;
1212         u16 status;
1213
1214         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1215         subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1216         if (!subsys) {
1217                 pr_warn("connect request for invalid subsystem %s!\n",
1218                         subsysnqn);
1219                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1220                 goto out;
1221         }
1222
1223         status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1224         down_read(&nvmet_config_sem);
1225         if (!nvmet_host_allowed(subsys, hostnqn)) {
1226                 pr_info("connect by host %s for subsystem %s not allowed\n",
1227                         hostnqn, subsysnqn);
1228                 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1229                 up_read(&nvmet_config_sem);
1230                 status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
1231                 goto out_put_subsystem;
1232         }
1233         up_read(&nvmet_config_sem);
1234
1235         status = NVME_SC_INTERNAL;
1236         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1237         if (!ctrl)
1238                 goto out_put_subsystem;
1239         mutex_init(&ctrl->lock);
1240
1241         nvmet_init_cap(ctrl);
1242
1243         ctrl->port = req->port;
1244
1245         INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
1246         INIT_LIST_HEAD(&ctrl->async_events);
1247         INIT_RADIX_TREE(&ctrl->p2p_ns_map, GFP_KERNEL);
1248         INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1249
1250         memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1251         memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1252
1253         kref_init(&ctrl->ref);
1254         ctrl->subsys = subsys;
1255         WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1256
1257         ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1258                         sizeof(__le32), GFP_KERNEL);
1259         if (!ctrl->changed_ns_list)
1260                 goto out_free_ctrl;
1261
1262         ctrl->cqs = kcalloc(subsys->max_qid + 1,
1263                         sizeof(struct nvmet_cq *),
1264                         GFP_KERNEL);
1265         if (!ctrl->cqs)
1266                 goto out_free_changed_ns_list;
1267
1268         ctrl->sqs = kcalloc(subsys->max_qid + 1,
1269                         sizeof(struct nvmet_sq *),
1270                         GFP_KERNEL);
1271         if (!ctrl->sqs)
1272                 goto out_free_cqs;
1273
1274         ret = ida_simple_get(&cntlid_ida,
1275                              NVME_CNTLID_MIN, NVME_CNTLID_MAX,
1276                              GFP_KERNEL);
1277         if (ret < 0) {
1278                 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
1279                 goto out_free_sqs;
1280         }
1281         ctrl->cntlid = ret;
1282
1283         ctrl->ops = req->ops;
1284
1285         /*
1286          * Discovery controllers may use some arbitrary high value
1287          * in order to cleanup stale discovery sessions
1288          */
1289         if ((ctrl->subsys->type == NVME_NQN_DISC) && !kato)
1290                 kato = NVMET_DISC_KATO_MS;
1291
1292         /* keep-alive timeout in seconds */
1293         ctrl->kato = DIV_ROUND_UP(kato, 1000);
1294
1295         ctrl->err_counter = 0;
1296         spin_lock_init(&ctrl->error_lock);
1297
1298         nvmet_start_keep_alive_timer(ctrl);
1299
1300         mutex_lock(&subsys->lock);
1301         list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1302         nvmet_setup_p2p_ns_map(ctrl, req);
1303         mutex_unlock(&subsys->lock);
1304
1305         *ctrlp = ctrl;
1306         return 0;
1307
1308 out_free_sqs:
1309         kfree(ctrl->sqs);
1310 out_free_cqs:
1311         kfree(ctrl->cqs);
1312 out_free_changed_ns_list:
1313         kfree(ctrl->changed_ns_list);
1314 out_free_ctrl:
1315         kfree(ctrl);
1316 out_put_subsystem:
1317         nvmet_subsys_put(subsys);
1318 out:
1319         return status;
1320 }
1321
1322 static void nvmet_ctrl_free(struct kref *ref)
1323 {
1324         struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1325         struct nvmet_subsys *subsys = ctrl->subsys;
1326
1327         mutex_lock(&subsys->lock);
1328         nvmet_release_p2p_ns_map(ctrl);
1329         list_del(&ctrl->subsys_entry);
1330         mutex_unlock(&subsys->lock);
1331
1332         nvmet_stop_keep_alive_timer(ctrl);
1333
1334         flush_work(&ctrl->async_event_work);
1335         cancel_work_sync(&ctrl->fatal_err_work);
1336
1337         ida_simple_remove(&cntlid_ida, ctrl->cntlid);
1338
1339         kfree(ctrl->sqs);
1340         kfree(ctrl->cqs);
1341         kfree(ctrl->changed_ns_list);
1342         kfree(ctrl);
1343
1344         nvmet_subsys_put(subsys);
1345 }
1346
1347 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1348 {
1349         kref_put(&ctrl->ref, nvmet_ctrl_free);
1350 }
1351
1352 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1353 {
1354         mutex_lock(&ctrl->lock);
1355         if (!(ctrl->csts & NVME_CSTS_CFS)) {
1356                 ctrl->csts |= NVME_CSTS_CFS;
1357                 schedule_work(&ctrl->fatal_err_work);
1358         }
1359         mutex_unlock(&ctrl->lock);
1360 }
1361 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1362
1363 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1364                 const char *subsysnqn)
1365 {
1366         struct nvmet_subsys_link *p;
1367
1368         if (!port)
1369                 return NULL;
1370
1371         if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1372                 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1373                         return NULL;
1374                 return nvmet_disc_subsys;
1375         }
1376
1377         down_read(&nvmet_config_sem);
1378         list_for_each_entry(p, &port->subsystems, entry) {
1379                 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1380                                 NVMF_NQN_SIZE)) {
1381                         if (!kref_get_unless_zero(&p->subsys->ref))
1382                                 break;
1383                         up_read(&nvmet_config_sem);
1384                         return p->subsys;
1385                 }
1386         }
1387         up_read(&nvmet_config_sem);
1388         return NULL;
1389 }
1390
1391 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1392                 enum nvme_subsys_type type)
1393 {
1394         struct nvmet_subsys *subsys;
1395
1396         subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1397         if (!subsys)
1398                 return ERR_PTR(-ENOMEM);
1399
1400         subsys->ver = NVME_VS(1, 3, 0); /* NVMe 1.3.0 */
1401         /* generate a random serial number as our controllers are ephemeral: */
1402         get_random_bytes(&subsys->serial, sizeof(subsys->serial));
1403
1404         switch (type) {
1405         case NVME_NQN_NVME:
1406                 subsys->max_qid = NVMET_NR_QUEUES;
1407                 break;
1408         case NVME_NQN_DISC:
1409                 subsys->max_qid = 0;
1410                 break;
1411         default:
1412                 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1413                 kfree(subsys);
1414                 return ERR_PTR(-EINVAL);
1415         }
1416         subsys->type = type;
1417         subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1418                         GFP_KERNEL);
1419         if (!subsys->subsysnqn) {
1420                 kfree(subsys);
1421                 return ERR_PTR(-ENOMEM);
1422         }
1423
1424         kref_init(&subsys->ref);
1425
1426         mutex_init(&subsys->lock);
1427         INIT_LIST_HEAD(&subsys->namespaces);
1428         INIT_LIST_HEAD(&subsys->ctrls);
1429         INIT_LIST_HEAD(&subsys->hosts);
1430
1431         return subsys;
1432 }
1433
1434 static void nvmet_subsys_free(struct kref *ref)
1435 {
1436         struct nvmet_subsys *subsys =
1437                 container_of(ref, struct nvmet_subsys, ref);
1438
1439         WARN_ON_ONCE(!list_empty(&subsys->namespaces));
1440
1441         kfree(subsys->subsysnqn);
1442         kfree(subsys);
1443 }
1444
1445 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1446 {
1447         struct nvmet_ctrl *ctrl;
1448
1449         mutex_lock(&subsys->lock);
1450         list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1451                 ctrl->ops->delete_ctrl(ctrl);
1452         mutex_unlock(&subsys->lock);
1453 }
1454
1455 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1456 {
1457         kref_put(&subsys->ref, nvmet_subsys_free);
1458 }
1459
1460 static int __init nvmet_init(void)
1461 {
1462         int error;
1463
1464         nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1465
1466         buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1467                         WQ_MEM_RECLAIM, 0);
1468         if (!buffered_io_wq) {
1469                 error = -ENOMEM;
1470                 goto out;
1471         }
1472
1473         error = nvmet_init_discovery();
1474         if (error)
1475                 goto out_free_work_queue;
1476
1477         error = nvmet_init_configfs();
1478         if (error)
1479                 goto out_exit_discovery;
1480         return 0;
1481
1482 out_exit_discovery:
1483         nvmet_exit_discovery();
1484 out_free_work_queue:
1485         destroy_workqueue(buffered_io_wq);
1486 out:
1487         return error;
1488 }
1489
1490 static void __exit nvmet_exit(void)
1491 {
1492         nvmet_exit_configfs();
1493         nvmet_exit_discovery();
1494         ida_destroy(&cntlid_ida);
1495         destroy_workqueue(buffered_io_wq);
1496
1497         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1498         BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1499 }
1500
1501 module_init(nvmet_init);
1502 module_exit(nvmet_exit);
1503
1504 MODULE_LICENSE("GPL v2");